[nodes] Add cancelation to the API

- 86932469e76f1315ee18bfa2fc52b588241dace1 add image_to_dataURL util
- 0c2611059711b45bb6142d30b1d1343ac24268f3 make fast latents method
static
- this method doesn't really need `self` and should be able to be called
without instantiating `Generator`
- 2360bfb6558ea511e9c9576f3d4b5535870d84b4 fix schema gen for
GraphExecutionState
- `GraphExecutionState` uses `default_factory` in its fields; the result
is the OpenAPI schema marks those fields as optional, which propagates
to the generated API client, which means we need a lot of unnecessary
type guards to use this data type. the [simple
fix](https://github.com/pydantic/pydantic/discussions/4577) is to add
config to explicitly say all class properties are required. looks this
this will be resolved in a future pydantic release
- 3cd7319cfdb0f07c6bb12d62d7d02efe1ab12675 fix step callback and fast
latent generation on nodes. have this working in UI. depends on the
small change in #2957

.coveragerc

@@ -0,0 +1,6 @@
+[run]
+omit='.env/*'
+source='.'
+
+[report]
+show_missing = true

.dockerignore

@@ -4,22 +4,22 @@
 !ldm
 !pyproject.toml
 
-# Guard against pulling in any models that might exist in the directory tree
-**/*.pt*
-**/*.ckpt
-
-# ignore frontend but whitelist dist
-invokeai/frontend/
-!invokeai/frontend/dist/
+# ignore frontend/web but whitelist dist
+invokeai/frontend/web/
+!invokeai/frontend/web/dist/
 
 # ignore invokeai/assets but whitelist invokeai/assets/web
 invokeai/assets/
 !invokeai/assets/web/
 
+# Guard against pulling in any models that might exist in the directory tree
+**/*.pt*
+**/*.ckpt
+
 # Byte-compiled / optimized / DLL files
 **/__pycache__/
 **/*.py[cod]
 
 # Distribution / packaging
-*.egg-info/
-*.egg
+**/*.egg-info/
+**/*.egg

.git-blame-ignore-revs

@@ -0,0 +1 @@
+b3dccfaeb636599c02effc377cdd8a87d658256c

.github/CODEOWNERS

@@ -1,51 +1,34 @@
 # continuous integration
-/.github/workflows/ @mauwii @lstein @blessedcoolant
+/.github/workflows/ @mauwii @lstein
 
 # documentation
-/docs/ @lstein @mauwii @tildebyte @blessedcoolant
-mkdocs.yml @lstein @mauwii @blessedcoolant
+/docs/ @lstein @mauwii @tildebyte
+/mkdocs.yml @lstein @mauwii
+
+# nodes
+/invokeai/app/ @Kyle0654 @blessedcoolant
 
 # installation and configuration
-/pyproject.toml @mauwii @lstein @ebr @blessedcoolant
-/docker/ @mauwii @lstein @blessedcoolant
-/scripts/ @ebr @lstein @blessedcoolant
-/installer/ @ebr @lstein @tildebyte  @blessedcoolant
-ldm/invoke/config @lstein @ebr  @blessedcoolant
-invokeai/assets @lstein @ebr  @blessedcoolant
-invokeai/configs @lstein @ebr  @blessedcoolant
-/ldm/invoke/_version.py @lstein @blessedcoolant
+/pyproject.toml @mauwii @lstein @blessedcoolant
+/docker/ @mauwii @lstein
+/scripts/ @ebr @lstein
+/installer/ @lstein @ebr
+/invokeai/assets @lstein @ebr
+/invokeai/configs @lstein
+/invokeai/version @lstein @blessedcoolant
 
 # web ui
 /invokeai/frontend @blessedcoolant @psychedelicious @lstein
 /invokeai/backend @blessedcoolant @psychedelicious @lstein
 
-# generation and model management
-/ldm/*.py @lstein @blessedcoolant
-/ldm/generate.py @lstein @keturn @blessedcoolant
-/ldm/invoke/args.py @lstein @blessedcoolant
-/ldm/invoke/ckpt* @lstein @blessedcoolant
-/ldm/invoke/ckpt_generator @lstein @blessedcoolant
-/ldm/invoke/CLI.py @lstein @blessedcoolant
-/ldm/invoke/config @lstein @ebr @mauwii @blessedcoolant
-/ldm/invoke/generator @keturn @damian0815 @blessedcoolant
-/ldm/invoke/globals.py @lstein @blessedcoolant 
-/ldm/invoke/merge_diffusers.py @lstein @blessedcoolant
-/ldm/invoke/model_manager.py @lstein @blessedcoolant
-/ldm/invoke/txt2mask.py @lstein @blessedcoolant
-/ldm/invoke/patchmatch.py @Kyle0654 @blessedcoolant @lstein
-/ldm/invoke/restoration @lstein @blessedcoolant
+# generation, model management, postprocessing
+/invokeai/backend @keturn @damian0815 @lstein @blessedcoolant @jpphoto
 
-# attention, textual inversion, model configuration
-/ldm/models @damian0815 @keturn @lstein @blessedcoolant
-/ldm/modules @damian0815 @keturn @lstein @blessedcoolant
+# front ends
+/invokeai/frontend/CLI @lstein
+/invokeai/frontend/install @lstein @ebr @mauwii 
+/invokeai/frontend/merge @lstein @blessedcoolant @hipsterusername
+/invokeai/frontend/training @lstein @blessedcoolant @hipsterusername
+/invokeai/frontend/web @psychedelicious @blessedcoolant
 
-# Nodes
-apps/ @Kyle0654 @lstein @blessedcoolant
-
-# legacy REST API
-# is CapableWeb still engaged?
-/ldm/invoke/pngwriter.py @CapableWeb @lstein @blessedcoolant
-/ldm/invoke/server_legacy.py @CapableWeb @lstein @blessedcoolant
-/scripts/legacy_api.py @CapableWeb @lstein @blessedcoolant
-/tests/legacy_tests.sh @CapableWeb @lstein @blessedcoolant
 

.github/ISSUE_TEMPLATE/BUG_REPORT.yml

@@ -65,6 +65,16 @@ body:
       placeholder: 8GB
     validations:
       required: false
+      
+  - type: input
+    id: version-number
+    attributes:
+      label: What version did you experience this issue on?
+      description: |
+        Please share the version of Invoke AI that you experienced the issue on. If this is not the latest version, please update first to confirm the issue still exists. If you are testing main, please include the commit hash instead.
+      placeholder: X.X.X
+    validations:
+      required: true
 
   - type: textarea
     id: what-happened

.github/workflows/build-container.yml

@@ -5,18 +5,17 @@ on:
       - 'main'
       - 'update/ci/docker/*'
       - 'update/docker/*'
+      - 'dev/ci/docker/*'
+      - 'dev/docker/*'
     paths:
       - 'pyproject.toml'
-      - 'ldm/**'
-      - 'invokeai/backend/**'
-      - 'invokeai/configs/**'
-      - 'invokeai/frontend/dist/**'
+      - '.dockerignore'
+      - 'invokeai/**'
       - 'docker/Dockerfile'
     tags:
       - 'v*.*.*'
   workflow_dispatch:
 
-
 jobs:
   docker:
     if: github.event.pull_request.draft == false
@@ -24,11 +23,11 @@ jobs:
       fail-fast: false
       matrix:
         flavor:
-          - amd
+          - rocm
           - cuda
           - cpu
         include:
-          - flavor: amd
+          - flavor: rocm
             pip-extra-index-url: 'https://download.pytorch.org/whl/rocm5.2'
           - flavor: cuda
             pip-extra-index-url: ''
@@ -54,9 +53,9 @@ jobs:
           tags: |
             type=ref,event=branch
             type=ref,event=tag
-            type=semver,pattern={{version}}
-            type=semver,pattern={{major}}.{{minor}}
-            type=semver,pattern={{major}}
+            type=pep440,pattern={{version}}
+            type=pep440,pattern={{major}}.{{minor}}
+            type=pep440,pattern={{major}}
             type=sha,enable=true,prefix=sha-,format=short
           flavor: |
             latest=${{ matrix.flavor == 'cuda' && github.ref == 'refs/heads/main' }}
@@ -92,7 +91,7 @@ jobs:
           context: .
           file: ${{ env.DOCKERFILE }}
           platforms: ${{ env.PLATFORMS }}
-          push: ${{ github.ref == 'refs/heads/main' || github.ref == 'refs/tags/*' }}
+          push: ${{ github.ref == 'refs/heads/main' || github.ref_type == 'tag' }}
           tags: ${{ steps.meta.outputs.tags }}
           labels: ${{ steps.meta.outputs.labels }}
           build-args: PIP_EXTRA_INDEX_URL=${{ matrix.pip-extra-index-url }}

.github/workflows/close-inactive-issues.yml

@@ -0,0 +1,27 @@
+name: Close inactive issues
+on:
+  schedule:
+    - cron: "00 6 * * *"
+
+env:
+  DAYS_BEFORE_ISSUE_STALE: 14
+  DAYS_BEFORE_ISSUE_CLOSE: 28
+
+jobs:
+  close-issues:
+    runs-on: ubuntu-latest
+    permissions:
+      issues: write
+      pull-requests: write
+    steps:
+      - uses: actions/stale@v5
+        with:
+          days-before-issue-stale: ${{ env.DAYS_BEFORE_ISSUE_STALE }}
+          days-before-issue-close: ${{ env.DAYS_BEFORE_ISSUE_CLOSE }}
+          stale-issue-label: "Inactive Issue"
+          stale-issue-message: "There has been no activity in this issue for ${{ env.DAYS_BEFORE_ISSUE_STALE }} days. If this issue is still being experienced, please reply with an updated confirmation that the issue is still being experienced with the latest release."
+          close-issue-message: "Due to inactivity, this issue was automatically closed. If you are still experiencing the issue, please recreate the issue."
+          days-before-pr-stale: -1
+          days-before-pr-close: -1
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
+          operations-per-run: 500

.github/workflows/lint-frontend.yml

@@ -3,14 +3,22 @@ name: Lint frontend
 on:
   pull_request:
     paths:
-      - 'invokeai/frontend/**'
+      - 'invokeai/frontend/web/**'
+    types:
+      - 'ready_for_review'
+      - 'opened'
+      - 'synchronize'
   push:
+    branches:
+      - 'main'
     paths:
-      - 'invokeai/frontend/**'
+      - 'invokeai/frontend/web/**'
+  merge_group:
+  workflow_dispatch:
 
 defaults:
   run:
-    working-directory: invokeai/frontend
+    working-directory: invokeai/frontend/web
 
 jobs:
   lint-frontend:
@@ -23,7 +31,7 @@ jobs:
           node-version: '18'
       - uses: actions/checkout@v3
       - run: 'yarn install --frozen-lockfile'
-      - run: 'yarn tsc'
-      - run: 'yarn run madge'
-      - run: 'yarn run lint --max-warnings=0'
-      - run: 'yarn run prettier --check'
+      - run: 'yarn run lint:tsc'
+      - run: 'yarn run lint:madge'
+      - run: 'yarn run lint:eslint'
+      - run: 'yarn run lint:prettier'

.github/workflows/pypi-release.yml

@@ -3,7 +3,7 @@ name: PyPI Release
 on:
   push:
     paths:
-      - 'ldm/invoke/_version.py'
+      - 'invokeai/version/invokeai_version.py'
   workflow_dispatch:
 
 jobs:

.github/workflows/test-invoke-pip-skip.yml

@@ -1,12 +1,12 @@
 name: Test invoke.py pip
 on:
   pull_request:
-    paths-ignore:
-      - 'pyproject.toml'
-      - 'ldm/**'
-      - 'invokeai/backend/**'
-      - 'invokeai/configs/**'
-      - 'invokeai/frontend/dist/**'
+    paths:
+      - '**'
+      - '!pyproject.toml'
+      - '!invokeai/**'
+      - 'invokeai/frontend/web/**'
+      - '!invokeai/frontend/web/dist/**'
   merge_group:
   workflow_dispatch:
 

.github/workflows/test-invoke-pip.yml

@@ -5,17 +5,15 @@ on:
       - 'main'
     paths:
       - 'pyproject.toml'
-      - 'ldm/**'
-      - 'invokeai/backend/**'
-      - 'invokeai/configs/**'
-      - 'invokeai/frontend/dist/**'
+      - 'invokeai/**'
+      - '!invokeai/frontend/web/**'
+      - 'invokeai/frontend/web/dist/**'
   pull_request:
     paths:
       - 'pyproject.toml'
-      - 'ldm/**'
-      - 'invokeai/backend/**'
-      - 'invokeai/configs/**'
-      - 'invokeai/frontend/dist/**'
+      - 'invokeai/**'
+      - '!invokeai/frontend/web/**'
+      - 'invokeai/frontend/web/dist/**'
     types:
       - 'ready_for_review'
       - 'opened'
@@ -112,7 +110,7 @@ jobs:
       - name: set INVOKEAI_OUTDIR
         run: >
           python -c
-          "import os;from ldm.invoke.globals import Globals;OUTDIR=os.path.join(Globals.root,str('outputs'));print(f'INVOKEAI_OUTDIR={OUTDIR}')"
+          "import os;from invokeai.backend.globals import Globals;OUTDIR=os.path.join(Globals.root,str('outputs'));print(f'INVOKEAI_OUTDIR={OUTDIR}')"
           >> ${{ matrix.github-env }}
 
       - name: run invokeai-configure

.gitignore

@@ -68,6 +68,7 @@ htmlcov/
 .cache
 nosetests.xml
 coverage.xml
+cov.xml
 *.cover
 *.py,cover
 .hypothesis/
@@ -197,7 +198,7 @@ checkpoints
 .DS_Store
 
 # Let the frontend manage its own gitignore
-!invokeai/frontend/*
+!invokeai/frontend/web/*
 
 # Scratch folder
 .scratch/
@@ -212,11 +213,6 @@ gfpgan/
 # config file (will be created by installer)
 configs/models.yaml
 
-# weights (will be created by installer)
-models/ldm/stable-diffusion-v1/*.ckpt
-models/clipseg
-models/gfpgan
-
 # ignore initfile
 .invokeai
 
@@ -231,6 +227,3 @@ installer/install.bat
 installer/install.sh
 installer/update.bat
 installer/update.sh
-
-# no longer stored in source directory
-models

.pytest.ini

@@ -0,0 +1,5 @@
+[pytest]
+DJANGO_SETTINGS_MODULE = webtas.settings
+; python_files = tests.py test_*.py *_tests.py
+
+addopts = --cov=. --cov-config=.coveragerc --cov-report xml:cov.xml

docker/Dockerfile

@@ -4,15 +4,15 @@ ARG PYTHON_VERSION=3.9
 ##################
 ##  base image  ##
 ##################
-FROM python:${PYTHON_VERSION}-slim AS python-base
+FROM --platform=${TARGETPLATFORM} python:${PYTHON_VERSION}-slim AS python-base
 
 LABEL org.opencontainers.image.authors="mauwii@outlook.de"
 
-# prepare for buildkit cache
+# Prepare apt for buildkit cache
 RUN rm -f /etc/apt/apt.conf.d/docker-clean \
   && echo 'Binary::apt::APT::Keep-Downloaded-Packages "true";' >/etc/apt/apt.conf.d/keep-cache
 
-# Install necessary packages
+# Install dependencies
 RUN \
   --mount=type=cache,target=/var/cache/apt,sharing=locked \
   --mount=type=cache,target=/var/lib/apt,sharing=locked \
@@ -23,7 +23,7 @@ RUN \
     libglib2.0-0=2.66.* \
     libopencv-dev=4.5.*
 
-# set working directory and env
+# Set working directory and env
 ARG APPDIR=/usr/src
 ARG APPNAME=InvokeAI
 WORKDIR ${APPDIR}
@@ -32,7 +32,7 @@ ENV PATH ${APPDIR}/${APPNAME}/bin:$PATH
 ENV PYTHONDONTWRITEBYTECODE 1
 # Turns off buffering for easier container logging
 ENV PYTHONUNBUFFERED 1
-# don't fall back to legacy build system
+# Don't fall back to legacy build system
 ENV PIP_USE_PEP517=1
 
 #######################
@@ -40,7 +40,7 @@ ENV PIP_USE_PEP517=1
 #######################
 FROM python-base AS pyproject-builder
 
-# Install dependencies
+# Install build dependencies
 RUN \
   --mount=type=cache,target=/var/cache/apt,sharing=locked \
   --mount=type=cache,target=/var/lib/apt,sharing=locked \
@@ -51,26 +51,30 @@ RUN \
     gcc=4:10.2.* \
     python3-dev=3.9.*
 
-# prepare pip for buildkit cache
+# Prepare pip for buildkit cache
 ARG PIP_CACHE_DIR=/var/cache/buildkit/pip
 ENV PIP_CACHE_DIR ${PIP_CACHE_DIR}
 RUN mkdir -p ${PIP_CACHE_DIR}
 
-# create virtual environment
-RUN --mount=type=cache,target=${PIP_CACHE_DIR},sharing=locked \
+# Create virtual environment
+RUN --mount=type=cache,target=${PIP_CACHE_DIR} \
   python3 -m venv "${APPNAME}" \
   --upgrade-deps
 
-# copy sources
-COPY --link . .
-
-# install pyproject.toml
+# Install requirements
+COPY --link pyproject.toml .
+COPY --link invokeai/version/invokeai_version.py invokeai/version/__init__.py invokeai/version/
 ARG PIP_EXTRA_INDEX_URL
 ENV PIP_EXTRA_INDEX_URL ${PIP_EXTRA_INDEX_URL}
-RUN --mount=type=cache,target=${PIP_CACHE_DIR},sharing=locked \
+RUN --mount=type=cache,target=${PIP_CACHE_DIR} \
+  "${APPNAME}"/bin/pip install .
+
+# Install pyproject.toml
+COPY --link . .
+RUN --mount=type=cache,target=${PIP_CACHE_DIR} \
   "${APPNAME}/bin/pip" install .
 
-# build patchmatch
+# Build patchmatch
 RUN python3 -c "from patchmatch import patch_match"
 
 #####################
@@ -86,14 +90,14 @@ RUN useradd \
   -U \
   "${UNAME}"
 
-# create volume directory
+# Create volume directory
 ARG VOLUME_DIR=/data
 RUN mkdir -p "${VOLUME_DIR}" \
-  && chown -R "${UNAME}" "${VOLUME_DIR}"
+  && chown -hR "${UNAME}:${UNAME}" "${VOLUME_DIR}"
 
-# setup runtime environment
-USER ${UNAME}
-COPY --chown=${UNAME} --from=pyproject-builder ${APPDIR}/${APPNAME} ${APPNAME}
+# Setup runtime environment
+USER ${UNAME}:${UNAME}
+COPY --chown=${UNAME}:${UNAME} --from=pyproject-builder ${APPDIR}/${APPNAME} ${APPNAME}
 ENV INVOKEAI_ROOT ${VOLUME_DIR}
 ENV TRANSFORMERS_CACHE ${VOLUME_DIR}/.cache
 ENV INVOKE_MODEL_RECONFIGURE "--yes --default_only"

docker/build.sh

@@ -41,7 +41,7 @@ else
 fi
 
 # Build Container
-DOCKER_BUILDKIT=1 docker build \
+docker build \
     --platform="${PLATFORM:-linux/amd64}" \
     --tag="${CONTAINER_IMAGE:-invokeai}" \
     ${CONTAINER_FLAVOR:+--build-arg="CONTAINER_FLAVOR=${CONTAINER_FLAVOR}"} \

docker/env.sh

@@ -49,3 +49,6 @@ CONTAINER_FLAVOR="${CONTAINER_FLAVOR-cuda}"
 CONTAINER_TAG="${CONTAINER_TAG-"${INVOKEAI_BRANCH##*/}-${CONTAINER_FLAVOR}"}"
 CONTAINER_IMAGE="${CONTAINER_REGISTRY}/${CONTAINER_REPOSITORY}:${CONTAINER_TAG}"
 CONTAINER_IMAGE="${CONTAINER_IMAGE,,}"
+
+# enable docker buildkit
+export DOCKER_BUILDKIT=1

docker/run.sh

@@ -21,10 +21,10 @@ docker run \
   --tty \
   --rm \
   --platform="${PLATFORM}" \
-  --name="${REPOSITORY_NAME,,}" \
-  --hostname="${REPOSITORY_NAME,,}" \
-  --mount=source="${VOLUMENAME}",target=/data \
-  --mount type=bind,source="$(pwd)"/outputs,target=/data/outputs \
+  --name="${REPOSITORY_NAME}" \
+  --hostname="${REPOSITORY_NAME}" \
+  --mount type=volume,volume-driver=local,source="${VOLUMENAME}",target=/data \
+  --mount type=bind,source="$(pwd)"/outputs/,target=/data/outputs/ \
   ${MODELSPATH:+--mount="type=bind,source=${MODELSPATH},target=/data/models"} \
   ${HUGGING_FACE_HUB_TOKEN:+--env="HUGGING_FACE_HUB_TOKEN=${HUGGING_FACE_HUB_TOKEN}"} \
   --publish=9090:9090 \
@@ -32,7 +32,7 @@ docker run \
   ${GPU_FLAGS:+--gpus="${GPU_FLAGS}"} \
   "${CONTAINER_IMAGE}" ${@:+$@}
 
-# Remove Trash folder
+echo -e "\nCleaning trash folder ..."
 for f in outputs/.Trash*; do
   if [ -e "$f" ]; then
     rm -Rf "$f"

docs/contributing/ARCHITECTURE.md

@@ -0,0 +1,93 @@
+# Invoke.AI Architecture
+
+```mermaid
+flowchart TB
+
+  subgraph apps[Applications]
+    webui[WebUI]
+    cli[CLI]
+
+  subgraph webapi[Web API]
+    api[HTTP API]
+    sio[Socket.IO]
+  end
+
+  end
+
+  subgraph invoke[Invoke]
+    direction LR
+    invoker
+    services
+    sessions
+    invocations
+  end
+
+  subgraph core[AI Core]
+    Generate
+  end
+
+  webui --> webapi
+  webapi --> invoke
+  cli --> invoke
+
+  invoker --> services & sessions
+  invocations --> services
+  sessions --> invocations
+
+  services --> core
+
+  %% Styles
+  classDef sg fill:#5028C8,font-weight:bold,stroke-width:2,color:#fff,stroke:#14141A
+  classDef default stroke-width:2px,stroke:#F6B314,color:#fff,fill:#14141A
+
+  class apps,webapi,invoke,core sg
+
+```
+
+## Applications
+
+Applications are built on top of the invoke framework. They should construct `invoker` and then interact through it. They should avoid interacting directly with core code in order to support a variety of configurations.
+
+### Web UI
+
+The Web UI is built on top of an HTTP API built with [FastAPI](https://fastapi.tiangolo.com/) and [Socket.IO](https://socket.io/). The frontend code is found in `/frontend` and the backend code is found in `/ldm/invoke/app/api_app.py` and `/ldm/invoke/app/api/`. The code is further organized as such:
+
+| Component | Description |
+| --- | --- |
+| api_app.py | Sets up the API app, annotates the OpenAPI spec with additional data, and runs the API |
+| dependencies | Creates all invoker services and the invoker, and provides them to the API |
+| events | An eventing system that could in the future be adapted to support horizontal scale-out |
+| sockets | The Socket.IO interface - handles listening to and emitting session events (events are defined in the events service module) |
+| routers | API definitions for different areas of API functionality |
+
+### CLI
+
+The CLI is built automatically from invocation metadata, and also supports invocation piping and auto-linking. Code is available in `/ldm/invoke/app/cli_app.py`.
+
+## Invoke
+
+The Invoke framework provides the interface to the underlying AI systems and is built with flexibility and extensibility in mind. There are four major concepts: invoker, sessions, invocations, and services.
+
+### Invoker
+
+The invoker (`/ldm/invoke/app/services/invoker.py`) is the primary interface through which applications interact with the framework. Its primary purpose is to create, manage, and invoke sessions. It also maintains two sets of services:
+- **invocation services**, which are used by invocations to interact with core functionality.
+- **invoker services**, which are used by the invoker to manage sessions and manage the invocation queue.
+
+### Sessions
+
+Invocations and links between them form a graph, which is maintained in a session. Sessions can be queued for invocation, which will execute their graph (either the next ready invocation, or all invocations). Sessions also maintain execution history for the graph (including storage of any outputs). An invocation may be added to a session at any time, and there is capability to add and entire graph at once, as well as to automatically link new invocations to previous invocations. Invocations can not be deleted or modified once added.
+
+The session graph does not support looping. This is left as an application problem to prevent additional complexity in the graph.
+
+### Invocations
+
+Invocations represent individual units of execution, with inputs and outputs. All invocations are located in `/ldm/invoke/app/invocations`, and are all automatically discovered and made available in the applications. These are the primary way to expose new functionality in Invoke.AI, and the [implementation guide](INVOCATIONS.md) explains how to add new invocations.
+
+### Services
+
+Services provide invocations access AI Core functionality and other necessary functionality (e.g. image storage). These are available in `/ldm/invoke/app/services`. As a general rule, new services should provide an interface as an abstract base class, and may provide a lightweight local implementation by default in their module. The goal for all services should be to enable the usage of different implementations (e.g. using cloud storage for image storage), but should not load any module dependencies unless that implementation has been used (i.e. don't import anything that won't be used, especially if it's expensive to import).
+
+## AI Core
+
+The AI Core is represented by the rest of the code base (i.e. the code outside of `/ldm/invoke/app/`).

docs/contributing/INVOCATIONS.md

@@ -0,0 +1,105 @@
+# Invocations
+
+Invocations represent a single operation, its inputs, and its outputs. These operations and their outputs can be chained together to generate and modify images.
+
+## Creating a new invocation
+
+To create a new invocation, either find the appropriate module file in `/ldm/invoke/app/invocations` to add your invocation to, or create a new one in that folder. All invocations in that folder will be discovered and made available to the CLI and API automatically. Invocations make use of [typing](https://docs.python.org/3/library/typing.html) and [pydantic](https://pydantic-docs.helpmanual.io/) for validation and integration into the CLI and API.
+
+An invocation looks like this:
+
+```py
+class UpscaleInvocation(BaseInvocation):
+    """Upscales an image."""
+    type: Literal['upscale'] = 'upscale'
+
+    # Inputs
+    image: Union[ImageField,None] = Field(description="The input image")
+    strength: float               = Field(default=0.75, gt=0, le=1, description="The strength")
+    level: Literal[2,4]           = Field(default=2, description = "The upscale level")
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get(self.image.image_type, self.image.image_name)
+        results = context.services.generate.upscale_and_reconstruct(
+            image_list     = [[image, 0]],
+            upscale        = (self.level, self.strength),
+            strength       = 0.0, # GFPGAN strength
+            save_original  = False,
+            image_callback = None,
+        )
+
+        # Results are image and seed, unwrap for now
+        # TODO: can this return multiple results?
+        image_type = ImageType.RESULT
+        image_name = context.services.images.create_name(context.graph_execution_state_id, self.id)
+        context.services.images.save(image_type, image_name, results[0][0])
+        return ImageOutput(
+            image = ImageField(image_type = image_type, image_name = image_name)
+        )
+```
+
+Each portion is important to implement correctly.
+
+### Class definition and type
+```py
+class UpscaleInvocation(BaseInvocation):
+    """Upscales an image."""
+    type: Literal['upscale'] = 'upscale'
+```
+All invocations must derive from `BaseInvocation`. They should have a docstring that declares what they do in a single, short line. They should also have a `type` with a type hint that's `Literal["command_name"]`, where `command_name` is what the user will type on the CLI or use in the API to create this invocation. The `command_name` must be unique. The `type` must be assigned to the value of the literal in the type hint.
+
+### Inputs
+```py
+    # Inputs
+    image: Union[ImageField,None] = Field(description="The input image")
+    strength: float               = Field(default=0.75, gt=0, le=1, description="The strength")
+    level: Literal[2,4]           = Field(default=2, description="The upscale level")
+```
+Inputs consist of three parts: a name, a type hint, and a `Field` with default, description, and validation information. For example:
+| Part | Value | Description |
+| ---- | ----- | ----------- |
+| Name | `strength` | This field is referred to as `strength` |
+| Type Hint | `float` | This field must be of type `float` |
+| Field | `Field(default=0.75, gt=0, le=1, description="The strength")` | The default value is `0.75`, the value must be in the range (0,1], and help text will show "The strength" for this field. |
+
+Notice that `image` has type `Union[ImageField,None]`. The `Union` allows this field to be parsed with `None` as a value, which enables linking to previous invocations. All fields should either provide a default value or allow `None` as a value, so that they can be overwritten with a linked output from another invocation.
+
+The special type `ImageField` is also used here. All images are passed as `ImageField`, which protects them from pydantic validation errors (since images only ever come from links).
+
+Finally, note that for all linking, the `type` of the linked fields must match. If the `name` also matches, then the field can be **automatically linked** to a previous invocation by name and matching.
+
+### Invoke Function
+```py
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get(self.image.image_type, self.image.image_name)
+        results = context.services.generate.upscale_and_reconstruct(
+            image_list     = [[image, 0]],
+            upscale        = (self.level, self.strength),
+            strength       = 0.0, # GFPGAN strength
+            save_original  = False,
+            image_callback = None,
+        )
+
+        # Results are image and seed, unwrap for now
+        image_type = ImageType.RESULT
+        image_name = context.services.images.create_name(context.graph_execution_state_id, self.id)
+        context.services.images.save(image_type, image_name, results[0][0])
+        return ImageOutput(
+            image = ImageField(image_type = image_type, image_name = image_name)
+        )
+```
+The `invoke` function is the last portion of an invocation. It is provided an `InvocationContext` which contains services to perform work as well as a `session_id` for use as needed. It should return a class with output values that derives from `BaseInvocationOutput`.
+
+Before being called, the invocation will have all of its fields set from defaults, inputs, and finally links (overriding in that order).
+
+Assume that this invocation may be running simultaneously with other invocations, may be running on another machine, or in other interesting scenarios. If you need functionality, please provide it as a service in the `InvocationServices` class, and make sure it can be overridden.
+
+### Outputs
+```py
+class ImageOutput(BaseInvocationOutput):
+    """Base class for invocations that output an image"""
+    type: Literal['image'] = 'image'
+
+    image: ImageField = Field(default=None, description="The output image")
+```
+Output classes look like an invocation class without the invoke method. Prefer to use an existing output class if available, and prefer to name inputs the same as outputs when possible, to promote automatic invocation linking.

docs/installation/020_INSTALL_MANUAL.md

@@ -148,7 +148,7 @@ manager, please follow these steps:
     === "CUDA (NVidia)"
 
         ```bash
-        pip install InvokeAI[xformers] --use-pep517 --extra-index-url https://download.pytorch.org/whl/cu117
+        pip install "InvokeAI[xformers]" --use-pep517 --extra-index-url https://download.pytorch.org/whl/cu117
         ```
 
     === "ROCm (AMD)"

installer/create_installer.sh

@@ -11,10 +11,10 @@ if [[ -v "VIRTUAL_ENV" ]]; then
     exit -1
 fi
 
-VERSION=$(cd ..; python -c "from ldm.invoke import __version__ as version; print(version)")
+VERSION=$(cd ..; python -c "from invokeai.version import __version__ as version; print(version)")
 PATCH=""
 VERSION="v${VERSION}${PATCH}"
-LATEST_TAG="v2.3-latest"
+LATEST_TAG="v3.0-latest"
 
 echo Building installer for version $VERSION
 echo "Be certain that you're in the 'installer' directory before continuing."

installer/lib/installer.py

@@ -291,7 +291,7 @@ class InvokeAiInstance:
                     src = Path(__file__).parents[1].expanduser().resolve()
                     # if the above directory contains one of these files, we'll do a source install
                     next(src.glob("pyproject.toml"))
-                    next(src.glob("ldm"))
+                    next(src.glob("invokeai"))
                 except StopIteration:
                     print("Unable to find a wheel or perform a source install. Giving up.")
 
@@ -342,14 +342,14 @@ class InvokeAiInstance:
 
         introduction()
 
-        from ldm.invoke.config import invokeai_configure
+        from invokeai.frontend.install import invokeai_configure
 
         # NOTE: currently the config script does its own arg parsing! this means the command-line switches
         # from the installer will also automatically propagate down to the config script.
         # this may change in the future with config refactoring!
         succeeded = False
         try:
-            invokeai_configure.main()
+            invokeai_configure()
             succeeded = True
         except requests.exceptions.ConnectionError as e:
             print(f'\nA network error was encountered during configuration and download: {str(e)}')

invokeai/README

@@ -1,3 +1,11 @@
-After version 2.3 is released, the ldm/invoke modules will be migrated to this location
-so that we have a proper invokeai distribution. Currently it is only being used for
-data files.
+Organization of the source tree:
+
+app -- Home of nodes invocations and services
+assets -- Images and other data files used by InvokeAI
+backend -- Non-user facing libraries, including the rendering
+	core.
+configs -- Configuration files used at install and run times
+frontend -- User-facing scripts, including the CLI and the WebUI
+version -- Current InvokeAI version string, stored
+	in version/invokeai_version.py
+	

invokeai/app/api/dependencies.py

@@ -0,0 +1,79 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+import os
+from argparse import Namespace
+
+from ...backend import Globals
+from ..services.model_manager_initializer import get_model_manager
+from ..services.restoration_services import RestorationServices
+from ..services.graph import GraphExecutionState
+from ..services.image_storage import DiskImageStorage
+from ..services.invocation_queue import MemoryInvocationQueue
+from ..services.invocation_services import InvocationServices
+from ..services.invoker import Invoker
+from ..services.processor import DefaultInvocationProcessor
+from ..services.sqlite import SqliteItemStorage
+from .events import FastAPIEventService
+
+
+# TODO: is there a better way to achieve this?
+def check_internet() -> bool:
+    """
+    Return true if the internet is reachable.
+    It does this by pinging huggingface.co.
+    """
+    import urllib.request
+
+    host = "http://huggingface.co"
+    try:
+        urllib.request.urlopen(host, timeout=1)
+        return True
+    except:
+        return False
+
+
+class ApiDependencies:
+    """Contains and initializes all dependencies for the API"""
+
+    invoker: Invoker = None
+
+    @staticmethod
+    def initialize(config, event_handler_id: int):
+        Globals.try_patchmatch = config.patchmatch
+        Globals.always_use_cpu = config.always_use_cpu
+        Globals.internet_available = config.internet_available and check_internet()
+        Globals.disable_xformers = not config.xformers
+        Globals.ckpt_convert = config.ckpt_convert
+
+        # TODO: Use a logger
+        print(f">> Internet connectivity is {Globals.internet_available}")
+
+        events = FastAPIEventService(event_handler_id)
+
+        output_folder = os.path.abspath(
+            os.path.join(os.path.dirname(__file__), "../../../../outputs")
+        )
+
+        images = DiskImageStorage(output_folder)
+
+        # TODO: build a file/path manager?
+        db_location = os.path.join(output_folder, "invokeai.db")
+
+        services = InvocationServices(
+            model_manager=get_model_manager(config),
+            events=events,
+            images=images,
+            queue=MemoryInvocationQueue(),
+            graph_execution_manager=SqliteItemStorage[GraphExecutionState](
+                filename=db_location, table_name="graph_executions"
+            ),
+            processor=DefaultInvocationProcessor(),
+            restoration=RestorationServices(config),
+        )
+
+        ApiDependencies.invoker = Invoker(services)
+
+    @staticmethod
+    def shutdown():
+        if ApiDependencies.invoker:
+            ApiDependencies.invoker.stop()

invokeai/app/api/events.py

@@ -0,0 +1,52 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+import asyncio
+import threading
+from queue import Empty, Queue
+from typing import Any
+
+from fastapi_events.dispatcher import dispatch
+
+from ..services.events import EventServiceBase
+
+
+class FastAPIEventService(EventServiceBase):
+    event_handler_id: int
+    __queue: Queue
+    __stop_event: threading.Event
+
+    def __init__(self, event_handler_id: int) -> None:
+        self.event_handler_id = event_handler_id
+        self.__queue = Queue()
+        self.__stop_event = threading.Event()
+        asyncio.create_task(self.__dispatch_from_queue(stop_event=self.__stop_event))
+
+        super().__init__()
+
+    def stop(self, *args, **kwargs):
+        self.__stop_event.set()
+        self.__queue.put(None)
+
+    def dispatch(self, event_name: str, payload: Any) -> None:
+        self.__queue.put(dict(event_name=event_name, payload=payload))
+
+    async def __dispatch_from_queue(self, stop_event: threading.Event):
+        """Get events on from the queue and dispatch them, from the correct thread"""
+        while not stop_event.is_set():
+            try:
+                event = self.__queue.get(block=False)
+                if not event:  # Probably stopping
+                    continue
+
+                dispatch(
+                    event.get("event_name"),
+                    payload=event.get("payload"),
+                    middleware_id=self.event_handler_id,
+                )
+
+            except Empty:
+                await asyncio.sleep(0.001)
+                pass
+
+            except asyncio.CancelledError as e:
+                raise e  # Raise a proper error

invokeai/app/api/routers/images.py

@@ -0,0 +1,56 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from datetime import datetime, timezone
+
+from fastapi import Path, Request, UploadFile
+from fastapi.responses import FileResponse, Response
+from fastapi.routing import APIRouter
+from PIL import Image
+
+from ...services.image_storage import ImageType
+from ..dependencies import ApiDependencies
+
+images_router = APIRouter(prefix="/v1/images", tags=["images"])
+
+
+@images_router.get("/{image_type}/{image_name}", operation_id="get_image")
+async def get_image(
+    image_type: ImageType = Path(description="The type of image to get"),
+    image_name: str = Path(description="The name of the image to get"),
+):
+    """Gets a result"""
+    # TODO: This is not really secure at all. At least make sure only output results are served
+    filename = ApiDependencies.invoker.services.images.get_path(image_type, image_name)
+    return FileResponse(filename)
+
+
+@images_router.post(
+    "/uploads/",
+    operation_id="upload_image",
+    responses={
+        201: {"description": "The image was uploaded successfully"},
+        404: {"description": "Session not found"},
+    },
+)
+async def upload_image(file: UploadFile, request: Request):
+    if not file.content_type.startswith("image"):
+        return Response(status_code=415)
+
+    contents = await file.read()
+    try:
+        im = Image.open(contents)
+    except:
+        # Error opening the image
+        return Response(status_code=415)
+
+    filename = f"{str(int(datetime.now(timezone.utc).timestamp()))}.png"
+    ApiDependencies.invoker.services.images.save(ImageType.UPLOAD, filename, im)
+
+    return Response(
+        status_code=201,
+        headers={
+            "Location": request.url_for(
+                "get_image", image_type=ImageType.UPLOAD, image_name=filename
+            )
+        },
+    )

invokeai/app/api/routers/sessions.py

@@ -0,0 +1,287 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from typing import Annotated, List, Optional, Union
+
+from fastapi import Body, Path, Query
+from fastapi.responses import Response
+from fastapi.routing import APIRouter
+from pydantic.fields import Field
+
+from ...invocations import *
+from ...invocations.baseinvocation import BaseInvocation
+from ...services.graph import (
+    Edge,
+    EdgeConnection,
+    Graph,
+    GraphExecutionState,
+    NodeAlreadyExecutedError,
+)
+from ...services.item_storage import PaginatedResults
+from ..dependencies import ApiDependencies
+
+session_router = APIRouter(prefix="/v1/sessions", tags=["sessions"])
+
+
+@session_router.post(
+    "/",
+    operation_id="create_session",
+    responses={
+        200: {"model": GraphExecutionState},
+        400: {"description": "Invalid json"},
+    },
+)
+async def create_session(
+    graph: Optional[Graph] = Body(
+        default=None, description="The graph to initialize the session with"
+    )
+) -> GraphExecutionState:
+    """Creates a new session, optionally initializing it with an invocation graph"""
+    session = ApiDependencies.invoker.create_execution_state(graph)
+    return session
+
+
+@session_router.get(
+    "/",
+    operation_id="list_sessions",
+    responses={200: {"model": PaginatedResults[GraphExecutionState]}},
+)
+async def list_sessions(
+    page: int = Query(default=0, description="The page of results to get"),
+    per_page: int = Query(default=10, description="The number of results per page"),
+    query: str = Query(default="", description="The query string to search for"),
+) -> PaginatedResults[GraphExecutionState]:
+    """Gets a list of sessions, optionally searching"""
+    if filter == "":
+        result = ApiDependencies.invoker.services.graph_execution_manager.list(
+            page, per_page
+        )
+    else:
+        result = ApiDependencies.invoker.services.graph_execution_manager.search(
+            query, page, per_page
+        )
+    return result
+
+
+@session_router.get(
+    "/{session_id}",
+    operation_id="get_session",
+    responses={
+        200: {"model": GraphExecutionState},
+        404: {"description": "Session not found"},
+    },
+)
+async def get_session(
+    session_id: str = Path(description="The id of the session to get"),
+) -> GraphExecutionState:
+    """Gets a session"""
+    session = ApiDependencies.invoker.services.graph_execution_manager.get(session_id)
+    if session is None:
+        return Response(status_code=404)
+    else:
+        return session
+
+
+@session_router.post(
+    "/{session_id}/nodes",
+    operation_id="add_node",
+    responses={
+        200: {"model": str},
+        400: {"description": "Invalid node or link"},
+        404: {"description": "Session not found"},
+    },
+)
+async def add_node(
+    session_id: str = Path(description="The id of the session"),
+    node: Annotated[
+        Union[BaseInvocation.get_invocations()], Field(discriminator="type") # type: ignore
+    ] = Body(description="The node to add"),
+) -> str:
+    """Adds a node to the graph"""
+    session = ApiDependencies.invoker.services.graph_execution_manager.get(session_id)
+    if session is None:
+        return Response(status_code=404)
+
+    try:
+        session.add_node(node)
+        ApiDependencies.invoker.services.graph_execution_manager.set(
+            session
+        )  # TODO: can this be done automatically, or add node through an API?
+        return session.id
+    except NodeAlreadyExecutedError:
+        return Response(status_code=400)
+    except IndexError:
+        return Response(status_code=400)
+
+
+@session_router.put(
+    "/{session_id}/nodes/{node_path}",
+    operation_id="update_node",
+    responses={
+        200: {"model": GraphExecutionState},
+        400: {"description": "Invalid node or link"},
+        404: {"description": "Session not found"},
+    },
+)
+async def update_node(
+    session_id: str = Path(description="The id of the session"),
+    node_path: str = Path(description="The path to the node in the graph"),
+    node: Annotated[
+        Union[BaseInvocation.get_invocations()], Field(discriminator="type") # type: ignore
+    ] = Body(description="The new node"),
+) -> GraphExecutionState:
+    """Updates a node in the graph and removes all linked edges"""
+    session = ApiDependencies.invoker.services.graph_execution_manager.get(session_id)
+    if session is None:
+        return Response(status_code=404)
+
+    try:
+        session.update_node(node_path, node)
+        ApiDependencies.invoker.services.graph_execution_manager.set(
+            session
+        )  # TODO: can this be done automatically, or add node through an API?
+        return session
+    except NodeAlreadyExecutedError:
+        return Response(status_code=400)
+    except IndexError:
+        return Response(status_code=400)
+
+
+@session_router.delete(
+    "/{session_id}/nodes/{node_path}",
+    operation_id="delete_node",
+    responses={
+        200: {"model": GraphExecutionState},
+        400: {"description": "Invalid node or link"},
+        404: {"description": "Session not found"},
+    },
+)
+async def delete_node(
+    session_id: str = Path(description="The id of the session"),
+    node_path: str = Path(description="The path to the node to delete"),
+) -> GraphExecutionState:
+    """Deletes a node in the graph and removes all linked edges"""
+    session = ApiDependencies.invoker.services.graph_execution_manager.get(session_id)
+    if session is None:
+        return Response(status_code=404)
+
+    try:
+        session.delete_node(node_path)
+        ApiDependencies.invoker.services.graph_execution_manager.set(
+            session
+        )  # TODO: can this be done automatically, or add node through an API?
+        return session
+    except NodeAlreadyExecutedError:
+        return Response(status_code=400)
+    except IndexError:
+        return Response(status_code=400)
+
+
+@session_router.post(
+    "/{session_id}/edges",
+    operation_id="add_edge",
+    responses={
+        200: {"model": GraphExecutionState},
+        400: {"description": "Invalid node or link"},
+        404: {"description": "Session not found"},
+    },
+)
+async def add_edge(
+    session_id: str = Path(description="The id of the session"),
+    edge: Edge = Body(description="The edge to add"),
+) -> GraphExecutionState:
+    """Adds an edge to the graph"""
+    session = ApiDependencies.invoker.services.graph_execution_manager.get(session_id)
+    if session is None:
+        return Response(status_code=404)
+
+    try:
+        session.add_edge(edge)
+        ApiDependencies.invoker.services.graph_execution_manager.set(
+            session
+        )  # TODO: can this be done automatically, or add node through an API?
+        return session
+    except NodeAlreadyExecutedError:
+        return Response(status_code=400)
+    except IndexError:
+        return Response(status_code=400)
+
+
+# TODO: the edge being in the path here is really ugly, find a better solution
+@session_router.delete(
+    "/{session_id}/edges/{from_node_id}/{from_field}/{to_node_id}/{to_field}",
+    operation_id="delete_edge",
+    responses={
+        200: {"model": GraphExecutionState},
+        400: {"description": "Invalid node or link"},
+        404: {"description": "Session not found"},
+    },
+)
+async def delete_edge(
+    session_id: str = Path(description="The id of the session"),
+    from_node_id: str = Path(description="The id of the node the edge is coming from"),
+    from_field: str = Path(description="The field of the node the edge is coming from"),
+    to_node_id: str = Path(description="The id of the node the edge is going to"),
+    to_field: str = Path(description="The field of the node the edge is going to"),
+) -> GraphExecutionState:
+    """Deletes an edge from the graph"""
+    session = ApiDependencies.invoker.services.graph_execution_manager.get(session_id)
+    if session is None:
+        return Response(status_code=404)
+
+    try:
+        edge = Edge(
+            source=EdgeConnection(node_id=from_node_id, field=from_field),
+            destination=EdgeConnection(node_id=to_node_id, field=to_field)
+        )
+        session.delete_edge(edge)
+        ApiDependencies.invoker.services.graph_execution_manager.set(
+            session
+        )  # TODO: can this be done automatically, or add node through an API?
+        return session
+    except NodeAlreadyExecutedError:
+        return Response(status_code=400)
+    except IndexError:
+        return Response(status_code=400)
+
+
+@session_router.put(
+    "/{session_id}/invoke",
+    operation_id="invoke_session",
+    responses={
+        200: {"model": None},
+        202: {"description": "The invocation is queued"},
+        400: {"description": "The session has no invocations ready to invoke"},
+        404: {"description": "Session not found"},
+    },
+)
+async def invoke_session(
+    session_id: str = Path(description="The id of the session to invoke"),
+    all: bool = Query(
+        default=False, description="Whether or not to invoke all remaining invocations"
+    ),
+) -> None:
+    """Invokes a session"""
+    session = ApiDependencies.invoker.services.graph_execution_manager.get(session_id)
+    if session is None:
+        return Response(status_code=404)
+
+    if session.is_complete():
+        return Response(status_code=400)
+
+    ApiDependencies.invoker.invoke(session, invoke_all=all)
+    return Response(status_code=202)
+
+
+@session_router.delete(
+    "/{session_id}/invoke",
+    operation_id="cancel_session_invoke",
+    responses={
+        202: {"description": "The invocation is canceled"}
+    },
+)
+async def cancel_session_invoke(
+    session_id: str = Path(description="The id of the session to cancel"),
+) -> None:
+    """Invokes a session"""
+    ApiDependencies.invoker.cancel(session_id)
+    return Response(status_code=202)

invokeai/app/api/sockets.py

@@ -0,0 +1,38 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from fastapi import FastAPI
+from fastapi_events.handlers.local import local_handler
+from fastapi_events.typing import Event
+from fastapi_socketio import SocketManager
+
+from ..services.events import EventServiceBase
+
+
+class SocketIO:
+    __sio: SocketManager
+
+    def __init__(self, app: FastAPI):
+        self.__sio = SocketManager(app=app)
+        self.__sio.on("subscribe", handler=self._handle_sub)
+        self.__sio.on("unsubscribe", handler=self._handle_unsub)
+
+        local_handler.register(
+            event_name=EventServiceBase.session_event, _func=self._handle_session_event
+        )
+
+    async def _handle_session_event(self, event: Event):
+        await self.__sio.emit(
+            event=event[1]["event"],
+            data=event[1]["data"],
+            room=event[1]["data"]["graph_execution_state_id"],
+        )
+
+    async def _handle_sub(self, sid, data, *args, **kwargs):
+        if "session" in data:
+            self.__sio.enter_room(sid, data["session"])
+
+        # @app.sio.on('unsubscribe')
+
+    async def _handle_unsub(self, sid, data, *args, **kwargs):
+        if "session" in data:
+            self.__sio.leave_room(sid, data["session"])

invokeai/app/api_app.py

@@ -0,0 +1,158 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+import asyncio
+from inspect import signature
+
+import uvicorn
+from fastapi import FastAPI
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.openapi.docs import get_redoc_html, get_swagger_ui_html
+from fastapi.openapi.utils import get_openapi
+from fastapi.staticfiles import StaticFiles
+from fastapi_events.handlers.local import local_handler
+from fastapi_events.middleware import EventHandlerASGIMiddleware
+from pydantic.schema import schema
+
+from ..backend import Args
+from .api.dependencies import ApiDependencies
+from .api.routers import images, sessions
+from .api.sockets import SocketIO
+from .invocations import *
+from .invocations.baseinvocation import BaseInvocation
+
+# Create the app
+# TODO: create this all in a method so configuration/etc. can be passed in?
+app = FastAPI(title="Invoke AI", docs_url=None, redoc_url=None)
+
+# Add event handler
+event_handler_id: int = id(app)
+app.add_middleware(
+    EventHandlerASGIMiddleware,
+    handlers=[
+        local_handler
+    ],  # TODO: consider doing this in services to support different configurations
+    middleware_id=event_handler_id,
+)
+
+# Add CORS
+# TODO: use configuration for this
+origins = []
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=origins,
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+socket_io = SocketIO(app)
+
+config = {}
+
+
+# Add startup event to load dependencies
+@app.on_event("startup")
+async def startup_event():
+    config = Args()
+    config.parse_args()
+
+    ApiDependencies.initialize(
+        config=config, event_handler_id=event_handler_id
+    )
+
+
+# Shut down threads
+@app.on_event("shutdown")
+async def shutdown_event():
+    ApiDependencies.shutdown()
+
+
+# Include all routers
+# TODO: REMOVE
+# app.include_router(
+#     invocation.invocation_router,
+#     prefix = '/api')
+
+app.include_router(sessions.session_router, prefix="/api")
+
+app.include_router(images.images_router, prefix="/api")
+
+
+# Build a custom OpenAPI to include all outputs
+# TODO: can outputs be included on metadata of invocation schemas somehow?
+def custom_openapi():
+    if app.openapi_schema:
+        return app.openapi_schema
+    openapi_schema = get_openapi(
+        title=app.title,
+        description="An API for invoking AI image operations",
+        version="1.0.0",
+        routes=app.routes,
+    )
+
+    # Add all outputs
+    all_invocations = BaseInvocation.get_invocations()
+    output_types = set()
+    output_type_titles = dict()
+    for invoker in all_invocations:
+        output_type = signature(invoker.invoke).return_annotation
+        output_types.add(output_type)
+
+    output_schemas = schema(output_types, ref_prefix="#/components/schemas/")
+    for schema_key, output_schema in output_schemas["definitions"].items():
+        openapi_schema["components"]["schemas"][schema_key] = output_schema
+
+        # TODO: note that we assume the schema_key here is the TYPE.__name__
+        # This could break in some cases, figure out a better way to do it
+        output_type_titles[schema_key] = output_schema["title"]
+
+    # Add a reference to the output type to additionalProperties of the invoker schema
+    for invoker in all_invocations:
+        invoker_name = invoker.__name__
+        output_type = signature(invoker.invoke).return_annotation
+        output_type_title = output_type_titles[output_type.__name__]
+        invoker_schema = openapi_schema["components"]["schemas"][invoker_name]
+        outputs_ref = {"$ref": f"#/components/schemas/{output_type_title}"}
+
+        invoker_schema["output"] = outputs_ref
+
+    app.openapi_schema = openapi_schema
+    return app.openapi_schema
+
+
+app.openapi = custom_openapi
+
+# Override API doc favicons
+app.mount("/static", StaticFiles(directory="static/dream_web"), name="static")
+
+
+@app.get("/docs", include_in_schema=False)
+def overridden_swagger():
+    return get_swagger_ui_html(
+        openapi_url=app.openapi_url,
+        title=app.title,
+        swagger_favicon_url="/static/favicon.ico",
+    )
+
+
+@app.get("/redoc", include_in_schema=False)
+def overridden_redoc():
+    return get_redoc_html(
+        openapi_url=app.openapi_url,
+        title=app.title,
+        redoc_favicon_url="/static/favicon.ico",
+    )
+
+
+def invoke_api():
+    # Start our own event loop for eventing usage
+    # TODO: determine if there's a better way to do this
+    loop = asyncio.new_event_loop()
+    config = uvicorn.Config(app=app, host="0.0.0.0", port=9090, loop=loop)
+    # Use access_log to turn off logging
+
+    server = uvicorn.Server(config)
+    loop.run_until_complete(server.serve())
+
+
+if __name__ == "__main__":
+    invoke_api()

invokeai/app/cli/commands.py

@@ -0,0 +1,202 @@
+# Copyright (c) 2023 Kyle Schouviller (https://github.com/kyle0654)
+
+from abc import ABC, abstractmethod
+import argparse
+from typing import Any, Callable, Iterable, Literal, get_args, get_origin, get_type_hints
+from pydantic import BaseModel, Field
+
+from ..invocations.image import ImageField
+from ..services.graph import GraphExecutionState
+from ..services.invoker import Invoker
+
+
+def add_parsers(
+    subparsers,
+    commands: list[type],
+    command_field: str = "type",
+    exclude_fields: list[str] = ["id", "type"],
+    add_arguments: Callable[[argparse.ArgumentParser], None]|None = None
+    ):
+    """Adds parsers for each command to the subparsers"""
+
+    # Create subparsers for each command
+    for command in commands:
+        hints = get_type_hints(command)
+        cmd_name = get_args(hints[command_field])[0]
+        command_parser = subparsers.add_parser(cmd_name, help=command.__doc__)
+
+        if add_arguments is not None:
+            add_arguments(command_parser)
+
+        # Convert all fields to arguments
+        fields = command.__fields__ # type: ignore
+        for name, field in fields.items():
+            if name in exclude_fields:
+                continue
+
+            if get_origin(field.type_) == Literal:
+                allowed_values = get_args(field.type_)
+                allowed_types = set()
+                for val in allowed_values:
+                    allowed_types.add(type(val))
+                allowed_types_list = list(allowed_types)
+                field_type = allowed_types_list[0] if len(allowed_types) == 1 else Union[allowed_types_list]  # type: ignore
+
+                command_parser.add_argument(
+                    f"--{name}",
+                    dest=name,
+                    type=field_type,
+                    default=field.default,
+                    choices=allowed_values,
+                    help=field.field_info.description,
+                )
+            else:
+                command_parser.add_argument(
+                    f"--{name}",
+                    dest=name,
+                    type=field.type_,
+                    default=field.default,
+                    help=field.field_info.description,
+                )
+
+
+class CliContext:
+    invoker: Invoker
+    session: GraphExecutionState
+    parser: argparse.ArgumentParser
+    defaults: dict[str, Any]
+
+    def __init__(self, invoker: Invoker, session: GraphExecutionState, parser: argparse.ArgumentParser):
+        self.invoker = invoker
+        self.session = session
+        self.parser = parser
+        self.defaults = dict()
+
+    def get_session(self):
+        self.session = self.invoker.services.graph_execution_manager.get(self.session.id)
+        return self.session
+
+
+class ExitCli(Exception):
+    """Exception to exit the CLI"""
+    pass
+
+
+class BaseCommand(ABC, BaseModel):
+    """A CLI command"""
+
+    # All commands must include a type name like this:
+    # type: Literal['your_command_name'] = 'your_command_name'
+
+    @classmethod
+    def get_all_subclasses(cls):
+        subclasses = []
+        toprocess = [cls]
+        while len(toprocess) > 0:
+            next = toprocess.pop(0)
+            next_subclasses = next.__subclasses__()
+            subclasses.extend(next_subclasses)
+            toprocess.extend(next_subclasses)
+        return subclasses
+
+    @classmethod
+    def get_commands(cls):
+        return tuple(BaseCommand.get_all_subclasses())
+
+    @classmethod
+    def get_commands_map(cls):
+        # Get the type strings out of the literals and into a dictionary
+        return dict(map(lambda t: (get_args(get_type_hints(t)['type'])[0], t),BaseCommand.get_all_subclasses()))
+
+    @abstractmethod
+    def run(self, context: CliContext) -> None:
+        """Run the command. Raise ExitCli to exit."""
+        pass
+
+
+class ExitCommand(BaseCommand):
+    """Exits the CLI"""
+    type: Literal['exit'] = 'exit'
+
+    def run(self, context: CliContext) -> None:
+        raise ExitCli()
+
+
+class HelpCommand(BaseCommand):
+    """Shows help"""
+    type: Literal['help'] = 'help'
+
+    def run(self, context: CliContext) -> None:
+        context.parser.print_help()
+
+
+def get_graph_execution_history(
+    graph_execution_state: GraphExecutionState,
+) -> Iterable[str]:
+    """Gets the history of fully-executed invocations for a graph execution"""
+    return (
+        n
+        for n in reversed(graph_execution_state.executed_history)
+        if n in graph_execution_state.graph.nodes
+    )
+
+
+def get_invocation_command(invocation) -> str:
+    fields = invocation.__fields__.items()
+    type_hints = get_type_hints(type(invocation))
+    command = [invocation.type]
+    for name, field in fields:
+        if name in ["id", "type"]:
+            continue
+
+        # TODO: add links
+
+        # Skip image fields when serializing command
+        type_hint = type_hints.get(name) or None
+        if type_hint is ImageField or ImageField in get_args(type_hint):
+            continue
+
+        field_value = getattr(invocation, name)
+        field_default = field.default
+        if field_value != field_default:
+            if type_hint is str or str in get_args(type_hint):
+                command.append(f'--{name} "{field_value}"')
+            else:
+                command.append(f"--{name} {field_value}")
+
+    return " ".join(command)
+
+
+class HistoryCommand(BaseCommand):
+    """Shows the invocation history"""
+    type: Literal['history'] = 'history'
+
+    # Inputs
+    # fmt: off
+    count: int = Field(default=5, gt=0, description="The number of history entries to show")
+    # fmt: on
+
+    def run(self, context: CliContext) -> None:
+        history = list(get_graph_execution_history(context.get_session()))
+        for i in range(min(self.count, len(history))):
+            entry_id = history[-1 - i]
+            entry = context.get_session().graph.get_node(entry_id)
+            print(f"{entry_id}: {get_invocation_command(entry)}")
+
+
+class SetDefaultCommand(BaseCommand):
+    """Sets a default value for a field"""
+    type: Literal['default'] = 'default'
+
+    # Inputs
+    # fmt: off
+    field: str = Field(description="The field to set the default for")
+    value: str = Field(description="The value to set the default to, or None to clear the default")
+    # fmt: on
+
+    def run(self, context: CliContext) -> None:
+        if self.value is None:
+            if self.field in context.defaults:
+                del context.defaults[self.field]
+        else:
+            context.defaults[self.field] = self.value

invokeai/app/cli_app.py

@@ -0,0 +1,275 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+import argparse
+import os
+import shlex
+import time
+from typing import (
+    Union,
+    get_type_hints,
+)
+
+from pydantic import BaseModel
+from pydantic.fields import Field
+
+from ..backend import Args
+from .cli.commands import BaseCommand, CliContext, ExitCli, add_parsers, get_graph_execution_history
+from .invocations import *
+from .invocations.baseinvocation import BaseInvocation
+from .services.events import EventServiceBase
+from .services.model_manager_initializer import get_model_manager
+from .services.restoration_services import RestorationServices
+from .services.graph import Edge, EdgeConnection, GraphExecutionState
+from .services.image_storage import DiskImageStorage
+from .services.invocation_queue import MemoryInvocationQueue
+from .services.invocation_services import InvocationServices
+from .services.invoker import Invoker
+from .services.processor import DefaultInvocationProcessor
+from .services.sqlite import SqliteItemStorage
+
+
+class CliCommand(BaseModel):
+    command: Union[BaseCommand.get_commands() + BaseInvocation.get_invocations()] = Field(discriminator="type")  # type: ignore
+
+
+class InvalidArgs(Exception):
+    pass
+
+
+def add_invocation_args(command_parser):
+    # Add linking capability
+    command_parser.add_argument(
+        "--link",
+        "-l",
+        action="append",
+        nargs=3,
+        help="A link in the format 'dest_field source_node source_field'. source_node can be relative to history (e.g. -1)",
+    )
+
+    command_parser.add_argument(
+        "--link_node",
+        "-ln",
+        action="append",
+        help="A link from all fields in the specified node. Node can be relative to history (e.g. -1)",
+    )
+
+
+def get_command_parser() -> argparse.ArgumentParser:
+    # Create invocation parser
+    parser = argparse.ArgumentParser()
+
+    def exit(*args, **kwargs):
+        raise InvalidArgs
+
+    parser.exit = exit
+    subparsers = parser.add_subparsers(dest="type")
+
+    # Create subparsers for each invocation
+    invocations = BaseInvocation.get_all_subclasses()
+    add_parsers(subparsers, invocations, add_arguments=add_invocation_args)
+
+    # Create subparsers for each command
+    commands = BaseCommand.get_all_subclasses()
+    add_parsers(subparsers, commands, exclude_fields=["type"])
+
+    return parser
+
+
+def generate_matching_edges(
+    a: BaseInvocation, b: BaseInvocation
+) -> list[Edge]:
+    """Generates all possible edges between two invocations"""
+    atype = type(a)
+    btype = type(b)
+
+    aoutputtype = atype.get_output_type()
+
+    afields = get_type_hints(aoutputtype)
+    bfields = get_type_hints(btype)
+
+    matching_fields = set(afields.keys()).intersection(bfields.keys())
+
+    # Remove invalid fields
+    invalid_fields = set(["type", "id"])
+    matching_fields = matching_fields.difference(invalid_fields)
+
+    edges = [
+        Edge(
+            source=EdgeConnection(node_id=a.id, field=field),
+            destination=EdgeConnection(node_id=b.id, field=field)
+        )
+        for field in matching_fields
+    ]
+    return edges
+
+
+class SessionError(Exception):
+    """Raised when a session error has occurred"""
+    pass
+
+
+def invoke_all(context: CliContext):
+    """Runs all invocations in the specified session"""
+    context.invoker.invoke(context.session, invoke_all=True)
+    while not context.get_session().is_complete():
+        # Wait some time
+        time.sleep(0.1)
+
+    # Print any errors
+    if context.session.has_error():
+        for n in context.session.errors:
+            print(
+                f"Error in node {n} (source node {context.session.prepared_source_mapping[n]}): {context.session.errors[n]}"
+            )
+        
+        raise SessionError()
+
+
+def invoke_cli():
+    config = Args()
+    config.parse_args()
+    model_manager = get_model_manager(config)
+
+    events = EventServiceBase()
+
+    output_folder = os.path.abspath(
+        os.path.join(os.path.dirname(__file__), "../../../outputs")
+    )
+
+    # TODO: build a file/path manager?
+    db_location = os.path.join(output_folder, "invokeai.db")
+
+    services = InvocationServices(
+        model_manager=model_manager,
+        events=events,
+        images=DiskImageStorage(output_folder),
+        queue=MemoryInvocationQueue(),
+        graph_execution_manager=SqliteItemStorage[GraphExecutionState](
+            filename=db_location, table_name="graph_executions"
+        ),
+        processor=DefaultInvocationProcessor(),
+        restoration=RestorationServices(config),
+    )
+
+    invoker = Invoker(services)
+    session: GraphExecutionState = invoker.create_execution_state()
+    parser = get_command_parser()
+
+    # Uncomment to print out previous sessions at startup
+    # print(services.session_manager.list())
+
+    context = CliContext(invoker, session, parser)
+
+    while True:
+        try:
+            cmd_input = input("> ")
+        except KeyboardInterrupt:
+            # Ctrl-c exits
+            break
+
+        try:
+            # Refresh the state of the session
+            history = list(get_graph_execution_history(context.session))
+
+            # Split the command for piping
+            cmds = cmd_input.split("|")
+            start_id = len(history)
+            current_id = start_id
+            new_invocations = list()
+            for cmd in cmds:
+                if cmd is None or cmd.strip() == "":
+                    raise InvalidArgs("Empty command")
+
+                # Parse args to create invocation
+                args = vars(context.parser.parse_args(shlex.split(cmd.strip())))
+
+                # Override defaults
+                for field_name, field_default in context.defaults.items():
+                    if field_name in args:
+                        args[field_name] = field_default
+
+                # Parse invocation
+                args["id"] = current_id
+                command = CliCommand(command=args)
+
+                # Run any CLI commands immediately
+                if isinstance(command.command, BaseCommand):
+                    # Invoke all current nodes to preserve operation order
+                    invoke_all(context)
+
+                    # Run the command
+                    command.command.run(context)
+                    continue
+
+                # Pipe previous command output (if there was a previous command)
+                edges: list[Edge] = list()
+                if len(history) > 0 or current_id != start_id:
+                    from_id = (
+                        history[0] if current_id == start_id else str(current_id - 1)
+                    )
+                    from_node = (
+                        next(filter(lambda n: n[0].id == from_id, new_invocations))[0]
+                        if current_id != start_id
+                        else context.session.graph.get_node(from_id)
+                    )
+                    matching_edges = generate_matching_edges(
+                        from_node, command.command
+                    )
+                    edges.extend(matching_edges)
+
+                # Parse provided links
+                if "link_node" in args and args["link_node"]:
+                    for link in args["link_node"]:
+                        link_node = context.session.graph.get_node(link)
+                        matching_edges = generate_matching_edges(
+                            link_node, command.command
+                        )
+                        matching_destinations = [e.destination for e in matching_edges]
+                        edges = [e for e in edges if e.destination not in matching_destinations]
+                        edges.extend(matching_edges)
+
+                if "link" in args and args["link"]:
+                    for link in args["link"]:
+                        edges = [e for e in edges if e.destination.node_id != command.command.id and e.destination.field != link[2]]
+                        edges.append(
+                            Edge(
+                                source=EdgeConnection(node_id=link[1], field=link[0]),
+                                destination=EdgeConnection(
+                                    node_id=command.command.id, field=link[2]
+                                )
+                            )
+                        )
+
+                new_invocations.append((command.command, edges))
+
+                current_id = current_id + 1
+
+                # Add the node to the session
+                context.session.add_node(command.command)
+                for edge in edges:
+                    print(edge)
+                    context.session.add_edge(edge)
+
+            # Execute all remaining nodes
+            invoke_all(context)
+
+        except InvalidArgs:
+            print('Invalid command, use "help" to list commands')
+            continue
+
+        except SessionError:
+            # Start a new session
+            print("Session error: creating a new session")
+            context.session = context.invoker.create_execution_state()
+
+        except ExitCli:
+            break
+
+        except SystemExit:
+            continue
+
+    invoker.stop()
+
+
+if __name__ == "__main__":
+    invoke_cli()

invokeai/app/invocations/__init__.py

@@ -0,0 +1,12 @@
+import os
+
+__all__ = []
+
+dirname = os.path.dirname(os.path.abspath(__file__))
+for f in os.listdir(dirname):
+    if (
+        f != "__init__.py"
+        and os.path.isfile("%s/%s" % (dirname, f))
+        and f[-3:] == ".py"
+    ):
+        __all__.append(f[:-3])

invokeai/app/invocations/baseinvocation.py

@@ -0,0 +1,78 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from abc import ABC, abstractmethod
+from inspect import signature
+from typing import get_args, get_type_hints
+
+from pydantic import BaseModel, Field
+
+from ..services.invocation_services import InvocationServices
+
+
+class InvocationContext:
+    services: InvocationServices
+    graph_execution_state_id: str
+
+    def __init__(self, services: InvocationServices, graph_execution_state_id: str):
+        self.services = services
+        self.graph_execution_state_id = graph_execution_state_id
+
+
+class BaseInvocationOutput(BaseModel):
+    """Base class for all invocation outputs"""
+
+    # All outputs must include a type name like this:
+    # type: Literal['your_output_name']
+
+    @classmethod
+    def get_all_subclasses_tuple(cls):
+        subclasses = []
+        toprocess = [cls]
+        while len(toprocess) > 0:
+            next = toprocess.pop(0)
+            next_subclasses = next.__subclasses__()
+            subclasses.extend(next_subclasses)
+            toprocess.extend(next_subclasses)
+        return tuple(subclasses)
+
+
+class BaseInvocation(ABC, BaseModel):
+    """A node to process inputs and produce outputs.
+    May use dependency injection in __init__ to receive providers.
+    """
+
+    # All invocations must include a type name like this:
+    # type: Literal['your_output_name']
+
+    @classmethod
+    def get_all_subclasses(cls):
+        subclasses = []
+        toprocess = [cls]
+        while len(toprocess) > 0:
+            next = toprocess.pop(0)
+            next_subclasses = next.__subclasses__()
+            subclasses.extend(next_subclasses)
+            toprocess.extend(next_subclasses)
+        return subclasses
+
+    @classmethod
+    def get_invocations(cls):
+        return tuple(BaseInvocation.get_all_subclasses())
+
+    @classmethod
+    def get_invocations_map(cls):
+        # Get the type strings out of the literals and into a dictionary
+        return dict(map(lambda t: (get_args(get_type_hints(t)['type'])[0], t),BaseInvocation.get_all_subclasses()))
+    
+    @classmethod
+    def get_output_type(cls):
+        return signature(cls.invoke).return_annotation
+
+    @abstractmethod
+    def invoke(self, context: InvocationContext) -> BaseInvocationOutput:
+        """Invoke with provided context and return outputs."""
+        pass
+    
+    #fmt: off
+    id: str = Field(description="The id of this node. Must be unique among all nodes.")
+    #fmt: on

invokeai/app/invocations/cv.py

@@ -0,0 +1,50 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from typing import Literal
+
+import cv2 as cv
+import numpy
+from PIL import Image, ImageOps
+from pydantic import Field
+
+from ..services.image_storage import ImageType
+from .baseinvocation import BaseInvocation, InvocationContext
+from .image import ImageField, ImageOutput
+
+
+class CvInpaintInvocation(BaseInvocation):
+    """Simple inpaint using opencv."""
+    #fmt: off
+    type: Literal["cv_inpaint"] = "cv_inpaint"
+
+    # Inputs
+    image: ImageField = Field(default=None, description="The image to inpaint")
+    mask: ImageField = Field(default=None, description="The mask to use when inpainting")
+    #fmt: on
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get(
+            self.image.image_type, self.image.image_name
+        )
+        mask = context.services.images.get(self.mask.image_type, self.mask.image_name)
+
+        # Convert to cv image/mask
+        # TODO: consider making these utility functions
+        cv_image = cv.cvtColor(numpy.array(image.convert("RGB")), cv.COLOR_RGB2BGR)
+        cv_mask = numpy.array(ImageOps.invert(mask))
+
+        # Inpaint
+        cv_inpainted = cv.inpaint(cv_image, cv_mask, 3, cv.INPAINT_TELEA)
+
+        # Convert back to Pillow
+        # TODO: consider making a utility function
+        image_inpainted = Image.fromarray(cv.cvtColor(cv_inpainted, cv.COLOR_BGR2RGB))
+
+        image_type = ImageType.INTERMEDIATE
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, image_inpainted)
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )

invokeai/app/invocations/generate.py

@@ -0,0 +1,221 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from datetime import datetime, timezone
+from typing import Any, Literal, Optional, Union
+
+import numpy as np
+
+from torch import Tensor
+from PIL import Image
+from pydantic import Field
+from skimage.exposure.histogram_matching import match_histograms
+
+from ..services.image_storage import ImageType
+from ..services.invocation_services import InvocationServices
+from .baseinvocation import BaseInvocation, InvocationContext
+from .image import ImageField, ImageOutput
+from ...backend.generator import Txt2Img, Img2Img, Inpaint, InvokeAIGenerator, Generator
+from ...backend.stable_diffusion import PipelineIntermediateState
+from ...backend.util.util import image_to_dataURL
+
+SAMPLER_NAME_VALUES = Literal[
+    tuple(InvokeAIGenerator.schedulers())
+]
+
+# Text to image
+class TextToImageInvocation(BaseInvocation):
+    """Generates an image using text2img."""
+
+    type: Literal["txt2img"] = "txt2img"
+
+    # Inputs
+    # TODO: consider making prompt optional to enable providing prompt through a link
+    # fmt: off
+    prompt: Optional[str] = Field(description="The prompt to generate an image from")
+    seed:        int = Field(default=-1,ge=-1, le=np.iinfo(np.uint32).max, description="The seed to use (-1 for a random seed)", )
+    steps:       int = Field(default=10, gt=0, description="The number of steps to use to generate the image")
+    width:       int = Field(default=512, multiple_of=64, gt=0, description="The width of the resulting image", )
+    height:      int = Field(default=512, multiple_of=64, gt=0, description="The height of the resulting image", )
+    cfg_scale: float = Field(default=7.5, gt=0, description="The Classifier-Free Guidance, higher values may result in a result closer to the prompt", )
+    sampler_name: SAMPLER_NAME_VALUES = Field(default="k_lms", description="The sampler to use" )
+    seamless:   bool = Field(default=False, description="Whether or not to generate an image that can tile without seams", )
+    model:       str = Field(default="", description="The model to use (currently ignored)")
+    progress_images: bool = Field(default=False, description="Whether or not to produce progress images during generation",  )
+    # fmt: on
+
+    # TODO: pass this an emitter method or something? or a session for dispatching?
+    def dispatch_progress(
+        self, context: InvocationContext, sample: Tensor, step: int
+    ) -> None:  
+        # TODO: only output a preview image when requested
+        image = Generator.sample_to_lowres_estimated_image(sample)
+
+        (width, height) = image.size
+        width *= 8
+        height *= 8
+
+        dataURL = image_to_dataURL(image, image_format="JPEG")
+
+        context.services.events.emit_generator_progress(
+            context.graph_execution_state_id,
+            self.id,
+            {
+                "width": width,
+                "height": height,
+                "dataURL": dataURL
+            },
+            step,
+            self.steps,
+        )
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        def step_callback(state: PipelineIntermediateState):
+            self.dispatch_progress(context, state.latents, state.step)
+
+        # Handle invalid model parameter
+        # TODO: figure out if this can be done via a validator that uses the model_cache
+        # TODO: How to get the default model name now?
+        #       (right now uses whatever current model is set in model manager)
+        model= context.services.model_manager.get_model()
+        outputs = Txt2Img(model).generate(
+            prompt=self.prompt,
+            step_callback=step_callback,
+            **self.dict(
+                exclude={"prompt"}
+            ),  # Shorthand for passing all of the parameters above manually
+        )
+        # Outputs is an infinite iterator that will return a new InvokeAIGeneratorOutput object
+        # each time it is called. We only need the first one.
+        generate_output = next(outputs)
+
+        # Results are image and seed, unwrap for now and ignore the seed
+        # TODO: pre-seed?
+        # TODO: can this return multiple results? Should it?
+        image_type = ImageType.RESULT
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, generate_output.image)
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )
+
+
+class ImageToImageInvocation(TextToImageInvocation):
+    """Generates an image using img2img."""
+
+    type: Literal["img2img"] = "img2img"
+
+    # Inputs
+    image: Union[ImageField, None] = Field(description="The input image")
+    strength: float = Field(
+        default=0.75, gt=0, le=1, description="The strength of the original image"
+    )
+    fit: bool = Field(
+        default=True,
+        description="Whether or not the result should be fit to the aspect ratio of the input image",
+    )
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = (
+            None
+            if self.image is None
+            else context.services.images.get(
+                self.image.image_type, self.image.image_name
+            )
+        )
+        mask = None
+
+        def step_callback(sample, step=0):
+            self.dispatch_progress(context, sample, step)
+
+        # Handle invalid model parameter
+        # TODO: figure out if this can be done via a validator that uses the model_cache
+        # TODO: How to get the default model name now?
+        model = context.services.model_manager.get_model()
+        generator_output = next(
+            Img2Img(model).generate(
+                prompt=self.prompt,
+                init_image=image,
+                init_mask=mask,
+                step_callback=step_callback,
+                **self.dict(
+                    exclude={"prompt", "image", "mask"}
+                ),  # Shorthand for passing all of the parameters above manually
+            )
+        )
+
+        result_image = generator_output.image
+
+        # Results are image and seed, unwrap for now and ignore the seed
+        # TODO: pre-seed?
+        # TODO: can this return multiple results? Should it?
+        image_type = ImageType.RESULT
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, result_image)
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )
+
+class InpaintInvocation(ImageToImageInvocation):
+    """Generates an image using inpaint."""
+
+    type: Literal["inpaint"] = "inpaint"
+
+    # Inputs
+    mask: Union[ImageField, None] = Field(description="The mask")
+    inpaint_replace: float = Field(
+        default=0.0,
+        ge=0.0,
+        le=1.0,
+        description="The amount by which to replace masked areas with latent noise",
+    )
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = (
+            None
+            if self.image is None
+            else context.services.images.get(
+                self.image.image_type, self.image.image_name
+            )
+        )
+        mask = (
+            None
+            if self.mask is None
+            else context.services.images.get(self.mask.image_type, self.mask.image_name)
+        )
+
+        def step_callback(sample, step=0):
+            self.dispatch_progress(context, sample, step)
+
+        # Handle invalid model parameter
+        # TODO: figure out if this can be done via a validator that uses the model_cache
+        # TODO: How to get the default model name now?
+        manager = context.services.model_manager.get_model()
+        generator_output = next(
+            Inpaint(model).generate(
+                prompt=self.prompt,
+                init_image=image,
+                mask_image=mask,
+                step_callback=step_callback,
+                **self.dict(
+                    exclude={"prompt", "image", "mask"}
+                ),  # Shorthand for passing all of the parameters above manually
+            )
+        )
+
+        result_image = generator_output.image
+
+        # Results are image and seed, unwrap for now and ignore the seed
+        # TODO: pre-seed?
+        # TODO: can this return multiple results? Should it?
+        image_type = ImageType.RESULT
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, result_image)
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )

invokeai/app/invocations/image.py

@@ -0,0 +1,287 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from datetime import datetime, timezone
+from typing import Literal, Optional
+
+import numpy
+from PIL import Image, ImageFilter, ImageOps
+from pydantic import BaseModel, Field
+
+from ..services.image_storage import ImageType
+from ..services.invocation_services import InvocationServices
+from .baseinvocation import BaseInvocation, BaseInvocationOutput, InvocationContext
+
+
+class ImageField(BaseModel):
+    """An image field used for passing image objects between invocations"""
+
+    image_type: str = Field(
+        default=ImageType.RESULT, description="The type of the image"
+    )
+    image_name: Optional[str] = Field(default=None, description="The name of the image")
+
+
+class ImageOutput(BaseInvocationOutput):
+    """Base class for invocations that output an image"""
+    #fmt: off
+    type: Literal["image"] = "image"
+    image:      ImageField = Field(default=None, description="The output image")
+    #fmt: on
+
+class MaskOutput(BaseInvocationOutput):
+    """Base class for invocations that output a mask"""
+    #fmt: off
+    type: Literal["mask"] = "mask"
+    mask:      ImageField = Field(default=None, description="The output mask")
+    #fomt: on
+
+# TODO: this isn't really necessary anymore
+class LoadImageInvocation(BaseInvocation):
+    """Load an image from a filename and provide it as output."""
+    #fmt: off
+    type: Literal["load_image"] = "load_image"
+
+    # Inputs
+    image_type: ImageType = Field(description="The type of the image")
+    image_name:       str = Field(description="The name of the image")
+    #fmt: on
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        return ImageOutput(
+            image=ImageField(image_type=self.image_type, image_name=self.image_name)
+        )
+
+
+class ShowImageInvocation(BaseInvocation):
+    """Displays a provided image, and passes it forward in the pipeline."""
+
+    type: Literal["show_image"] = "show_image"
+
+    # Inputs
+    image: ImageField = Field(default=None, description="The image to show")
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get(
+            self.image.image_type, self.image.image_name
+        )
+        if image:
+            image.show()
+
+        # TODO: how to handle failure?
+
+        return ImageOutput(
+            image=ImageField(
+                image_type=self.image.image_type, image_name=self.image.image_name
+            )
+        )
+
+
+class CropImageInvocation(BaseInvocation):
+    """Crops an image to a specified box. The box can be outside of the image."""
+    #fmt: off
+    type: Literal["crop"] = "crop"
+
+    # Inputs
+    image: ImageField = Field(default=None, description="The image to crop")
+    x:      int = Field(default=0, description="The left x coordinate of the crop rectangle")
+    y:      int = Field(default=0, description="The top y coordinate of the crop rectangle")
+    width:  int = Field(default=512, gt=0, description="The width of the crop rectangle")
+    height: int = Field(default=512, gt=0, description="The height of the crop rectangle")
+    #fmt: on
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get(
+            self.image.image_type, self.image.image_name
+        )
+
+        image_crop = Image.new(
+            mode="RGBA", size=(self.width, self.height), color=(0, 0, 0, 0)
+        )
+        image_crop.paste(image, (-self.x, -self.y))
+
+        image_type = ImageType.INTERMEDIATE
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, image_crop)
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )
+
+
+class PasteImageInvocation(BaseInvocation):
+    """Pastes an image into another image."""
+    #fmt: off
+    type: Literal["paste"] = "paste"
+
+    # Inputs
+    base_image:     ImageField = Field(default=None, description="The base image")
+    image:          ImageField = Field(default=None, description="The image to paste")
+    mask: Optional[ImageField] = Field(default=None, description="The mask to use when pasting")
+    x:                     int = Field(default=0, description="The left x coordinate at which to paste the image")
+    y:                     int = Field(default=0, description="The top y coordinate at which to paste the image")
+    #fmt: on
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        base_image = context.services.images.get(
+            self.base_image.image_type, self.base_image.image_name
+        )
+        image = context.services.images.get(
+            self.image.image_type, self.image.image_name
+        )
+        mask = (
+            None
+            if self.mask is None
+            else ImageOps.invert(
+                services.images.get(self.mask.image_type, self.mask.image_name)
+            )
+        )
+        # TODO: probably shouldn't invert mask here... should user be required to do it?
+
+        min_x = min(0, self.x)
+        min_y = min(0, self.y)
+        max_x = max(base_image.width, image.width + self.x)
+        max_y = max(base_image.height, image.height + self.y)
+
+        new_image = Image.new(
+            mode="RGBA", size=(max_x - min_x, max_y - min_y), color=(0, 0, 0, 0)
+        )
+        new_image.paste(base_image, (abs(min_x), abs(min_y)))
+        new_image.paste(image, (max(0, self.x), max(0, self.y)), mask=mask)
+
+        image_type = ImageType.RESULT
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, new_image)
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )
+
+
+class MaskFromAlphaInvocation(BaseInvocation):
+    """Extracts the alpha channel of an image as a mask."""
+    #fmt: off
+    type: Literal["tomask"] = "tomask"
+
+    # Inputs
+    image: ImageField = Field(default=None, description="The image to create the mask from")
+    invert:      bool = Field(default=False, description="Whether or not to invert the mask")
+    #fmt: on
+
+    def invoke(self, context: InvocationContext) -> MaskOutput:
+        image = context.services.images.get(
+            self.image.image_type, self.image.image_name
+        )
+
+        image_mask = image.split()[-1]
+        if self.invert:
+            image_mask = ImageOps.invert(image_mask)
+
+        image_type = ImageType.INTERMEDIATE
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, image_mask)
+        return MaskOutput(mask=ImageField(image_type=image_type, image_name=image_name))
+
+
+class BlurInvocation(BaseInvocation):
+    """Blurs an image"""
+
+    #fmt: off
+    type: Literal["blur"] = "blur"
+
+    # Inputs
+    image: ImageField = Field(default=None, description="The image to blur")
+    radius:     float = Field(default=8.0, ge=0, description="The blur radius")
+    blur_type: Literal["gaussian", "box"] = Field(default="gaussian", description="The type of blur")
+    #fmt: on
+    
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get(
+            self.image.image_type, self.image.image_name
+        )
+
+        blur = (
+            ImageFilter.GaussianBlur(self.radius)
+            if self.blur_type == "gaussian"
+            else ImageFilter.BoxBlur(self.radius)
+        )
+        blur_image = image.filter(blur)
+
+        image_type = ImageType.INTERMEDIATE
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, blur_image)
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )
+
+
+class LerpInvocation(BaseInvocation):
+    """Linear interpolation of all pixels of an image"""
+    #fmt: off
+    type: Literal["lerp"] = "lerp"
+
+    # Inputs
+    image: ImageField = Field(default=None, description="The image to lerp")
+    min: int = Field(default=0, ge=0, le=255, description="The minimum output value")
+    max: int = Field(default=255, ge=0, le=255, description="The maximum output value")
+    #fmt: on
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get(
+            self.image.image_type, self.image.image_name
+        )
+
+        image_arr = numpy.asarray(image, dtype=numpy.float32) / 255
+        image_arr = image_arr * (self.max - self.min) + self.max
+
+        lerp_image = Image.fromarray(numpy.uint8(image_arr))
+
+        image_type = ImageType.INTERMEDIATE
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, lerp_image)
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )
+
+
+class InverseLerpInvocation(BaseInvocation):
+    """Inverse linear interpolation of all pixels of an image"""
+    #fmt: off
+    type: Literal["ilerp"] = "ilerp"
+
+    # Inputs
+    image: ImageField = Field(default=None, description="The image to lerp")
+    min: int = Field(default=0, ge=0, le=255, description="The minimum input value")
+    max: int = Field(default=255, ge=0, le=255, description="The maximum input value")
+    #fmt: on
+    
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get(
+            self.image.image_type, self.image.image_name
+        )
+
+        image_arr = numpy.asarray(image, dtype=numpy.float32)
+        image_arr = (
+            numpy.minimum(
+                numpy.maximum(image_arr - self.min, 0) / float(self.max - self.min), 1
+            )
+            * 255
+        )
+
+        ilerp_image = Image.fromarray(numpy.uint8(image_arr))
+
+        image_type = ImageType.INTERMEDIATE
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, ilerp_image)
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )

invokeai/app/invocations/prompt.py

@@ -0,0 +1,14 @@
+from typing import Literal
+
+from pydantic.fields import Field
+
+from .baseinvocation import BaseInvocationOutput
+
+
+class PromptOutput(BaseInvocationOutput):
+    """Base class for invocations that output a prompt"""
+    #fmt: off
+    type: Literal["prompt"] = "prompt"
+
+    prompt: str = Field(default=None, description="The output prompt")
+    #fmt: on

invokeai/app/invocations/reconstruct.py

@@ -0,0 +1,42 @@
+from datetime import datetime, timezone
+from typing import Literal, Union
+
+from pydantic import Field
+
+from ..services.image_storage import ImageType
+from ..services.invocation_services import InvocationServices
+from .baseinvocation import BaseInvocation, InvocationContext
+from .image import ImageField, ImageOutput
+
+class RestoreFaceInvocation(BaseInvocation):
+    """Restores faces in an image."""
+    #fmt: off
+    type:  Literal["restore_face"] = "restore_face"
+
+    # Inputs
+    image: Union[ImageField, None] = Field(description="The input image")
+    strength:                float = Field(default=0.75, gt=0, le=1, description="The strength of the restoration"  )
+    #fmt: on
+    
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get(
+            self.image.image_type, self.image.image_name
+        )
+        results = context.services.restoration.upscale_and_reconstruct(
+            image_list=[[image, 0]],
+            upscale=None,
+            strength=self.strength,  # GFPGAN strength
+            save_original=False,
+            image_callback=None,
+        )
+
+        # Results are image and seed, unwrap for now
+        # TODO: can this return multiple results?
+        image_type = ImageType.RESULT
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, results[0][0])
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )

invokeai/app/invocations/upscale.py

@@ -0,0 +1,46 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from datetime import datetime, timezone
+from typing import Literal, Union
+
+from pydantic import Field
+
+from ..services.image_storage import ImageType
+from ..services.invocation_services import InvocationServices
+from .baseinvocation import BaseInvocation, InvocationContext
+from .image import ImageField, ImageOutput
+
+
+class UpscaleInvocation(BaseInvocation):
+    """Upscales an image."""
+    #fmt: off
+    type: Literal["upscale"] = "upscale"
+
+    # Inputs
+    image: Union[ImageField, None] = Field(description="The input image", default=None)
+    strength: float = Field(default=0.75, gt=0, le=1, description="The strength")
+    level: Literal[2, 4] = Field(default=2, description="The upscale level")
+    #fmt: on
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get(
+            self.image.image_type, self.image.image_name
+        )
+        results = context.services.restoration.upscale_and_reconstruct(
+            image_list=[[image, 0]],
+            upscale=(self.level, self.strength),
+            strength=0.0,  # GFPGAN strength
+            save_original=False,
+            image_callback=None,
+        )
+
+        # Results are image and seed, unwrap for now
+        # TODO: can this return multiple results?
+        image_type = ImageType.RESULT
+        image_name = context.services.images.create_name(
+            context.graph_execution_state_id, self.id
+        )
+        context.services.images.save(image_type, image_name, results[0][0])
+        return ImageOutput(
+            image=ImageField(image_type=image_type, image_name=image_name)
+        )

invokeai/app/services/events.py

@@ -0,0 +1,88 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from typing import Any, Dict, TypedDict
+
+ProgressImage = TypedDict(
+    "ProgressImage", {"dataURL": str, "width": int, "height": int}
+)
+
+class EventServiceBase:
+    session_event: str = "session_event"
+
+    """Basic event bus, to have an empty stand-in when not needed"""
+
+    def dispatch(self, event_name: str, payload: Any) -> None:
+        pass
+
+    def __emit_session_event(self, event_name: str, payload: Dict) -> None:
+        self.dispatch(
+            event_name=EventServiceBase.session_event,
+            payload=dict(event=event_name, data=payload),
+        )
+
+    # Define events here for every event in the system.
+    # This will make them easier to integrate until we find a schema generator.
+    def emit_generator_progress(
+        self,
+        graph_execution_state_id: str,
+        invocation_id: str,
+        progress_image: ProgressImage | None,
+        step: int,
+        total_steps: int,
+    ) -> None:
+        """Emitted when there is generation progress"""
+        self.__emit_session_event(
+            event_name="generator_progress",
+            payload=dict(
+                graph_execution_state_id=graph_execution_state_id,
+                invocation_id=invocation_id,
+                progress_image=progress_image,
+                step=step,
+                total_steps=total_steps,
+            ),
+        )
+
+    def emit_invocation_complete(
+        self, graph_execution_state_id: str, invocation_id: str, result: Dict
+    ) -> None:
+        """Emitted when an invocation has completed"""
+        self.__emit_session_event(
+            event_name="invocation_complete",
+            payload=dict(
+                graph_execution_state_id=graph_execution_state_id,
+                invocation_id=invocation_id,
+                result=result,
+            ),
+        )
+
+    def emit_invocation_error(
+        self, graph_execution_state_id: str, invocation_id: str, error: str
+    ) -> None:
+        """Emitted when an invocation has completed"""
+        self.__emit_session_event(
+            event_name="invocation_error",
+            payload=dict(
+                graph_execution_state_id=graph_execution_state_id,
+                invocation_id=invocation_id,
+                error=error,
+            ),
+        )
+
+    def emit_invocation_started(
+        self, graph_execution_state_id: str, invocation_id: str
+    ) -> None:
+        """Emitted when an invocation has started"""
+        self.__emit_session_event(
+            event_name="invocation_started",
+            payload=dict(
+                graph_execution_state_id=graph_execution_state_id,
+                invocation_id=invocation_id,
+            ),
+        )
+
+    def emit_graph_execution_complete(self, graph_execution_state_id: str) -> None:
+        """Emitted when a session has completed all invocations"""
+        self.__emit_session_event(
+            event_name="graph_execution_state_complete",
+            payload=dict(graph_execution_state_id=graph_execution_state_id),
+        )

invokeai/app/services/image_storage.py

@@ -0,0 +1,113 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+import datetime
+import os
+from abc import ABC, abstractmethod
+from enum import Enum
+from pathlib import Path
+from queue import Queue
+from typing import Dict
+
+from PIL.Image import Image
+
+from invokeai.backend.image_util import PngWriter
+
+
+class ImageType(str, Enum):
+    RESULT = "results"
+    INTERMEDIATE = "intermediates"
+    UPLOAD = "uploads"
+
+
+class ImageStorageBase(ABC):
+    """Responsible for storing and retrieving images."""
+
+    @abstractmethod
+    def get(self, image_type: ImageType, image_name: str) -> Image:
+        pass
+
+    # TODO: make this a bit more flexible for e.g. cloud storage
+    @abstractmethod
+    def get_path(self, image_type: ImageType, image_name: str) -> str:
+        pass
+
+    @abstractmethod
+    def save(self, image_type: ImageType, image_name: str, image: Image) -> None:
+        pass
+
+    @abstractmethod
+    def delete(self, image_type: ImageType, image_name: str) -> None:
+        pass
+
+    def create_name(self, context_id: str, node_id: str) -> str:
+        return f"{context_id}_{node_id}_{str(int(datetime.datetime.now(datetime.timezone.utc).timestamp()))}.png"
+
+
+class DiskImageStorage(ImageStorageBase):
+    """Stores images on disk"""
+
+    __output_folder: str
+    __pngWriter: PngWriter
+    __cache_ids: Queue  # TODO: this is an incredibly naive cache
+    __cache: Dict[str, Image]
+    __max_cache_size: int
+
+    def __init__(self, output_folder: str):
+        self.__output_folder = output_folder
+        self.__pngWriter = PngWriter(output_folder)
+        self.__cache = dict()
+        self.__cache_ids = Queue()
+        self.__max_cache_size = 10  # TODO: get this from config
+
+        Path(output_folder).mkdir(parents=True, exist_ok=True)
+
+        # TODO: don't hard-code. get/save/delete should maybe take subpath?
+        for image_type in ImageType:
+            Path(os.path.join(output_folder, image_type)).mkdir(
+                parents=True, exist_ok=True
+            )
+
+    def get(self, image_type: ImageType, image_name: str) -> Image:
+        image_path = self.get_path(image_type, image_name)
+        cache_item = self.__get_cache(image_path)
+        if cache_item:
+            return cache_item
+
+        image = Image.open(image_path)
+        self.__set_cache(image_path, image)
+        return image
+
+    # TODO: make this a bit more flexible for e.g. cloud storage
+    def get_path(self, image_type: ImageType, image_name: str) -> str:
+        path = os.path.join(self.__output_folder, image_type, image_name)
+        return path
+
+    def save(self, image_type: ImageType, image_name: str, image: Image) -> None:
+        image_subpath = os.path.join(image_type, image_name)
+        self.__pngWriter.save_image_and_prompt_to_png(
+            image, "", image_subpath, None
+        )  # TODO: just pass full path to png writer
+
+        image_path = self.get_path(image_type, image_name)
+        self.__set_cache(image_path, image)
+
+    def delete(self, image_type: ImageType, image_name: str) -> None:
+        image_path = self.get_path(image_type, image_name)
+        if os.path.exists(image_path):
+            os.remove(image_path)
+
+        if image_path in self.__cache:
+            del self.__cache[image_path]
+
+    def __get_cache(self, image_name: str) -> Image:
+        return None if image_name not in self.__cache else self.__cache[image_name]
+
+    def __set_cache(self, image_name: str, image: Image):
+        if not image_name in self.__cache:
+            self.__cache[image_name] = image
+            self.__cache_ids.put(
+                image_name
+            )  # TODO: this should refresh position for LRU cache
+            if len(self.__cache) > self.__max_cache_size:
+                cache_id = self.__cache_ids.get()
+                del self.__cache[cache_id]

invokeai/app/services/invocation_queue.py

@@ -0,0 +1,81 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from abc import ABC, abstractmethod
+from queue import Queue
+import time
+
+
+# TODO: make this serializable
+class InvocationQueueItem:
+    # session_id: str
+    graph_execution_state_id: str
+    invocation_id: str
+    invoke_all: bool
+    timestamp: float
+
+    def __init__(
+        self,
+        # session_id: str,
+        graph_execution_state_id: str,
+        invocation_id: str,
+        invoke_all: bool = False,
+    ):
+        # self.session_id = session_id
+        self.graph_execution_state_id = graph_execution_state_id
+        self.invocation_id = invocation_id
+        self.invoke_all = invoke_all
+        self.timestamp = time.time()
+
+
+class InvocationQueueABC(ABC):
+    """Abstract base class for all invocation queues"""
+
+    @abstractmethod
+    def get(self) -> InvocationQueueItem:
+        pass
+
+    @abstractmethod
+    def put(self, item: InvocationQueueItem | None) -> None:
+        pass
+
+    @abstractmethod
+    def cancel(self, graph_execution_state_id: str) -> None:
+        pass
+
+    @abstractmethod
+    def is_canceled(self, graph_execution_state_id: str) -> bool:
+        pass
+
+
+class MemoryInvocationQueue(InvocationQueueABC):
+    __queue: Queue
+    __cancellations: dict[str, float]
+
+    def __init__(self):
+        self.__queue = Queue()
+        self.__cancellations = dict()
+
+    def get(self) -> InvocationQueueItem:
+        item = self.__queue.get()
+
+        while isinstance(item, InvocationQueueItem) \
+            and item.graph_execution_state_id in self.__cancellations \
+            and self.__cancellations[item.graph_execution_state_id] > item.timestamp:
+            item = self.__queue.get()
+
+        # Clear old items
+        for graph_execution_state_id in list(self.__cancellations.keys()):
+            if self.__cancellations[graph_execution_state_id] < item.timestamp:
+                del self.__cancellations[graph_execution_state_id]
+
+        return item
+
+    def put(self, item: InvocationQueueItem | None) -> None:
+        self.__queue.put(item)
+
+    def cancel(self, graph_execution_state_id: str) -> None:
+        if graph_execution_state_id not in self.__cancellations:
+            self.__cancellations[graph_execution_state_id] = time.time()
+
+    def is_canceled(self, graph_execution_state_id: str) -> bool:
+        return graph_execution_state_id in self.__cancellations

invokeai/app/services/invocation_services.py

@@ -0,0 +1,39 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+from invokeai.backend import ModelManager
+
+from .events import EventServiceBase
+from .image_storage import ImageStorageBase
+from .restoration_services import RestorationServices
+from .invocation_queue import InvocationQueueABC
+from .item_storage import ItemStorageABC
+
+class InvocationServices:
+    """Services that can be used by invocations"""
+
+    events: EventServiceBase
+    images: ImageStorageBase
+    queue: InvocationQueueABC
+    model_manager: ModelManager
+    restoration: RestorationServices
+
+    # NOTE: we must forward-declare any types that include invocations, since invocations can use services
+    graph_execution_manager: ItemStorageABC["GraphExecutionState"]
+    processor: "InvocationProcessorABC"
+
+    def __init__(
+            self,
+            model_manager: ModelManager,
+            events: EventServiceBase,
+            images: ImageStorageBase,
+            queue: InvocationQueueABC,
+            graph_execution_manager: ItemStorageABC["GraphExecutionState"],
+            processor: "InvocationProcessorABC",
+            restoration: RestorationServices,
+    ):
+        self.model_manager = model_manager
+        self.events = events
+        self.images = images
+        self.queue = queue
+        self.graph_execution_manager = graph_execution_manager
+        self.processor = processor
+        self.restoration = restoration

invokeai/app/services/invoker.py

@@ -0,0 +1,91 @@
+# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
+
+from abc import ABC
+from threading import Event, Thread
+
+from ..invocations.baseinvocation import InvocationContext
+from .graph import Graph, GraphExecutionState
+from .invocation_queue import InvocationQueueABC, InvocationQueueItem
+from .invocation_services import InvocationServices
+from .item_storage import ItemStorageABC
+
+
+class Invoker:
+    """The invoker, used to execute invocations"""
+
+    services: InvocationServices
+
+    def __init__(self, services: InvocationServices):
+        self.services = services
+        self._start()
+
+    def invoke(
+        self, graph_execution_state: GraphExecutionState, invoke_all: bool = False
+    ) -> str | None:
+        """Determines the next node to invoke and returns the id of the invoked node, or None if there are no nodes to execute"""
+
+        # Get the next invocation
+        invocation = graph_execution_state.next()
+        if not invocation:
+            return None
+
+        # Save the execution state
+        self.services.graph_execution_manager.set(graph_execution_state)
+
+        # Queue the invocation
+        print(f"queueing item {invocation.id}")
+        self.services.queue.put(
+            InvocationQueueItem(
+                # session_id    = session.id,
+                graph_execution_state_id=graph_execution_state.id,
+                invocation_id=invocation.id,
+                invoke_all=invoke_all,
+            )
+        )
+
+        return invocation.id
+
+    def create_execution_state(self, graph: Graph | None = None) -> GraphExecutionState:
+        """Creates a new execution state for the given graph"""
+        new_state = GraphExecutionState(graph=Graph() if graph is None else graph)
+        self.services.graph_execution_manager.set(new_state)
+        return new_state
+    
+    def cancel(self, graph_execution_state_id: str) -> None:
+        """Cancels the given execution state"""
+        self.services.queue.cancel(graph_execution_state_id)
+
+    def __start_service(self, service) -> None:
+        # Call start() method on any services that have it
+        start_op = getattr(service, "start", None)
+        if callable(start_op):
+            start_op(self)
+
+    def __stop_service(self, service) -> None:
+        # Call stop() method on any services that have it
+        stop_op = getattr(service, "stop", None)
+        if callable(stop_op):
+            stop_op(self)
+
+    def _start(self) -> None:
+        """Starts the invoker. This is called automatically when the invoker is created."""
+        for service in vars(self.services):
+            self.__start_service(getattr(self.services, service))
+
+        for service in vars(self.services):
+            self.__start_service(getattr(self.services, service))
+
+    def stop(self) -> None:
+        """Stops the invoker. A new invoker will have to be created to execute further."""
+        # First stop all services
+        for service in vars(self.services):
+            self.__stop_service(getattr(self.services, service))
+
+        for service in vars(self.services):
+            self.__stop_service(getattr(self.services, service))
+
+        self.services.queue.put(None)
+
+
+class InvocationProcessorABC(ABC):
+    pass

invokeai/app/services/item_storage.py

@@ -0,0 +1,62 @@
+from abc import ABC, abstractmethod
+from typing import Callable, Generic, TypeVar
+
+from pydantic import BaseModel, Field
+from pydantic.generics import GenericModel
+
+T = TypeVar("T", bound=BaseModel)
+
+
+class PaginatedResults(GenericModel, Generic[T]):
+    """Paginated results"""
+    #fmt: off
+    items: list[T] = Field(description="Items")
+    page: int = Field(description="Current Page")
+    pages: int = Field(description="Total number of pages")
+    per_page: int = Field(description="Number of items per page")
+    total: int = Field(description="Total number of items in result")
+    #fmt: on
+
+class ItemStorageABC(ABC, Generic[T]):
+    _on_changed_callbacks: list[Callable[[T], None]]
+    _on_deleted_callbacks: list[Callable[[str], None]]
+
+    def __init__(self) -> None:
+        self._on_changed_callbacks = list()
+        self._on_deleted_callbacks = list()
+
+    """Base item storage class"""
+
+    @abstractmethod
+    def get(self, item_id: str) -> T:
+        pass
+
+    @abstractmethod
+    def set(self, item: T) -> None:
+        pass
+
+    @abstractmethod
+    def list(self, page: int = 0, per_page: int = 10) -> PaginatedResults[T]:
+        pass
+
+    @abstractmethod
+    def search(
+        self, query: str, page: int = 0, per_page: int = 10
+    ) -> PaginatedResults[T]:
+        pass
+
+    def on_changed(self, on_changed: Callable[[T], None]) -> None:
+        """Register a callback for when an item is changed"""
+        self._on_changed_callbacks.append(on_changed)
+
+    def on_deleted(self, on_deleted: Callable[[str], None]) -> None:
+        """Register a callback for when an item is deleted"""
+        self._on_deleted_callbacks.append(on_deleted)
+
+    def _on_changed(self, item: T) -> None:
+        for callback in self._on_changed_callbacks:
+            callback(item)
+
+    def _on_deleted(self, item_id: str) -> None:
+        for callback in self._on_deleted_callbacks:
+            callback(item_id)

invokeai/app/services/model_manager_initializer.py

@@ -0,0 +1,120 @@
+import os
+import sys
+import torch
+from argparse import Namespace
+from invokeai.backend import Args
+from omegaconf import OmegaConf
+from pathlib import Path
+
+import invokeai.version
+from ...backend import ModelManager
+from ...backend.util import choose_precision, choose_torch_device
+from ...backend import Globals
+
+# TODO: Replace with an abstract class base ModelManagerBase
+def get_model_manager(config: Args) -> ModelManager:
+    if not config.conf:
+        config_file = os.path.join(Globals.root, "configs", "models.yaml")
+        if not os.path.exists(config_file):
+            report_model_error(
+                config, FileNotFoundError(f"The file {config_file} could not be found.")
+            )
+
+    print(f">> {invokeai.version.__app_name__}, version {invokeai.version.__version__}")
+    print(f'>> InvokeAI runtime directory is "{Globals.root}"')
+
+    # these two lines prevent a horrible warning message from appearing
+    # when the frozen CLIP tokenizer is imported
+    import transformers  # type: ignore
+
+    transformers.logging.set_verbosity_error()
+    import diffusers
+
+    diffusers.logging.set_verbosity_error()
+
+    # normalize the config directory relative to root
+    if not os.path.isabs(config.conf):
+        config.conf = os.path.normpath(os.path.join(Globals.root, config.conf))
+
+    if config.embeddings:
+        if not os.path.isabs(config.embedding_path):
+            embedding_path = os.path.normpath(
+                os.path.join(Globals.root, config.embedding_path)
+            )
+        else:
+            embedding_path = config.embedding_path
+    else:
+        embedding_path = None
+
+    # migrate legacy models
+    ModelManager.migrate_models()
+
+    # creating the model manager
+    try:
+        device = torch.device(choose_torch_device())
+        precision = 'float16' if config.precision=='float16' \
+        else 'float32' if config.precision=='float32' \
+        else choose_precision(device)
+        
+        model_manager = ModelManager(
+            OmegaConf.load(config.conf),
+            precision=precision,
+            device_type=device,
+            max_loaded_models=config.max_loaded_models,
+            embedding_path = Path(embedding_path),
+        )
+    except (FileNotFoundError, TypeError, AssertionError) as e:
+        report_model_error(config, e)
+    except (IOError, KeyError) as e:
+        print(f"{e}. Aborting.")
+        sys.exit(-1)
+
+    # try to autoconvert new models
+    # autoimport new .ckpt files
+    if path := config.autoconvert:
+        model_manager.autoconvert_weights(
+            conf_path=config.conf,
+            weights_directory=path,
+        )
+
+    return model_manager
+
+def report_model_error(opt: Namespace, e: Exception):
+    print(f'** An error occurred while attempting to initialize the model: "{str(e)}"')
+    print(
+        "** This can be caused by a missing or corrupted models file, and can sometimes be fixed by (re)installing the models."
+    )
+    yes_to_all = os.environ.get("INVOKE_MODEL_RECONFIGURE")
+    if yes_to_all:
+        print(
+            "** Reconfiguration is being forced by environment variable INVOKE_MODEL_RECONFIGURE"
+        )
+    else:
+        response = input(
+            "Do you want to run invokeai-configure script to select and/or reinstall models? [y] "
+        )
+        if response.startswith(("n", "N")):
+            return
+
+    print("invokeai-configure is launching....\n")
+
+    # Match arguments that were set on the CLI
+    # only the arguments accepted by the configuration script are parsed
+    root_dir = ["--root", opt.root_dir] if opt.root_dir is not None else []
+    config = ["--config", opt.conf] if opt.conf is not None else []
+    previous_config = sys.argv
+    sys.argv = ["invokeai-configure"]
+    sys.argv.extend(root_dir)
+    sys.argv.extend(config.to_dict())
+    if yes_to_all is not None:
+        for arg in yes_to_all.split():
+            sys.argv.append(arg)
+
+    from invokeai.frontend.install import invokeai_configure
+
+    invokeai_configure()
+    # TODO: Figure out how to restart
+    # print('** InvokeAI will now restart')
+    # sys.argv = previous_args
+    # main() # would rather do a os.exec(), but doesn't exist?
+    # sys.exit(0)

invokeai/app/services/processor.py

@@ -0,0 +1,121 @@
+import traceback
+from threading import Event, Thread
+
+from ..invocations.baseinvocation import InvocationContext
+from .invocation_queue import InvocationQueueItem
+from .invoker import InvocationProcessorABC, Invoker
+
+
+class DefaultInvocationProcessor(InvocationProcessorABC):
+    __invoker_thread: Thread
+    __stop_event: Event
+    __invoker: Invoker
+
+    def start(self, invoker) -> None:
+        self.__invoker = invoker
+        self.__stop_event = Event()
+        self.__invoker_thread = Thread(
+            name="invoker_processor",
+            target=self.__process,
+            kwargs=dict(stop_event=self.__stop_event),
+        )
+        self.__invoker_thread.daemon = (
+            True  # TODO: probably better to just not use threads?
+        )
+        self.__invoker_thread.start()
+
+    def stop(self, *args, **kwargs) -> None:
+        self.__stop_event.set()
+
+    def __process(self, stop_event: Event):
+        try:
+            while not stop_event.is_set():
+                queue_item: InvocationQueueItem = self.__invoker.services.queue.get()
+                if not queue_item:  # Probably stopping
+                    continue
+
+                graph_execution_state = (
+                    self.__invoker.services.graph_execution_manager.get(
+                        queue_item.graph_execution_state_id
+                    )
+                )
+                invocation = graph_execution_state.execution_graph.get_node(
+                    queue_item.invocation_id
+                )
+
+                # Send starting event
+                self.__invoker.services.events.emit_invocation_started(
+                    graph_execution_state_id=graph_execution_state.id,
+                    invocation_id=invocation.id,
+                )
+
+                # Invoke
+                try:
+                    outputs = invocation.invoke(
+                        InvocationContext(
+                            services=self.__invoker.services,
+                            graph_execution_state_id=graph_execution_state.id,
+                        )
+                    )
+
+                    # Check queue to see if this is canceled, and skip if so
+                    if self.__invoker.services.queue.is_canceled(
+                        graph_execution_state.id
+                    ):
+                        continue
+
+                    # Save outputs and history
+                    graph_execution_state.complete(invocation.id, outputs)
+
+                    # Save the state changes
+                    self.__invoker.services.graph_execution_manager.set(
+                        graph_execution_state
+                    )
+
+                    # Send complete event
+                    self.__invoker.services.events.emit_invocation_complete(
+                        graph_execution_state_id=graph_execution_state.id,
+                        invocation_id=invocation.id,
+                        result=outputs.dict(),
+                    )
+
+                except KeyboardInterrupt:
+                    pass
+
+                except Exception as e:
+                    error = traceback.format_exc()
+
+                    # Save error
+                    graph_execution_state.set_node_error(invocation.id, error)
+
+                    # Save the state changes
+                    self.__invoker.services.graph_execution_manager.set(
+                        graph_execution_state
+                    )
+
+                    # Send error event
+                    self.__invoker.services.events.emit_invocation_error(
+                        graph_execution_state_id=graph_execution_state.id,
+                        invocation_id=invocation.id,
+                        error=error,
+                    )
+
+                    pass
+                
+                # Check queue to see if this is canceled, and skip if so
+                if self.__invoker.services.queue.is_canceled(
+                    graph_execution_state.id
+                ):
+                    continue
+
+                # Queue any further commands if invoking all
+                is_complete = graph_execution_state.is_complete()
+                if queue_item.invoke_all and not is_complete:
+                    self.__invoker.invoke(graph_execution_state, invoke_all=True)
+                elif is_complete:
+                    self.__invoker.services.events.emit_graph_execution_complete(
+                        graph_execution_state.id
+                    )
+
+        except KeyboardInterrupt:
+            ...  # Log something?

invokeai/app/services/restoration_services.py

@@ -0,0 +1,109 @@
+import sys
+import traceback
+import torch
+from ...backend.restoration import Restoration
+from ...backend.util import choose_torch_device, CPU_DEVICE, MPS_DEVICE
+
+# This should be a real base class for postprocessing functions,
+# but right now we just instantiate the existing gfpgan, esrgan
+# and codeformer functions.
+class RestorationServices:
+    '''Face restoration and upscaling'''
+    
+    def __init__(self,args):
+        try:
+            gfpgan, codeformer, esrgan = None, None, None
+            if args.restore or args.esrgan:
+                restoration = Restoration()
+                if args.restore:
+                    gfpgan, codeformer = restoration.load_face_restore_models(
+                        args.gfpgan_model_path
+                    )
+                else:
+                    print(">> Face restoration disabled")
+                    if args.esrgan:
+                        esrgan = restoration.load_esrgan(args.esrgan_bg_tile)
+                    else:
+                        print(">> Upscaling disabled")
+            else:
+                print(">> Face restoration and upscaling disabled")
+        except (ModuleNotFoundError, ImportError):
+            print(traceback.format_exc(), file=sys.stderr)
+            print(">> You may need to install the ESRGAN and/or GFPGAN modules")
+        self.device = torch.device(choose_torch_device())
+        self.gfpgan = gfpgan
+        self.codeformer = codeformer
+        self.esrgan = esrgan
+
+    # note that this one method does gfpgan and codepath reconstruction, as well as
+    # esrgan upscaling
+    # TO DO: refactor into separate methods
+    def upscale_and_reconstruct(
+        self,
+        image_list,
+        facetool="gfpgan",
+        upscale=None,
+        upscale_denoise_str=0.75,
+        strength=0.0,
+        codeformer_fidelity=0.75,
+        save_original=False,
+        image_callback=None,
+        prefix=None,
+    ):
+        results = []
+        for r in image_list:
+            image, seed = r
+            try:
+                if strength > 0:
+                    if self.gfpgan is not None or self.codeformer is not None:
+                        if facetool == "gfpgan":
+                            if self.gfpgan is None:
+                                print(
+                                    ">> GFPGAN not found. Face restoration is disabled."
+                                )
+                            else:
+                                image = self.gfpgan.process(image, strength, seed)
+                        if facetool == "codeformer":
+                            if self.codeformer is None:
+                                print(
+                                    ">> CodeFormer not found. Face restoration is disabled."
+                                )
+                            else:
+                                cf_device = (
+                                    CPU_DEVICE if self.device == MPS_DEVICE else self.device
+                                )
+                                image = self.codeformer.process(
+                                    image=image,
+                                    strength=strength,
+                                    device=cf_device,
+                                    seed=seed,
+                                    fidelity=codeformer_fidelity,
+                                )
+                    else:
+                        print(">> Face Restoration is disabled.")
+                if upscale is not None:
+                    if self.esrgan is not None:
+                        if len(upscale) < 2:
+                            upscale.append(0.75)
+                        image = self.esrgan.process(
+                            image,
+                            upscale[1],
+                            seed,
+                            int(upscale[0]),
+                            denoise_str=upscale_denoise_str,
+                        )
+                    else:
+                        print(">> ESRGAN is disabled. Image not upscaled.")
+            except Exception as e:
+                print(
+                    f">> Error running RealESRGAN or GFPGAN. Your image was not upscaled.\n{e}"
+                )
+
+            if image_callback is not None:
+                image_callback(image, seed, upscaled=True, use_prefix=prefix)
+            else:
+                r[0] = image
+
+            results.append([image, seed])
+
+        return results

invokeai/app/services/sqlite.py

@@ -0,0 +1,138 @@
+import sqlite3
+from threading import Lock
+from typing import Generic, TypeVar, Union, get_args
+
+from pydantic import BaseModel, parse_raw_as
+
+from .item_storage import ItemStorageABC, PaginatedResults
+
+T = TypeVar("T", bound=BaseModel)
+
+sqlite_memory = ":memory:"
+
+
+class SqliteItemStorage(ItemStorageABC, Generic[T]):
+    _filename: str
+    _table_name: str
+    _conn: sqlite3.Connection
+    _cursor: sqlite3.Cursor
+    _id_field: str
+    _lock: Lock
+
+    def __init__(self, filename: str, table_name: str, id_field: str = "id"):
+        super().__init__()
+
+        self._filename = filename
+        self._table_name = table_name
+        self._id_field = id_field  # TODO: validate that T has this field
+        self._lock = Lock()
+
+        self._conn = sqlite3.connect(
+            self._filename, check_same_thread=False
+        )  # TODO: figure out a better threading solution
+        self._cursor = self._conn.cursor()
+
+        self._create_table()
+
+    def _create_table(self):
+        try:
+            self._lock.acquire()
+            self._cursor.execute(
+                f"""CREATE TABLE IF NOT EXISTS {self._table_name} (
+                item TEXT,
+                id TEXT GENERATED ALWAYS AS (json_extract(item, '$.{self._id_field}')) VIRTUAL NOT NULL);"""
+            )
+            self._cursor.execute(
+                f"""CREATE UNIQUE INDEX IF NOT EXISTS {self._table_name}_id ON {self._table_name}(id);"""
+            )
+        finally:
+            self._lock.release()
+
+    def _parse_item(self, item: str) -> T:
+        item_type = get_args(self.__orig_class__)[0]
+        return parse_raw_as(item_type, item)
+
+    def set(self, item: T):
+        try:
+            self._lock.acquire()
+            self._cursor.execute(
+                f"""INSERT OR REPLACE INTO {self._table_name} (item) VALUES (?);""",
+                (item.json(),),
+            )
+        finally:
+            self._lock.release()
+        self._on_changed(item)
+
+    def get(self, id: str) -> Union[T, None]:
+        try:
+            self._lock.acquire()
+            self._cursor.execute(
+                f"""SELECT item FROM {self._table_name} WHERE id = ?;""", (str(id),)
+            )
+            result = self._cursor.fetchone()
+        finally:
+            self._lock.release()
+
+        if not result:
+            return None
+
+        return self._parse_item(result[0])
+
+    def delete(self, id: str):
+        try:
+            self._lock.acquire()
+            self._cursor.execute(
+                f"""DELETE FROM {self._table_name} WHERE id = ?;""", (str(id),)
+            )
+        finally:
+            self._lock.release()
+        self._on_deleted(id)
+
+    def list(self, page: int = 0, per_page: int = 10) -> PaginatedResults[T]:
+        try:
+            self._lock.acquire()
+            self._cursor.execute(
+                f"""SELECT item FROM {self._table_name} LIMIT ? OFFSET ?;""",
+                (per_page, page * per_page),
+            )
+            result = self._cursor.fetchall()
+
+            items = list(map(lambda r: self._parse_item(r[0]), result))
+
+            self._cursor.execute(f"""SELECT count(*) FROM {self._table_name};""")
+            count = self._cursor.fetchone()[0]
+        finally:
+            self._lock.release()
+
+        pageCount = int(count / per_page) + 1
+
+        return PaginatedResults[T](
+            items=items, page=page, pages=pageCount, per_page=per_page, total=count
+        )
+
+    def search(
+        self, query: str, page: int = 0, per_page: int = 10
+    ) -> PaginatedResults[T]:
+        try:
+            self._lock.acquire()
+            self._cursor.execute(
+                f"""SELECT item FROM {self._table_name} WHERE item LIKE ? LIMIT ? OFFSET ?;""",
+                (f"%{query}%", per_page, page * per_page),
+            )
+            result = self._cursor.fetchall()
+
+            items = list(map(lambda r: self._parse_item(r[0]), result))
+
+            self._cursor.execute(
+                f"""SELECT count(*) FROM {self._table_name} WHERE item LIKE ?;""",
+                (f"%{query}%",),
+            )
+            count = self._cursor.fetchone()[0]
+        finally:
+            self._lock.release()
+
+        pageCount = int(count / per_page) + 1
+
+        return PaginatedResults[T](
+            items=items, page=page, pages=pageCount, per_page=per_page, total=count
+        )

invokeai/backend/__init__.py

@@ -1,5 +1,16 @@
-'''
+"""
 Initialization file for invokeai.backend
-'''
-from .invoke_ai_web_server import InvokeAIWebServer
-
+"""
+from .generate import Generate
+from .generator import (
+    InvokeAIGeneratorBasicParams,
+    InvokeAIGenerator,
+    InvokeAIGeneratorOutput,
+    Txt2Img,
+    Img2Img,
+    Inpaint
+)
+from .model_management import ModelManager
+from .safety_checker import SafetyChecker
+from .args import Args
+from .globals import Globals

invokeai/backend/config/invokeai_configure.py

@@ -6,19 +6,20 @@
 #
 # Coauthor: Kevin Turner http://github.com/keturn
 #
-print("Loading Python libraries...\n")
+import sys
+print("Loading Python libraries...\n",file=sys.stderr)
+
 import argparse
 import io
 import os
 import re
 import shutil
-import sys
 import traceback
 import warnings
 from argparse import Namespace
 from pathlib import Path
-from urllib import request
 from shutil import get_terminal_size
+from urllib import request
 
 import npyscreen
 import torch
@@ -37,17 +38,20 @@ from transformers import (
 
 import invokeai.configs as configs
 
+from ...frontend.install.model_install import addModelsForm, process_and_execute
+from ...frontend.install.widgets import (
+    CenteredButtonPress,
+    IntTitleSlider,
+    set_min_terminal_size,
+)
 from ..args import PRECISION_CHOICES, Args
-from ..globals import Globals, global_config_dir, global_config_file, global_cache_dir
-from .model_install import addModelsForm, process_and_execute
+from ..globals import Globals, global_cache_dir, global_config_dir, global_config_file
 from .model_install_backend import (
     default_dataset,
     download_from_hf,
-    recommended_datasets,
     hf_download_with_resume,
+    recommended_datasets,
 )
-from .widgets import IntTitleSlider, CenteredButtonPress, set_min_terminal_size
-
 
 warnings.filterwarnings("ignore")
 
@@ -82,6 +86,7 @@ INIT_FILE_PREAMBLE = """# InvokeAI initialization file
 # -Ak_euler_a -C10.0
 """
 
+
 # --------------------------------------------
 def postscript(errors: None):
     if not any(errors):
@@ -180,13 +185,11 @@ def download_with_progress_bar(model_url: str, model_dest: str, label: str = "th
 # ---------------------------------------------
 # this will preload the Bert tokenizer fles
 def download_bert():
-    print(
-        "Installing bert tokenizer...",
-        file=sys.stderr
-    )
+    print("Installing bert tokenizer...", file=sys.stderr)
     with warnings.catch_warnings():
         warnings.filterwarnings("ignore", category=DeprecationWarning)
         from transformers import BertTokenizerFast
+
         download_from_hf(BertTokenizerFast, "bert-base-uncased")
 
 
@@ -197,12 +200,14 @@ def download_sd1_clip():
     download_from_hf(CLIPTokenizer, version)
     download_from_hf(CLIPTextModel, version)
 
+
 # ---------------------------------------------
 def download_sd2_clip():
-    version = 'stabilityai/stable-diffusion-2'
+    version = "stabilityai/stable-diffusion-2"
     print("Installing SD2 clip model...", file=sys.stderr)
-    download_from_hf(CLIPTokenizer, version, subfolder='tokenizer')
-    download_from_hf(CLIPTextModel, version, subfolder='text_encoder')
+    download_from_hf(CLIPTokenizer, version, subfolder="tokenizer")
+    download_from_hf(CLIPTextModel, version, subfolder="text_encoder")
+
 
 # ---------------------------------------------
 def download_realesrgan():
@@ -290,7 +295,7 @@ def download_vaes():
         # first the diffusers version
         repo_id = "stabilityai/sd-vae-ft-mse"
         args = dict(
-            cache_dir=global_cache_dir("diffusers"),
+            cache_dir=global_cache_dir("hub"),
         )
         if not AutoencoderKL.from_pretrained(repo_id, **args):
             raise Exception(f"download of {repo_id} failed")
@@ -323,13 +328,13 @@ def get_root(root: str = None) -> str:
 class editOptsForm(npyscreen.FormMultiPage):
     # for responsive resizing - disabled
     # FIX_MINIMUM_SIZE_WHEN_CREATED = False
-    
+
     def create(self):
         program_opts = self.parentApp.program_opts
         old_opts = self.parentApp.invokeai_opts
         first_time = not (Globals.root / Globals.initfile).exists()
         access_token = HfFolder.get_token()
-        window_width,window_height = get_terminal_size()
+        window_width, window_height = get_terminal_size()
         for i in [
             "Configure startup settings. You can come back and change these later.",
             "Use ctrl-N and ctrl-P to move to the <N>ext and <P>revious fields.",
@@ -681,6 +686,7 @@ def run_console_ui(
     else:
         return (editApp.new_opts, editApp.user_selections)
 
+
 # -------------------------------------
 def write_opts(opts: Namespace, init_file: Path):
     """
@@ -701,8 +707,8 @@ def write_opts(opts: Namespace, init_file: Path):
         "^--?(o|out|no-xformer|xformer|no-ckpt|ckpt|free|no-nsfw|nsfw|prec|max_load|embed|always|ckpt|free_gpu)"
     )
     # fix windows paths
-    opts.outdir = opts.outdir.replace('\\','/')
-    opts.embedding_path = opts.embedding_path.replace('\\','/')
+    opts.outdir = opts.outdir.replace("\\", "/")
+    opts.embedding_path = opts.embedding_path.replace("\\", "/")
     new_file = f"{init_file}.new"
     try:
         lines = [x.strip() for x in open(init_file, "r").readlines()]
@@ -855,6 +861,7 @@ def main():
     except KeyboardInterrupt:
         print("\nGoodbye! Come back soon.")
 
+
 # -------------------------------------
 if __name__ == "__main__":
     main()

invokeai/backend/config/model_install_backend.py

@@ -8,6 +8,7 @@ import sys
 import warnings
 from pathlib import Path
 from tempfile import TemporaryFile
+from typing import List
 
 import requests
 from diffusers import AutoencoderKL
@@ -15,12 +16,12 @@ from huggingface_hub import hf_hub_url
 from omegaconf import OmegaConf
 from omegaconf.dictconfig import DictConfig
 from tqdm import tqdm
-from typing import List
 
 import invokeai.configs as configs
-from ..generator.diffusers_pipeline import StableDiffusionGeneratorPipeline
+
 from ..globals import Globals, global_cache_dir, global_config_dir
-from ..model_manager import ModelManager
+from ..model_management import ModelManager
+from ..stable_diffusion import StableDiffusionGeneratorPipeline
 
 warnings.filterwarnings("ignore")
 
@@ -44,45 +45,49 @@ Config_preamble = """
 # was trained on.
 """
 
+
 def default_config_file():
     return Path(global_config_dir()) / "models.yaml"
 
+
 def sd_configs():
     return Path(global_config_dir()) / "stable-diffusion"
 
+
 def initial_models():
     global Datasets
     if Datasets:
         return Datasets
     return (Datasets := OmegaConf.load(Dataset_path))
 
+
 def install_requested_models(
-        install_initial_models: List[str] = None,
-        remove_models: List[str] = None,
-        scan_directory: Path = None,
-        external_models: List[str] = None,
-        scan_at_startup: bool = False,
-        convert_to_diffusers: bool = False,
-        precision: str = "float16",
-        purge_deleted: bool = False,
-        config_file_path: Path = None,
+    install_initial_models: List[str] = None,
+    remove_models: List[str] = None,
+    scan_directory: Path = None,
+    external_models: List[str] = None,
+    scan_at_startup: bool = False,
+    convert_to_diffusers: bool = False,
+    precision: str = "float16",
+    purge_deleted: bool = False,
+    config_file_path: Path = None,
 ):
-    '''
+    """
     Entry point for installing/deleting starter models, or installing external models.
-    '''
-    config_file_path=config_file_path or default_config_file()
+    """
+    config_file_path = config_file_path or default_config_file()
     if not config_file_path.exists():
-        open(config_file_path,'w')
-            
-    model_manager= ModelManager(OmegaConf.load(config_file_path),precision=precision)
-    
+        open(config_file_path, "w")
+
+    model_manager = ModelManager(OmegaConf.load(config_file_path), precision=precision)
+
     if remove_models and len(remove_models) > 0:
         print("== DELETING UNCHECKED STARTER MODELS ==")
         for model in remove_models:
-            print(f'{model}...')
+            print(f"{model}...")
             model_manager.del_model(model, delete_files=purge_deleted)
         model_manager.commit(config_file_path)
-    
+
     if install_initial_models and len(install_initial_models) > 0:
         print("== INSTALLING SELECTED STARTER MODELS ==")
         successfully_downloaded = download_weight_datasets(
@@ -96,20 +101,20 @@ def install_requested_models(
 
     # due to above, we have to reload the model manager because conf file
     # was changed behind its back
-    model_manager= ModelManager(OmegaConf.load(config_file_path),precision=precision)
+    model_manager = ModelManager(OmegaConf.load(config_file_path), precision=precision)
 
     external_models = external_models or list()
     if scan_directory:
         external_models.append(str(scan_directory))
 
-    if len(external_models)>0:
+    if len(external_models) > 0:
         print("== INSTALLING EXTERNAL MODELS ==")
         for path_url_or_repo in external_models:
             try:
                 model_manager.heuristic_import(
                     path_url_or_repo,
                     convert=convert_to_diffusers,
-                    commit_to_conf=config_file_path
+                    commit_to_conf=config_file_path,
                 )
             except KeyboardInterrupt:
                 sys.exit(-1)
@@ -117,17 +122,18 @@ def install_requested_models(
                 pass
 
     if scan_at_startup and scan_directory.is_dir():
-        argument = '--autoconvert' if convert_to_diffusers else '--autoimport'
+        argument = "--autoconvert" if convert_to_diffusers else "--autoimport"
         initfile = Path(Globals.root, Globals.initfile)
-        replacement = Path(Globals.root, f'{Globals.initfile}.new')
-        directory = str(scan_directory).replace('\\','/')
-        with open(initfile,'r') as input:
-            with open(replacement,'w') as output:
+        replacement = Path(Globals.root, f"{Globals.initfile}.new")
+        directory = str(scan_directory).replace("\\", "/")
+        with open(initfile, "r") as input:
+            with open(replacement, "w") as output:
                 while line := input.readline():
                     if not line.startswith(argument):
                         output.writelines([line])
-                output.writelines([f'{argument} {directory}'])
-        os.replace(replacement,initfile)
+                output.writelines([f"{argument} {directory}"])
+        os.replace(replacement, initfile)
+
 
 # -------------------------------------
 def yes_or_no(prompt: str, default_yes=True):
@@ -183,7 +189,9 @@ def migrate_models_ckpt():
     if not os.path.exists(os.path.join(model_path, "model.ckpt")):
         return
     new_name = initial_models()["stable-diffusion-1.4"]["file"]
-    print('The Stable Diffusion v4.1 "model.ckpt" is already installed. The name will be changed to {new_name} to avoid confusion.')
+    print(
+        'The Stable Diffusion v4.1 "model.ckpt" is already installed. The name will be changed to {new_name} to avoid confusion.'
+    )
     print(f"model.ckpt => {new_name}")
     os.replace(
         os.path.join(model_path, "model.ckpt"), os.path.join(model_path, new_name)
@@ -262,7 +270,6 @@ def _download_diffusion_weights(
             path = download_from_hf(
                 model_class,
                 repo_id,
-                cache_subdir="diffusers",
                 safety_checker=None,
                 **extra_args,
             )
@@ -383,7 +390,8 @@ def update_config_file(successfully_downloaded: dict, config_file: Path):
 
 # ---------------------------------------------
 def new_config_file_contents(
-        successfully_downloaded: dict, config_file: Path,
+    successfully_downloaded: dict,
+    config_file: Path,
 ) -> str:
     if config_file.exists():
         conf = OmegaConf.load(str(config_file.expanduser().resolve()))
@@ -413,7 +421,9 @@ def new_config_file_contents(
             stanza["weights"] = os.path.relpath(
                 successfully_downloaded[model], start=Globals.root
             )
-            stanza["config"] = os.path.normpath(os.path.join(sd_configs(), mod["config"]))
+            stanza["config"] = os.path.normpath(
+                os.path.join(sd_configs(), mod["config"])
+            )
         if "vae" in mod:
             if "file" in mod["vae"]:
                 stanza["vae"] = os.path.normpath(
@@ -445,7 +455,7 @@ def delete_weights(model_name: str, conf_stanza: dict):
     print(
         f"\n** The checkpoint version of {model_name} is superseded by the diffusers version. Deleting the original file {weights}?"
     )
-          
+
     weights = Path(weights)
     if not weights.is_absolute():
         weights = Path(Globals.root) / weights

invokeai/backend/generate.py

@@ -5,6 +5,7 @@
 
 import gc
 import importlib
+import logging
 import os
 import random
 import re
@@ -19,28 +20,23 @@ import numpy as np
 import skimage
 import torch
 import transformers
+from PIL import Image, ImageOps
+from accelerate.utils import set_seed
 from diffusers.pipeline_utils import DiffusionPipeline
 from diffusers.utils.import_utils import is_xformers_available
 from omegaconf import OmegaConf
-from PIL import Image, ImageOps
-from pytorch_lightning import logging, seed_everything
-
-import ldm.invoke.conditioning
-from ldm.invoke.args import metadata_from_png
-from ldm.invoke.concepts_lib import HuggingFaceConceptsLibrary
-from ldm.invoke.conditioning import get_uc_and_c_and_ec
-from ldm.invoke.devices import choose_precision, choose_torch_device
-from ldm.invoke.generator.inpaint import infill_methods
-from ldm.invoke.globals import Globals, global_cache_dir
-from ldm.invoke.image_util import InitImageResizer
-from ldm.invoke.model_manager import ModelManager
-from ldm.invoke.pngwriter import PngWriter
-from ldm.invoke.seamless import configure_model_padding
-from ldm.invoke.txt2mask import Txt2Mask
-from ldm.models.diffusion.ddim import DDIMSampler
-from ldm.models.diffusion.ksampler import KSampler
-from ldm.models.diffusion.plms import PLMSSampler
+from pathlib import Path
 
+from .args import metadata_from_png
+from .generator import infill_methods
+from .globals import Globals, global_cache_dir
+from .image_util import InitImageResizer, PngWriter, Txt2Mask, configure_model_padding
+from .model_management import ModelManager
+from .safety_checker import SafetyChecker
+from .prompting import get_uc_and_c_and_ec
+from .prompting.conditioning import log_tokenization
+from .stable_diffusion import HuggingFaceConceptsLibrary
+from .util import choose_precision, choose_torch_device
 
 def fix_func(orig):
     if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
@@ -227,6 +223,7 @@ class Generate:
             self.precision,
             max_loaded_models=max_loaded_models,
             sequential_offload=self.free_gpu_mem,
+            embedding_path=Path(self.embedding_path),
         )
         # don't accept invalid models
         fallback = self.model_manager.default_model() or FALLBACK_MODEL_NAME
@@ -249,31 +246,8 @@ class Generate:
 
         # load safety checker if requested
         if safety_checker:
-            try:
-                print(">> Initializing NSFW checker")
-                from diffusers.pipelines.stable_diffusion.safety_checker import (
-                    StableDiffusionSafetyChecker,
-                )
-                from transformers import AutoFeatureExtractor
-
-                safety_model_id = "CompVis/stable-diffusion-safety-checker"
-                safety_model_path = global_cache_dir("hub")
-                self.safety_checker = StableDiffusionSafetyChecker.from_pretrained(
-                    safety_model_id,
-                    local_files_only=True,
-                    cache_dir=safety_model_path,
-                )
-                self.safety_feature_extractor = AutoFeatureExtractor.from_pretrained(
-                    safety_model_id,
-                    local_files_only=True,
-                    cache_dir=safety_model_path,
-                )
-                self.safety_checker.to(self.device)
-            except Exception:
-                print(
-                    "** An error was encountered while installing the safety checker:"
-                )
-                print(traceback.format_exc())
+            print(">> Initializing NSFW checker")
+            self.safety_checker = SafetyChecker(self.device)
         else:
             print(">> NSFW checker is disabled")
 
@@ -328,8 +302,8 @@ class Generate:
         variation_amount=0.0,
         threshold=0.0,
         perlin=0.0,
-        h_symmetry_time_pct = None,
-        v_symmetry_time_pct = None,
+        h_symmetry_time_pct=None,
+        v_symmetry_time_pct=None,
         karras_max=None,
         outdir=None,
         # these are specific to img2img and inpaint
@@ -486,7 +460,7 @@ class Generate:
 
         if sampler_name and (sampler_name != self.sampler_name):
             self.sampler_name = sampler_name
-            self._set_sampler()
+            self._set_scheduler()
 
         # apply the concepts library to the prompt
         prompt = self.huggingface_concepts_library.replace_concepts_with_triggers(
@@ -495,28 +469,11 @@ class Generate:
             self.model.textual_inversion_manager.get_all_trigger_strings(),
         )
 
-        # bit of a hack to change the cached sampler's karras threshold to
-        # whatever the user asked for
-        if karras_max is not None and isinstance(self.sampler, KSampler):
-            self.sampler.adjust_settings(karras_max=karras_max)
-
         tic = time.time()
         if self._has_cuda():
             torch.cuda.reset_peak_memory_stats()
 
         results = list()
-        init_image = None
-        mask_image = None
-
-        try:
-            if (
-                self.free_gpu_mem
-                and self.model.cond_stage_model.device != self.model.device
-            ):
-                self.model.cond_stage_model.device = self.model.device
-                self.model.cond_stage_model.to(self.model.device)
-        except AttributeError:
-            pass
 
         try:
             uc, c, extra_conditioning_info = get_uc_and_c_and_ec(
@@ -545,15 +502,6 @@ class Generate:
             generator.set_variation(self.seed, variation_amount, with_variations)
             generator.use_mps_noise = use_mps_noise
 
-            checker = (
-                {
-                    "checker": self.safety_checker,
-                    "extractor": self.safety_feature_extractor,
-                }
-                if self.safety_checker
-                else None
-            )
-
             results = generator.generate(
                 prompt,
                 iterations=iterations,
@@ -580,7 +528,7 @@ class Generate:
                 embiggen_strength=embiggen_strength,
                 inpaint_replace=inpaint_replace,
                 mask_blur_radius=mask_blur_radius,
-                safety_checker=checker,
+                safety_checker=self.safety_checker,
                 seam_size=seam_size,
                 seam_blur=seam_blur,
                 seam_strength=seam_strength,
@@ -717,7 +665,7 @@ class Generate:
             prompt,
             model=self.model,
             skip_normalize_legacy_blend=opt.skip_normalize,
-            log_tokens=ldm.invoke.conditioning.log_tokenization,
+            log_tokens=log_tokenization,
         )
 
         if tool in ("gfpgan", "codeformer", "upscale"):
@@ -741,7 +689,7 @@ class Generate:
             )
 
         elif tool == "outcrop":
-            from ldm.invoke.restoration.outcrop import Outcrop
+            from .restoration.outcrop import Outcrop
 
             extend_instructions = {}
             for direction, pixels in _pairwise(opt.outcrop):
@@ -794,7 +742,7 @@ class Generate:
                 clear_cuda_cache=self.clear_cuda_cache,
             )
         elif tool == "outpaint":
-            from ldm.invoke.restoration.outpaint import Outpaint
+            from .restoration.outpaint import Outpaint
 
             restorer = Outpaint(image, self)
             return restorer.process(opt, args, image_callback=callback, prefix=prefix)
@@ -816,17 +764,12 @@ class Generate:
         hires_fix: bool = False,
         force_outpaint: bool = False,
     ):
-        inpainting_model_in_use = self.sampler.uses_inpainting_model()
-
         if hires_fix:
             return self._make_txt2img2img()
 
         if embiggen is not None:
             return self._make_embiggen()
 
-        if inpainting_model_in_use:
-            return self._make_omnibus()
-
         if ((init_image is not None) and (mask_image is not None)) or force_outpaint:
             return self._make_inpaint()
 
@@ -903,16 +846,9 @@ class Generate:
     def _make_inpaint(self):
         return self._load_generator(".inpaint", "Inpaint")
 
-    def _make_omnibus(self):
-        return self._load_generator(".omnibus", "Omnibus")
-
     def _load_generator(self, module, class_name):
-        if self.is_legacy_model(self.model_name):
-            mn = f"ldm.invoke.ckpt_generator{module}"
-            cn = f"Ckpt{class_name}"
-        else:
-            mn = f"ldm.invoke.generator{module}"
-            cn = class_name
+        mn = f"invokeai.backend.generator{module}"
+        cn = class_name
         module = importlib.import_module(mn)
         constructor = getattr(module, cn)
         return constructor(self.model, self.precision)
@@ -973,21 +909,9 @@ class Generate:
         # uncache generators so they pick up new models
         self.generators = {}
 
-        seed_everything(random.randrange(0, np.iinfo(np.uint32).max))
-        if self.embedding_path is not None:
-            print(f'>> Loading embeddings from {self.embedding_path}')
-            for root, _, files in os.walk(self.embedding_path):
-                for name in files:
-                    ti_path = os.path.join(root, name)
-                    self.model.textual_inversion_manager.load_textual_inversion(
-                        ti_path, defer_injecting_tokens=True
-                    )
-            print(
-                f'>> Textual inversion triggers: {", ".join(sorted(self.model.textual_inversion_manager.get_all_trigger_strings()))}'
-            )
-
+        set_seed(random.randrange(0, np.iinfo(np.uint32).max))
         self.model_name = model_name
-        self._set_sampler()  # requires self.model_name to be set first
+        self._set_scheduler()  # requires self.model_name to be set first
         return self.model
 
     def load_huggingface_concepts(self, concepts: list[str]):
@@ -1030,8 +954,9 @@ class Generate:
         image_callback=None,
         prefix=None,
     ):
+        results = []
         for r in image_list:
-            image, seed = r
+            image, seed, _ = r
             try:
                 if strength > 0:
                     if self.gfpgan is not None or self.codeformer is not None:
@@ -1083,6 +1008,10 @@ class Generate:
             else:
                 r[0] = image
 
+            results.append([image, seed])
+
+        return results
+
     def apply_textmask(
         self, image_path: str, prompt: str, callback, threshold: float = 0.5
     ):
@@ -1114,44 +1043,6 @@ class Generate:
     def is_legacy_model(self, model_name) -> bool:
         return self.model_manager.is_legacy(model_name)
 
-    def _set_sampler(self):
-        if isinstance(self.model, DiffusionPipeline):
-            return self._set_scheduler()
-        else:
-            return self._set_sampler_legacy()
-
-    # very repetitive code - can this be simplified? The KSampler names are
-    # consistent, at least
-    def _set_sampler_legacy(self):
-        msg = f">> Setting Sampler to {self.sampler_name}"
-        if self.sampler_name == "plms":
-            self.sampler = PLMSSampler(self.model, device=self.device)
-        elif self.sampler_name == "ddim":
-            self.sampler = DDIMSampler(self.model, device=self.device)
-        elif self.sampler_name == "k_dpm_2_a":
-            self.sampler = KSampler(self.model, "dpm_2_ancestral", device=self.device)
-        elif self.sampler_name == "k_dpm_2":
-            self.sampler = KSampler(self.model, "dpm_2", device=self.device)
-        elif self.sampler_name == "k_dpmpp_2_a":
-            self.sampler = KSampler(
-                self.model, "dpmpp_2s_ancestral", device=self.device
-            )
-        elif self.sampler_name == "k_dpmpp_2":
-            self.sampler = KSampler(self.model, "dpmpp_2m", device=self.device)
-        elif self.sampler_name == "k_euler_a":
-            self.sampler = KSampler(self.model, "euler_ancestral", device=self.device)
-        elif self.sampler_name == "k_euler":
-            self.sampler = KSampler(self.model, "euler", device=self.device)
-        elif self.sampler_name == "k_heun":
-            self.sampler = KSampler(self.model, "heun", device=self.device)
-        elif self.sampler_name == "k_lms":
-            self.sampler = KSampler(self.model, "lms", device=self.device)
-        else:
-            msg = f">> Unsupported Sampler: {self.sampler_name}, Defaulting to plms"
-            self.sampler = PLMSSampler(self.model, device=self.device)
-
-        print(msg)
-
     def _set_scheduler(self):
         default = self.model.scheduler
 

invokeai/backend/generator/__init__.py

@@ -0,0 +1,13 @@
+"""
+Initialization file for the invokeai.generator package
+"""
+from .base import (
+    InvokeAIGenerator,
+    InvokeAIGeneratorBasicParams,
+    InvokeAIGeneratorOutput,
+    Txt2Img,
+    Img2Img,
+    Inpaint,
+    Generator,
+)
+from .inpaint import infill_methods

invokeai/backend/generator/base.py

@@ -0,0 +1,648 @@
+"""
+Base class for invokeai.backend.generator.*
+including img2img, txt2img, and inpaint
+"""
+from __future__ import annotations
+
+import itertools
+import dataclasses
+import diffusers
+import os
+import random
+import traceback
+from abc import ABCMeta
+from argparse import Namespace
+from contextlib import nullcontext
+
+import cv2
+import numpy as np
+import torch
+from PIL import Image, ImageChops, ImageFilter
+from accelerate.utils import set_seed
+from diffusers import DiffusionPipeline
+from tqdm import trange
+from typing import List, Iterator, Type
+from dataclasses import dataclass, field
+from diffusers.schedulers import SchedulerMixin as Scheduler
+
+from ..image_util import configure_model_padding
+from ..util.util import rand_perlin_2d
+from ..safety_checker import SafetyChecker
+from ..prompting.conditioning import get_uc_and_c_and_ec
+from ..stable_diffusion.diffusers_pipeline import StableDiffusionGeneratorPipeline
+
+downsampling = 8
+
+@dataclass
+class InvokeAIGeneratorBasicParams:
+    seed: int=None
+    width: int=512
+    height: int=512
+    cfg_scale: int=7.5
+    steps: int=20
+    ddim_eta: float=0.0
+    scheduler: int='ddim'
+    precision: str='float16'
+    perlin: float=0.0
+    threshold: int=0.0
+    seamless: bool=False
+    seamless_axes: List[str]=field(default_factory=lambda: ['x', 'y'])
+    h_symmetry_time_pct: float=None
+    v_symmetry_time_pct: float=None
+    variation_amount: float = 0.0
+    with_variations: list=field(default_factory=list)
+    safety_checker: SafetyChecker=None
+
+@dataclass
+class InvokeAIGeneratorOutput:
+    '''
+    InvokeAIGeneratorOutput is a dataclass that contains the outputs of a generation
+    operation, including the image, its seed, the model name used to generate the image
+    and the model hash, as well as all the generate() parameters that went into
+    generating the image (in .params, also available as attributes)
+    '''
+    image: Image
+    seed: int
+    model_hash: str
+    attention_maps_images: List[Image]
+    params: Namespace
+
+# we are interposing a wrapper around the original Generator classes so that
+# old code that calls Generate will continue to work.
+class InvokeAIGenerator(metaclass=ABCMeta):
+    scheduler_map = dict(
+        ddim=diffusers.DDIMScheduler,
+        dpmpp_2=diffusers.DPMSolverMultistepScheduler,
+        k_dpm_2=diffusers.KDPM2DiscreteScheduler,
+        k_dpm_2_a=diffusers.KDPM2AncestralDiscreteScheduler,
+        k_dpmpp_2=diffusers.DPMSolverMultistepScheduler,
+        k_euler=diffusers.EulerDiscreteScheduler,
+        k_euler_a=diffusers.EulerAncestralDiscreteScheduler,
+        k_heun=diffusers.HeunDiscreteScheduler,
+        k_lms=diffusers.LMSDiscreteScheduler,
+        plms=diffusers.PNDMScheduler,
+    )
+
+    def __init__(self,
+                 model_info: dict,
+                 params: InvokeAIGeneratorBasicParams=InvokeAIGeneratorBasicParams(),
+                 ):
+        self.model_info=model_info
+        self.params=params
+
+    def generate(self,
+                 prompt: str='',
+                 callback: callable=None,
+                 step_callback: callable=None,
+                 iterations: int=1,
+                 **keyword_args,
+                 )->Iterator[InvokeAIGeneratorOutput]:
+        '''
+        Return an iterator across the indicated number of generations.
+        Each time the iterator is called it will return an InvokeAIGeneratorOutput
+        object. Use like this:
+
+           outputs = txt2img.generate(prompt='banana sushi', iterations=5)
+           for result in outputs:
+               print(result.image, result.seed)
+
+        In the typical case of wanting to get just a single image, iterations
+        defaults to 1 and do:
+
+           output = next(txt2img.generate(prompt='banana sushi')
+
+        Pass None to get an infinite iterator.
+
+           outputs = txt2img.generate(prompt='banana sushi', iterations=None)
+           for o in outputs:
+               print(o.image, o.seed)
+
+        '''
+        generator_args = dataclasses.asdict(self.params)
+        generator_args.update(keyword_args)
+
+        model_info = self.model_info
+        model_name = model_info['model_name']
+        model:StableDiffusionGeneratorPipeline = model_info['model']
+        model_hash = model_info['hash']
+        scheduler: Scheduler = self.get_scheduler(
+            model=model,
+            scheduler_name=generator_args.get('scheduler')
+        )
+        uc, c, extra_conditioning_info = get_uc_and_c_and_ec(prompt,model=model)
+        gen_class = self._generator_class()
+        generator = gen_class(model, self.params.precision)
+        if self.params.variation_amount > 0:
+            generator.set_variation(generator_args.get('seed'),
+                                    generator_args.get('variation_amount'),
+                                    generator_args.get('with_variations')
+                                    )
+
+        if isinstance(model, DiffusionPipeline):
+            for component in [model.unet, model.vae]:
+                configure_model_padding(component,
+                                        generator_args.get('seamless',False),
+                                        generator_args.get('seamless_axes')
+                                        )
+        else:
+            configure_model_padding(model,
+                                    generator_args.get('seamless',False),
+                                    generator_args.get('seamless_axes')
+                                    )
+
+        iteration_count = range(iterations) if iterations else itertools.count(start=0, step=1)
+        for i in iteration_count:
+            results = generator.generate(prompt,
+                                         conditioning=(uc, c, extra_conditioning_info),
+                                         sampler=scheduler,
+                                         **generator_args,
+                                         )
+            output = InvokeAIGeneratorOutput(
+                image=results[0][0],
+                seed=results[0][1],
+                attention_maps_images=results[0][2],
+                model_hash = model_hash,
+                params=Namespace(model_name=model_name,**generator_args),
+            )
+            if callback:
+                callback(output)
+            yield output
+
+    @classmethod
+    def schedulers(self)->List[str]:
+        '''
+        Return list of all the schedulers that we currently handle.
+        '''
+        return list(self.scheduler_map.keys())
+
+    def load_generator(self, model: StableDiffusionGeneratorPipeline, generator_class: Type[Generator]):
+        return generator_class(model, self.params.precision)
+
+    def get_scheduler(self, scheduler_name:str, model: StableDiffusionGeneratorPipeline)->Scheduler:
+        scheduler_class = self.scheduler_map.get(scheduler_name,'ddim')
+        scheduler = scheduler_class.from_config(model.scheduler.config)
+        # hack copied over from generate.py
+        if not hasattr(scheduler, 'uses_inpainting_model'):
+            scheduler.uses_inpainting_model = lambda: False
+        return scheduler
+
+    @classmethod
+    def _generator_class(cls)->Type[Generator]:
+        '''
+        In derived classes return the name of the generator to apply.
+        If you don't override will return the name of the derived
+        class, which nicely parallels the generator class names.
+        '''
+        return Generator
+
+# ------------------------------------
+class Txt2Img(InvokeAIGenerator):
+    @classmethod
+    def _generator_class(cls):
+        from .txt2img import Txt2Img
+        return Txt2Img
+
+# ------------------------------------
+class Img2Img(InvokeAIGenerator):
+    def generate(self,
+               init_image: Image | torch.FloatTensor,
+               strength: float=0.75,
+               **keyword_args
+               )->List[InvokeAIGeneratorOutput]:
+        return super().generate(init_image=init_image,
+                                strength=strength,
+                                **keyword_args
+                                )
+    @classmethod
+    def _generator_class(cls):
+        from .img2img import Img2Img
+        return Img2Img
+
+# ------------------------------------
+# Takes all the arguments of Img2Img and adds the mask image and the seam/infill stuff
+class Inpaint(Img2Img):
+    def generate(self,
+                 mask_image: Image | torch.FloatTensor,
+                 # Seam settings - when 0, doesn't fill seam
+                 seam_size: int = 0,
+                 seam_blur: int = 0,
+                 seam_strength: float = 0.7,
+                 seam_steps: int = 10,
+                 tile_size: int = 32,
+                 inpaint_replace=False,
+                 infill_method=None,
+                 inpaint_width=None,
+                 inpaint_height=None,
+                 inpaint_fill: tuple(int) = (0x7F, 0x7F, 0x7F, 0xFF),
+                 **keyword_args
+                 )->List[InvokeAIGeneratorOutput]:
+        return super().generate(
+            mask_image=mask_image,
+            seam_size=seam_size,
+            seam_blur=seam_blur,
+            seam_strength=seam_strength,
+            seam_steps=seam_steps,
+            tile_size=tile_size,
+            inpaint_replace=inpaint_replace,
+            infill_method=infill_method,
+            inpaint_width=inpaint_width,
+            inpaint_height=inpaint_height,
+            inpaint_fill=inpaint_fill,
+            **keyword_args
+        )
+    @classmethod
+    def _generator_class(cls):
+        from .inpaint import Inpaint
+        return Inpaint
+
+# ------------------------------------
+class Embiggen(Txt2Img):
+    def generate(
+            self,
+            embiggen: list=None,
+            embiggen_tiles: list = None,
+            strength: float=0.75,
+            **kwargs)->List[InvokeAIGeneratorOutput]:
+        return super().generate(embiggen=embiggen,
+                                embiggen_tiles=embiggen_tiles,
+                                strength=strength,
+                                **kwargs)
+
+    @classmethod
+    def _generator_class(cls):
+        from .embiggen import Embiggen
+        return Embiggen
+
+
+class Generator:
+    downsampling_factor: int
+    latent_channels: int
+    precision: str
+    model: DiffusionPipeline
+
+    def __init__(self, model: DiffusionPipeline, precision: str):
+        self.model = model
+        self.precision = precision
+        self.seed = None
+        self.latent_channels = model.channels
+        self.downsampling_factor = downsampling  # BUG: should come from model or config
+        self.safety_checker = None
+        self.perlin = 0.0
+        self.threshold = 0
+        self.variation_amount = 0
+        self.with_variations = []
+        self.use_mps_noise = False
+        self.free_gpu_mem = None
+
+    # this is going to be overridden in img2img.py, txt2img.py and inpaint.py
+    def get_make_image(self, prompt, **kwargs):
+        """
+        Returns a function returning an image derived from the prompt and the initial image
+        Return value depends on the seed at the time you call it
+        """
+        raise NotImplementedError(
+            "image_iterator() must be implemented in a descendent class"
+        )
+
+    def set_variation(self, seed, variation_amount, with_variations):
+        self.seed = seed
+        self.variation_amount = variation_amount
+        self.with_variations = with_variations
+
+    def generate(
+        self,
+        prompt,
+        width,
+        height,
+        sampler,
+        init_image=None,
+        iterations=1,
+        seed=None,
+        image_callback=None,
+        step_callback=None,
+        threshold=0.0,
+        perlin=0.0,
+        h_symmetry_time_pct=None,
+        v_symmetry_time_pct=None,
+        safety_checker: SafetyChecker=None,
+        free_gpu_mem: bool = False,
+        **kwargs,
+    ):
+        scope = nullcontext
+        self.safety_checker = safety_checker
+        self.free_gpu_mem = free_gpu_mem
+        attention_maps_images = []
+        attention_maps_callback = lambda saver: attention_maps_images.append(
+            saver.get_stacked_maps_image()
+        )
+        make_image = self.get_make_image(
+            prompt,
+            sampler=sampler,
+            init_image=init_image,
+            width=width,
+            height=height,
+            step_callback=step_callback,
+            threshold=threshold,
+            perlin=perlin,
+            h_symmetry_time_pct=h_symmetry_time_pct,
+            v_symmetry_time_pct=v_symmetry_time_pct,
+            attention_maps_callback=attention_maps_callback,
+            **kwargs,
+        )
+        results = []
+        seed = seed if seed is not None and seed >= 0 else self.new_seed()
+        first_seed = seed
+        seed, initial_noise = self.generate_initial_noise(seed, width, height)
+
+        # There used to be an additional self.model.ema_scope() here, but it breaks
+        # the inpaint-1.5 model. Not sure what it did.... ?
+        with scope(self.model.device.type):
+            for n in trange(iterations, desc="Generating"):
+                x_T = None
+                if self.variation_amount > 0:
+                    set_seed(seed)
+                    target_noise = self.get_noise(width, height)
+                    x_T = self.slerp(self.variation_amount, initial_noise, target_noise)
+                elif initial_noise is not None:
+                    # i.e. we specified particular variations
+                    x_T = initial_noise
+                else:
+                    set_seed(seed)
+                    try:
+                        x_T = self.get_noise(width, height)
+                    except:
+                        print("** An error occurred while getting initial noise **")
+                        print(traceback.format_exc())
+
+                # Pass on the seed in case a layer beneath us needs to generate noise on its own.
+                image = make_image(x_T, seed)
+
+                if self.safety_checker is not None:
+                    image = self.safety_checker.check(image)
+
+                results.append([image, seed, attention_maps_images])
+
+                if image_callback is not None:
+                    attention_maps_image = (
+                        None
+                        if len(attention_maps_images) == 0
+                        else attention_maps_images[-1]
+                    )
+                    image_callback(
+                        image,
+                        seed,
+                        first_seed=first_seed,
+                        attention_maps_image=attention_maps_image,
+                    )
+
+                seed = self.new_seed()
+
+                # Free up memory from the last generation.
+                clear_cuda_cache = (
+                    kwargs["clear_cuda_cache"] if "clear_cuda_cache" in kwargs else None
+                )
+                if clear_cuda_cache is not None:
+                    clear_cuda_cache()
+
+        return results
+
+    def sample_to_image(self, samples) -> Image.Image:
+        """
+        Given samples returned from a sampler, converts
+        it into a PIL Image
+        """
+        with torch.inference_mode():
+            image = self.model.decode_latents(samples)
+        return self.model.numpy_to_pil(image)[0]
+
+    def repaste_and_color_correct(
+        self,
+        result: Image.Image,
+        init_image: Image.Image,
+        init_mask: Image.Image,
+        mask_blur_radius: int = 8,
+    ) -> Image.Image:
+        if init_image is None or init_mask is None:
+            return result
+
+        # Get the original alpha channel of the mask if there is one.
+        # Otherwise it is some other black/white image format ('1', 'L' or 'RGB')
+        pil_init_mask = (
+            init_mask.getchannel("A")
+            if init_mask.mode == "RGBA"
+            else init_mask.convert("L")
+        )
+        pil_init_image = init_image.convert(
+            "RGBA"
+        )  # Add an alpha channel if one doesn't exist
+
+        # Build an image with only visible pixels from source to use as reference for color-matching.
+        init_rgb_pixels = np.asarray(init_image.convert("RGB"), dtype=np.uint8)
+        init_a_pixels = np.asarray(pil_init_image.getchannel("A"), dtype=np.uint8)
+        init_mask_pixels = np.asarray(pil_init_mask, dtype=np.uint8)
+
+        # Get numpy version of result
+        np_image = np.asarray(result, dtype=np.uint8)
+
+        # Mask and calculate mean and standard deviation
+        mask_pixels = init_a_pixels * init_mask_pixels > 0
+        np_init_rgb_pixels_masked = init_rgb_pixels[mask_pixels, :]
+        np_image_masked = np_image[mask_pixels, :]
+
+        if np_init_rgb_pixels_masked.size > 0:
+            init_means = np_init_rgb_pixels_masked.mean(axis=0)
+            init_std = np_init_rgb_pixels_masked.std(axis=0)
+            gen_means = np_image_masked.mean(axis=0)
+            gen_std = np_image_masked.std(axis=0)
+
+            # Color correct
+            np_matched_result = np_image.copy()
+            np_matched_result[:, :, :] = (
+                (
+                    (
+                        (
+                            np_matched_result[:, :, :].astype(np.float32)
+                            - gen_means[None, None, :]
+                        )
+                        / gen_std[None, None, :]
+                    )
+                    * init_std[None, None, :]
+                    + init_means[None, None, :]
+                )
+                .clip(0, 255)
+                .astype(np.uint8)
+            )
+            matched_result = Image.fromarray(np_matched_result, mode="RGB")
+        else:
+            matched_result = Image.fromarray(np_image, mode="RGB")
+
+        # Blur the mask out (into init image) by specified amount
+        if mask_blur_radius > 0:
+            nm = np.asarray(pil_init_mask, dtype=np.uint8)
+            nmd = cv2.erode(
+                nm,
+                kernel=np.ones((3, 3), dtype=np.uint8),
+                iterations=int(mask_blur_radius / 2),
+            )
+            pmd = Image.fromarray(nmd, mode="L")
+            blurred_init_mask = pmd.filter(ImageFilter.BoxBlur(mask_blur_radius))
+        else:
+            blurred_init_mask = pil_init_mask
+
+        multiplied_blurred_init_mask = ImageChops.multiply(
+            blurred_init_mask, self.pil_image.split()[-1]
+        )
+
+        # Paste original on color-corrected generation (using blurred mask)
+        matched_result.paste(init_image, (0, 0), mask=multiplied_blurred_init_mask)
+        return matched_result
+
+    @staticmethod
+    def sample_to_lowres_estimated_image(samples):
+        # origingally adapted from code by @erucipe and @keturn here:
+        # https://discuss.huggingface.co/t/decoding-latents-to-rgb-without-upscaling/23204/7
+
+        # these updated numbers for v1.5 are from @torridgristle
+        v1_5_latent_rgb_factors = torch.tensor(
+            [
+                #    R        G        B
+                [0.3444, 0.1385, 0.0670],  # L1
+                [0.1247, 0.4027, 0.1494],  # L2
+                [-0.3192, 0.2513, 0.2103],  # L3
+                [-0.1307, -0.1874, -0.7445],  # L4
+            ],
+            dtype=samples.dtype,
+            device=samples.device,
+        )
+
+        latent_image = samples[0].permute(1, 2, 0) @ v1_5_latent_rgb_factors
+        latents_ubyte = (
+            ((latent_image + 1) / 2)
+            .clamp(0, 1)  # change scale from -1..1 to 0..1
+            .mul(0xFF)  # to 0..255
+            .byte()
+        ).cpu()
+
+        return Image.fromarray(latents_ubyte.numpy())
+
+    def generate_initial_noise(self, seed, width, height):
+        initial_noise = None
+        if self.variation_amount > 0 or len(self.with_variations) > 0:
+            # use fixed initial noise plus random noise per iteration
+            set_seed(seed)
+            initial_noise = self.get_noise(width, height)
+            for v_seed, v_weight in self.with_variations:
+                seed = v_seed
+                set_seed(seed)
+                next_noise = self.get_noise(width, height)
+                initial_noise = self.slerp(v_weight, initial_noise, next_noise)
+            if self.variation_amount > 0:
+                random.seed()  # reset RNG to an actually random state, so we can get a random seed for variations
+                seed = random.randrange(0, np.iinfo(np.uint32).max)
+        return (seed, initial_noise)
+
+    def get_perlin_noise(self, width, height):
+        fixdevice = "cpu" if (self.model.device.type == "mps") else self.model.device
+        # limit noise to only the diffusion image channels, not the mask channels
+        input_channels = min(self.latent_channels, 4)
+        # round up to the nearest block of 8
+        temp_width = int((width + 7) / 8) * 8
+        temp_height = int((height + 7) / 8) * 8
+        noise = torch.stack(
+            [
+                rand_perlin_2d(
+                    (temp_height, temp_width), (8, 8), device=self.model.device
+                ).to(fixdevice)
+                for _ in range(input_channels)
+            ],
+            dim=0,
+        ).to(self.model.device)
+        return noise[0:4, 0:height, 0:width]
+
+    def new_seed(self):
+        self.seed = random.randrange(0, np.iinfo(np.uint32).max)
+        return self.seed
+
+    def slerp(self, t, v0, v1, DOT_THRESHOLD=0.9995):
+        """
+        Spherical linear interpolation
+        Args:
+            t (float/np.ndarray): Float value between 0.0 and 1.0
+            v0 (np.ndarray): Starting vector
+            v1 (np.ndarray): Final vector
+            DOT_THRESHOLD (float): Threshold for considering the two vectors as
+                                colineal. Not recommended to alter this.
+        Returns:
+            v2 (np.ndarray): Interpolation vector between v0 and v1
+        """
+        inputs_are_torch = False
+        if not isinstance(v0, np.ndarray):
+            inputs_are_torch = True
+            v0 = v0.detach().cpu().numpy()
+        if not isinstance(v1, np.ndarray):
+            inputs_are_torch = True
+            v1 = v1.detach().cpu().numpy()
+
+        dot = np.sum(v0 * v1 / (np.linalg.norm(v0) * np.linalg.norm(v1)))
+        if np.abs(dot) > DOT_THRESHOLD:
+            v2 = (1 - t) * v0 + t * v1
+        else:
+            theta_0 = np.arccos(dot)
+            sin_theta_0 = np.sin(theta_0)
+            theta_t = theta_0 * t
+            sin_theta_t = np.sin(theta_t)
+            s0 = np.sin(theta_0 - theta_t) / sin_theta_0
+            s1 = sin_theta_t / sin_theta_0
+            v2 = s0 * v0 + s1 * v1
+
+        if inputs_are_torch:
+            v2 = torch.from_numpy(v2).to(self.model.device)
+
+        return v2
+
+    # this is a handy routine for debugging use. Given a generated sample,
+    # convert it into a PNG image and store it at the indicated path
+    def save_sample(self, sample, filepath):
+        image = self.sample_to_image(sample)
+        dirname = os.path.dirname(filepath) or "."
+        if not os.path.exists(dirname):
+            print(f"** creating directory {dirname}")
+            os.makedirs(dirname, exist_ok=True)
+        image.save(filepath, "PNG")
+
+    def torch_dtype(self) -> torch.dtype:
+        return torch.float16 if self.precision == "float16" else torch.float32
+
+    # returns a tensor filled with random numbers from a normal distribution
+    def get_noise(self, width, height):
+        device = self.model.device
+        # limit noise to only the diffusion image channels, not the mask channels
+        input_channels = min(self.latent_channels, 4)
+        if self.use_mps_noise or device.type == "mps":
+            x = torch.randn(
+                [
+                    1,
+                    input_channels,
+                    height // self.downsampling_factor,
+                    width // self.downsampling_factor,
+                ],
+                dtype=self.torch_dtype(),
+                device="cpu",
+            ).to(device)
+        else:
+            x = torch.randn(
+                [
+                    1,
+                    input_channels,
+                    height // self.downsampling_factor,
+                    width // self.downsampling_factor,
+                ],
+                dtype=self.torch_dtype(),
+                device=device,
+            )
+        if self.perlin > 0.0:
+            perlin_noise = self.get_perlin_noise(
+                width // self.downsampling_factor, height // self.downsampling_factor
+            )
+            x = (1 - self.perlin) * x + self.perlin * perlin_noise
+        return x

invokeai/backend/generator/embiggen.py

@@ -1,37 +1,38 @@
-'''
-ldm.invoke.generator.embiggen descends from ldm.invoke.generator
-and generates with ldm.invoke.generator.img2img
-'''
+"""
+invokeai.backend.generator.embiggen descends from .generator
+and generates with .generator.img2img
+"""
 
 import numpy as np
 import torch
 from PIL import Image
 from tqdm import trange
 
-from ldm.invoke.generator.base import Generator
-from ldm.invoke.generator.img2img import Img2Img
+from .base import Generator
+from .img2img import Img2Img
 
 
 class Embiggen(Generator):
     def __init__(self, model, precision):
         super().__init__(model, precision)
-        self.init_latent         = None
+        self.init_latent = None
 
     # Replace generate because Embiggen doesn't need/use most of what it does normallly
-    def generate(self,prompt,iterations=1,seed=None,
-                 image_callback=None, step_callback=None,
-                 **kwargs):
-
-        make_image = self.get_make_image(
-            prompt,
-            step_callback = step_callback,
-            **kwargs
-        )
-        results             = []
-        seed                = seed if seed else self.new_seed()
+    def generate(
+        self,
+        prompt,
+        iterations=1,
+        seed=None,
+        image_callback=None,
+        step_callback=None,
+        **kwargs,
+    ):
+        make_image = self.get_make_image(prompt, step_callback=step_callback, **kwargs)
+        results = []
+        seed = seed if seed else self.new_seed()
 
         # Noise will be generated by the Img2Img generator when called
-        for _ in trange(iterations, desc='Generating'):
+        for _ in trange(iterations, desc="Generating"):
             # make_image will call Img2Img which will do the equivalent of get_noise itself
             image = make_image()
             results.append([image, seed])
@@ -56,13 +57,15 @@ class Embiggen(Generator):
         embiggen,
         embiggen_tiles,
         step_callback=None,
-        **kwargs
+        **kwargs,
     ):
         """
         Returns a function returning an image derived from the prompt and multi-stage twice-baked potato layering over the img2img on the initial image
         Return value depends on the seed at the time you call it
         """
-        assert not sampler.uses_inpainting_model(), "--embiggen is not supported by inpainting models"
+        assert (
+            not sampler.uses_inpainting_model()
+        ), "--embiggen is not supported by inpainting models"
 
         # Construct embiggen arg array, and sanity check arguments
         if embiggen == None:  # embiggen can also be called with just embiggen_tiles
@@ -70,48 +73,57 @@ class Embiggen(Generator):
         elif embiggen[0] < 0:
             embiggen[0] = 1.0
             print(
-                '>> Embiggen scaling factor cannot be negative, fell back to the default of 1.0 !')
+                ">> Embiggen scaling factor cannot be negative, fell back to the default of 1.0 !"
+            )
         if len(embiggen) < 2:
             embiggen.append(0.75)
         elif embiggen[1] > 1.0 or embiggen[1] < 0:
             embiggen[1] = 0.75
-            print('>> Embiggen upscaling strength for ESRGAN must be between 0 and 1, fell back to the default of 0.75 !')
+            print(
+                ">> Embiggen upscaling strength for ESRGAN must be between 0 and 1, fell back to the default of 0.75 !"
+            )
         if len(embiggen) < 3:
             embiggen.append(0.25)
         elif embiggen[2] < 0:
             embiggen[2] = 0.25
-            print('>> Overlap size for Embiggen must be a positive ratio between 0 and 1 OR a number of pixels, fell back to the default of 0.25 !')
+            print(
+                ">> Overlap size for Embiggen must be a positive ratio between 0 and 1 OR a number of pixels, fell back to the default of 0.25 !"
+            )
 
         # Convert tiles from their user-freindly count-from-one to count-from-zero, because we need to do modulo math
         # and then sort them, because... people.
         if embiggen_tiles:
-            embiggen_tiles = list(map(lambda n: n-1, embiggen_tiles))
+            embiggen_tiles = list(map(lambda n: n - 1, embiggen_tiles))
             embiggen_tiles.sort()
 
         if strength >= 0.5:
-            print(f'* WARNING: Embiggen may produce mirror motifs if the strength (-f) is too high (currently {strength}). Try values between 0.35-0.45.')
+            print(
+                f"* WARNING: Embiggen may produce mirror motifs if the strength (-f) is too high (currently {strength}). Try values between 0.35-0.45."
+            )
 
         # Prep img2img generator, since we wrap over it
-        gen_img2img = Img2Img(self.model,self.precision)
+        gen_img2img = Img2Img(self.model, self.precision)
 
         # Open original init image (not a tensor) to manipulate
         initsuperimage = Image.open(init_img)
 
         with Image.open(init_img) as img:
-            initsuperimage = img.convert('RGB')
+            initsuperimage = img.convert("RGB")
 
         # Size of the target super init image in pixels
         initsuperwidth, initsuperheight = initsuperimage.size
 
         # Increase by scaling factor if not already resized, using ESRGAN as able
         if embiggen[0] != 1.0:
-            initsuperwidth = round(initsuperwidth*embiggen[0])
-            initsuperheight = round(initsuperheight*embiggen[0])
+            initsuperwidth = round(initsuperwidth * embiggen[0])
+            initsuperheight = round(initsuperheight * embiggen[0])
             if embiggen[1] > 0:  # No point in ESRGAN upscaling if strength is set zero
-                from ldm.invoke.restoration.realesrgan import ESRGAN
+                from ..restoration.realesrgan import ESRGAN
+
                 esrgan = ESRGAN()
                 print(
-                    f'>> ESRGAN upscaling init image prior to cutting with Embiggen with strength {embiggen[1]}')
+                    f">> ESRGAN upscaling init image prior to cutting with Embiggen with strength {embiggen[1]}"
+                )
                 if embiggen[0] > 2:
                     initsuperimage = esrgan.process(
                         initsuperimage,
@@ -130,7 +142,8 @@ class Embiggen(Generator):
             #   but from personal experiance it doesn't greatly improve anything after 4x
             # Resize to target scaling factor resolution
             initsuperimage = initsuperimage.resize(
-                (initsuperwidth, initsuperheight), Image.Resampling.LANCZOS)
+                (initsuperwidth, initsuperheight), Image.Resampling.LANCZOS
+            )
 
         # Use width and height as tile widths and height
         # Determine buffer size in pixels
@@ -153,23 +166,24 @@ class Embiggen(Generator):
         emb_tiles_x = 1
         emb_tiles_y = 1
         if (initsuperwidth - width) > 0:
-            emb_tiles_x = ceildiv(initsuperwidth - width,
-                                  width - overlap_size_x) + 1
+            emb_tiles_x = ceildiv(initsuperwidth - width, width - overlap_size_x) + 1
         if (initsuperheight - height) > 0:
-            emb_tiles_y = ceildiv(initsuperheight - height,
-                                  height - overlap_size_y) + 1
+            emb_tiles_y = ceildiv(initsuperheight - height, height - overlap_size_y) + 1
         # Sanity
-        assert emb_tiles_x > 1 or emb_tiles_y > 1, f'ERROR: Based on the requested dimensions of {initsuperwidth}x{initsuperheight} and tiles of {width}x{height} you don\'t need to Embiggen! Check your arguments.'
+        assert (
+            emb_tiles_x > 1 or emb_tiles_y > 1
+        ), f"ERROR: Based on the requested dimensions of {initsuperwidth}x{initsuperheight} and tiles of {width}x{height} you don't need to Embiggen! Check your arguments."
 
         # Prep alpha layers --------------
         # https://stackoverflow.com/questions/69321734/how-to-create-different-transparency-like-gradient-with-python-pil
         # agradientL is Left-side transparent
-        agradientL = Image.linear_gradient('L').rotate(
-            90).resize((overlap_size_x, height))
+        agradientL = (
+            Image.linear_gradient("L").rotate(90).resize((overlap_size_x, height))
+        )
         # agradientT is Top-side transparent
-        agradientT = Image.linear_gradient('L').resize((width, overlap_size_y))
+        agradientT = Image.linear_gradient("L").resize((width, overlap_size_y))
         # radial corner is the left-top corner, made full circle then cut to just the left-top quadrant
-        agradientC = Image.new('L', (256, 256))
+        agradientC = Image.new("L", (256, 256))
         for y in range(256):
             for x in range(256):
                 # Find distance to lower right corner (numpy takes arrays)
@@ -177,16 +191,16 @@ class Embiggen(Generator):
                 # Clamp values to max 255
                 if distanceToLR > 255:
                     distanceToLR = 255
-                #Place the pixel as invert of distance
+                # Place the pixel as invert of distance
                 agradientC.putpixel((x, y), round(255 - distanceToLR))
 
         # Create alternative asymmetric diagonal corner to use on "tailing" intersections to prevent hard edges
         # Fits for a left-fading gradient on the bottom side and full opacity on the right side.
-        agradientAsymC = Image.new('L', (256, 256))
+        agradientAsymC = Image.new("L", (256, 256))
         for y in range(256):
             for x in range(256):
-                value = round(max(0, x-(255-y)) * (255 / max(1,y)))
-                #Clamp values
+                value = round(max(0, x - (255 - y)) * (255 / max(1, y)))
+                # Clamp values
                 value = max(0, value)
                 value = min(255, value)
                 agradientAsymC.putpixel((x, y), value)
@@ -204,80 +218,91 @@ class Embiggen(Generator):
         # make masks with an asymmetric upper-right corner so when the curved transparent corner of the next tile
         # to its right is placed it doesn't reveal a hard trailing semi-transparent edge in the overlapping space
         alphaLayerTaC = alphaLayerT.copy()
-        alphaLayerTaC.paste(agradientAsymC.rotate(270).resize((overlap_size_x, overlap_size_y)), (width - overlap_size_x, 0))
+        alphaLayerTaC.paste(
+            agradientAsymC.rotate(270).resize((overlap_size_x, overlap_size_y)),
+            (width - overlap_size_x, 0),
+        )
         alphaLayerLTaC = alphaLayerLTC.copy()
-        alphaLayerLTaC.paste(agradientAsymC.rotate(270).resize((overlap_size_x, overlap_size_y)), (width - overlap_size_x, 0))
+        alphaLayerLTaC.paste(
+            agradientAsymC.rotate(270).resize((overlap_size_x, overlap_size_y)),
+            (width - overlap_size_x, 0),
+        )
 
         if embiggen_tiles:
             # Individual unconnected sides
             alphaLayerR = Image.new("L", (width, height), 255)
-            alphaLayerR.paste(agradientL.rotate(
-                180), (width - overlap_size_x, 0))
+            alphaLayerR.paste(agradientL.rotate(180), (width - overlap_size_x, 0))
             alphaLayerB = Image.new("L", (width, height), 255)
-            alphaLayerB.paste(agradientT.rotate(
-                180), (0, height - overlap_size_y))
+            alphaLayerB.paste(agradientT.rotate(180), (0, height - overlap_size_y))
             alphaLayerTB = Image.new("L", (width, height), 255)
             alphaLayerTB.paste(agradientT, (0, 0))
-            alphaLayerTB.paste(agradientT.rotate(
-                180), (0, height - overlap_size_y))
+            alphaLayerTB.paste(agradientT.rotate(180), (0, height - overlap_size_y))
             alphaLayerLR = Image.new("L", (width, height), 255)
             alphaLayerLR.paste(agradientL, (0, 0))
-            alphaLayerLR.paste(agradientL.rotate(
-                180), (width - overlap_size_x, 0))
+            alphaLayerLR.paste(agradientL.rotate(180), (width - overlap_size_x, 0))
 
             # Sides and corner Layers
             alphaLayerRBC = Image.new("L", (width, height), 255)
-            alphaLayerRBC.paste(agradientL.rotate(
-                180), (width - overlap_size_x, 0))
-            alphaLayerRBC.paste(agradientT.rotate(
-                180), (0, height - overlap_size_y))
-            alphaLayerRBC.paste(agradientC.rotate(180).resize(
-                (overlap_size_x, overlap_size_y)), (width - overlap_size_x, height - overlap_size_y))
+            alphaLayerRBC.paste(agradientL.rotate(180), (width - overlap_size_x, 0))
+            alphaLayerRBC.paste(agradientT.rotate(180), (0, height - overlap_size_y))
+            alphaLayerRBC.paste(
+                agradientC.rotate(180).resize((overlap_size_x, overlap_size_y)),
+                (width - overlap_size_x, height - overlap_size_y),
+            )
             alphaLayerLBC = Image.new("L", (width, height), 255)
             alphaLayerLBC.paste(agradientL, (0, 0))
-            alphaLayerLBC.paste(agradientT.rotate(
-                180), (0, height - overlap_size_y))
-            alphaLayerLBC.paste(agradientC.rotate(90).resize(
-                (overlap_size_x, overlap_size_y)), (0, height - overlap_size_y))
+            alphaLayerLBC.paste(agradientT.rotate(180), (0, height - overlap_size_y))
+            alphaLayerLBC.paste(
+                agradientC.rotate(90).resize((overlap_size_x, overlap_size_y)),
+                (0, height - overlap_size_y),
+            )
             alphaLayerRTC = Image.new("L", (width, height), 255)
-            alphaLayerRTC.paste(agradientL.rotate(
-                180), (width - overlap_size_x, 0))
+            alphaLayerRTC.paste(agradientL.rotate(180), (width - overlap_size_x, 0))
             alphaLayerRTC.paste(agradientT, (0, 0))
-            alphaLayerRTC.paste(agradientC.rotate(270).resize(
-                (overlap_size_x, overlap_size_y)), (width - overlap_size_x, 0))
+            alphaLayerRTC.paste(
+                agradientC.rotate(270).resize((overlap_size_x, overlap_size_y)),
+                (width - overlap_size_x, 0),
+            )
 
             # All but X layers
             alphaLayerABT = Image.new("L", (width, height), 255)
             alphaLayerABT.paste(alphaLayerLBC, (0, 0))
-            alphaLayerABT.paste(agradientL.rotate(
-                180), (width - overlap_size_x, 0))
-            alphaLayerABT.paste(agradientC.rotate(180).resize(
-                (overlap_size_x, overlap_size_y)), (width - overlap_size_x, height - overlap_size_y))
+            alphaLayerABT.paste(agradientL.rotate(180), (width - overlap_size_x, 0))
+            alphaLayerABT.paste(
+                agradientC.rotate(180).resize((overlap_size_x, overlap_size_y)),
+                (width - overlap_size_x, height - overlap_size_y),
+            )
             alphaLayerABL = Image.new("L", (width, height), 255)
             alphaLayerABL.paste(alphaLayerRTC, (0, 0))
-            alphaLayerABL.paste(agradientT.rotate(
-                180), (0, height - overlap_size_y))
-            alphaLayerABL.paste(agradientC.rotate(180).resize(
-                (overlap_size_x, overlap_size_y)), (width - overlap_size_x, height - overlap_size_y))
+            alphaLayerABL.paste(agradientT.rotate(180), (0, height - overlap_size_y))
+            alphaLayerABL.paste(
+                agradientC.rotate(180).resize((overlap_size_x, overlap_size_y)),
+                (width - overlap_size_x, height - overlap_size_y),
+            )
             alphaLayerABR = Image.new("L", (width, height), 255)
             alphaLayerABR.paste(alphaLayerLBC, (0, 0))
             alphaLayerABR.paste(agradientT, (0, 0))
-            alphaLayerABR.paste(agradientC.resize(
-                (overlap_size_x, overlap_size_y)), (0, 0))
+            alphaLayerABR.paste(
+                agradientC.resize((overlap_size_x, overlap_size_y)), (0, 0)
+            )
             alphaLayerABB = Image.new("L", (width, height), 255)
             alphaLayerABB.paste(alphaLayerRTC, (0, 0))
             alphaLayerABB.paste(agradientL, (0, 0))
-            alphaLayerABB.paste(agradientC.resize(
-                (overlap_size_x, overlap_size_y)), (0, 0))
+            alphaLayerABB.paste(
+                agradientC.resize((overlap_size_x, overlap_size_y)), (0, 0)
+            )
 
             # All-around layer
             alphaLayerAA = Image.new("L", (width, height), 255)
             alphaLayerAA.paste(alphaLayerABT, (0, 0))
             alphaLayerAA.paste(agradientT, (0, 0))
-            alphaLayerAA.paste(agradientC.resize(
-                (overlap_size_x, overlap_size_y)), (0, 0))
-            alphaLayerAA.paste(agradientC.rotate(270).resize(
-                (overlap_size_x, overlap_size_y)), (width - overlap_size_x, 0))
+            alphaLayerAA.paste(
+                agradientC.resize((overlap_size_x, overlap_size_y)), (0, 0)
+            )
+            alphaLayerAA.paste(
+                agradientC.rotate(270).resize((overlap_size_x, overlap_size_y)),
+                (width - overlap_size_x, 0),
+            )
 
         # Clean up temporary gradients
         del agradientL
@@ -287,17 +312,20 @@ class Embiggen(Generator):
         def make_image():
             # Make main tiles -------------------------------------------------
             if embiggen_tiles:
-                print(f'>> Making {len(embiggen_tiles)} Embiggen tiles...')
+                print(f">> Making {len(embiggen_tiles)} Embiggen tiles...")
             else:
                 print(
-                    f'>> Making {(emb_tiles_x * emb_tiles_y)} Embiggen tiles ({emb_tiles_x}x{emb_tiles_y})...')
+                    f">> Making {(emb_tiles_x * emb_tiles_y)} Embiggen tiles ({emb_tiles_x}x{emb_tiles_y})..."
+                )
 
             emb_tile_store = []
             # Although we could use the same seed for every tile for determinism, at higher strengths this may
             # produce duplicated structures for each tile and make the tiling effect more obvious
             # instead track and iterate a local seed we pass to Img2Img
             seed = self.seed
-            seedintlimit = np.iinfo(np.uint32).max - 1 # only retreive this one from numpy
+            seedintlimit = (
+                np.iinfo(np.uint32).max - 1
+            )  # only retreive this one from numpy
 
             for tile in range(emb_tiles_x * emb_tiles_y):
                 # Don't iterate on first tile
@@ -334,37 +362,38 @@ class Embiggen(Generator):
 
                 if embiggen_tiles:
                     print(
-                        f'Making tile #{tile + 1} ({embiggen_tiles.index(tile) + 1} of {len(embiggen_tiles)} requested)')
+                        f"Making tile #{tile + 1} ({embiggen_tiles.index(tile) + 1} of {len(embiggen_tiles)} requested)"
+                    )
                 else:
-                    print(
-                        f'Starting {tile + 1} of {(emb_tiles_x * emb_tiles_y)} tiles')
+                    print(f"Starting {tile + 1} of {(emb_tiles_x * emb_tiles_y)} tiles")
 
                 # create a torch tensor from an Image
-                newinitimage = np.array(
-                    newinitimage).astype(np.float32) / 255.0
+                newinitimage = np.array(newinitimage).astype(np.float32) / 255.0
                 newinitimage = newinitimage[None].transpose(0, 3, 1, 2)
                 newinitimage = torch.from_numpy(newinitimage)
                 newinitimage = 2.0 * newinitimage - 1.0
                 newinitimage = newinitimage.to(self.model.device)
-                clear_cuda_cache = kwargs['clear_cuda_cache'] if 'clear_cuda_cache' in kwargs else None
+                clear_cuda_cache = (
+                    kwargs["clear_cuda_cache"] if "clear_cuda_cache" in kwargs else None
+                )
 
                 tile_results = gen_img2img.generate(
                     prompt,
-                    iterations     = 1,
-                    seed           = seed,
-                    sampler        = sampler,
-                    steps          = steps,
-                    cfg_scale      = cfg_scale,
-                    conditioning   = conditioning,
-                    ddim_eta       = ddim_eta,
-                    image_callback = None,  # called only after the final image is generated
-                    step_callback  = step_callback,   # called after each intermediate image is generated
-                    width          = width,
-                    height         = height,
-                    init_image     = newinitimage,    # notice that init_image is different from init_img
-                    mask_image     = None,
-                    strength       = strength,
-                    clear_cuda_cache     = clear_cuda_cache
+                    iterations=1,
+                    seed=seed,
+                    sampler=sampler,
+                    steps=steps,
+                    cfg_scale=cfg_scale,
+                    conditioning=conditioning,
+                    ddim_eta=ddim_eta,
+                    image_callback=None,  # called only after the final image is generated
+                    step_callback=step_callback,  # called after each intermediate image is generated
+                    width=width,
+                    height=height,
+                    init_image=newinitimage,  # notice that init_image is different from init_img
+                    mask_image=None,
+                    strength=strength,
+                    clear_cuda_cache=clear_cuda_cache,
                 )
 
                 emb_tile_store.append(tile_results[0][0])
@@ -373,12 +402,14 @@ class Embiggen(Generator):
                 del newinitimage
 
             # Sanity check we have them all
-            if len(emb_tile_store) == (emb_tiles_x * emb_tiles_y) or (embiggen_tiles != [] and len(emb_tile_store) == len(embiggen_tiles)):
-                outputsuperimage = Image.new(
-                    "RGBA", (initsuperwidth, initsuperheight))
+            if len(emb_tile_store) == (emb_tiles_x * emb_tiles_y) or (
+                embiggen_tiles != [] and len(emb_tile_store) == len(embiggen_tiles)
+            ):
+                outputsuperimage = Image.new("RGBA", (initsuperwidth, initsuperheight))
                 if embiggen_tiles:
                     outputsuperimage.alpha_composite(
-                        initsuperimage.convert('RGBA'), (0, 0))
+                        initsuperimage.convert("RGBA"), (0, 0)
+                    )
                 for tile in range(emb_tiles_x * emb_tiles_y):
                     if embiggen_tiles:
                         if tile in embiggen_tiles:
@@ -387,7 +418,7 @@ class Embiggen(Generator):
                             continue
                     else:
                         intileimage = emb_tile_store[tile]
-                    intileimage = intileimage.convert('RGBA')
+                    intileimage = intileimage.convert("RGBA")
                     # Get row and column entries
                     emb_row_i = tile // emb_tiles_x
                     emb_column_i = tile % emb_tiles_x
@@ -399,8 +430,7 @@ class Embiggen(Generator):
                         if emb_column_i + 1 == emb_tiles_x:
                             left = initsuperwidth - width
                         else:
-                            left = round(emb_column_i *
-                                         (width - overlap_size_x))
+                            left = round(emb_column_i * (width - overlap_size_x))
                         if emb_row_i + 1 == emb_tiles_y:
                             top = initsuperheight - height
                         else:
@@ -411,33 +441,43 @@ class Embiggen(Generator):
                             # top of image
                             if emb_row_i == 0:
                                 if emb_column_i == 0:
-                                    if (tile+1) in embiggen_tiles:  # Look-ahead right
-                                        if (tile+emb_tiles_x) not in embiggen_tiles:  # Look-ahead down
+                                    if (tile + 1) in embiggen_tiles:  # Look-ahead right
+                                        if (
+                                            tile + emb_tiles_x
+                                        ) not in embiggen_tiles:  # Look-ahead down
                                             intileimage.putalpha(alphaLayerB)
                                         # Otherwise do nothing on this tile
-                                    elif (tile+emb_tiles_x) in embiggen_tiles:  # Look-ahead down only
+                                    elif (
+                                        tile + emb_tiles_x
+                                    ) in embiggen_tiles:  # Look-ahead down only
                                         intileimage.putalpha(alphaLayerR)
                                     else:
                                         intileimage.putalpha(alphaLayerRBC)
                                 elif emb_column_i == emb_tiles_x - 1:
-                                    if (tile+emb_tiles_x) in embiggen_tiles:  # Look-ahead down
+                                    if (
+                                        tile + emb_tiles_x
+                                    ) in embiggen_tiles:  # Look-ahead down
                                         intileimage.putalpha(alphaLayerL)
                                     else:
                                         intileimage.putalpha(alphaLayerLBC)
                                 else:
-                                    if (tile+1) in embiggen_tiles:  # Look-ahead right
-                                        if (tile+emb_tiles_x) in embiggen_tiles:  # Look-ahead down
+                                    if (tile + 1) in embiggen_tiles:  # Look-ahead right
+                                        if (
+                                            tile + emb_tiles_x
+                                        ) in embiggen_tiles:  # Look-ahead down
                                             intileimage.putalpha(alphaLayerL)
                                         else:
                                             intileimage.putalpha(alphaLayerLBC)
-                                    elif (tile+emb_tiles_x) in embiggen_tiles:  # Look-ahead down only
+                                    elif (
+                                        tile + emb_tiles_x
+                                    ) in embiggen_tiles:  # Look-ahead down only
                                         intileimage.putalpha(alphaLayerLR)
                                     else:
                                         intileimage.putalpha(alphaLayerABT)
                             # bottom of image
                             elif emb_row_i == emb_tiles_y - 1:
                                 if emb_column_i == 0:
-                                    if (tile+1) in embiggen_tiles: # Look-ahead right
+                                    if (tile + 1) in embiggen_tiles:  # Look-ahead right
                                         intileimage.putalpha(alphaLayerTaC)
                                     else:
                                         intileimage.putalpha(alphaLayerRTC)
@@ -445,34 +485,44 @@ class Embiggen(Generator):
                                     # No tiles to look ahead to
                                     intileimage.putalpha(alphaLayerLTC)
                                 else:
-                                    if (tile+1) in embiggen_tiles: # Look-ahead right
+                                    if (tile + 1) in embiggen_tiles:  # Look-ahead right
                                         intileimage.putalpha(alphaLayerLTaC)
                                     else:
                                         intileimage.putalpha(alphaLayerABB)
                             # vertical middle of image
                             else:
                                 if emb_column_i == 0:
-                                    if (tile+1) in embiggen_tiles: # Look-ahead right
-                                        if (tile+emb_tiles_x) in embiggen_tiles: # Look-ahead down
+                                    if (tile + 1) in embiggen_tiles:  # Look-ahead right
+                                        if (
+                                            tile + emb_tiles_x
+                                        ) in embiggen_tiles:  # Look-ahead down
                                             intileimage.putalpha(alphaLayerTaC)
                                         else:
                                             intileimage.putalpha(alphaLayerTB)
-                                    elif (tile+emb_tiles_x) in embiggen_tiles:  # Look-ahead down only
+                                    elif (
+                                        tile + emb_tiles_x
+                                    ) in embiggen_tiles:  # Look-ahead down only
                                         intileimage.putalpha(alphaLayerRTC)
                                     else:
                                         intileimage.putalpha(alphaLayerABL)
                                 elif emb_column_i == emb_tiles_x - 1:
-                                    if (tile+emb_tiles_x) in embiggen_tiles:  # Look-ahead down
+                                    if (
+                                        tile + emb_tiles_x
+                                    ) in embiggen_tiles:  # Look-ahead down
                                         intileimage.putalpha(alphaLayerLTC)
                                     else:
                                         intileimage.putalpha(alphaLayerABR)
                                 else:
-                                    if (tile+1) in embiggen_tiles: # Look-ahead right
-                                        if (tile+emb_tiles_x) in embiggen_tiles: # Look-ahead down
+                                    if (tile + 1) in embiggen_tiles:  # Look-ahead right
+                                        if (
+                                            tile + emb_tiles_x
+                                        ) in embiggen_tiles:  # Look-ahead down
                                             intileimage.putalpha(alphaLayerLTaC)
                                         else:
                                             intileimage.putalpha(alphaLayerABR)
-                                    elif (tile+emb_tiles_x) in embiggen_tiles:  # Look-ahead down only
+                                    elif (
+                                        tile + emb_tiles_x
+                                    ) in embiggen_tiles:  # Look-ahead down only
                                         intileimage.putalpha(alphaLayerABB)
                                     else:
                                         intileimage.putalpha(alphaLayerAA)
@@ -481,21 +531,28 @@ class Embiggen(Generator):
                             if emb_row_i == 0 and emb_column_i >= 1:
                                 intileimage.putalpha(alphaLayerL)
                             elif emb_row_i >= 1 and emb_column_i == 0:
-                                if emb_column_i + 1 == emb_tiles_x: # If we don't have anything that can be placed to the right
+                                if (
+                                    emb_column_i + 1 == emb_tiles_x
+                                ):  # If we don't have anything that can be placed to the right
                                     intileimage.putalpha(alphaLayerT)
                                 else:
                                     intileimage.putalpha(alphaLayerTaC)
                             else:
-                                if emb_column_i + 1 == emb_tiles_x: # If we don't have anything that can be placed to the right
+                                if (
+                                    emb_column_i + 1 == emb_tiles_x
+                                ):  # If we don't have anything that can be placed to the right
                                     intileimage.putalpha(alphaLayerLTC)
                                 else:
                                     intileimage.putalpha(alphaLayerLTaC)
                     # Layer tile onto final image
                     outputsuperimage.alpha_composite(intileimage, (left, top))
             else:
-                print('Error: could not find all Embiggen output tiles in memory? Something must have gone wrong with img2img generation.')
+                print(
+                    "Error: could not find all Embiggen output tiles in memory? Something must have gone wrong with img2img generation."
+                )
 
             # after internal loops and patching up return Embiggen image
             return outputsuperimage
+
         # end of function declaration
         return make_image

invokeai/backend/generator/img2img.py

@@ -0,0 +1,101 @@
+"""
+invokeai.backend.generator.img2img descends from .generator
+"""
+from typing import Optional
+
+import torch
+from accelerate.utils import set_seed
+from diffusers import logging
+
+from ..stable_diffusion import (
+    ConditioningData,
+    PostprocessingSettings,
+    StableDiffusionGeneratorPipeline,
+)
+from .base import Generator
+
+
+class Img2Img(Generator):
+    def __init__(self, model, precision):
+        super().__init__(model, precision)
+        self.init_latent = None  # by get_noise()
+
+    def get_make_image(
+        self,
+        prompt,
+        sampler,
+        steps,
+        cfg_scale,
+        ddim_eta,
+        conditioning,
+        init_image,
+        strength,
+        step_callback=None,
+        threshold=0.0,
+        warmup=0.2,
+        perlin=0.0,
+        h_symmetry_time_pct=None,
+        v_symmetry_time_pct=None,
+        attention_maps_callback=None,
+        **kwargs,
+    ):
+        """
+        Returns a function returning an image derived from the prompt and the initial image
+        Return value depends on the seed at the time you call it.
+        """
+        self.perlin = perlin
+
+        # noinspection PyTypeChecker
+        pipeline: StableDiffusionGeneratorPipeline = self.model
+        pipeline.scheduler = sampler
+
+        uc, c, extra_conditioning_info = conditioning
+        conditioning_data = ConditioningData(
+            uc,
+            c,
+            cfg_scale,
+            extra_conditioning_info,
+            postprocessing_settings=PostprocessingSettings(
+                threshold=threshold,
+                warmup=warmup,
+                h_symmetry_time_pct=h_symmetry_time_pct,
+                v_symmetry_time_pct=v_symmetry_time_pct,
+            ),
+        ).add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta)
+
+        def make_image(x_T: torch.Tensor, seed: int):
+            # FIXME: use x_T for initial seeded noise
+            # We're not at the moment because the pipeline automatically resizes init_image if
+            # necessary, which the x_T input might not match.
+            # In the meantime, reset the seed prior to generating pipeline output so we at least get the same result.
+            logging.set_verbosity_error()  # quench safety check warnings
+            pipeline_output = pipeline.img2img_from_embeddings(
+                init_image,
+                strength,
+                steps,
+                conditioning_data,
+                noise_func=self.get_noise_like,
+                callback=step_callback,
+                seed=seed,
+            )
+            if (
+                pipeline_output.attention_map_saver is not None
+                and attention_maps_callback is not None
+            ):
+                attention_maps_callback(pipeline_output.attention_map_saver)
+            return pipeline.numpy_to_pil(pipeline_output.images)[0]
+
+        return make_image
+
+    def get_noise_like(self, like: torch.Tensor):
+        device = like.device
+        if device.type == "mps":
+            x = torch.randn_like(like, device="cpu").to(device)
+        else:
+            x = torch.randn_like(like, device=device)
+        if self.perlin > 0.0:
+            shape = like.shape
+            x = (1 - self.perlin) * x + self.perlin * self.get_perlin_noise(
+                shape[3], shape[2]
+            )
+        return x

invokeai/backend/generator/inpaint.py

@@ -1,32 +1,35 @@
-'''
-ldm.invoke.generator.inpaint descends from ldm.invoke.generator
-'''
+"""
+invokeai.backend.generator.inpaint descends from .generator
+"""
 from __future__ import annotations
 
 import math
 
-import PIL
 import cv2
 import numpy as np
+import PIL
 import torch
-from PIL import Image, ImageFilter, ImageOps, ImageChops
+from PIL import Image, ImageChops, ImageFilter, ImageOps
 
-from ldm.invoke.generator.diffusers_pipeline import image_resized_to_grid_as_tensor, StableDiffusionGeneratorPipeline, \
-    ConditioningData
-from ldm.invoke.generator.img2img import Img2Img
-from ldm.invoke.patchmatch import PatchMatch
-from ldm.util import debug_image
+from ..image_util import PatchMatch, debug_image
+from ..stable_diffusion.diffusers_pipeline import (
+    ConditioningData,
+    StableDiffusionGeneratorPipeline,
+    image_resized_to_grid_as_tensor,
+)
+from .img2img import Img2Img
 
 
-def infill_methods()->list[str]:
+def infill_methods() -> list[str]:
     methods = [
         "tile",
         "solid",
     ]
     if PatchMatch.patchmatch_available():
-        methods.insert(0, 'patchmatch')
+        methods.insert(0, "patchmatch")
     return methods
 
+
 class Inpaint(Img2Img):
     def __init__(self, model, precision):
         self.inpaint_height = 0
@@ -53,11 +56,11 @@ class Inpaint(Img2Img):
             np.ravel(image),
             shape=(nrows, ncols, height, width, depth),
             strides=(height * _strides[0], width * _strides[1], *_strides),
-            writeable=False
+            writeable=False,
         )
 
     def infill_patchmatch(self, im: Image.Image) -> Image:
-        if im.mode != 'RGBA':
+        if im.mode != "RGBA":
             return im
 
         # Skip patchmatch if patchmatch isn't available
@@ -65,13 +68,17 @@ class Inpaint(Img2Img):
             return im
 
         # Patchmatch (note, we may want to expose patch_size? Increasing it significantly impacts performance though)
-        im_patched_np = PatchMatch.inpaint(im.convert('RGB'), ImageOps.invert(im.split()[-1]), patch_size = 3)
-        im_patched = Image.fromarray(im_patched_np, mode = 'RGB')
+        im_patched_np = PatchMatch.inpaint(
+            im.convert("RGB"), ImageOps.invert(im.split()[-1]), patch_size=3
+        )
+        im_patched = Image.fromarray(im_patched_np, mode="RGB")
         return im_patched
 
-    def tile_fill_missing(self, im: Image.Image, tile_size: int = 16, seed: int = None) -> Image:
+    def tile_fill_missing(
+        self, im: Image.Image, tile_size: int = 16, seed: int = None
+    ) -> Image:
         # Only fill if there's an alpha layer
-        if im.mode != 'RGBA':
+        if im.mode != "RGBA":
             return im
 
         a = np.asarray(im, dtype=np.uint8)
@@ -79,21 +86,21 @@ class Inpaint(Img2Img):
         tile_size = (tile_size, tile_size)
 
         # Get the image as tiles of a specified size
-        tiles = self.get_tile_images(a,*tile_size).copy()
+        tiles = self.get_tile_images(a, *tile_size).copy()
 
         # Get the mask as tiles
-        tiles_mask = tiles[:,:,:,:,3]
+        tiles_mask = tiles[:, :, :, :, 3]
 
         # Find any mask tiles with any fully transparent pixels (we will be replacing these later)
         tmask_shape = tiles_mask.shape
         tiles_mask = tiles_mask.reshape(math.prod(tiles_mask.shape))
-        n,ny = (math.prod(tmask_shape[0:2])), math.prod(tmask_shape[2:])
-        tiles_mask = (tiles_mask > 0)
-        tiles_mask = tiles_mask.reshape((n,ny)).all(axis = 1)
+        n, ny = (math.prod(tmask_shape[0:2])), math.prod(tmask_shape[2:])
+        tiles_mask = tiles_mask > 0
+        tiles_mask = tiles_mask.reshape((n, ny)).all(axis=1)
 
         # Get RGB tiles in single array and filter by the mask
         tshape = tiles.shape
-        tiles_all = tiles.reshape((math.prod(tiles.shape[0:2]), * tiles.shape[2:]))
+        tiles_all = tiles.reshape((math.prod(tiles.shape[0:2]), *tiles.shape[2:]))
         filtered_tiles = tiles_all[tiles_mask]
 
         if len(filtered_tiles) == 0:
@@ -101,23 +108,32 @@ class Inpaint(Img2Img):
 
         # Find all invalid tiles and replace with a random valid tile
         replace_count = (tiles_mask == False).sum()
-        rng = np.random.default_rng(seed = seed)
-        tiles_all[np.logical_not(tiles_mask)] = filtered_tiles[rng.choice(filtered_tiles.shape[0], replace_count),:,:,:]
+        rng = np.random.default_rng(seed=seed)
+        tiles_all[np.logical_not(tiles_mask)] = filtered_tiles[
+            rng.choice(filtered_tiles.shape[0], replace_count), :, :, :
+        ]
 
         # Convert back to an image
         tiles_all = tiles_all.reshape(tshape)
-        tiles_all = tiles_all.swapaxes(1,2)
-        st = tiles_all.reshape((math.prod(tiles_all.shape[0:2]), math.prod(tiles_all.shape[2:4]), tiles_all.shape[4]))
-        si = Image.fromarray(st, mode='RGBA')
+        tiles_all = tiles_all.swapaxes(1, 2)
+        st = tiles_all.reshape(
+            (
+                math.prod(tiles_all.shape[0:2]),
+                math.prod(tiles_all.shape[2:4]),
+                tiles_all.shape[4],
+            )
+        )
+        si = Image.fromarray(st, mode="RGBA")
 
         return si
 
-
     def mask_edge(self, mask: Image, edge_size: int, edge_blur: int) -> Image:
         npimg = np.asarray(mask, dtype=np.uint8)
 
         # Detect any partially transparent regions
-        npgradient = np.uint8(255 * (1.0 - np.floor(np.abs(0.5 - np.float32(npimg) / 255.0) * 2.0)))
+        npgradient = np.uint8(
+            255 * (1.0 - np.floor(np.abs(0.5 - np.float32(npimg) / 255.0) * 2.0))
+        )
 
         # Detect hard edges
         npedge = cv2.Canny(npimg, threshold1=100, threshold2=200)
@@ -126,7 +142,9 @@ class Inpaint(Img2Img):
         npmask = npgradient + npedge
 
         # Expand
-        npmask = cv2.dilate(npmask, np.ones((3,3), np.uint8), iterations = int(edge_size / 2))
+        npmask = cv2.dilate(
+            npmask, np.ones((3, 3), np.uint8), iterations=int(edge_size / 2)
+        )
 
         new_mask = Image.fromarray(npmask)
 
@@ -135,9 +153,23 @@ class Inpaint(Img2Img):
 
         return ImageOps.invert(new_mask)
 
-
-    def seam_paint(self, im: Image.Image, seam_size: int, seam_blur: int, prompt, sampler, steps, cfg_scale, ddim_eta,
-                   conditioning, strength, noise, infill_method, step_callback) -> Image.Image:
+    def seam_paint(
+        self,
+        im: Image.Image,
+        seam_size: int,
+        seam_blur: int,
+        prompt,
+        seed,
+        sampler,
+        steps,
+        cfg_scale,
+        ddim_eta,
+        conditioning,
+        strength,
+        noise,
+        infill_method,
+        step_callback,
+    ) -> Image.Image:
         hard_mask = self.pil_image.split()[-1].copy()
         mask = self.mask_edge(hard_mask, seam_size, seam_blur)
 
@@ -148,45 +180,52 @@ class Inpaint(Img2Img):
             cfg_scale,
             ddim_eta,
             conditioning,
-            init_image = im.copy().convert('RGBA'),
-            mask_image = mask,
-            strength = strength,
-            mask_blur_radius = 0,
-            seam_size = 0,
-            step_callback = step_callback,
-            inpaint_width = im.width,
-            inpaint_height = im.height,
-            infill_method = infill_method
+            init_image=im.copy().convert("RGBA"),
+            mask_image=mask,
+            strength=strength,
+            mask_blur_radius=0,
+            seam_size=0,
+            step_callback=step_callback,
+            inpaint_width=im.width,
+            inpaint_height=im.height,
+            infill_method=infill_method,
         )
 
         seam_noise = self.get_noise(im.width, im.height)
 
-        result = make_image(seam_noise)
+        result = make_image(seam_noise, seed)
 
         return result
 
-
     @torch.no_grad()
-    def get_make_image(self,prompt,sampler,steps,cfg_scale,ddim_eta,
-                       conditioning,
-                       init_image: PIL.Image.Image | torch.FloatTensor,
-                       mask_image: PIL.Image.Image | torch.FloatTensor,
-                       strength: float,
-                       mask_blur_radius: int = 8,
-                       # Seam settings - when 0, doesn't fill seam
-                       seam_size: int = 0,
-                       seam_blur: int = 0,
-                       seam_strength: float = 0.7,
-                       seam_steps: int = 10,
-                       tile_size: int = 32,
-                       step_callback=None,
-                       inpaint_replace=False, enable_image_debugging=False,
-                       infill_method = None,
-                       inpaint_width=None,
-                       inpaint_height=None,
-                       inpaint_fill:tuple(int)=(0x7F, 0x7F, 0x7F, 0xFF),
-                       attention_maps_callback=None,
-                       **kwargs):
+    def get_make_image(
+        self,
+        prompt,
+        sampler,
+        steps,
+        cfg_scale,
+        ddim_eta,
+        conditioning,
+        init_image: PIL.Image.Image | torch.FloatTensor,
+        mask_image: PIL.Image.Image | torch.FloatTensor,
+        strength: float,
+        mask_blur_radius: int = 8,
+        # Seam settings - when 0, doesn't fill seam
+        seam_size: int = 0,
+        seam_blur: int = 0,
+        seam_strength: float = 0.7,
+        seam_steps: int = 10,
+        tile_size: int = 32,
+        step_callback=None,
+        inpaint_replace=False,
+        enable_image_debugging=False,
+        infill_method=None,
+        inpaint_width=None,
+        inpaint_height=None,
+        inpaint_fill: tuple(int) = (0x7F, 0x7F, 0x7F, 0xFF),
+        attention_maps_callback=None,
+        **kwargs,
+    ):
         """
         Returns a function returning an image derived from the prompt and
         the initial image + mask.  Return value depends on the seed at
@@ -204,33 +243,39 @@ class Inpaint(Img2Img):
             self.pil_image = init_image.copy()
 
             # Do infill
-            if infill_method == 'patchmatch' and PatchMatch.patchmatch_available():
+            if infill_method == "patchmatch" and PatchMatch.patchmatch_available():
                 init_filled = self.infill_patchmatch(self.pil_image.copy())
-            elif infill_method == 'tile':
+            elif infill_method == "tile":
                 init_filled = self.tile_fill_missing(
-                    self.pil_image.copy(),
-                    seed = self.seed,
-                    tile_size = tile_size
+                    self.pil_image.copy(), seed=self.seed, tile_size=tile_size
                 )
-            elif infill_method == 'solid':
+            elif infill_method == "solid":
                 solid_bg = PIL.Image.new("RGBA", init_image.size, inpaint_fill)
                 init_filled = PIL.Image.alpha_composite(solid_bg, init_image)
             else:
-                raise ValueError(f"Non-supported infill type {infill_method}", infill_method)
-            init_filled.paste(init_image, (0,0), init_image.split()[-1])
+                raise ValueError(
+                    f"Non-supported infill type {infill_method}", infill_method
+                )
+            init_filled.paste(init_image, (0, 0), init_image.split()[-1])
 
             # Resize if requested for inpainting
             if inpaint_width and inpaint_height:
                 init_filled = init_filled.resize((inpaint_width, inpaint_height))
 
-            debug_image(init_filled, "init_filled", debug_status=self.enable_image_debugging)
+            debug_image(
+                init_filled, "init_filled", debug_status=self.enable_image_debugging
+            )
 
             # Create init tensor
-            init_image = image_resized_to_grid_as_tensor(init_filled.convert('RGB'))
+            init_image = image_resized_to_grid_as_tensor(init_filled.convert("RGB"))
 
         if isinstance(mask_image, PIL.Image.Image):
             self.pil_mask = mask_image.copy()
-            debug_image(mask_image, "mask_image BEFORE multiply with pil_image", debug_status=self.enable_image_debugging)
+            debug_image(
+                mask_image,
+                "mask_image BEFORE multiply with pil_image",
+                debug_status=self.enable_image_debugging,
+            )
 
             init_alpha = self.pil_image.getchannel("A")
             if mask_image.mode != "L":
@@ -243,8 +288,14 @@ class Inpaint(Img2Img):
             if inpaint_width and inpaint_height:
                 mask_image = mask_image.resize((inpaint_width, inpaint_height))
 
-            debug_image(mask_image, "mask_image AFTER multiply with pil_image", debug_status=self.enable_image_debugging)
-            mask: torch.FloatTensor = image_resized_to_grid_as_tensor(mask_image, normalize=False)
+            debug_image(
+                mask_image,
+                "mask_image AFTER multiply with pil_image",
+                debug_status=self.enable_image_debugging,
+            )
+            mask: torch.FloatTensor = image_resized_to_grid_as_tensor(
+                mask_image, normalize=False
+            )
         else:
             mask: torch.FloatTensor = mask_image
 
@@ -256,11 +307,11 @@ class Inpaint(Img2Img):
 
         # todo: support cross-attention control
         uc, c, _ = conditioning
-        conditioning_data = (ConditioningData(uc, c, cfg_scale)
-                             .add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta))
+        conditioning_data = ConditioningData(
+            uc, c, cfg_scale
+        ).add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta)
 
-
-        def make_image(x_T):
+        def make_image(x_T: torch.Tensor, seed: int):
             pipeline_output = pipeline.inpaint_from_embeddings(
                 init_image=init_image,
                 mask=1 - mask,  # expects white means "paint here."
@@ -269,45 +320,75 @@ class Inpaint(Img2Img):
                 conditioning_data=conditioning_data,
                 noise_func=self.get_noise_like,
                 callback=step_callback,
+                seed=seed,
             )
 
-            if pipeline_output.attention_map_saver is not None and attention_maps_callback is not None:
+            if (
+                pipeline_output.attention_map_saver is not None
+                and attention_maps_callback is not None
+            ):
                 attention_maps_callback(pipeline_output.attention_map_saver)
 
-            result = self.postprocess_size_and_mask(pipeline.numpy_to_pil(pipeline_output.images)[0])
+            result = self.postprocess_size_and_mask(
+                pipeline.numpy_to_pil(pipeline_output.images)[0]
+            )
 
             # Seam paint if this is our first pass (seam_size set to 0 during seam painting)
             if seam_size > 0:
                 old_image = self.pil_image or init_image
                 old_mask = self.pil_mask or mask_image
 
-                result = self.seam_paint(result, seam_size, seam_blur, prompt, sampler, seam_steps, cfg_scale, ddim_eta,
-                                         conditioning, seam_strength, x_T, infill_method, step_callback)
+                result = self.seam_paint(
+                    result,
+                    seam_size,
+                    seam_blur,
+                    prompt,
+                    seed,
+                    sampler,
+                    seam_steps,
+                    cfg_scale,
+                    ddim_eta,
+                    conditioning,
+                    seam_strength,
+                    x_T,
+                    infill_method,
+                    step_callback,
+                )
 
                 # Restore original settings
-                self.get_make_image(prompt,sampler,steps,cfg_scale,ddim_eta,
-                       conditioning,
-                       old_image,
-                       old_mask,
-                       strength,
-                       mask_blur_radius, seam_size, seam_blur, seam_strength,
-                       seam_steps, tile_size, step_callback,
-                       inpaint_replace, enable_image_debugging,
-                       inpaint_width = inpaint_width,
-                       inpaint_height = inpaint_height,
-                       infill_method = infill_method,
-                       **kwargs)
+                self.get_make_image(
+                    prompt,
+                    sampler,
+                    steps,
+                    cfg_scale,
+                    ddim_eta,
+                    conditioning,
+                    old_image,
+                    old_mask,
+                    strength,
+                    mask_blur_radius,
+                    seam_size,
+                    seam_blur,
+                    seam_strength,
+                    seam_steps,
+                    tile_size,
+                    step_callback,
+                    inpaint_replace,
+                    enable_image_debugging,
+                    inpaint_width=inpaint_width,
+                    inpaint_height=inpaint_height,
+                    infill_method=infill_method,
+                    **kwargs,
+                )
 
             return result
 
         return make_image
 
-
-    def sample_to_image(self, samples)->Image.Image:
-        gen_result = super().sample_to_image(samples).convert('RGB')
+    def sample_to_image(self, samples) -> Image.Image:
+        gen_result = super().sample_to_image(samples).convert("RGB")
         return self.postprocess_size_and_mask(gen_result)
 
-
     def postprocess_size_and_mask(self, gen_result: Image.Image) -> Image.Image:
         debug_image(gen_result, "gen_result", debug_status=self.enable_image_debugging)
 
@@ -318,7 +399,13 @@ class Inpaint(Img2Img):
         if self.pil_image is None or self.pil_mask is None:
             return gen_result
 
-        corrected_result = self.repaste_and_color_correct(gen_result, self.pil_image, self.pil_mask, self.mask_blur_radius)
-        debug_image(corrected_result, "corrected_result", debug_status=self.enable_image_debugging)
+        corrected_result = self.repaste_and_color_correct(
+            gen_result, self.pil_image, self.pil_mask, self.mask_blur_radius
+        )
+        debug_image(
+            corrected_result,
+            "corrected_result",
+            debug_status=self.enable_image_debugging,
+        )
 
         return corrected_result

invokeai/backend/generator/txt2img.py

@@ -0,0 +1,81 @@
+"""
+invokeai.backend.generator.txt2img inherits from invokeai.backend.generator
+"""
+import PIL.Image
+import torch
+
+from ..stable_diffusion import (
+    ConditioningData,
+    PostprocessingSettings,
+    StableDiffusionGeneratorPipeline,
+)
+from .base import Generator
+
+
+class Txt2Img(Generator):
+    def __init__(self, model, precision):
+        super().__init__(model, precision)
+
+    @torch.no_grad()
+    def get_make_image(
+        self,
+        prompt,
+        sampler,
+        steps,
+        cfg_scale,
+        ddim_eta,
+        conditioning,
+        width,
+        height,
+        step_callback=None,
+        threshold=0.0,
+        warmup=0.2,
+        perlin=0.0,
+        h_symmetry_time_pct=None,
+        v_symmetry_time_pct=None,
+        attention_maps_callback=None,
+        **kwargs,
+    ):
+        """
+        Returns a function returning an image derived from the prompt and the initial image
+        Return value depends on the seed at the time you call it
+        kwargs are 'width' and 'height'
+        """
+        self.perlin = perlin
+
+        # noinspection PyTypeChecker
+        pipeline: StableDiffusionGeneratorPipeline = self.model
+        pipeline.scheduler = sampler
+
+        uc, c, extra_conditioning_info = conditioning
+        conditioning_data = ConditioningData(
+            uc,
+            c,
+            cfg_scale,
+            extra_conditioning_info,
+            postprocessing_settings=PostprocessingSettings(
+                threshold=threshold,
+                warmup=warmup,
+                h_symmetry_time_pct=h_symmetry_time_pct,
+                v_symmetry_time_pct=v_symmetry_time_pct,
+            ),
+        ).add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta)
+
+        def make_image(x_T: torch.Tensor, _: int) -> PIL.Image.Image:
+            pipeline_output = pipeline.image_from_embeddings(
+                latents=torch.zeros_like(x_T, dtype=self.torch_dtype()),
+                noise=x_T,
+                num_inference_steps=steps,
+                conditioning_data=conditioning_data,
+                callback=step_callback,
+            )
+
+            if (
+                pipeline_output.attention_map_saver is not None
+                and attention_maps_callback is not None
+            ):
+                attention_maps_callback(pipeline_output.attention_map_saver)
+
+            return pipeline.numpy_to_pil(pipeline_output.images)[0]
+
+        return make_image

invokeai/backend/generator/txt2img2img.py

@@ -1,6 +1,6 @@
-'''
-ldm.invoke.generator.txt2img inherits from ldm.invoke.generator
-'''
+"""
+invokeai.backend.generator.txt2img inherits from invokeai.backend.generator
+"""
 
 import math
 from typing import Callable, Optional
@@ -8,21 +8,37 @@ from typing import Callable, Optional
 import torch
 from diffusers.utils.logging import get_verbosity, set_verbosity, set_verbosity_error
 
-from ldm.invoke.generator.base import Generator
-from ldm.invoke.generator.diffusers_pipeline import trim_to_multiple_of, StableDiffusionGeneratorPipeline, \
-    ConditioningData
-from ldm.models.diffusion.shared_invokeai_diffusion import PostprocessingSettings
-
+from ..stable_diffusion import PostprocessingSettings
+from .base import Generator
+from ..stable_diffusion.diffusers_pipeline import StableDiffusionGeneratorPipeline
+from ..stable_diffusion.diffusers_pipeline import ConditioningData
+from ..stable_diffusion.diffusers_pipeline import trim_to_multiple_of
 
 class Txt2Img2Img(Generator):
     def __init__(self, model, precision):
         super().__init__(model, precision)
-        self.init_latent = None    # for get_noise()
+        self.init_latent = None  # for get_noise()
 
-    def get_make_image(self, prompt:str, sampler, steps:int, cfg_scale:float, ddim_eta,
-                       conditioning, width:int, height:int, strength:float,
-                       step_callback:Optional[Callable]=None, threshold=0.0, warmup=0.2, perlin=0.0,
-                       h_symmetry_time_pct=None, v_symmetry_time_pct=None, attention_maps_callback=None, **kwargs):
+    def get_make_image(
+        self,
+        prompt: str,
+        sampler,
+        steps: int,
+        cfg_scale: float,
+        ddim_eta,
+        conditioning,
+        width: int,
+        height: int,
+        strength: float,
+        step_callback: Optional[Callable] = None,
+        threshold=0.0,
+        warmup=0.2,
+        perlin=0.0,
+        h_symmetry_time_pct=None,
+        v_symmetry_time_pct=None,
+        attention_maps_callback=None,
+        **kwargs,
+    ):
         """
         Returns a function returning an image derived from the prompt and the initial image
         Return value depends on the seed at the time you call it
@@ -35,19 +51,20 @@ class Txt2Img2Img(Generator):
         pipeline.scheduler = sampler
 
         uc, c, extra_conditioning_info = conditioning
-        conditioning_data = (
-            ConditioningData(
-                uc, c, cfg_scale, extra_conditioning_info,
-                postprocessing_settings = PostprocessingSettings(
-                    threshold=threshold,
-                    warmup=0.2,
-                    h_symmetry_time_pct=h_symmetry_time_pct,
-                    v_symmetry_time_pct=v_symmetry_time_pct
-                )
-            ).add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta))
-
-        def make_image(x_T):
+        conditioning_data = ConditioningData(
+            uc,
+            c,
+            cfg_scale,
+            extra_conditioning_info,
+            postprocessing_settings=PostprocessingSettings(
+                threshold=threshold,
+                warmup=0.2,
+                h_symmetry_time_pct=h_symmetry_time_pct,
+                v_symmetry_time_pct=v_symmetry_time_pct,
+            ),
+        ).add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta)
 
+        def make_image(x_T: torch.Tensor, _: int):
             first_pass_latent_output, _ = pipeline.latents_from_embeddings(
                 latents=torch.zeros_like(x_T),
                 num_inference_steps=steps,
@@ -61,28 +78,40 @@ class Txt2Img2Img(Generator):
             init_width = first_pass_latent_output.size()[3] * self.downsampling_factor
             init_height = first_pass_latent_output.size()[2] * self.downsampling_factor
             print(
-                  f"\n>> Interpolating from {init_width}x{init_height} to {width}x{height} using DDIM sampling"
-                 )
+                f"\n>> Interpolating from {init_width}x{init_height} to {width}x{height} using DDIM sampling"
+            )
 
             # resizing
             resized_latents = torch.nn.functional.interpolate(
                 first_pass_latent_output,
-                size=(height // self.downsampling_factor, width // self.downsampling_factor),
-                mode="bilinear"
+                size=(
+                    height // self.downsampling_factor,
+                    width // self.downsampling_factor,
+                ),
+                mode="bilinear",
             )
 
             # Free up memory from the last generation.
-            clear_cuda_cache = kwargs['clear_cuda_cache'] or None
+            clear_cuda_cache = kwargs["clear_cuda_cache"] or None
             if clear_cuda_cache is not None:
                 clear_cuda_cache()
 
-            second_pass_noise = self.get_noise_like(resized_latents, override_perlin=True)
+            second_pass_noise = self.get_noise_like(
+                resized_latents, override_perlin=True
+            )
 
             # Clear symmetry for the second pass
             from dataclasses import replace
-            new_postprocessing_settings = replace(conditioning_data.postprocessing_settings, h_symmetry_time_pct=None)
-            new_postprocessing_settings = replace(new_postprocessing_settings, v_symmetry_time_pct=None)
-            new_conditioning_data = replace(conditioning_data, postprocessing_settings=new_postprocessing_settings)
+
+            new_postprocessing_settings = replace(
+                conditioning_data.postprocessing_settings, h_symmetry_time_pct=None
+            )
+            new_postprocessing_settings = replace(
+                new_postprocessing_settings, v_symmetry_time_pct=None
+            )
+            new_conditioning_data = replace(
+                conditioning_data, postprocessing_settings=new_postprocessing_settings
+            )
 
             verbosity = get_verbosity()
             set_verbosity_error()
@@ -92,15 +121,18 @@ class Txt2Img2Img(Generator):
                 conditioning_data=new_conditioning_data,
                 strength=strength,
                 noise=second_pass_noise,
-                callback=step_callback)
+                callback=step_callback,
+            )
             set_verbosity(verbosity)
 
-            if pipeline_output.attention_map_saver is not None and attention_maps_callback is not None:
+            if (
+                pipeline_output.attention_map_saver is not None
+                and attention_maps_callback is not None
+            ):
                 attention_maps_callback(pipeline_output.attention_map_saver)
 
             return pipeline.numpy_to_pil(pipeline_output.images)[0]
 
-
         # FIXME: do we really need something entirely different for the inpainting model?
 
         # in the case of the inpainting model being loaded, the trick of
@@ -111,19 +143,23 @@ class Txt2Img2Img(Generator):
 
         return make_image
 
-    def get_noise_like(self, like: torch.Tensor, override_perlin: bool=False):
+    def get_noise_like(self, like: torch.Tensor, override_perlin: bool = False):
         device = like.device
-        if device.type == 'mps':
-            x = torch.randn_like(like, device='cpu', dtype=self.torch_dtype()).to(device)
+        if device.type == "mps":
+            x = torch.randn_like(like, device="cpu", dtype=self.torch_dtype()).to(
+                device
+            )
         else:
             x = torch.randn_like(like, device=device, dtype=self.torch_dtype())
         if self.perlin > 0.0 and override_perlin == False:
             shape = like.shape
-            x = (1-self.perlin)*x + self.perlin*self.get_perlin_noise(shape[3], shape[2])
+            x = (1 - self.perlin) * x + self.perlin * self.get_perlin_noise(
+                shape[3], shape[2]
+            )
         return x
 
     # returns a tensor filled with random numbers from a normal distribution
-    def get_noise(self,width,height,scale = True):
+    def get_noise(self, width, height, scale=True):
         # print(f"Get noise: {width}x{height}")
         if scale:
             # Scale the input width and height for the initial generation
@@ -133,7 +169,9 @@ class Txt2Img2Img(Generator):
             aspect = width / height
             dimension = self.model.unet.config.sample_size * self.model.vae_scale_factor
             min_dimension = math.floor(dimension * 0.5)
-            model_area = dimension * dimension # hardcoded for now since all models are trained on square images
+            model_area = (
+                dimension * dimension
+            )  # hardcoded for now since all models are trained on square images
 
             if aspect > 1.0:
                 init_height = max(min_dimension, math.sqrt(model_area / aspect))
@@ -142,7 +180,9 @@ class Txt2Img2Img(Generator):
                 init_width = max(min_dimension, math.sqrt(model_area * aspect))
                 init_height = init_width / aspect
 
-            scaled_width, scaled_height = trim_to_multiple_of(math.floor(init_width), math.floor(init_height))
+            scaled_width, scaled_height = trim_to_multiple_of(
+                math.floor(init_width), math.floor(init_height)
+            )
 
         else:
             scaled_width = width
@@ -152,10 +192,14 @@ class Txt2Img2Img(Generator):
         channels = self.latent_channels
         if channels == 9:
             channels = 4  # we don't really want noise for all the mask channels
-        shape = (1, channels,
-                 scaled_height // self.downsampling_factor, scaled_width // self.downsampling_factor)
-        if self.use_mps_noise or device.type == 'mps':
-            tensor = torch.empty(size=shape, device='cpu')
+        shape = (
+            1,
+            channels,
+            scaled_height // self.downsampling_factor,
+            scaled_width // self.downsampling_factor,
+        )
+        if self.use_mps_noise or device.type == "mps":
+            tensor = torch.empty(size=shape, device="cpu")
             tensor = self.get_noise_like(like=tensor).to(device)
         else:
             tensor = torch.empty(size=shape, device=device)

invokeai/backend/globals.py

@@ -1,5 +1,5 @@
-'''
-ldm.invoke.globals defines a small number of global variables that would
+"""
+invokeai.backend.globals defines a small number of global variables that would
 otherwise have to be passed through long and complex call chains.
 
 It defines a Namespace object named "Globals" that contains
@@ -9,7 +9,7 @@ the attributes:
   - initfile       - path to the initialization file
   - try_patchmatch - option to globally disable loading of 'patchmatch' module
   - always_use_cpu - force use of CPU even if GPU is available
-'''
+"""
 
 import os
 import os.path as osp
@@ -20,12 +20,12 @@ from typing import Union
 Globals = Namespace()
 
 # Where to look for the initialization file and other key components
-Globals.initfile = 'invokeai.init'
-Globals.models_file = 'models.yaml'
-Globals.models_dir = 'models'
-Globals.config_dir = 'configs'
-Globals.autoscan_dir = 'weights'
-Globals.converted_ckpts_dir = 'converted_ckpts'
+Globals.initfile = "invokeai.init"
+Globals.models_file = "models.yaml"
+Globals.models_dir = "models"
+Globals.config_dir = "configs"
+Globals.autoscan_dir = "weights"
+Globals.converted_ckpts_dir = "converted_ckpts"
 
 # Set the default root directory. This can be overwritten by explicitly
 # passing the `--root <directory>` argument on the command line.
@@ -34,12 +34,15 @@ Globals.converted_ckpts_dir = 'converted_ckpts'
 # 2) use VIRTUAL_ENV environment variable, with a check for initfile being there
 # 3) use ~/invokeai
 
-if os.environ.get('INVOKEAI_ROOT'):
-    Globals.root = osp.abspath(os.environ.get('INVOKEAI_ROOT'))
-elif os.environ.get('VIRTUAL_ENV') and Path(os.environ.get('VIRTUAL_ENV'),'..',Globals.initfile).exists():
-    Globals.root = osp.abspath(osp.join(os.environ.get('VIRTUAL_ENV'), '..'))
+if os.environ.get("INVOKEAI_ROOT"):
+    Globals.root = osp.abspath(os.environ.get("INVOKEAI_ROOT"))
+elif (
+    os.environ.get("VIRTUAL_ENV")
+    and Path(os.environ.get("VIRTUAL_ENV"), "..", Globals.initfile).exists()
+):
+    Globals.root = osp.abspath(osp.join(os.environ.get("VIRTUAL_ENV"), ".."))
 else:
-    Globals.root = osp.abspath(osp.expanduser('~/invokeai'))
+    Globals.root = osp.abspath(osp.expanduser("~/invokeai"))
 
 # Try loading patchmatch
 Globals.try_patchmatch = True
@@ -61,55 +64,59 @@ Globals.sequential_guidance = False
 Globals.full_precision = False
 
 # whether we should convert ckpt files into diffusers models on the fly
-Globals.ckpt_convert = False
+Globals.ckpt_convert = True
 
 # logging tokenization everywhere
 Globals.log_tokenization = False
 
-def global_config_file()->Path:
+
+def global_config_file() -> Path:
     return Path(Globals.root, Globals.config_dir, Globals.models_file)
 
-def global_config_dir()->Path:
+
+def global_config_dir() -> Path:
     return Path(Globals.root, Globals.config_dir)
 
-def global_models_dir()->Path:
+
+def global_models_dir() -> Path:
     return Path(Globals.root, Globals.models_dir)
 
-def global_autoscan_dir()->Path:
+
+def global_autoscan_dir() -> Path:
     return Path(Globals.root, Globals.autoscan_dir)
 
-def global_converted_ckpts_dir()->Path:
+
+def global_converted_ckpts_dir() -> Path:
     return Path(global_models_dir(), Globals.converted_ckpts_dir)
 
-def global_set_root(root_dir:Union[str,Path]):
+
+def global_set_root(root_dir: Union[str, Path]):
     Globals.root = root_dir
 
-def global_cache_dir(subdir:Union[str,Path]='')->Path:
-    '''
+
+def global_cache_dir(subdir: Union[str, Path] = "") -> Path:
+    """
     Returns Path to the model cache directory. If a subdirectory
     is provided, it will be appended to the end of the path, allowing
-    for huggingface-style conventions:
-         global_cache_dir('diffusers')
+    for Hugging Face-style conventions. Currently, Hugging Face has
+    moved all models into the "hub" subfolder, so for any pretrained
+    HF model, use:
          global_cache_dir('hub')
-    Current HuggingFace documentation (mid-Jan 2023) indicates that
-    transformers models will be cached into a "transformers" subdirectory,
-    but in practice they seem to go into "hub". But if needed:
-         global_cache_dir('transformers')
-    One other caveat is that HuggingFace is moving some diffusers models
-    into the "hub" subdirectory as well, so this will need to be revisited
-    from time to time.
-    '''
-    home: str = os.getenv('HF_HOME')
+
+    The legacy location for transformers used to be global_cache_dir('transformers')
+    and global_cache_dir('diffusers') for diffusers.
+    """
+    home: str = os.getenv("HF_HOME")
 
     if home is None:
-        home = os.getenv('XDG_CACHE_HOME')
+        home = os.getenv("XDG_CACHE_HOME")
 
         if home is not None:
-            # Set `home` to $XDG_CACHE_HOME/huggingface, which is the default location mentioned in HuggingFace Hub Client Library.
+            # Set `home` to $XDG_CACHE_HOME/huggingface, which is the default location mentioned in Hugging Face Hub Client Library.
             # See: https://huggingface.co/docs/huggingface_hub/main/en/package_reference/environment_variables#xdgcachehome
-            home += os.sep + 'huggingface'
+            home += os.sep + "huggingface"
 
     if home is not None:
-        return Path(home,subdir)
+        return Path(home, subdir)
     else:
-        return Path(Globals.root,'models',subdir)
+        return Path(Globals.root, "models", subdir)

invokeai/backend/image_util/__init__.py

@@ -0,0 +1,24 @@
+"""
+Initialization file for invokeai.backend.image_util methods.
+"""
+from .patchmatch import PatchMatch
+from .pngwriter import PngWriter, PromptFormatter, retrieve_metadata, write_metadata
+from .seamless import configure_model_padding
+from .txt2mask import Txt2Mask
+from .util import InitImageResizer, make_grid
+
+
+def debug_image(
+    debug_image, debug_text, debug_show=True, debug_result=False, debug_status=False
+):
+    if not debug_status:
+        return
+
+    image_copy = debug_image.copy().convert("RGBA")
+    ImageDraw.Draw(image_copy).text((5, 5), debug_text, (255, 0, 0))
+
+    if debug_show:
+        image_copy.show()
+
+    if debug_result:
+        return image_copy

invokeai/backend/image_util/patchmatch.py

@@ -1,20 +1,22 @@
-'''
+"""
 This module defines a singleton object, "patchmatch" that
 wraps the actual patchmatch object. It respects the global
 "try_patchmatch" attribute, so that patchmatch loading can
 be suppressed or deferred
-'''
-from ldm.invoke.globals import Globals
-import numpy as  np
+"""
+import numpy as np
+
+from invokeai.backend.globals import Globals
+
 
 class PatchMatch:
-    '''
+    """
     Thin class wrapper around the patchmatch function.
-    '''
+    """
 
     patch_match = None
-    tried_load:bool = False
-    
+    tried_load: bool = False
+
     def __init__(self):
         super().__init__()
 
@@ -24,21 +26,22 @@ class PatchMatch:
             return
         if Globals.try_patchmatch:
             from patchmatch import patch_match as pm
+
             if pm.patchmatch_available:
-                print('>> Patchmatch initialized')
+                print(">> Patchmatch initialized")
             else:
-                print('>> Patchmatch not loaded (nonfatal)')
+                print(">> Patchmatch not loaded (nonfatal)")
             self.patch_match = pm
         else:
-            print('>> Patchmatch loading disabled')
+            print(">> Patchmatch loading disabled")
         self.tried_load = True
 
     @classmethod
-    def patchmatch_available(self)->bool:
+    def patchmatch_available(self) -> bool:
         self._load_patch_match()
         return self.patch_match and self.patch_match.patchmatch_available
 
     @classmethod
-    def inpaint(self,*args,**kwargs)->np.ndarray:
+    def inpaint(self, *args, **kwargs) -> np.ndarray:
         if self.patchmatch_available():
-            return self.patch_match.inpaint(*args,**kwargs)
+            return self.patch_match.inpaint(*args, **kwargs)

invokeai/backend/image_util/pngwriter.py

@@ -6,10 +6,11 @@ PngWriter -- Converts Images generated by T2I into PNGs, finds
 
 Exports function retrieve_metadata(path)
 """
+import json
 import os
 import re
-import json
-from PIL import PngImagePlugin, Image
+
+from PIL import Image, PngImagePlugin
 
 # -------------------image generation utils-----
 
@@ -25,52 +26,57 @@ class PngWriter:
         dirlist = sorted(os.listdir(self.outdir), reverse=True)
         # find the first filename that matches our pattern or return 000000.0.png
         existing_name = next(
-            (f for f in dirlist if re.match('^(\d+)\..*\.png', f)),
-            '0000000.0.png',
+            (f for f in dirlist if re.match("^(\d+)\..*\.png", f)),
+            "0000000.0.png",
         )
-        basecount = int(existing_name.split('.', 1)[0]) + 1
-        return f'{basecount:06}'
+        basecount = int(existing_name.split(".", 1)[0]) + 1
+        return f"{basecount:06}"
 
     # saves image named _image_ to outdir/name, writing metadata from prompt
     # returns full path of output
-    def save_image_and_prompt_to_png(self, image, dream_prompt, name, metadata=None, compress_level=6):
+    def save_image_and_prompt_to_png(
+        self, image, dream_prompt, name, metadata=None, compress_level=6
+    ):
         path = os.path.join(self.outdir, name)
         info = PngImagePlugin.PngInfo()
-        info.add_text('Dream', dream_prompt)
+        info.add_text("Dream", dream_prompt)
         if metadata:
-            info.add_text('sd-metadata', json.dumps(metadata))
-        image.save(path, 'PNG', pnginfo=info, compress_level=compress_level)
+            info.add_text("sd-metadata", json.dumps(metadata))
+        image.save(path, "PNG", pnginfo=info, compress_level=compress_level)
         return path
 
-    def retrieve_metadata(self,img_basename):
-        '''
+    def retrieve_metadata(self, img_basename):
+        """
         Given a PNG filename stored in outdir, returns the "sd-metadata"
         metadata stored there, as a dict
-        '''
-        path = os.path.join(self.outdir,img_basename)
+        """
+        path = os.path.join(self.outdir, img_basename)
         all_metadata = retrieve_metadata(path)
-        return all_metadata['sd-metadata']
+        return all_metadata["sd-metadata"]
+
 
 def retrieve_metadata(img_path):
-    '''
+    """
     Given a path to a PNG image, returns the "sd-metadata"
     metadata stored there, as a dict
-    '''
+    """
     im = Image.open(img_path)
-    if hasattr(im, 'text'):
-        md = im.text.get('sd-metadata', '{}')
-        dream_prompt = im.text.get('Dream', '')
+    if hasattr(im, "text"):
+        md = im.text.get("sd-metadata", "{}")
+        dream_prompt = im.text.get("Dream", "")
     else:
         # When trying to retrieve metadata from images without a 'text' payload, such as JPG images.
-        md = '{}'
-        dream_prompt = ''
-    return {'sd-metadata': json.loads(md), 'Dream': dream_prompt}
+        md = "{}"
+        dream_prompt = ""
+    return {"sd-metadata": json.loads(md), "Dream": dream_prompt}
 
-def write_metadata(img_path:str, meta:dict):
+
+def write_metadata(img_path: str, meta: dict):
     im = Image.open(img_path)
     info = PngImagePlugin.PngInfo()
-    info.add_text('sd-metadata', json.dumps(meta))
-    im.save(img_path,'PNG',pnginfo=info)
+    info.add_text("sd-metadata", json.dumps(meta))
+    im.save(img_path, "PNG", pnginfo=info)
+
 
 class PromptFormatter:
     def __init__(self, t2i, opt):
@@ -86,28 +92,30 @@ class PromptFormatter:
 
         switches = list()
         switches.append(f'"{opt.prompt}"')
-        switches.append(f'-s{opt.steps        or t2i.steps}')
-        switches.append(f'-W{opt.width        or t2i.width}')
-        switches.append(f'-H{opt.height       or t2i.height}')
-        switches.append(f'-C{opt.cfg_scale    or t2i.cfg_scale}')
-        switches.append(f'-A{opt.sampler_name or t2i.sampler_name}')
-# to do: put model name into the t2i object
-#        switches.append(f'--model{t2i.model_name}')
+        switches.append(f"-s{opt.steps        or t2i.steps}")
+        switches.append(f"-W{opt.width        or t2i.width}")
+        switches.append(f"-H{opt.height       or t2i.height}")
+        switches.append(f"-C{opt.cfg_scale    or t2i.cfg_scale}")
+        switches.append(f"-A{opt.sampler_name or t2i.sampler_name}")
+        # to do: put model name into the t2i object
+        #        switches.append(f'--model{t2i.model_name}')
         if opt.seamless or t2i.seamless:
-            switches.append(f'--seamless')
+            switches.append(f"--seamless")
         if opt.init_img:
-            switches.append(f'-I{opt.init_img}')
+            switches.append(f"-I{opt.init_img}")
         if opt.fit:
-            switches.append(f'--fit')
+            switches.append(f"--fit")
         if opt.strength and opt.init_img is not None:
-            switches.append(f'-f{opt.strength or t2i.strength}')
+            switches.append(f"-f{opt.strength or t2i.strength}")
         if opt.gfpgan_strength:
-            switches.append(f'-G{opt.gfpgan_strength}')
+            switches.append(f"-G{opt.gfpgan_strength}")
         if opt.upscale:
             switches.append(f'-U {" ".join([str(u) for u in opt.upscale])}')
         if opt.variation_amount > 0:
-            switches.append(f'-v{opt.variation_amount}')
+            switches.append(f"-v{opt.variation_amount}")
         if opt.with_variations:
-            formatted_variations = ','.join(f'{seed}:{weight}' for seed, weight in opt.with_variations)
-            switches.append(f'-V{formatted_variations}')
-        return ' '.join(switches)
+            formatted_variations = ",".join(
+                f"{seed}:{weight}" for seed, weight in opt.with_variations
+            )
+            switches.append(f"-V{formatted_variations}")
+        return " ".join(switches)

invokeai/backend/image_util/seamless.py

@@ -0,0 +1,59 @@
+import torch.nn as nn
+
+
+def _conv_forward_asymmetric(self, input, weight, bias):
+    """
+    Patch for Conv2d._conv_forward that supports asymmetric padding
+    """
+    working = nn.functional.pad(
+        input, self.asymmetric_padding["x"], mode=self.asymmetric_padding_mode["x"]
+    )
+    working = nn.functional.pad(
+        working, self.asymmetric_padding["y"], mode=self.asymmetric_padding_mode["y"]
+    )
+    return nn.functional.conv2d(
+        working,
+        weight,
+        bias,
+        self.stride,
+        nn.modules.utils._pair(0),
+        self.dilation,
+        self.groups,
+    )
+
+
+def configure_model_padding(model, seamless, seamless_axes):
+    """
+    Modifies the 2D convolution layers to use a circular padding mode based on the `seamless` and `seamless_axes` options.
+    """
+    # TODO: get an explicit interface for this in diffusers: https://github.com/huggingface/diffusers/issues/556
+    for m in model.modules():
+        if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
+            if seamless:
+                m.asymmetric_padding_mode = {}
+                m.asymmetric_padding = {}
+                m.asymmetric_padding_mode["x"] = (
+                    "circular" if ("x" in seamless_axes) else "constant"
+                )
+                m.asymmetric_padding["x"] = (
+                    m._reversed_padding_repeated_twice[0],
+                    m._reversed_padding_repeated_twice[1],
+                    0,
+                    0,
+                )
+                m.asymmetric_padding_mode["y"] = (
+                    "circular" if ("y" in seamless_axes) else "constant"
+                )
+                m.asymmetric_padding["y"] = (
+                    0,
+                    0,
+                    m._reversed_padding_repeated_twice[2],
+                    m._reversed_padding_repeated_twice[3],
+                )
+                m._conv_forward = _conv_forward_asymmetric.__get__(m, nn.Conv2d)
+            else:
+                m._conv_forward = nn.Conv2d._conv_forward.__get__(m, nn.Conv2d)
+                if hasattr(m, "asymmetric_padding_mode"):
+                    del m.asymmetric_padding_mode
+                if hasattr(m, "asymmetric_padding"):
+                    del m.asymmetric_padding

invokeai/backend/image_util/txt2mask.py

@@ -1,9 +1,9 @@
-'''Makes available the Txt2Mask class, which assists in the automatic
+"""Makes available the Txt2Mask class, which assists in the automatic
 assignment of masks via text prompt using clipseg.
 
 Here is typical usage:
 
-    from ldm.invoke.txt2mask import Txt2Mask, SegmentedGrayscale
+    from invokeai.backend.image_util.txt2mask import Txt2Mask, SegmentedGrayscale
     from PIL import Image
 
     txt2mask = Txt2Mask(self.device)
@@ -25,31 +25,39 @@ the mask that exceed the indicated confidence threshold. Values range
 from 0.0 to 1.0. The higher the threshold, the more confident the
 algorithm is. In limited testing, I have found that values around 0.5
 work fine.
-'''
+"""
 
+import numpy as np
 import torch
-import numpy as  np
-from transformers import AutoProcessor, CLIPSegForImageSegmentation
 from PIL import Image, ImageOps
 from torchvision import transforms
-from ldm.invoke.globals import global_cache_dir
+from transformers import AutoProcessor, CLIPSegForImageSegmentation
 
-CLIPSEG_MODEL = 'CIDAS/clipseg-rd64-refined'
+from invokeai.backend.globals import global_cache_dir
+
+CLIPSEG_MODEL = "CIDAS/clipseg-rd64-refined"
 CLIPSEG_SIZE = 352
 
+
 class SegmentedGrayscale(object):
-    def __init__(self, image:Image, heatmap:torch.Tensor):
+    def __init__(self, image: Image, heatmap: torch.Tensor):
         self.heatmap = heatmap
         self.image = image
 
-    def to_grayscale(self,invert:bool=False)->Image:
-        return self._rescale(Image.fromarray(np.uint8(255 - self.heatmap * 255 if invert else self.heatmap * 255)))
+    def to_grayscale(self, invert: bool = False) -> Image:
+        return self._rescale(
+            Image.fromarray(
+                np.uint8(255 - self.heatmap * 255 if invert else self.heatmap * 255)
+            )
+        )
 
-    def to_mask(self,threshold:float=0.5)->Image:
+    def to_mask(self, threshold: float = 0.5) -> Image:
         discrete_heatmap = self.heatmap.lt(threshold).int()
-        return self._rescale(Image.fromarray(np.uint8(discrete_heatmap*255),mode='L'))
+        return self._rescale(
+            Image.fromarray(np.uint8(discrete_heatmap * 255), mode="L")
+        )
 
-    def to_transparent(self,invert:bool=False)->Image:
+    def to_transparent(self, invert: bool = False) -> Image:
         transparent_image = self.image.copy()
         # For img2img, we want the selected regions to be transparent,
         # but to_grayscale() returns the opposite. Thus invert.
@@ -58,70 +66,77 @@ class SegmentedGrayscale(object):
         return transparent_image
 
     # unscales and uncrops the 352x352 heatmap so that it matches the image again
-    def _rescale(self, heatmap:Image)->Image:
-        size = self.image.width if (self.image.width > self.image.height) else self.image.height
-        resized_image = heatmap.resize(
-            (size,size),
-            resample=Image.Resampling.LANCZOS
+    def _rescale(self, heatmap: Image) -> Image:
+        size = (
+            self.image.width
+            if (self.image.width > self.image.height)
+            else self.image.height
         )
-        return resized_image.crop((0,0,self.image.width,self.image.height))
+        resized_image = heatmap.resize((size, size), resample=Image.Resampling.LANCZOS)
+        return resized_image.crop((0, 0, self.image.width, self.image.height))
+
 
 class Txt2Mask(object):
-    '''
+    """
     Create new Txt2Mask object. The optional device argument can be one of
     'cuda', 'mps' or 'cpu'.
-    '''
-    def __init__(self,device='cpu',refined=False):
-        print('>> Initializing clipseg model for text to mask inference')
+    """
+
+    def __init__(self, device="cpu", refined=False):
+        print(">> Initializing clipseg model for text to mask inference")
 
         # BUG: we are not doing anything with the device option at this time
         self.device = device
-        self.processor = AutoProcessor.from_pretrained(CLIPSEG_MODEL,
-                                                       cache_dir=global_cache_dir('hub')
-                                                       )
-        self.model = CLIPSegForImageSegmentation.from_pretrained(CLIPSEG_MODEL,
-                                                                 cache_dir=global_cache_dir('hub')
-                                                                 )
+        self.processor = AutoProcessor.from_pretrained(
+            CLIPSEG_MODEL, cache_dir=global_cache_dir("hub")
+        )
+        self.model = CLIPSegForImageSegmentation.from_pretrained(
+            CLIPSEG_MODEL, cache_dir=global_cache_dir("hub")
+        )
 
     @torch.no_grad()
-    def segment(self, image, prompt:str) -> SegmentedGrayscale:
-        '''
+    def segment(self, image, prompt: str) -> SegmentedGrayscale:
+        """
         Given a prompt string such as "a bagel", tries to identify the object in the
         provided image and returns a SegmentedGrayscale object in which the brighter
         pixels indicate where the object is inferred to be.
-        '''
-        transform = transforms.Compose([
-            transforms.ToTensor(),
-            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-            transforms.Resize((CLIPSEG_SIZE, CLIPSEG_SIZE)), # must be multiple of 64...
-        ])
+        """
+        transform = transforms.Compose(
+            [
+                transforms.ToTensor(),
+                transforms.Normalize(
+                    mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+                ),
+                transforms.Resize(
+                    (CLIPSEG_SIZE, CLIPSEG_SIZE)
+                ),  # must be multiple of 64...
+            ]
+        )
 
         if type(image) is str:
-            image = Image.open(image).convert('RGB')
+            image = Image.open(image).convert("RGB")
 
         image = ImageOps.exif_transpose(image)
         img = self._scale_and_crop(image)
 
-        inputs = self.processor(text=[prompt],
-                                images=[img],
-                                padding=True,
-                                return_tensors='pt')
+        inputs = self.processor(
+            text=[prompt], images=[img], padding=True, return_tensors="pt"
+        )
         outputs = self.model(**inputs)
         heatmap = torch.sigmoid(outputs.logits)
         return SegmentedGrayscale(image, heatmap)
 
-    def _scale_and_crop(self, image:Image)->Image:
-        scaled_image = Image.new('RGB',(CLIPSEG_SIZE,CLIPSEG_SIZE))
-        if image.width > image.height: # width is constraint
+    def _scale_and_crop(self, image: Image) -> Image:
+        scaled_image = Image.new("RGB", (CLIPSEG_SIZE, CLIPSEG_SIZE))
+        if image.width > image.height:  # width is constraint
             scale = CLIPSEG_SIZE / image.width
         else:
             scale = CLIPSEG_SIZE / image.height
         scaled_image.paste(
             image.resize(
-                (int(scale * image.width),
-                 int(scale * image.height)
-                ),
-                resample=Image.Resampling.LANCZOS
-            ),box=(0,0)
+                (int(scale * image.width), int(scale * image.height)),
+                resample=Image.Resampling.LANCZOS,
+            ),
+            box=(0, 0),
         )
         return scaled_image

invokeai/backend/image_util/util.py

@@ -1,12 +1,15 @@
-from math import sqrt, floor, ceil
+from math import ceil, floor, sqrt
+
 from PIL import Image
 
-class InitImageResizer():
+
+class InitImageResizer:
     """Simple class to create resized copies of an Image while preserving the aspect ratio."""
-    def __init__(self,Image):
+
+    def __init__(self, Image):
         self.image = Image
 
-    def resize(self,width=None,height=None) -> Image:
+    def resize(self, width=None, height=None) -> Image:
         """
         Return a copy of the image resized to fit within
         a box width x height. The aspect ratio is
@@ -18,37 +21,36 @@ class InitImageResizer():
         Everything is floored to the nearest multiple of 64 so
         that it can be passed to img2img()
         """
-        im    = self.image
+        im = self.image
 
-        ar = im.width/float(im.height)
+        ar = im.width / float(im.height)
 
         # Infer missing values from aspect ratio
-        if not(width or height): # both missing
-            width  = im.width
+        if not (width or height):  # both missing
+            width = im.width
             height = im.height
-        elif not height:           # height missing
-            height = int(width/ar)
-        elif not width:            # width missing
-            width  = int(height*ar)
+        elif not height:  # height missing
+            height = int(width / ar)
+        elif not width:  # width missing
+            width = int(height * ar)
 
-        w_scale = width/im.width
-        h_scale = height/im.height
-        scale = min(w_scale,h_scale)
-        (rw,rh) = (int(scale*im.width),int(scale*im.height))
+        w_scale = width / im.width
+        h_scale = height / im.height
+        scale = min(w_scale, h_scale)
+        (rw, rh) = (int(scale * im.width), int(scale * im.height))
 
-        #round everything to multiples of 64
-        width,height,rw,rh = map(
-            lambda x: x-x%64, (width,height,rw,rh)
-        )
+        # round everything to multiples of 64
+        width, height, rw, rh = map(lambda x: x - x % 64, (width, height, rw, rh))
 
         # no resize necessary, but return a copy
         if im.width == width and im.height == height:
             return im.copy()
 
         # otherwise resize the original image so that it fits inside the bounding box
-        resized_image = self.image.resize((rw,rh),resample=Image.Resampling.LANCZOS)
+        resized_image = self.image.resize((rw, rh), resample=Image.Resampling.LANCZOS)
         return resized_image
 
+
 def make_grid(image_list, rows=None, cols=None):
     image_cnt = len(image_list)
     if None in (rows, cols):
@@ -57,7 +59,7 @@ def make_grid(image_list, rows=None, cols=None):
     width = image_list[0].width
     height = image_list[0].height
 
-    grid_img = Image.new('RGB', (width * cols, height * rows))
+    grid_img = Image.new("RGB", (width * cols, height * rows))
     i = 0
     for r in range(0, rows):
         for c in range(0, cols):
@@ -67,4 +69,3 @@ def make_grid(image_list, rows=None, cols=None):
             i = i + 1
 
     return grid_img
-

invokeai/backend/model_management/__init__.py

@@ -0,0 +1,9 @@
+"""
+Initialization file for invokeai.backend.model_management
+"""
+from .convert_ckpt_to_diffusers import (
+    convert_ckpt_to_diffusers,
+    load_pipeline_from_original_stable_diffusion_ckpt,
+)
+from .model_manager import ModelManager
+

invokeai/backend/model_management/model_manager.py

@@ -9,7 +9,6 @@ from __future__ import annotations
 import contextlib
 import gc
 import hashlib
-import io
 import os
 import re
 import sys
@@ -32,48 +31,48 @@ from omegaconf import OmegaConf
 from omegaconf.dictconfig import DictConfig
 from picklescan.scanner import scan_file_path
 
-from ldm.invoke.devices import CPU_DEVICE
-from ldm.invoke.generator.diffusers_pipeline import StableDiffusionGeneratorPipeline
-from ldm.invoke.globals import Globals, global_cache_dir
-from ldm.util import (
-    ask_user,
-    download_with_resume,
-    instantiate_from_config,
-    url_attachment_name,
-)
+from invokeai.backend.globals import Globals, global_cache_dir
 
+from ..stable_diffusion import StableDiffusionGeneratorPipeline
+from ..util import CUDA_DEVICE, CPU_DEVICE, ask_user, download_with_resume
 
 class SDLegacyType(Enum):
     V1 = 1
     V1_INPAINT = 2
-    V2 = 3
+    V2   = 3
+    V2_e = 4
+    V2_v = 5
     UNKNOWN = 99
 
-
 DEFAULT_MAX_MODELS = 2
 VAE_TO_REPO_ID = {  # hack, see note in convert_and_import()
     "vae-ft-mse-840000-ema-pruned": "stabilityai/sd-vae-ft-mse",
 }
 
-
 class ModelManager(object):
+    '''
+    Model manager handles loading, caching, importing, deleting, converting, and editing models.
+    '''
     def __init__(
-        self,
-        config: OmegaConf,
-        device_type: torch.device = CPU_DEVICE,
-        precision: str = "float16",
-        max_loaded_models=DEFAULT_MAX_MODELS,
-        sequential_offload=False,
+            self,
+            config: OmegaConf|Path,
+            device_type: torch.device = CUDA_DEVICE,
+            precision: str = "float16",
+            max_loaded_models=DEFAULT_MAX_MODELS,
+            sequential_offload=False,
+            embedding_path: Path=None,
     ):
         """
-        Initialize with the path to the models.yaml config file,
-        the torch device type, and precision. The optional
-        min_avail_mem argument specifies how much unused system
-        (CPU) memory to preserve. The cache of models in RAM will
-        grow until this value is approached. Default is 2G.
+        Initialize with the path to the models.yaml config file or
+        an initialized OmegaConf dictionary. Optional parameters
+        are the torch device type, precision, max_loaded_models,
+        and sequential_offload boolean. Note that the default device 
+        type and precision are set up for a CUDA system running at half precision.
         """
         # prevent nasty-looking CLIP log message
         transformers.logging.set_verbosity_error()
+        if not isinstance(config, DictConfig):
+            config = OmegaConf.load(config)
         self.config = config
         self.precision = precision
         self.device = torch.device(device_type)
@@ -82,6 +81,7 @@ class ModelManager(object):
         self.stack = []  # this is an LRU FIFO
         self.current_model = None
         self.sequential_offload = sequential_offload
+        self.embedding_path = embedding_path
 
     def valid_model(self, model_name: str) -> bool:
         """
@@ -90,12 +90,15 @@ class ModelManager(object):
         """
         return model_name in self.config
 
-    def get_model(self, model_name: str):
+    def get_model(self, model_name: str=None)->dict:
         """
         Given a model named identified in models.yaml, return
         the model object. If in RAM will load into GPU VRAM.
         If on disk, will load from there.
         """
+        if not model_name:
+            return self.get_model(self.current_model) if self.current_model else self.get_model(self.default_model())
+        
         if not self.valid_model(model_name):
             print(
                 f'** "{model_name}" is not a known model name. Please check your models.yaml file'
@@ -110,7 +113,7 @@ class ModelManager(object):
         if model_name in self.models:
             requested_model = self.models[model_name]["model"]
             print(f">> Retrieving model {model_name} from system RAM cache")
-            self.models[model_name]["model"] = self._model_from_cpu(requested_model)
+            requested_model.ready()
             width = self.models[model_name]["width"]
             height = self.models[model_name]["height"]
             hash = self.models[model_name]["hash"]
@@ -118,6 +121,7 @@ class ModelManager(object):
         else:  # we're about to load a new model, so potentially offload the least recently used one
             requested_model, width, height, hash = self._load_model(model_name)
             self.models[model_name] = {
+                "model_name": model_name,
                 "model": requested_model,
                 "width": width,
                 "height": height,
@@ -127,6 +131,7 @@ class ModelManager(object):
         self.current_model = model_name
         self._push_newest_model(model_name)
         return {
+            "model_name": model_name,
             "model": requested_model,
             "width": width,
             "height": height,
@@ -357,7 +362,7 @@ class ModelManager(object):
             raise NotImplementedError(
                 f"Unknown model format {model_name}: {model_format}"
             )
-
+        
         # usage statistics
         toc = time.time()
         print(">> Model loaded in", "%4.2fs" % (toc - tic))
@@ -370,125 +375,6 @@ class ModelManager(object):
             )
         return model, width, height, model_hash
 
-    def _load_ckpt_model(self, model_name, mconfig):
-        config = mconfig.config
-        weights = mconfig.weights
-        vae = mconfig.get("vae")
-        width = mconfig.width
-        height = mconfig.height
-
-        if not os.path.isabs(config):
-            config = os.path.join(Globals.root, config)
-        if not os.path.isabs(weights):
-            weights = os.path.normpath(os.path.join(Globals.root, weights))
-
-        # if converting automatically to diffusers, then we do the conversion and return
-        # a diffusers pipeline
-        if Globals.ckpt_convert:
-            print(
-                f">> Converting legacy checkpoint {model_name} into a diffusers model..."
-            )
-            from ldm.invoke.ckpt_to_diffuser import (
-                load_pipeline_from_original_stable_diffusion_ckpt,
-            )
-
-            self.offload_model(self.current_model)
-            if vae_config := self._choose_diffusers_vae(model_name):
-                vae = self._load_vae(vae_config)
-            if self._has_cuda():
-                torch.cuda.empty_cache()
-            pipeline = load_pipeline_from_original_stable_diffusion_ckpt(
-                checkpoint_path=weights,
-                original_config_file=config,
-                vae=vae,
-                return_generator_pipeline=True,
-                precision=torch.float16
-                if self.precision == "float16"
-                else torch.float32,
-            )
-            if self.sequential_offload:
-                pipeline.enable_offload_submodels(self.device)
-            else:
-                pipeline.to(self.device)
-
-            return (
-                pipeline,
-                width,
-                height,
-                "NOHASH",
-            )
-
-        # scan model
-        self.scan_model(model_name, weights)
-
-        print(f">> Loading {model_name} from {weights}")
-
-        # for usage statistics
-        if self._has_cuda():
-            torch.cuda.reset_peak_memory_stats()
-            torch.cuda.empty_cache()
-
-        # this does the work
-        if not os.path.isabs(config):
-            config = os.path.join(Globals.root, config)
-        omega_config = OmegaConf.load(config)
-        with open(weights, "rb") as f:
-            weight_bytes = f.read()
-        model_hash = self._cached_sha256(weights, weight_bytes)
-        sd = None
-        if weights.endswith(".safetensors"):
-            sd = safetensors.torch.load(weight_bytes)
-        else:
-            sd = torch.load(io.BytesIO(weight_bytes), map_location="cpu")
-        del weight_bytes
-        # merged models from auto11 merge board are flat for some reason
-        if "state_dict" in sd:
-            sd = sd["state_dict"]
-
-        print("   | Forcing garbage collection prior to loading new model")
-        gc.collect()
-        model = instantiate_from_config(omega_config.model)
-        model.load_state_dict(sd, strict=False)
-
-        if self.precision == "float16":
-            print("   | Using faster float16 precision")
-            model = model.to(torch.float16)
-        else:
-            print("   | Using more accurate float32 precision")
-
-        # look and load a matching vae file. Code borrowed from AUTOMATIC1111 modules/sd_models.py
-        if vae:
-            if not os.path.isabs(vae):
-                vae = os.path.normpath(os.path.join(Globals.root, vae))
-            if os.path.exists(vae):
-                print(f"   | Loading VAE weights from: {vae}")
-                vae_ckpt = None
-                vae_dict = None
-                if vae.endswith(".safetensors"):
-                    vae_ckpt = safetensors.torch.load_file(vae)
-                    vae_dict = {k: v for k, v in vae_ckpt.items() if k[0:4] != "loss"}
-                else:
-                    vae_ckpt = torch.load(vae, map_location="cpu")
-                    vae_dict = {
-                        k: v
-                        for k, v in vae_ckpt["state_dict"].items()
-                        if k[0:4] != "loss"
-                    }
-                model.first_stage_model.load_state_dict(vae_dict, strict=False)
-            else:
-                print(f"   | VAE file {vae} not found. Skipping.")
-
-        model.to(self.device)
-        # model.to doesn't change the cond_stage_model.device used to move the tokenizer output, so set it here
-        model.cond_stage_model.device = self.device
-
-        model.eval()
-
-        for module in model.modules():
-            if isinstance(module, (torch.nn.Conv2d, torch.nn.ConvTranspose2d)):
-                module._orig_padding_mode = module.padding_mode
-        return model, width, height, model_hash
-
     def _load_diffusers_model(self, mconfig):
         name_or_path = self.model_name_or_path(mconfig)
         using_fp16 = self.precision == "float16"
@@ -507,7 +393,7 @@ class ModelManager(object):
             if vae := self._load_vae(mconfig["vae"]):
                 pipeline_args.update(vae=vae)
         if not isinstance(name_or_path, Path):
-            pipeline_args.update(cache_dir=global_cache_dir("diffusers"))
+            pipeline_args.update(cache_dir=global_cache_dir("hub"))
         if using_fp16:
             pipeline_args.update(torch_dtype=torch.float16)
             fp_args_list = [{"revision": "fp16"}, {}]
@@ -550,9 +436,51 @@ class ModelManager(object):
         height = width
 
         print(f"  | Default image dimensions = {width} x {height}")
+        self._add_embeddings_to_model(pipeline)
 
         return pipeline, width, height, model_hash
 
+    def _load_ckpt_model(self, model_name, mconfig):
+        config = mconfig.config
+        weights = mconfig.weights
+        vae = mconfig.get("vae")
+        width = mconfig.width
+        height = mconfig.height
+
+        if not os.path.isabs(config):
+            config = os.path.join(Globals.root, config)
+        if not os.path.isabs(weights):
+            weights = os.path.normpath(os.path.join(Globals.root, weights))
+
+        # Convert to diffusers and return a diffusers pipeline
+        print(f">> Converting legacy checkpoint {model_name} into a diffusers model...")
+
+        from . import load_pipeline_from_original_stable_diffusion_ckpt
+
+        self.offload_model(self.current_model)
+        if vae_config := self._choose_diffusers_vae(model_name):
+            vae = self._load_vae(vae_config)
+        if self._has_cuda():
+            torch.cuda.empty_cache()
+        pipeline = load_pipeline_from_original_stable_diffusion_ckpt(
+            checkpoint_path=weights,
+            original_config_file=config,
+            vae=vae,
+            return_generator_pipeline=True,
+            precision=torch.float16 if self.precision == "float16" else torch.float32,
+        )
+        if self.sequential_offload:
+            pipeline.enable_offload_submodels(self.device)
+        else:
+            pipeline.to(self.device)
+
+        return (
+            pipeline,
+            width,
+            height,
+            "NOHASH",
+        )
+
     def model_name_or_path(self, model_name: Union[str, DictConfig]) -> str | Path:
         if isinstance(model_name, DictConfig) or isinstance(model_name, dict):
             mconfig = model_name
@@ -583,7 +511,7 @@ class ModelManager(object):
 
         print(f">> Offloading {model_name} to CPU")
         model = self.models[model_name]["model"]
-        self.models[model_name]["model"] = self._model_to_cpu(model)
+        model.offload_all()
 
         gc.collect()
         if self._has_cuda():
@@ -624,7 +552,7 @@ class ModelManager(object):
         self,
         repo_or_path: Union[str, Path],
         model_name: str = None,
-        model_description: str = None,
+        description: str = None,
         vae: dict = None,
         commit_to_conf: Path = None,
     ) -> bool:
@@ -640,7 +568,9 @@ class ModelManager(object):
         models.yaml file.
         """
         model_name = model_name or Path(repo_or_path).stem
-        model_description = model_description or f"Imported diffusers model {model_name}"
+        model_description = (
+            description or f"Imported diffusers model {model_name}"
+        )
         new_config = dict(
             description=model_description,
             vae=vae,
@@ -656,66 +586,6 @@ class ModelManager(object):
             self.commit(commit_to_conf)
         return model_name
 
-    def import_ckpt_model(
-        self,
-        weights: Union[str, Path],
-        config: Union[str, Path] = "configs/stable-diffusion/v1-inference.yaml",
-        vae: Union[str, Path] = None,
-        model_name: str = None,
-        model_description: str = None,
-        commit_to_conf: Path = None,
-    ) -> str:
-        """
-        Attempts to install the indicated ckpt file and returns True if successful.
-
-        "weights" can be either a path-like object corresponding to a local .ckpt file
-        or a http/https URL pointing to a remote model.
-
-        "vae" is a Path or str object pointing to a ckpt or safetensors file to be used
-        as the VAE for this model.
-
-        "config" is the model config file to use with this ckpt file. It defaults to
-        v1-inference.yaml. If a URL is provided, the config will be downloaded.
-
-        You can optionally provide a model name and/or description. If not provided,
-        then these will be derived from the weight file name. If you provide a commit_to_conf
-        path to the configuration file, then the new entry will be committed to the
-        models.yaml file.
-
-        Return value is the name of the imported file, or None if an error occurred.
-        """
-        if str(weights).startswith(("http:", "https:")):
-            model_name = model_name or url_attachment_name(weights)
-
-        weights_path = self._resolve_path(weights, "models/ldm/stable-diffusion-v1")
-        config_path = self._resolve_path(config, "configs/stable-diffusion")
-
-        if weights_path is None or not weights_path.exists():
-            return
-        if config_path is None or not config_path.exists():
-            return
-
-        model_name = (
-            model_name or Path(weights).stem
-        )  # note this gives ugly pathnames if used on a URL without a Content-Disposition header
-        model_description = (
-            model_description or f"Imported stable diffusion weights file {model_name}"
-        )
-        new_config = dict(
-            weights=str(weights_path),
-            config=str(config_path),
-            description=model_description,
-            format="ckpt",
-            width=512,
-            height=512,
-        )
-        if vae:
-            new_config["vae"] = vae
-        self.add_model(model_name, new_config, True)
-        if commit_to_conf:
-            self.commit(commit_to_conf)
-        return model_name
-
     @classmethod
     def probe_model_type(self, checkpoint: dict) -> SDLegacyType:
         """
@@ -724,15 +594,24 @@ class ModelManager(object):
         format. Valid return values include:
         SDLegacyType.V1
         SDLegacyType.V1_INPAINT
-        SDLegacyType.V2
+        SDLegacyType.V2     (V2 prediction type unknown)
+        SDLegacyType.V2_e   (V2 using 'epsilon' prediction type)
+        SDLegacyType.V2_v   (V2 using 'v_prediction' prediction type)
         SDLegacyType.UNKNOWN
         """
-        key_name = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight"
-        if key_name in checkpoint and checkpoint[key_name].shape[-1] == 1024:
-            return SDLegacyType.V2
+        global_step = checkpoint.get('global_step')
+        state_dict = checkpoint.get("state_dict") or checkpoint
 
         try:
-            state_dict = checkpoint.get("state_dict") or checkpoint
+            key_name = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight"
+            if key_name in state_dict and state_dict[key_name].shape[-1] == 1024:
+                if global_step == 220000:
+                    return SDLegacyType.V2_e
+                elif global_step == 110000:
+                    return SDLegacyType.V2_v
+                else:
+                    return SDLegacyType.V2
+            # otherwise we assume a V1 file
             in_channels = state_dict[
                 "model.diffusion_model.input_blocks.0.0.weight"
             ].shape[1]
@@ -746,12 +625,13 @@ class ModelManager(object):
             return SDLegacyType.UNKNOWN
 
     def heuristic_import(
-        self,
-        path_url_or_repo: str,
-        convert: bool = False,
-        model_name: str = None,
-        description: str = None,
-        commit_to_conf: Path = None,
+            self,
+            path_url_or_repo: str,
+            convert: bool = True,
+            model_name: str = None,
+            description: str = None,
+            model_config_file: Path = None,
+            commit_to_conf: Path = None,
     ) -> str:
         """
         Accept a string which could be:
@@ -848,70 +728,67 @@ class ModelManager(object):
 
         if model_path.stem in self.config:  # already imported
             print("  | Already imported. Skipping")
-            return
+            return model_path.stem
 
         # another round of heuristics to guess the correct config file.
         checkpoint = (
-            safetensors.torch.load_file(model_path)
-            if model_path.suffix == ".safetensors"
-            else torch.load(model_path)
+            torch.load(model_path)
+            if model_path.suffix == ".ckpt"
+            else safetensors.torch.load_file(model_path)
         )
-        model_type = self.probe_model_type(checkpoint)
 
-        model_config_file = None
-        if model_type == SDLegacyType.V1:
-            print("  | SD-v1 model detected")
-            model_config_file = Path(
-                Globals.root, "configs/stable-diffusion/v1-inference.yaml"
-            )
-        elif model_type == SDLegacyType.V1_INPAINT:
-            print("  | SD-v1 inpainting model detected")
-            model_config_file = Path(
-                Globals.root, "configs/stable-diffusion/v1-inpainting-inference.yaml"
-            )
-        elif model_type == SDLegacyType.V2:
-            print(
-                "  | SD-v2 model detected; model will be converted to diffusers format"
-            )
-            model_config_file = Path(
-                Globals.root, "configs/stable-diffusion/v2-inference-v.yaml"
-            )
-            convert = True
-        else:
-            print(
-                f"** {thing} is a legacy checkpoint file but not in a known Stable Diffusion model. Skipping import"
-            )
-            return
+        # additional probing needed if no config file provided
+        if model_config_file is None:
+            model_type = self.probe_model_type(checkpoint)
+            if model_type == SDLegacyType.V1:
+                print("  | SD-v1 model detected")
+                model_config_file = Path(
+                    Globals.root, "configs/stable-diffusion/v1-inference.yaml"
+                )
+            elif model_type == SDLegacyType.V1_INPAINT:
+                print("  | SD-v1 inpainting model detected")
+                model_config_file = Path(
+                    Globals.root, "configs/stable-diffusion/v1-inpainting-inference.yaml"
+                )
+            elif model_type == SDLegacyType.V2_v:
+                print(
+                    "  | SD-v2-v model detected; model will be converted to diffusers format"
+                )
+                model_config_file = Path(
+                    Globals.root, "configs/stable-diffusion/v2-inference-v.yaml"
+                )
+                convert = True
+            elif model_type == SDLegacyType.V2_e:
+                print(
+                    "  | SD-v2-e model detected; model will be converted to diffusers format"
+                )
+                model_config_file = Path(
+                    Globals.root, "configs/stable-diffusion/v2-inference.yaml"
+                )
+                convert = True
+            elif model_type == SDLegacyType.V2:
+                print(
+                    f"** {thing} is a V2 checkpoint file, but its parameterization cannot be determined. Please provide configuration file path."
+                )
+                return
+            else:
+                print(
+                    f"** {thing} is a legacy checkpoint file but not a known Stable Diffusion model. Please provide configuration file path."
+                )
+                return
 
-        if convert:
-            diffuser_path = Path(
-                Globals.root, "models", Globals.converted_ckpts_dir, model_path.stem
-            )
-            model_name = self.convert_and_import(
-                model_path,
-                diffusers_path=diffuser_path,
-                vae=dict(repo_id="stabilityai/sd-vae-ft-mse"),
-                model_name=model_name,
-                model_description=description,
-                original_config_file=model_config_file,
-                commit_to_conf=commit_to_conf,
-            )
-        else:
-            model_name = self.import_ckpt_model(
-                model_path,
-                config=model_config_file,
-                model_name=model_name,
-                model_description=description,
-                vae=str(
-                    Path(
-                        Globals.root,
-                        "models/ldm/stable-diffusion-v1/vae-ft-mse-840000-ema-pruned.ckpt",
-                    )
-                ),
-                commit_to_conf=commit_to_conf,
-            )
-        if commit_to_conf:
-            self.commit(commit_to_conf)
+        diffuser_path = Path(
+            Globals.root, "models", Globals.converted_ckpts_dir, model_path.stem
+        )
+        model_name = self.convert_and_import(
+            model_path,
+            diffusers_path=diffuser_path,
+            vae=dict(repo_id="stabilityai/sd-vae-ft-mse"),
+            model_name=model_name,
+            model_description=description,
+            original_config_file=model_config_file,
+            commit_to_conf=commit_to_conf,
+        )
         return model_name
 
     def convert_and_import(
@@ -936,7 +813,7 @@ class ModelManager(object):
 
         new_config = None
 
-        from ldm.invoke.ckpt_to_diffuser import convert_ckpt_to_diffuser
+        from . import convert_ckpt_to_diffusers
 
         if diffusers_path.exists():
             print(
@@ -951,7 +828,7 @@ class ModelManager(object):
             # By passing the specified VAE to the conversion function, the autoencoder
             # will be built into the model rather than tacked on afterward via the config file
             vae_model = self._load_vae(vae) if vae else None
-            convert_ckpt_to_diffuser(
+            convert_ckpt_to_diffusers(
                 ckpt_path,
                 diffusers_path,
                 extract_ema=True,
@@ -1088,27 +965,40 @@ class ModelManager(object):
         to the 2.3.0 "diffusers" version. This should be a one-time operation, called at
         script startup time.
         """
-        # Three transformer models to check: bert, clip and safety checker
+        # Three transformer models to check: bert, clip and safety checker, and
+        # the diffusers as well
+        models_dir = Path(Globals.root, "models")
         legacy_locations = [
             Path(
+                models_dir,
                 "CompVis/stable-diffusion-safety-checker/models--CompVis--stable-diffusion-safety-checker"
             ),
-            Path("bert-base-uncased/models--bert-base-uncased"),
+            Path(models_dir, "bert-base-uncased/models--bert-base-uncased"),
             Path(
+                models_dir,
                 "openai/clip-vit-large-patch14/models--openai--clip-vit-large-patch14"
             ),
         ]
-        models_dir = Path(Globals.root, "models")
+        legacy_locations.extend(list(global_cache_dir("diffusers").glob('*')))
+        
         legacy_layout = False
         for model in legacy_locations:
-            legacy_layout = legacy_layout or Path(models_dir, model).exists()
+            legacy_layout = legacy_layout or model.exists()
         if not legacy_layout:
             return
 
         print(
-            "** Legacy version <= 2.2.5 model directory layout detected. Reorganizing."
+            """
+>> ALERT:
+>> The location of your previously-installed diffusers models needs to move from
+>> invokeai/models/diffusers to invokeai/models/hub due to a change introduced by
+>> diffusers version 0.14. InvokeAI will now move all models from the "diffusers" directory
+>> into "hub" and then remove the diffusers directory. This is a quick, safe, one-time
+>> operation. However if you have customized either of these directories and need to
+>> make adjustments, please press ctrl-C now to abort and relaunch InvokeAI when you are ready.
+>> Otherwise press <enter> to continue."""
         )
-        print("** This is a quick one-time operation.")
+        input('continue> ')
 
         # transformer files get moved into the hub directory
         if cls._is_huggingface_hub_directory_present():
@@ -1120,33 +1010,20 @@ class ModelManager(object):
         for model in legacy_locations:
             source = models_dir / model
             dest = hub / model.stem
+            if dest.exists() and not source.exists():
+                continue
             print(f"** {source} => {dest}")
             if source.exists():
-                if dest.exists():
-                    rmtree(source)
+                if dest.is_symlink():
+                    print(f"** Found symlink at {dest.name}. Not migrating.")
+                elif dest.exists():
+                    if source.is_dir():
+                        rmtree(source)
+                    else:
+                        source.unlink()
                 else:
                     move(source, dest)
 
-        # anything else gets moved into the diffusers directory
-        if cls._is_huggingface_hub_directory_present():
-            diffusers = global_cache_dir("diffusers")
-        else:
-            diffusers = models_dir / "diffusers"
-
-        os.makedirs(diffusers, exist_ok=True)
-        for root, dirs, _ in os.walk(models_dir, topdown=False):
-            for dir in dirs:
-                full_path = Path(root, dir)
-                if full_path.is_relative_to(hub) or full_path.is_relative_to(diffusers):
-                    continue
-                if Path(dir).match("models--*--*"):
-                    dest = diffusers / dir
-                    print(f"** {full_path} => {dest}")
-                    if dest.exists():
-                        rmtree(full_path)
-                    else:
-                        move(full_path, dest)
-
         # now clean up by removing any empty directories
         empty = [
             root
@@ -1179,43 +1056,6 @@ class ModelManager(object):
             self.stack.remove(model_name)
         self.models.pop(model_name, None)
 
-    def _model_to_cpu(self, model):
-        if self.device == CPU_DEVICE:
-            return model
-
-        if isinstance(model, StableDiffusionGeneratorPipeline):
-            model.offload_all()
-            return model
-
-        model.cond_stage_model.device = CPU_DEVICE
-        model.to(CPU_DEVICE)
-
-        for submodel in ("first_stage_model", "cond_stage_model", "model"):
-            try:
-                getattr(model, submodel).to(CPU_DEVICE)
-            except AttributeError:
-                pass
-        return model
-
-    def _model_from_cpu(self, model):
-        if self.device == CPU_DEVICE:
-            return model
-
-        if isinstance(model, StableDiffusionGeneratorPipeline):
-            model.ready()
-            return model
-
-        model.to(self.device)
-        model.cond_stage_model.device = self.device
-
-        for submodel in ("first_stage_model", "cond_stage_model", "model"):
-            try:
-                getattr(model, submodel).to(self.device)
-            except AttributeError:
-                pass
-
-        return model
-
     def _pop_oldest_model(self):
         """
         Remove the first element of the FIFO, which ought
@@ -1233,6 +1073,19 @@ class ModelManager(object):
             self.stack.remove(model_name)
         self.stack.append(model_name)
 
+    def _add_embeddings_to_model(self, model: StableDiffusionGeneratorPipeline):
+        if self.embedding_path is not None:
+            print(f">> Loading embeddings from {self.embedding_path}")
+            for root, _, files in os.walk(self.embedding_path):
+                for name in files:
+                    ti_path = os.path.join(root, name)
+                    model.textual_inversion_manager.load_textual_inversion(
+                        ti_path, defer_injecting_tokens=True
+                    )
+            print(
+                f'>> Textual inversion triggers: {", ".join(sorted(model.textual_inversion_manager.get_all_trigger_strings()))}'
+            )
+            
     def _has_cuda(self) -> bool:
         return self.device.type == "cuda"
 
@@ -1244,7 +1097,7 @@ class ModelManager(object):
             path = name_or_path
         else:
             owner, repo = name_or_path.split("/")
-            path = Path(global_cache_dir("diffusers") / f"models--{owner}--{repo}")
+            path = Path(global_cache_dir("hub") / f"models--{owner}--{repo}")
         if not path.exists():
             return None
         hashpath = path / "checksum.sha256"
@@ -1305,7 +1158,7 @@ class ModelManager(object):
         using_fp16 = self.precision == "float16"
 
         vae_args.update(
-            cache_dir=global_cache_dir("diffusers"),
+            cache_dir=global_cache_dir("hub"),
             local_files_only=not Globals.internet_available,
         )
 
@@ -1344,7 +1197,7 @@ class ModelManager(object):
 
     @staticmethod
     def _delete_model_from_cache(repo_id):
-        cache_info = scan_cache_dir(global_cache_dir("diffusers"))
+        cache_info = scan_cache_dir(global_cache_dir("hub"))
 
         # I'm sure there is a way to do this with comprehensions
         # but the code quickly became incomprehensible!

invokeai/backend/prompting/__init__.py

@@ -0,0 +1,9 @@
+"""
+Initialization file for invokeai.backend.prompting
+"""
+from .conditioning import (
+    get_prompt_structure,
+    get_tokens_for_prompt_object,
+    get_uc_and_c_and_ec,
+    split_weighted_subprompts,
+)

invokeai/backend/prompting/conditioning.py

@@ -1,110 +1,138 @@
-'''
+"""
 This module handles the generation of the conditioning tensors.
 
 Useful function exports:
 
 get_uc_and_c_and_ec()           get the conditioned and unconditioned latent, and edited conditioning if we're doing cross-attention control
 
-'''
+"""
 import re
-from typing import Union, Optional, Any
-
-from transformers import CLIPTokenizer, CLIPTextModel
+from typing import Optional, Union
 
 from compel import Compel
-from compel.prompt_parser import FlattenedPrompt, Blend, Fragment, CrossAttentionControlSubstitute, PromptParser
-from .devices import torch_dtype
-from ..models.diffusion.shared_invokeai_diffusion import InvokeAIDiffuserComponent
-from ldm.invoke.globals import Globals
+from compel.prompt_parser import (
+    Blend,
+    CrossAttentionControlSubstitute,
+    FlattenedPrompt,
+    Fragment,
+    PromptParser,
+)
 
-def get_tokenizer(model) -> CLIPTokenizer:
-    # TODO remove legacy ckpt fallback handling
-    return (getattr(model, 'tokenizer', None) # diffusers
-            or model.cond_stage_model.tokenizer) # ldm
+from invokeai.backend.globals import Globals
 
-def get_text_encoder(model) -> Any:
-    # TODO remove legacy ckpt fallback handling
-    return (getattr(model, 'text_encoder', None)  # diffusers
-            or UnsqueezingLDMTransformer(model.cond_stage_model.transformer)) # ldm
-
-class UnsqueezingLDMTransformer:
-    def __init__(self, ldm_transformer):
-        self.ldm_transformer = ldm_transformer
-
-    @property
-    def device(self):
-        return self.ldm_transformer.device
-
-    def __call__(self, *args, **kwargs):
-        insufficiently_unsqueezed_tensor = self.ldm_transformer(*args, **kwargs)
-        return insufficiently_unsqueezed_tensor.unsqueeze(0)
+from ..stable_diffusion import InvokeAIDiffuserComponent
+from ..util import torch_dtype
 
 
-def get_uc_and_c_and_ec(prompt_string, model, log_tokens=False, skip_normalize_legacy_blend=False):
+def get_uc_and_c_and_ec(
+    prompt_string, model, log_tokens=False, skip_normalize_legacy_blend=False
+):
     # lazy-load any deferred textual inversions.
     # this might take a couple of seconds the first time a textual inversion is used.
-    model.textual_inversion_manager.create_deferred_token_ids_for_any_trigger_terms(prompt_string)
+    model.textual_inversion_manager.create_deferred_token_ids_for_any_trigger_terms(
+        prompt_string
+    )
 
-    tokenizer = get_tokenizer(model)
-    text_encoder = get_text_encoder(model)
-    compel = Compel(tokenizer=tokenizer,
-                    text_encoder=text_encoder,
-                    textual_inversion_manager=model.textual_inversion_manager,
-                    dtype_for_device_getter=torch_dtype)
+    tokenizer = model.tokenizer
+    compel = Compel(
+        tokenizer=tokenizer,
+        text_encoder=model.text_encoder,
+        textual_inversion_manager=model.textual_inversion_manager,
+        dtype_for_device_getter=torch_dtype,
+        truncate_long_prompts=False
+    )
 
-    positive_prompt_string, negative_prompt_string = split_prompt_to_positive_and_negative(prompt_string)
-    legacy_blend = try_parse_legacy_blend(positive_prompt_string, skip_normalize_legacy_blend)
-    positive_prompt: FlattenedPrompt|Blend
+    # get rid of any newline characters
+    prompt_string = prompt_string.replace("\n", " ")
+    (
+        positive_prompt_string,
+        negative_prompt_string,
+    ) = split_prompt_to_positive_and_negative(prompt_string)
+    legacy_blend = try_parse_legacy_blend(
+        positive_prompt_string, skip_normalize_legacy_blend
+    )
+    positive_prompt: Union[FlattenedPrompt, Blend]
     if legacy_blend is not None:
         positive_prompt = legacy_blend
     else:
         positive_prompt = Compel.parse_prompt_string(positive_prompt_string)
-    negative_prompt: FlattenedPrompt|Blend = Compel.parse_prompt_string(negative_prompt_string)
+    negative_prompt: Union[FlattenedPrompt, Blend] = Compel.parse_prompt_string(
+        negative_prompt_string
+    )
 
     if log_tokens or getattr(Globals, "log_tokenization", False):
         log_tokenization(positive_prompt, negative_prompt, tokenizer=tokenizer)
 
     c, options = compel.build_conditioning_tensor_for_prompt_object(positive_prompt)
     uc, _ = compel.build_conditioning_tensor_for_prompt_object(negative_prompt)
+    [c, uc] = compel.pad_conditioning_tensors_to_same_length([c, uc])
 
     tokens_count = get_max_token_count(tokenizer, positive_prompt)
 
-    ec = InvokeAIDiffuserComponent.ExtraConditioningInfo(tokens_count_including_eos_bos=tokens_count,
-                                                         cross_attention_control_args=options.get(
-                                                             'cross_attention_control', None))
+    ec = InvokeAIDiffuserComponent.ExtraConditioningInfo(
+        tokens_count_including_eos_bos=tokens_count,
+        cross_attention_control_args=options.get("cross_attention_control", None),
+    )
     return uc, c, ec
 
 
-def get_prompt_structure(prompt_string, skip_normalize_legacy_blend: bool = False) -> (
-    Union[FlattenedPrompt, Blend], FlattenedPrompt):
-    positive_prompt_string, negative_prompt_string = split_prompt_to_positive_and_negative(prompt_string)
-    legacy_blend = try_parse_legacy_blend(positive_prompt_string, skip_normalize_legacy_blend)
-    positive_prompt: FlattenedPrompt|Blend
+def get_prompt_structure(
+    prompt_string, skip_normalize_legacy_blend: bool = False
+) -> (Union[FlattenedPrompt, Blend], FlattenedPrompt):
+    (
+        positive_prompt_string,
+        negative_prompt_string,
+    ) = split_prompt_to_positive_and_negative(prompt_string)
+    legacy_blend = try_parse_legacy_blend(
+        positive_prompt_string, skip_normalize_legacy_blend
+    )
+    positive_prompt: Union[FlattenedPrompt, Blend]
     if legacy_blend is not None:
         positive_prompt = legacy_blend
     else:
         positive_prompt = Compel.parse_prompt_string(positive_prompt_string)
-    negative_prompt: FlattenedPrompt|Blend = Compel.parse_prompt_string(negative_prompt_string)
+    negative_prompt: Union[FlattenedPrompt, Blend] = Compel.parse_prompt_string(
+        negative_prompt_string
+    )
 
     return positive_prompt, negative_prompt
 
-def get_max_token_count(tokenizer, prompt: Union[FlattenedPrompt, Blend], truncate_if_too_long=True) -> int:
+
+def get_max_token_count(
+    tokenizer, prompt: Union[FlattenedPrompt, Blend], truncate_if_too_long=False
+) -> int:
     if type(prompt) is Blend:
         blend: Blend = prompt
-        return max([get_max_token_count(tokenizer, c, truncate_if_too_long) for c in blend.prompts])
+        return max(
+            [
+                get_max_token_count(tokenizer, c, truncate_if_too_long)
+                for c in blend.prompts
+            ]
+        )
     else:
-        return len(get_tokens_for_prompt_object(tokenizer, prompt, truncate_if_too_long))
+        return len(
+            get_tokens_for_prompt_object(tokenizer, prompt, truncate_if_too_long)
+        )
 
 
-def get_tokens_for_prompt_object(tokenizer, parsed_prompt: FlattenedPrompt, truncate_if_too_long=True) -> [str]:
-
+def get_tokens_for_prompt_object(
+    tokenizer, parsed_prompt: FlattenedPrompt, truncate_if_too_long=True
+) -> [str]:
     if type(parsed_prompt) is Blend:
-        raise ValueError("Blend is not supported here - you need to get tokens for each of its .children")
+        raise ValueError(
+            "Blend is not supported here - you need to get tokens for each of its .children"
+        )
 
-    text_fragments = [x.text if type(x) is Fragment else
-                      (" ".join([f.text for f in x.original]) if type(x) is CrossAttentionControlSubstitute else
-                       str(x))
-                      for x in parsed_prompt.children]
+    text_fragments = [
+        x.text
+        if type(x) is Fragment
+        else (
+            " ".join([f.text for f in x.original])
+            if type(x) is CrossAttentionControlSubstitute
+            else str(x)
+        )
+        for x in parsed_prompt.children
+    ]
     text = " ".join(text_fragments)
     tokens = tokenizer.tokenize(text)
     if truncate_if_too_long:
@@ -113,40 +141,48 @@ def get_tokens_for_prompt_object(tokenizer, parsed_prompt: FlattenedPrompt, trun
     return tokens
 
 
-def split_prompt_to_positive_and_negative(prompt_string_uncleaned):
-    unconditioned_words = ''
-    unconditional_regex = r'\[(.*?)\]'
+def split_prompt_to_positive_and_negative(prompt_string_uncleaned: str):
+    unconditioned_words = ""
+    unconditional_regex = r"\[(.*?)\]"
     unconditionals = re.findall(unconditional_regex, prompt_string_uncleaned)
     if len(unconditionals) > 0:
-        unconditioned_words = ' '.join(unconditionals)
+        unconditioned_words = " ".join(unconditionals)
 
         # Remove Unconditioned Words From Prompt
         unconditional_regex_compile = re.compile(unconditional_regex)
-        clean_prompt = unconditional_regex_compile.sub(' ', prompt_string_uncleaned)
-        prompt_string_cleaned = re.sub(' +', ' ', clean_prompt)
+        clean_prompt = unconditional_regex_compile.sub(" ", prompt_string_uncleaned)
+        prompt_string_cleaned = re.sub(" +", " ", clean_prompt)
     else:
         prompt_string_cleaned = prompt_string_uncleaned
     return prompt_string_cleaned, unconditioned_words
 
 
-def log_tokenization(positive_prompt: Union[Blend, FlattenedPrompt],
-                     negative_prompt: Union[Blend, FlattenedPrompt],
-                     tokenizer):
+def log_tokenization(
+    positive_prompt: Union[Blend, FlattenedPrompt],
+    negative_prompt: Union[Blend, FlattenedPrompt],
+    tokenizer,
+):
     print(f"\n>> [TOKENLOG] Parsed Prompt: {positive_prompt}")
     print(f"\n>> [TOKENLOG] Parsed Negative Prompt: {negative_prompt}")
 
     log_tokenization_for_prompt_object(positive_prompt, tokenizer)
-    log_tokenization_for_prompt_object(negative_prompt, tokenizer, display_label_prefix="(negative prompt)")
+    log_tokenization_for_prompt_object(
+        negative_prompt, tokenizer, display_label_prefix="(negative prompt)"
+    )
 
 
-def log_tokenization_for_prompt_object(p: Union[Blend, FlattenedPrompt], tokenizer, display_label_prefix=None):
+def log_tokenization_for_prompt_object(
+    p: Union[Blend, FlattenedPrompt], tokenizer, display_label_prefix=None
+):
     display_label_prefix = display_label_prefix or ""
     if type(p) is Blend:
         blend: Blend = p
         for i, c in enumerate(blend.prompts):
             log_tokenization_for_prompt_object(
-                c, tokenizer,
-                display_label_prefix=f"{display_label_prefix}(blend part {i + 1}, weight={blend.weights[i]})")
+                c,
+                tokenizer,
+                display_label_prefix=f"{display_label_prefix}(blend part {i + 1}, weight={blend.weights[i]})",
+            )
     elif type(p) is FlattenedPrompt:
         flattened_prompt: FlattenedPrompt = p
         if flattened_prompt.wants_cross_attention_control:
@@ -161,18 +197,26 @@ def log_tokenization_for_prompt_object(p: Union[Blend, FlattenedPrompt], tokeniz
                     edited_fragments.append(f)
 
             original_text = " ".join([x.text for x in original_fragments])
-            log_tokenization_for_text(original_text, tokenizer,
-                                      display_label=f"{display_label_prefix}(.swap originals)")
+            log_tokenization_for_text(
+                original_text,
+                tokenizer,
+                display_label=f"{display_label_prefix}(.swap originals)",
+            )
             edited_text = " ".join([x.text for x in edited_fragments])
-            log_tokenization_for_text(edited_text, tokenizer,
-                                      display_label=f"{display_label_prefix}(.swap replacements)")
+            log_tokenization_for_text(
+                edited_text,
+                tokenizer,
+                display_label=f"{display_label_prefix}(.swap replacements)",
+            )
         else:
             text = " ".join([x.text for x in flattened_prompt.children])
-            log_tokenization_for_text(text, tokenizer, display_label=display_label_prefix)
+            log_tokenization_for_text(
+                text, tokenizer, display_label=display_label_prefix
+            )
 
 
-def log_tokenization_for_text(text, tokenizer, display_label=None):
-    """ shows how the prompt is tokenized
+def log_tokenization_for_text(text, tokenizer, display_label=None, truncate_if_too_long=False):
+    """shows how the prompt is tokenized
     # usually tokens have '</w>' to indicate end-of-word,
     # but for readability it has been replaced with ' '
     """
@@ -183,25 +227,25 @@ def log_tokenization_for_text(text, tokenizer, display_label=None):
     totalTokens = len(tokens)
 
     for i in range(0, totalTokens):
-        token = tokens[i].replace('</w>', ' ')
+        token = tokens[i].replace("</w>", " ")
         # alternate color
         s = (usedTokens % 6) + 1
-        if i < tokenizer.model_max_length:
+        if truncate_if_too_long and i >= tokenizer.model_max_length:
+            discarded = discarded + f"\x1b[0;3{s};40m{token}"
+        else:
             tokenized = tokenized + f"\x1b[0;3{s};40m{token}"
             usedTokens += 1
-        else:  # over max token length
-            discarded = discarded + f"\x1b[0;3{s};40m{token}"
 
     if usedTokens > 0:
         print(f'\n>> [TOKENLOG] Tokens {display_label or ""} ({usedTokens}):')
-        print(f'{tokenized}\x1b[0m')
+        print(f"{tokenized}\x1b[0m")
 
     if discarded != "":
-        print(f'\n>> [TOKENLOG] Tokens Discarded ({totalTokens - usedTokens}):')
-        print(f'{discarded}\x1b[0m')
+        print(f"\n>> [TOKENLOG] Tokens Discarded ({totalTokens - usedTokens}):")
+        print(f"{discarded}\x1b[0m")
 
 
-def try_parse_legacy_blend(text: str, skip_normalize: bool=False) -> Optional[Blend]:
+def try_parse_legacy_blend(text: str, skip_normalize: bool = False) -> Optional[Blend]:
     weighted_subprompts = split_weighted_subprompts(text, skip_normalize=skip_normalize)
     if len(weighted_subprompts) <= 1:
         return None
@@ -212,10 +256,12 @@ def try_parse_legacy_blend(text: str, skip_normalize: bool=False) -> Optional[Bl
     parsed_conjunctions = [pp.parse_conjunction(x) for x in strings]
     flattened_prompts = [x.prompts[0] for x in parsed_conjunctions]
 
-    return Blend(prompts=flattened_prompts, weights=weights, normalize_weights=not skip_normalize)
+    return Blend(
+        prompts=flattened_prompts, weights=weights, normalize_weights=not skip_normalize
+    )
 
 
-def split_weighted_subprompts(text, skip_normalize=False)->list:
+def split_weighted_subprompts(text, skip_normalize=False) -> list:
     """
     Legacy blend parsing.
 
@@ -224,7 +270,8 @@ def split_weighted_subprompts(text, skip_normalize=False)->list:
     if ':' has no value defined, defaults to 1.0
     repeats until no text remaining
     """
-    prompt_parser = re.compile("""
+    prompt_parser = re.compile(
+        """
             (?P<prompt>     # capture group for 'prompt'
             (?:\\\:|[^:])+  # match one or more non ':' characters or escaped colons '\:'
             )               # end 'prompt'
@@ -237,16 +284,20 @@ def split_weighted_subprompts(text, skip_normalize=False)->list:
             |               # OR
             $               # else, if no ':' then match end of line
             )               # end non-capture group
-            """, re.VERBOSE)
-    parsed_prompts = [(match.group("prompt").replace("\\:", ":"), float(
-        match.group("weight") or 1)) for match in re.finditer(prompt_parser, text)]
+            """,
+        re.VERBOSE,
+    )
+    parsed_prompts = [
+        (match.group("prompt").replace("\\:", ":"), float(match.group("weight") or 1))
+        for match in re.finditer(prompt_parser, text)
+    ]
     if skip_normalize:
         return parsed_prompts
     weight_sum = sum(map(lambda x: x[1], parsed_prompts))
     if weight_sum == 0:
         print(
-            "* Warning: Subprompt weights add up to zero. Discarding and using even weights instead.")
+            "* Warning: Subprompt weights add up to zero. Discarding and using even weights instead."
+        )
         equal_weight = 1 / max(len(parsed_prompts), 1)
         return [(x[0], equal_weight) for x in parsed_prompts]
     return [(x[0], x[1] / weight_sum) for x in parsed_prompts]
-

invokeai/backend/restoration/__init__.py

@@ -0,0 +1,4 @@
+"""
+Initialization file for the invokeai.backend.restoration package
+"""
+from .base import Restoration

invokeai/backend/restoration/base.py

@@ -1,38 +1,43 @@
-class Restoration():
+class Restoration:
     def __init__(self) -> None:
         pass
 
-    def load_face_restore_models(self, gfpgan_model_path='./models/gfpgan/GFPGANv1.4.pth'):
+    def load_face_restore_models(
+        self, gfpgan_model_path="./models/gfpgan/GFPGANv1.4.pth"
+    ):
         # Load GFPGAN
         gfpgan = self.load_gfpgan(gfpgan_model_path)
         if gfpgan.gfpgan_model_exists:
-            print('>> GFPGAN Initialized')
+            print(">> GFPGAN Initialized")
         else:
-            print('>> GFPGAN Disabled')
+            print(">> GFPGAN Disabled")
             gfpgan = None
 
         # Load CodeFormer
         codeformer = self.load_codeformer()
         if codeformer.codeformer_model_exists:
-            print('>> CodeFormer Initialized')
+            print(">> CodeFormer Initialized")
         else:
-            print('>> CodeFormer Disabled')
+            print(">> CodeFormer Disabled")
             codeformer = None
 
         return gfpgan, codeformer
 
     # Face Restore Models
     def load_gfpgan(self, gfpgan_model_path):
-        from ldm.invoke.restoration.gfpgan import GFPGAN
+        from .gfpgan import GFPGAN
+
         return GFPGAN(gfpgan_model_path)
 
     def load_codeformer(self):
-        from ldm.invoke.restoration.codeformer import CodeFormerRestoration
+        from .codeformer import CodeFormerRestoration
+
         return CodeFormerRestoration()
 
     # Upscale Models
     def load_esrgan(self, esrgan_bg_tile=400):
-        from ldm.invoke.restoration.realesrgan import ESRGAN
+        from .realesrgan import ESRGAN
+
         esrgan = ESRGAN(esrgan_bg_tile)
-        print('>> ESRGAN Initialized')
-        return esrgan;
+        print(">> ESRGAN Initialized")
+        return esrgan

invokeai/backend/restoration/codeformer.py

@@ -1,17 +1,21 @@
 import os
-import torch
-import numpy as np
-import warnings
 import sys
-from ldm.invoke.globals import Globals
+import warnings
 
-pretrained_model_url = 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer.pth'
+import numpy as np
+import torch
 
-class CodeFormerRestoration():
-    def __init__(self,
-            codeformer_dir='models/codeformer',
-            codeformer_model_path='codeformer.pth') -> None:
+from ..globals import Globals
 
+pretrained_model_url = (
+    "https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer.pth"
+)
+
+
+class CodeFormerRestoration:
+    def __init__(
+        self, codeformer_dir="models/codeformer", codeformer_model_path="codeformer.pth"
+    ) -> None:
         if not os.path.isabs(codeformer_dir):
             codeformer_dir = os.path.join(Globals.root, codeformer_dir)
 
@@ -19,22 +23,23 @@ class CodeFormerRestoration():
         self.codeformer_model_exists = os.path.isfile(self.model_path)
 
         if not self.codeformer_model_exists:
-            print('## NOT FOUND: CodeFormer model not found at ' + self.model_path)
+            print("## NOT FOUND: CodeFormer model not found at " + self.model_path)
         sys.path.append(os.path.abspath(codeformer_dir))
 
     def process(self, image, strength, device, seed=None, fidelity=0.75):
         if seed is not None:
-            print(f'>> CodeFormer - Restoring Faces for image seed:{seed}')
+            print(f">> CodeFormer - Restoring Faces for image seed:{seed}")
         with warnings.catch_warnings():
-            warnings.filterwarnings('ignore', category=DeprecationWarning)
-            warnings.filterwarnings('ignore', category=UserWarning)
+            warnings.filterwarnings("ignore", category=DeprecationWarning)
+            warnings.filterwarnings("ignore", category=UserWarning)
 
-            from basicsr.utils.download_util import load_file_from_url
             from basicsr.utils import img2tensor, tensor2img
+            from basicsr.utils.download_util import load_file_from_url
             from facexlib.utils.face_restoration_helper import FaceRestoreHelper
-            from ldm.invoke.restoration.codeformer_arch import CodeFormer
-            from torchvision.transforms.functional import normalize
             from PIL import Image
+            from torchvision.transforms.functional import normalize
+
+            from .codeformer_arch import CodeFormer
 
             cf_class = CodeFormer
 
@@ -43,28 +48,31 @@ class CodeFormerRestoration():
                 codebook_size=1024,
                 n_head=8,
                 n_layers=9,
-                connect_list=['32', '64', '128', '256']
+                connect_list=["32", "64", "128", "256"],
             ).to(device)
 
             # note that this file should already be downloaded and cached at
             # this point
-            checkpoint_path = load_file_from_url(url=pretrained_model_url,
-                                                 model_dir=os.path.abspath(os.path.dirname(self.model_path)),
-                                                 progress=True
+            checkpoint_path = load_file_from_url(
+                url=pretrained_model_url,
+                model_dir=os.path.abspath(os.path.dirname(self.model_path)),
+                progress=True,
             )
-            checkpoint = torch.load(checkpoint_path)['params_ema']
+            checkpoint = torch.load(checkpoint_path)["params_ema"]
             cf.load_state_dict(checkpoint)
             cf.eval()
 
-            image = image.convert('RGB')
+            image = image.convert("RGB")
             # Codeformer expects a BGR np array; make array and flip channels
-            bgr_image_array = np.array(image, dtype=np.uint8)[...,::-1]
+            bgr_image_array = np.array(image, dtype=np.uint8)[..., ::-1]
 
             face_helper = FaceRestoreHelper(
                 upscale_factor=1,
                 use_parse=True,
                 device=device,
-                model_rootpath=os.path.join(Globals.root,'models','gfpgan','weights'),
+                model_rootpath=os.path.join(
+                    Globals.root, "models", "gfpgan", "weights"
+                ),
             )
             face_helper.clean_all()
             face_helper.read_image(bgr_image_array)
@@ -72,30 +80,35 @@ class CodeFormerRestoration():
             face_helper.align_warp_face()
 
             for idx, cropped_face in enumerate(face_helper.cropped_faces):
-                cropped_face_t = img2tensor(cropped_face / 255., bgr2rgb=True, float32=True)
-                normalize(cropped_face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
+                cropped_face_t = img2tensor(
+                    cropped_face / 255.0, bgr2rgb=True, float32=True
+                )
+                normalize(
+                    cropped_face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True
+                )
                 cropped_face_t = cropped_face_t.unsqueeze(0).to(device)
 
                 try:
                     with torch.no_grad():
                         output = cf(cropped_face_t, w=fidelity, adain=True)[0]
-                        restored_face = tensor2img(output.squeeze(0), rgb2bgr=True, min_max=(-1, 1))
+                        restored_face = tensor2img(
+                            output.squeeze(0), rgb2bgr=True, min_max=(-1, 1)
+                        )
                     del output
                     torch.cuda.empty_cache()
                 except RuntimeError as error:
-                    print(f'\tFailed inference for CodeFormer: {error}.')
+                    print(f"\tFailed inference for CodeFormer: {error}.")
                     restored_face = cropped_face
 
-                restored_face = restored_face.astype('uint8')
+                restored_face = restored_face.astype("uint8")
                 face_helper.add_restored_face(restored_face)
 
-
             face_helper.get_inverse_affine(None)
 
             restored_img = face_helper.paste_faces_to_input_image()
 
             # Flip the channels back to RGB
-            res = Image.fromarray(restored_img[...,::-1])
+            res = Image.fromarray(restored_img[..., ::-1])
 
             if strength < 1.0:
                 # Resize the image to the new image if the sizes have changed

invokeai/backend/restoration/codeformer_arch.py

@@ -1,13 +1,15 @@
 import math
+from typing import List, Optional
+
 import numpy as np
 import torch
-from torch import nn, Tensor
 import torch.nn.functional as F
-from typing import Optional, List
-
-from ldm.invoke.restoration.vqgan_arch import *
 from basicsr.utils import get_root_logger
 from basicsr.utils.registry import ARCH_REGISTRY
+from torch import Tensor, nn
+
+from .vqgan_arch import *
+
 
 def calc_mean_std(feat, eps=1e-5):
     """Calculate mean and std for adaptive_instance_normalization.
@@ -18,7 +20,7 @@ def calc_mean_std(feat, eps=1e-5):
             divide-by-zero. Default: 1e-5.
     """
     size = feat.size()
-    assert len(size) == 4, 'The input feature should be 4D tensor.'
+    assert len(size) == 4, "The input feature should be 4D tensor."
     b, c = size[:2]
     feat_var = feat.view(b, c, -1).var(dim=2) + eps
     feat_std = feat_var.sqrt().view(b, c, 1, 1)
@@ -39,7 +41,9 @@ def adaptive_instance_normalization(content_feat, style_feat):
     size = content_feat.size()
     style_mean, style_std = calc_mean_std(style_feat)
     content_mean, content_std = calc_mean_std(content_feat)
-    normalized_feat = (content_feat - content_mean.expand(size)) / content_std.expand(size)
+    normalized_feat = (content_feat - content_mean.expand(size)) / content_std.expand(
+        size
+    )
     return normalized_feat * style_std.expand(size) + style_mean.expand(size)
 
 
@@ -49,7 +53,9 @@ class PositionEmbeddingSine(nn.Module):
     used by the Attention is all you need paper, generalized to work on images.
     """
 
-    def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None):
+    def __init__(
+        self, num_pos_feats=64, temperature=10000, normalize=False, scale=None
+    ):
         super().__init__()
         self.num_pos_feats = num_pos_feats
         self.temperature = temperature
@@ -62,7 +68,9 @@ class PositionEmbeddingSine(nn.Module):
 
     def forward(self, x, mask=None):
         if mask is None:
-            mask = torch.zeros((x.size(0), x.size(2), x.size(3)), device=x.device, dtype=torch.bool)
+            mask = torch.zeros(
+                (x.size(0), x.size(2), x.size(3)), device=x.device, dtype=torch.bool
+            )
         not_mask = ~mask
         y_embed = not_mask.cumsum(1, dtype=torch.float32)
         x_embed = not_mask.cumsum(2, dtype=torch.float32)
@@ -85,6 +93,7 @@ class PositionEmbeddingSine(nn.Module):
         pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
         return pos
 
+
 def _get_activation_fn(activation):
     """Return an activation function given a string"""
     if activation == "relu":
@@ -93,11 +102,13 @@ def _get_activation_fn(activation):
         return F.gelu
     if activation == "glu":
         return F.glu
-    raise RuntimeError(F"activation should be relu/gelu, not {activation}.")
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
 
 
 class TransformerSALayer(nn.Module):
-    def __init__(self, embed_dim, nhead=8, dim_mlp=2048, dropout=0.0, activation="gelu"):
+    def __init__(
+        self, embed_dim, nhead=8, dim_mlp=2048, dropout=0.0, activation="gelu"
+    ):
         super().__init__()
         self.self_attn = nn.MultiheadAttention(embed_dim, nhead, dropout=dropout)
         # Implementation of Feedforward model - MLP
@@ -115,16 +126,19 @@ class TransformerSALayer(nn.Module):
     def with_pos_embed(self, tensor, pos: Optional[Tensor]):
         return tensor if pos is None else tensor + pos
 
-    def forward(self, tgt,
-                tgt_mask: Optional[Tensor] = None,
-                tgt_key_padding_mask: Optional[Tensor] = None,
-                query_pos: Optional[Tensor] = None):
-
+    def forward(
+        self,
+        tgt,
+        tgt_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+    ):
         # self attention
         tgt2 = self.norm1(tgt)
         q = k = self.with_pos_embed(tgt2, query_pos)
-        tgt2 = self.self_attn(q, k, value=tgt2, attn_mask=tgt_mask,
-                              key_padding_mask=tgt_key_padding_mask)[0]
+        tgt2 = self.self_attn(
+            q, k, value=tgt2, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask
+        )[0]
         tgt = tgt + self.dropout1(tgt2)
 
         # ffn
@@ -133,20 +147,23 @@ class TransformerSALayer(nn.Module):
         tgt = tgt + self.dropout2(tgt2)
         return tgt
 
+
 class Fuse_sft_block(nn.Module):
     def __init__(self, in_ch, out_ch):
         super().__init__()
-        self.encode_enc = ResBlock(2*in_ch, out_ch)
+        self.encode_enc = ResBlock(2 * in_ch, out_ch)
 
         self.scale = nn.Sequential(
-                    nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1),
-                    nn.LeakyReLU(0.2, True),
-                    nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1))
+            nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1),
+            nn.LeakyReLU(0.2, True),
+            nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1),
+        )
 
         self.shift = nn.Sequential(
-                    nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1),
-                    nn.LeakyReLU(0.2, True),
-                    nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1))
+            nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1),
+            nn.LeakyReLU(0.2, True),
+            nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1),
+        )
 
     def forward(self, enc_feat, dec_feat, w=1):
         enc_feat = self.encode_enc(torch.cat([enc_feat, dec_feat], dim=1))
@@ -159,11 +176,19 @@ class Fuse_sft_block(nn.Module):
 
 @ARCH_REGISTRY.register()
 class CodeFormer(VQAutoEncoder):
-    def __init__(self, dim_embd=512, n_head=8, n_layers=9,
-                codebook_size=1024, latent_size=256,
-                connect_list=['32', '64', '128', '256'],
-                fix_modules=['quantize','generator']):
-        super(CodeFormer, self).__init__(512, 64, [1, 2, 2, 4, 4, 8], 'nearest',2, [16], codebook_size)
+    def __init__(
+        self,
+        dim_embd=512,
+        n_head=8,
+        n_layers=9,
+        codebook_size=1024,
+        latent_size=256,
+        connect_list=["32", "64", "128", "256"],
+        fix_modules=["quantize", "generator"],
+    ):
+        super(CodeFormer, self).__init__(
+            512, 64, [1, 2, 2, 4, 4, 8], "nearest", 2, [16], codebook_size
+        )
 
         if fix_modules is not None:
             for module in fix_modules:
@@ -173,33 +198,53 @@ class CodeFormer(VQAutoEncoder):
         self.connect_list = connect_list
         self.n_layers = n_layers
         self.dim_embd = dim_embd
-        self.dim_mlp = dim_embd*2
+        self.dim_mlp = dim_embd * 2
 
         self.position_emb = nn.Parameter(torch.zeros(latent_size, self.dim_embd))
         self.feat_emb = nn.Linear(256, self.dim_embd)
 
         # transformer
-        self.ft_layers = nn.Sequential(*[TransformerSALayer(embed_dim=dim_embd, nhead=n_head, dim_mlp=self.dim_mlp, dropout=0.0)
-                                    for _ in range(self.n_layers)])
+        self.ft_layers = nn.Sequential(
+            *[
+                TransformerSALayer(
+                    embed_dim=dim_embd, nhead=n_head, dim_mlp=self.dim_mlp, dropout=0.0
+                )
+                for _ in range(self.n_layers)
+            ]
+        )
 
         # logits_predict head
         self.idx_pred_layer = nn.Sequential(
-            nn.LayerNorm(dim_embd),
-            nn.Linear(dim_embd, codebook_size, bias=False))
+            nn.LayerNorm(dim_embd), nn.Linear(dim_embd, codebook_size, bias=False)
+        )
 
         self.channels = {
-            '16': 512,
-            '32': 256,
-            '64': 256,
-            '128': 128,
-            '256': 128,
-            '512': 64,
+            "16": 512,
+            "32": 256,
+            "64": 256,
+            "128": 128,
+            "256": 128,
+            "512": 64,
         }
 
         # after second residual block for > 16, before attn layer for ==16
-        self.fuse_encoder_block = {'512':2, '256':5, '128':8, '64':11, '32':14, '16':18}
+        self.fuse_encoder_block = {
+            "512": 2,
+            "256": 5,
+            "128": 8,
+            "64": 11,
+            "32": 14,
+            "16": 18,
+        }
         # after first residual block for > 16, before attn layer for ==16
-        self.fuse_generator_block = {'16':6, '32': 9, '64':12, '128':15, '256':18, '512':21}
+        self.fuse_generator_block = {
+            "16": 6,
+            "32": 9,
+            "64": 12,
+            "128": 15,
+            "256": 18,
+            "512": 21,
+        }
 
         # fuse_convs_dict
         self.fuse_convs_dict = nn.ModuleDict()
@@ -228,20 +273,20 @@ class CodeFormer(VQAutoEncoder):
         lq_feat = x
         # ################# Transformer ###################
         # quant_feat, codebook_loss, quant_stats = self.quantize(lq_feat)
-        pos_emb = self.position_emb.unsqueeze(1).repeat(1,x.shape[0],1)
+        pos_emb = self.position_emb.unsqueeze(1).repeat(1, x.shape[0], 1)
         # BCHW -> BC(HW) -> (HW)BC
-        feat_emb = self.feat_emb(lq_feat.flatten(2).permute(2,0,1))
+        feat_emb = self.feat_emb(lq_feat.flatten(2).permute(2, 0, 1))
         query_emb = feat_emb
         # Transformer encoder
         for layer in self.ft_layers:
             query_emb = layer(query_emb, query_pos=pos_emb)
 
         # output logits
-        logits = self.idx_pred_layer(query_emb) # (hw)bn
-        logits = logits.permute(1,0,2) # (hw)bn -> b(hw)n
+        logits = self.idx_pred_layer(query_emb)  # (hw)bn
+        logits = logits.permute(1, 0, 2)  # (hw)bn -> b(hw)n
 
-        if code_only: # for training stage II
-          # logits doesn't need softmax before cross_entropy loss
+        if code_only:  # for training stage II
+            # logits doesn't need softmax before cross_entropy loss
             return logits, lq_feat
 
         # ################# Quantization ###################
@@ -252,12 +297,14 @@ class CodeFormer(VQAutoEncoder):
         # ------------
         soft_one_hot = F.softmax(logits, dim=2)
         _, top_idx = torch.topk(soft_one_hot, 1, dim=2)
-        quant_feat = self.quantize.get_codebook_feat(top_idx, shape=[x.shape[0],16,16,256])
+        quant_feat = self.quantize.get_codebook_feat(
+            top_idx, shape=[x.shape[0], 16, 16, 256]
+        )
         # preserve gradients
         # quant_feat = lq_feat + (quant_feat - lq_feat).detach()
 
         if detach_16:
-            quant_feat = quant_feat.detach() # for training stage III
+            quant_feat = quant_feat.detach()  # for training stage III
         if adain:
             quant_feat = adaptive_instance_normalization(quant_feat, lq_feat)
 
@@ -267,10 +314,12 @@ class CodeFormer(VQAutoEncoder):
 
         for i, block in enumerate(self.generator.blocks):
             x = block(x)
-            if i in fuse_list: # fuse after i-th block
+            if i in fuse_list:  # fuse after i-th block
                 f_size = str(x.shape[-1])
-                if w>0:
-                    x = self.fuse_convs_dict[f_size](enc_feat_dict[f_size].detach(), x, w)
+                if w > 0:
+                    x = self.fuse_convs_dict[f_size](
+                        enc_feat_dict[f_size].detach(), x, w
+                    )
         out = x
         # logits doesn't need softmax before cross_entropy loss
         return out, logits, lq_feat

invokeai/backend/restoration/gfpgan.py

@@ -1,26 +1,25 @@
-import torch
-import warnings
 import os
 import sys
-import numpy as np
-from ldm.invoke.globals import Globals
+import warnings
 
+import numpy as np
+import torch
 from PIL import Image
 
+from invokeai.backend.globals import Globals
 
-class GFPGAN():
-    def __init__(
-            self,
-            gfpgan_model_path='models/gfpgan/GFPGANv1.4.pth'
-    ) -> None:
 
+class GFPGAN:
+    def __init__(self, gfpgan_model_path="models/gfpgan/GFPGANv1.4.pth") -> None:
         if not os.path.isabs(gfpgan_model_path):
-            gfpgan_model_path=os.path.abspath(os.path.join(Globals.root,gfpgan_model_path))
+            gfpgan_model_path = os.path.abspath(
+                os.path.join(Globals.root, gfpgan_model_path)
+            )
         self.model_path = gfpgan_model_path
         self.gfpgan_model_exists = os.path.isfile(self.model_path)
 
         if not self.gfpgan_model_exists:
-            print('## NOT FOUND: GFPGAN model not found at ' + self.model_path)
+            print("## NOT FOUND: GFPGAN model not found at " + self.model_path)
             return None
 
     def model_exists(self):
@@ -28,40 +27,40 @@ class GFPGAN():
 
     def process(self, image, strength: float, seed: str = None):
         if seed is not None:
-            print(f'>> GFPGAN - Restoring Faces for image seed:{seed}')
+            print(f">> GFPGAN - Restoring Faces for image seed:{seed}")
 
         with warnings.catch_warnings():
-            warnings.filterwarnings('ignore', category=DeprecationWarning)
-            warnings.filterwarnings('ignore', category=UserWarning)
+            warnings.filterwarnings("ignore", category=DeprecationWarning)
+            warnings.filterwarnings("ignore", category=UserWarning)
             cwd = os.getcwd()
-            os.chdir(os.path.join(Globals.root,'models'))
+            os.chdir(os.path.join(Globals.root, "models"))
             try:
                 from gfpgan import GFPGANer
+
                 self.gfpgan = GFPGANer(
                     model_path=self.model_path,
                     upscale=1,
-                    arch='clean',
+                    arch="clean",
                     channel_multiplier=2,
                     bg_upsampler=None,
                 )
             except Exception:
                 import traceback
-                print('>> Error loading GFPGAN:', file=sys.stderr)
+
+                print(">> Error loading GFPGAN:", file=sys.stderr)
                 print(traceback.format_exc(), file=sys.stderr)
             os.chdir(cwd)
 
         if self.gfpgan is None:
+            print(f">> WARNING: GFPGAN not initialized.")
             print(
-                f'>> WARNING: GFPGAN not initialized.'
-            )
-            print(
-                f'>> Download https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth to {self.model_path}'
+                f">> Download https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth to {self.model_path}"
             )
 
-        image = image.convert('RGB')
+        image = image.convert("RGB")
 
         # GFPGAN expects a BGR np array; make array and flip channels
-        bgr_image_array = np.array(image, dtype=np.uint8)[...,::-1]
+        bgr_image_array = np.array(image, dtype=np.uint8)[..., ::-1]
 
         _, _, restored_img = self.gfpgan.enhance(
             bgr_image_array,
@@ -71,7 +70,7 @@ class GFPGAN():
         )
 
         # Flip the channels back to RGB
-        res = Image.fromarray(restored_img[...,::-1])
+        res = Image.fromarray(restored_img[..., ::-1])
 
         if strength < 1.0:
             # Resize the image to the new image if the sizes have changed
@@ -79,7 +78,6 @@ class GFPGAN():
                 image = image.resize(res.size)
             res = Image.blend(image, res, strength)
 
-
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
         self.gfpgan = None

invokeai/backend/restoration/outcrop.py

@@ -0,0 +1,118 @@
+import math
+
+from PIL import Image
+
+
+class Outcrop(object):
+    def __init__(
+        self,
+        image,
+        generate,  # current generate object
+    ):
+        self.image = image
+        self.generate = generate
+
+    def process(
+        self,
+        extents: dict,
+        opt,  # current options
+        orig_opt,  # ones originally used to generate the image
+        image_callback=None,
+        prefix=None,
+    ):
+        # grow and mask the image
+        extended_image = self._extend_all(extents)
+
+        # switch samplers temporarily
+        curr_sampler = self.generate.sampler
+        self.generate.sampler_name = opt.sampler_name
+        self.generate._set_scheduler()
+
+        def wrapped_callback(img, seed, **kwargs):
+            preferred_seed = (
+                orig_opt.seed
+                if orig_opt.seed is not None and orig_opt.seed >= 0
+                else seed
+            )
+            image_callback(img, preferred_seed, use_prefix=prefix, **kwargs)
+
+        result = self.generate.prompt2image(
+            opt.prompt,
+            seed=opt.seed or orig_opt.seed,
+            sampler=self.generate.sampler,
+            steps=opt.steps,
+            cfg_scale=opt.cfg_scale,
+            ddim_eta=self.generate.ddim_eta,
+            width=extended_image.width,
+            height=extended_image.height,
+            init_img=extended_image,
+            strength=0.90,
+            image_callback=wrapped_callback if image_callback else None,
+            seam_size=opt.seam_size or 96,
+            seam_blur=opt.seam_blur or 16,
+            seam_strength=opt.seam_strength or 0.7,
+            seam_steps=20,
+            tile_size=32,
+            color_match=True,
+            force_outpaint=True,  # this just stops the warning about erased regions
+        )
+
+        # swap sampler back
+        self.generate.sampler = curr_sampler
+        return result
+
+    def _extend_all(
+        self,
+        extents: dict,
+    ) -> Image:
+        """
+        Extend the image in direction ('top','bottom','left','right') by
+        the indicated value. The image canvas is extended, and the empty
+        rectangular section will be filled with a blurred copy of the
+        adjacent image.
+        """
+        image = self.image
+        for direction in extents:
+            assert direction in [
+                "top",
+                "left",
+                "bottom",
+                "right",
+            ], 'Direction must be one of "top", "left", "bottom", "right"'
+            pixels = extents[direction]
+            # round pixels up to the nearest 64
+            pixels = math.ceil(pixels / 64) * 64
+            print(f">> extending image {direction}ward by {pixels} pixels")
+            image = self._rotate(image, direction)
+            image = self._extend(image, pixels)
+            image = self._rotate(image, direction, reverse=True)
+        return image
+
+    def _rotate(self, image: Image, direction: str, reverse=False) -> Image:
+        """
+        Rotates image so that the area to extend is always at the top top.
+        Simplifies logic later. The reverse argument, if true, will undo the
+        previous transpose.
+        """
+        transposes = {
+            "right": ["ROTATE_90", "ROTATE_270"],
+            "bottom": ["ROTATE_180", "ROTATE_180"],
+            "left": ["ROTATE_270", "ROTATE_90"],
+        }
+        if direction not in transposes:
+            return image
+        transpose = transposes[direction][1 if reverse else 0]
+        return image.transpose(Image.Transpose.__dict__[transpose])
+
+    def _extend(self, image: Image, pixels: int) -> Image:
+        extended_img = Image.new("RGBA", (image.width, image.height + pixels))
+
+        extended_img.paste((0, 0, 0), [0, 0, image.width, image.height + pixels])
+        extended_img.paste(image, box=(0, pixels))
+
+        # now make the top part transparent to use as a mask
+        alpha = extended_img.getchannel("A")
+        alpha.paste(0, (0, 0, extended_img.width, pixels))
+        extended_img.putalpha(alpha)
+
+        return extended_img

invokeai/backend/restoration/outpaint.py

@@ -1,39 +1,43 @@
-import warnings
 import math
+import warnings
+
 from PIL import Image, ImageFilter
 
+
 class Outpaint(object):
     def __init__(self, image, generate):
-        self.image     = image
-        self.generate  = generate
+        self.image = image
+        self.generate = generate
 
-    def process(self, opt, old_opt, image_callback = None, prefix = None):
+    def process(self, opt, old_opt, image_callback=None, prefix=None):
         image = self._create_outpaint_image(self.image, opt.out_direction)
 
-        seed   = old_opt.seed
+        seed = old_opt.seed
         prompt = old_opt.prompt
 
-        def wrapped_callback(img,seed,**kwargs):
-            image_callback(img,seed,use_prefix=prefix,**kwargs)
-
+        def wrapped_callback(img, seed, **kwargs):
+            image_callback(img, seed, use_prefix=prefix, **kwargs)
 
         return self.generate.prompt2image(
             prompt,
-            seed           = seed,
-            sampler        = self.generate.sampler,
-            steps          = opt.steps,
-            cfg_scale      = opt.cfg_scale,
-            ddim_eta       = self.generate.ddim_eta,
-            width          = opt.width,
-            height         = opt.height,
-            init_img       = image,
-            strength       = 0.83,
-            image_callback = wrapped_callback,
-            prefix         = prefix,
+            seed=seed,
+            sampler=self.generate.sampler,
+            steps=opt.steps,
+            cfg_scale=opt.cfg_scale,
+            ddim_eta=self.generate.ddim_eta,
+            width=opt.width,
+            height=opt.height,
+            init_img=image,
+            strength=0.83,
+            image_callback=wrapped_callback,
+            prefix=prefix,
         )
 
     def _create_outpaint_image(self, image, direction_args):
-        assert len(direction_args) in [1, 2], 'Direction (-D) must have exactly one or two arguments.'
+        assert len(direction_args) in [
+            1,
+            2,
+        ], "Direction (-D) must have exactly one or two arguments."
 
         if len(direction_args) == 1:
             direction = direction_args[0]
@@ -42,19 +46,26 @@ class Outpaint(object):
             direction = direction_args[0]
             pixels = int(direction_args[1])
 
-        assert direction in ['top', 'left', 'bottom', 'right'], 'Direction (-D) must be one of "top", "left", "bottom", "right"'
+        assert direction in [
+            "top",
+            "left",
+            "bottom",
+            "right",
+        ], 'Direction (-D) must be one of "top", "left", "bottom", "right"'
 
         image = image.convert("RGBA")
         # we always extend top, but rotate to extend along the requested side
-        if direction == 'left':
+        if direction == "left":
             image = image.transpose(Image.Transpose.ROTATE_270)
-        elif direction == 'bottom':
+        elif direction == "bottom":
             image = image.transpose(Image.Transpose.ROTATE_180)
-        elif direction == 'right':
+        elif direction == "right":
             image = image.transpose(Image.Transpose.ROTATE_90)
 
-        pixels = image.height//2 if pixels is None else int(pixels)
-        assert 0 < pixels < image.height, 'Direction (-D) pixels length must be in the range 0 - image.size'
+        pixels = image.height // 2 if pixels is None else int(pixels)
+        assert (
+            0 < pixels < image.height
+        ), "Direction (-D) pixels length must be in the range 0 - image.size"
 
         # the top part of the image is taken from the source image mirrored
         # coordinates (0,0) are the upper left corner of an image
@@ -74,19 +85,18 @@ class Outpaint(object):
         new_img.paste(bottom, (0, pixels))
 
         # create a 10% dither in the middle
-        dither = min(image.height//10, pixels)
+        dither = min(image.height // 10, pixels)
         for x in range(0, image.width, 2):
             for y in range(pixels - dither, pixels + dither):
                 (r, g, b, a) = new_img.getpixel((x, y))
                 new_img.putpixel((x, y), (r, g, b, 0))
 
         # let's rotate back again
-        if direction == 'left':
+        if direction == "left":
             new_img = new_img.transpose(Image.Transpose.ROTATE_90)
-        elif direction == 'bottom':
+        elif direction == "bottom":
             new_img = new_img.transpose(Image.Transpose.ROTATE_180)
-        elif direction == 'right':
+        elif direction == "right":
             new_img = new_img.transpose(Image.Transpose.ROTATE_270)
 
         return new_img
-

invokeai/backend/restoration/realesrgan.py

@@ -1,13 +1,15 @@
-import torch
-import warnings
-import numpy as np
 import os
+import warnings
 
-from ldm.invoke.globals import Globals
+import numpy as np
+import torch
 from PIL import Image
 from PIL.Image import Image as ImageType
 
-class ESRGAN():
+from invokeai.backend.globals import Globals
+
+
+class ESRGAN:
     def __init__(self, bg_tile_size=400) -> None:
         self.bg_tile_size = bg_tile_size
 
@@ -22,12 +24,23 @@ class ESRGAN():
         else:
             use_half_precision = True
 
-        from realesrgan.archs.srvgg_arch import SRVGGNetCompact
         from realesrgan import RealESRGANer
+        from realesrgan.archs.srvgg_arch import SRVGGNetCompact
 
-        model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=32, upscale=4, act_type='prelu')
-        model_path = os.path.join(Globals.root, 'models/realesrgan/realesr-general-x4v3.pth')
-        wdn_model_path = os.path.join(Globals.root, 'models/realesrgan/realesr-general-wdn-x4v3.pth')
+        model = SRVGGNetCompact(
+            num_in_ch=3,
+            num_out_ch=3,
+            num_feat=64,
+            num_conv=32,
+            upscale=4,
+            act_type="prelu",
+        )
+        model_path = os.path.join(
+            Globals.root, "models/realesrgan/realesr-general-x4v3.pth"
+        )
+        wdn_model_path = os.path.join(
+            Globals.root, "models/realesrgan/realesr-general-wdn-x4v3.pth"
+        )
         scale = 4
 
         bg_upsampler = RealESRGANer(
@@ -43,41 +56,49 @@ class ESRGAN():
 
         return bg_upsampler
 
-    def process(self, image: ImageType, strength: float, seed: str = None, upsampler_scale: int = 2, denoise_str: float = 0.75):
+    def process(
+        self,
+        image: ImageType,
+        strength: float,
+        seed: str = None,
+        upsampler_scale: int = 2,
+        denoise_str: float = 0.75,
+    ):
         with warnings.catch_warnings():
-            warnings.filterwarnings('ignore', category=DeprecationWarning)
-            warnings.filterwarnings('ignore', category=UserWarning)
+            warnings.filterwarnings("ignore", category=DeprecationWarning)
+            warnings.filterwarnings("ignore", category=UserWarning)
 
             try:
                 upsampler = self.load_esrgan_bg_upsampler(denoise_str)
             except Exception:
-                import traceback
                 import sys
-                print('>> Error loading Real-ESRGAN:', file=sys.stderr)
+                import traceback
+
+                print(">> Error loading Real-ESRGAN:", file=sys.stderr)
                 print(traceback.format_exc(), file=sys.stderr)
 
         if upsampler_scale == 0:
-            print('>> Real-ESRGAN: Invalid scaling option. Image not upscaled.')
+            print(">> Real-ESRGAN: Invalid scaling option. Image not upscaled.")
             return image
 
         if seed is not None:
             print(
-                f'>> Real-ESRGAN Upscaling seed:{seed}, scale:{upsampler_scale}x, tile:{self.bg_tile_size}, denoise:{denoise_str}'
+                f">> Real-ESRGAN Upscaling seed:{seed}, scale:{upsampler_scale}x, tile:{self.bg_tile_size}, denoise:{denoise_str}"
             )
         # ESRGAN outputs images with partial transparency if given RGBA images; convert to RGB
         image = image.convert("RGB")
 
         # REALSRGAN expects a BGR np array; make array and flip channels
-        bgr_image_array = np.array(image, dtype=np.uint8)[...,::-1]
+        bgr_image_array = np.array(image, dtype=np.uint8)[..., ::-1]
 
         output, _ = upsampler.enhance(
             bgr_image_array,
             outscale=upsampler_scale,
-            alpha_upsampler='realesrgan',
+            alpha_upsampler="realesrgan",
         )
 
         # Flip the channels back to RGB
-        res = Image.fromarray(output[...,::-1])
+        res = Image.fromarray(output[..., ::-1])
 
         if strength < 1.0:
             # Resize the image to the new image if the sizes have changed

invokeai/backend/restoration/vqgan_arch.py

@@ -1,23 +1,27 @@
-'''
+"""
 VQGAN code, adapted from the original created by the Unleashing Transformers authors:
 https://github.com/samb-t/unleashing-transformers/blob/master/models/vqgan.py
 
-'''
+"""
+import copy
+
 import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-import copy
 from basicsr.utils import get_root_logger
 from basicsr.utils.registry import ARCH_REGISTRY
 
+
 def normalize(in_channels):
-    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+    return torch.nn.GroupNorm(
+        num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
+    )
 
 
 @torch.jit.script
 def swish(x):
-    return x*torch.sigmoid(x)
+    return x * torch.sigmoid(x)
 
 
 #  Define VQVAE classes
@@ -28,7 +32,9 @@ class VectorQuantizer(nn.Module):
         self.emb_dim = emb_dim  # dimension of embedding
         self.beta = beta  # commitment cost used in loss term, beta * ||z_e(x)-sg[e]||^2
         self.embedding = nn.Embedding(self.codebook_size, self.emb_dim)
-        self.embedding.weight.data.uniform_(-1.0 / self.codebook_size, 1.0 / self.codebook_size)
+        self.embedding.weight.data.uniform_(
+            -1.0 / self.codebook_size, 1.0 / self.codebook_size
+        )
 
     def forward(self, z):
         # reshape z -> (batch, height, width, channel) and flatten
@@ -36,23 +42,32 @@ class VectorQuantizer(nn.Module):
         z_flattened = z.view(-1, self.emb_dim)
 
         # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
-        d = (z_flattened ** 2).sum(dim=1, keepdim=True) + (self.embedding.weight**2).sum(1) - \
-            2 * torch.matmul(z_flattened, self.embedding.weight.t())
+        d = (
+            (z_flattened**2).sum(dim=1, keepdim=True)
+            + (self.embedding.weight**2).sum(1)
+            - 2 * torch.matmul(z_flattened, self.embedding.weight.t())
+        )
 
         mean_distance = torch.mean(d)
         # find closest encodings
         # min_encoding_indices = torch.argmin(d, dim=1).unsqueeze(1)
-        min_encoding_scores, min_encoding_indices = torch.topk(d, 1, dim=1, largest=False)
+        min_encoding_scores, min_encoding_indices = torch.topk(
+            d, 1, dim=1, largest=False
+        )
         # [0-1], higher score, higher confidence
-        min_encoding_scores = torch.exp(-min_encoding_scores/10)
+        min_encoding_scores = torch.exp(-min_encoding_scores / 10)
 
-        min_encodings = torch.zeros(min_encoding_indices.shape[0], self.codebook_size).to(z)
+        min_encodings = torch.zeros(
+            min_encoding_indices.shape[0], self.codebook_size
+        ).to(z)
         min_encodings.scatter_(1, min_encoding_indices, 1)
 
         # get quantized latent vectors
         z_q = torch.matmul(min_encodings, self.embedding.weight).view(z.shape)
         # compute loss for embedding
-        loss = torch.mean((z_q.detach()-z)**2) + self.beta * torch.mean((z_q - z.detach()) ** 2)
+        loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * torch.mean(
+            (z_q - z.detach()) ** 2
+        )
         # preserve gradients
         z_q = z + (z_q - z).detach()
 
@@ -62,18 +77,22 @@ class VectorQuantizer(nn.Module):
         # reshape back to match original input shape
         z_q = z_q.permute(0, 3, 1, 2).contiguous()
 
-        return z_q, loss, {
-            "perplexity": perplexity,
-            "min_encodings": min_encodings,
-            "min_encoding_indices": min_encoding_indices,
-            "min_encoding_scores": min_encoding_scores,
-            "mean_distance": mean_distance
-            }
+        return (
+            z_q,
+            loss,
+            {
+                "perplexity": perplexity,
+                "min_encodings": min_encodings,
+                "min_encoding_indices": min_encoding_indices,
+                "min_encoding_scores": min_encoding_scores,
+                "mean_distance": mean_distance,
+            },
+        )
 
     def get_codebook_feat(self, indices, shape):
         # input indices: batch*token_num -> (batch*token_num)*1
         # shape: batch, height, width, channel
-        indices = indices.view(-1,1)
+        indices = indices.view(-1, 1)
         min_encodings = torch.zeros(indices.shape[0], self.codebook_size).to(indices)
         min_encodings.scatter_(1, indices, 1)
         # get quantized latent vectors
@@ -86,14 +105,24 @@ class VectorQuantizer(nn.Module):
 
 
 class GumbelQuantizer(nn.Module):
-    def __init__(self, codebook_size, emb_dim, num_hiddens, straight_through=False, kl_weight=5e-4, temp_init=1.0):
+    def __init__(
+        self,
+        codebook_size,
+        emb_dim,
+        num_hiddens,
+        straight_through=False,
+        kl_weight=5e-4,
+        temp_init=1.0,
+    ):
         super().__init__()
         self.codebook_size = codebook_size  # number of embeddings
         self.emb_dim = emb_dim  # dimension of embedding
         self.straight_through = straight_through
         self.temperature = temp_init
         self.kl_weight = kl_weight
-        self.proj = nn.Conv2d(num_hiddens, codebook_size, 1)  # projects last encoder layer to quantized logits
+        self.proj = nn.Conv2d(
+            num_hiddens, codebook_size, 1
+        )  # projects last encoder layer to quantized logits
         self.embed = nn.Embedding(codebook_size, emb_dim)
 
     def forward(self, z):
@@ -107,18 +136,21 @@ class GumbelQuantizer(nn.Module):
 
         # + kl divergence to the prior loss
         qy = F.softmax(logits, dim=1)
-        diff = self.kl_weight * torch.sum(qy * torch.log(qy * self.codebook_size + 1e-10), dim=1).mean()
+        diff = (
+            self.kl_weight
+            * torch.sum(qy * torch.log(qy * self.codebook_size + 1e-10), dim=1).mean()
+        )
         min_encoding_indices = soft_one_hot.argmax(dim=1)
 
-        return z_q, diff, {
-            "min_encoding_indices": min_encoding_indices
-        }
+        return z_q, diff, {"min_encoding_indices": min_encoding_indices}
 
 
 class Downsample(nn.Module):
     def __init__(self, in_channels):
         super().__init__()
-        self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
+        self.conv = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=3, stride=2, padding=0
+        )
 
     def forward(self, x):
         pad = (0, 1, 0, 1)
@@ -130,7 +162,9 @@ class Downsample(nn.Module):
 class Upsample(nn.Module):
     def __init__(self, in_channels):
         super().__init__()
-        self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1)
+        self.conv = nn.Conv2d(
+            in_channels, in_channels, kernel_size=3, stride=1, padding=1
+        )
 
     def forward(self, x):
         x = F.interpolate(x, scale_factor=2.0, mode="nearest")
@@ -145,11 +179,17 @@ class ResBlock(nn.Module):
         self.in_channels = in_channels
         self.out_channels = in_channels if out_channels is None else out_channels
         self.norm1 = normalize(in_channels)
-        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        self.conv1 = nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
         self.norm2 = normalize(out_channels)
-        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        self.conv2 = nn.Conv2d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
         if self.in_channels != self.out_channels:
-            self.conv_out = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+            self.conv_out = nn.Conv2d(
+                in_channels, out_channels, kernel_size=1, stride=1, padding=0
+            )
 
     def forward(self, x_in):
         x = x_in
@@ -172,32 +212,16 @@ class AttnBlock(nn.Module):
 
         self.norm = normalize(in_channels)
         self.q = torch.nn.Conv2d(
-            in_channels,
-            in_channels,
-            kernel_size=1,
-            stride=1,
-            padding=0
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
         )
         self.k = torch.nn.Conv2d(
-            in_channels,
-            in_channels,
-            kernel_size=1,
-            stride=1,
-            padding=0
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
         )
         self.v = torch.nn.Conv2d(
-            in_channels,
-            in_channels,
-            kernel_size=1,
-            stride=1,
-            padding=0
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
         )
         self.proj_out = torch.nn.Conv2d(
-            in_channels,
-            in_channels,
-            kernel_size=1,
-            stride=1,
-            padding=0
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
         )
 
     def forward(self, x):
@@ -209,26 +233,35 @@ class AttnBlock(nn.Module):
 
         # compute attention
         b, c, h, w = q.shape
-        q = q.reshape(b, c, h*w)
+        q = q.reshape(b, c, h * w)
         q = q.permute(0, 2, 1)
-        k = k.reshape(b, c, h*w)
+        k = k.reshape(b, c, h * w)
         w_ = torch.bmm(q, k)
-        w_ = w_ * (int(c)**(-0.5))
+        w_ = w_ * (int(c) ** (-0.5))
         w_ = F.softmax(w_, dim=2)
 
         # attend to values
-        v = v.reshape(b, c, h*w)
+        v = v.reshape(b, c, h * w)
         w_ = w_.permute(0, 2, 1)
         h_ = torch.bmm(v, w_)
         h_ = h_.reshape(b, c, h, w)
 
         h_ = self.proj_out(h_)
 
-        return x+h_
+        return x + h_
 
 
 class Encoder(nn.Module):
-    def __init__(self, in_channels, nf, emb_dim, ch_mult, num_res_blocks, resolution, attn_resolutions):
+    def __init__(
+        self,
+        in_channels,
+        nf,
+        emb_dim,
+        ch_mult,
+        num_res_blocks,
+        resolution,
+        attn_resolutions,
+    ):
         super().__init__()
         self.nf = nf
         self.num_resolutions = len(ch_mult)
@@ -237,7 +270,7 @@ class Encoder(nn.Module):
         self.attn_resolutions = attn_resolutions
 
         curr_res = self.resolution
-        in_ch_mult = (1,)+tuple(ch_mult)
+        in_ch_mult = (1,) + tuple(ch_mult)
 
         blocks = []
         # initial convultion
@@ -264,7 +297,9 @@ class Encoder(nn.Module):
 
         # normalise and convert to latent size
         blocks.append(normalize(block_in_ch))
-        blocks.append(nn.Conv2d(block_in_ch, emb_dim, kernel_size=3, stride=1, padding=1))
+        blocks.append(
+            nn.Conv2d(block_in_ch, emb_dim, kernel_size=3, stride=1, padding=1)
+        )
         self.blocks = nn.ModuleList(blocks)
 
     def forward(self, x):
@@ -286,11 +321,13 @@ class Generator(nn.Module):
         self.in_channels = emb_dim
         self.out_channels = 3
         block_in_ch = self.nf * self.ch_mult[-1]
-        curr_res = self.resolution // 2 ** (self.num_resolutions-1)
+        curr_res = self.resolution // 2 ** (self.num_resolutions - 1)
 
         blocks = []
         # initial conv
-        blocks.append(nn.Conv2d(self.in_channels, block_in_ch, kernel_size=3, stride=1, padding=1))
+        blocks.append(
+            nn.Conv2d(self.in_channels, block_in_ch, kernel_size=3, stride=1, padding=1)
+        )
 
         # non-local attention block
         blocks.append(ResBlock(block_in_ch, block_in_ch))
@@ -312,11 +349,14 @@ class Generator(nn.Module):
                 curr_res = curr_res * 2
 
         blocks.append(normalize(block_in_ch))
-        blocks.append(nn.Conv2d(block_in_ch, self.out_channels, kernel_size=3, stride=1, padding=1))
+        blocks.append(
+            nn.Conv2d(
+                block_in_ch, self.out_channels, kernel_size=3, stride=1, padding=1
+            )
+        )
 
         self.blocks = nn.ModuleList(blocks)
 
-
     def forward(self, x):
         for block in self.blocks:
             x = block(x)
@@ -326,8 +366,21 @@ class Generator(nn.Module):
 
 @ARCH_REGISTRY.register()
 class VQAutoEncoder(nn.Module):
-    def __init__(self, img_size, nf, ch_mult, quantizer="nearest", res_blocks=2, attn_resolutions=[16], codebook_size=1024, emb_dim=256,
-                beta=0.25, gumbel_straight_through=False, gumbel_kl_weight=1e-8, model_path=None):
+    def __init__(
+        self,
+        img_size,
+        nf,
+        ch_mult,
+        quantizer="nearest",
+        res_blocks=2,
+        attn_resolutions=[16],
+        codebook_size=1024,
+        emb_dim=256,
+        beta=0.25,
+        gumbel_straight_through=False,
+        gumbel_kl_weight=1e-8,
+        model_path=None,
+    ):
         super().__init__()
         logger = get_root_logger()
         self.in_channels = 3
@@ -346,11 +399,13 @@ class VQAutoEncoder(nn.Module):
             self.ch_mult,
             self.n_blocks,
             self.resolution,
-            self.attn_resolutions
+            self.attn_resolutions,
         )
         if self.quantizer_type == "nearest":
-            self.beta = beta #0.25
-            self.quantize = VectorQuantizer(self.codebook_size, self.embed_dim, self.beta)
+            self.beta = beta  # 0.25
+            self.quantize = VectorQuantizer(
+                self.codebook_size, self.embed_dim, self.beta
+            )
         elif self.quantizer_type == "gumbel":
             self.gumbel_num_hiddens = emb_dim
             self.straight_through = gumbel_straight_through
@@ -360,7 +415,7 @@ class VQAutoEncoder(nn.Module):
                 self.embed_dim,
                 self.gumbel_num_hiddens,
                 self.straight_through,
-                self.kl_weight
+                self.kl_weight,
             )
         self.generator = Generator(
             self.nf,
@@ -368,20 +423,23 @@ class VQAutoEncoder(nn.Module):
             self.ch_mult,
             self.n_blocks,
             self.resolution,
-            self.attn_resolutions
+            self.attn_resolutions,
         )
 
         if model_path is not None:
-            chkpt = torch.load(model_path, map_location='cpu')
-            if 'params_ema' in chkpt:
-                self.load_state_dict(torch.load(model_path, map_location='cpu')['params_ema'])
-                logger.info(f'vqgan is loaded from: {model_path} [params_ema]')
-            elif 'params' in chkpt:
-                self.load_state_dict(torch.load(model_path, map_location='cpu')['params'])
-                logger.info(f'vqgan is loaded from: {model_path} [params]')
+            chkpt = torch.load(model_path, map_location="cpu")
+            if "params_ema" in chkpt:
+                self.load_state_dict(
+                    torch.load(model_path, map_location="cpu")["params_ema"]
+                )
+                logger.info(f"vqgan is loaded from: {model_path} [params_ema]")
+            elif "params" in chkpt:
+                self.load_state_dict(
+                    torch.load(model_path, map_location="cpu")["params"]
+                )
+                logger.info(f"vqgan is loaded from: {model_path} [params]")
             else:
-                raise ValueError(f'Wrong params!')
-
+                raise ValueError(f"Wrong params!")
 
     def forward(self, x):
         x = self.encoder(x)
@@ -390,46 +448,67 @@ class VQAutoEncoder(nn.Module):
         return x, codebook_loss, quant_stats
 
 
-
 # patch based discriminator
 @ARCH_REGISTRY.register()
 class VQGANDiscriminator(nn.Module):
     def __init__(self, nc=3, ndf=64, n_layers=4, model_path=None):
         super().__init__()
 
-        layers = [nn.Conv2d(nc, ndf, kernel_size=4, stride=2, padding=1), nn.LeakyReLU(0.2, True)]
+        layers = [
+            nn.Conv2d(nc, ndf, kernel_size=4, stride=2, padding=1),
+            nn.LeakyReLU(0.2, True),
+        ]
         ndf_mult = 1
         ndf_mult_prev = 1
         for n in range(1, n_layers):  # gradually increase the number of filters
             ndf_mult_prev = ndf_mult
-            ndf_mult = min(2 ** n, 8)
+            ndf_mult = min(2**n, 8)
             layers += [
-                nn.Conv2d(ndf * ndf_mult_prev, ndf * ndf_mult, kernel_size=4, stride=2, padding=1, bias=False),
+                nn.Conv2d(
+                    ndf * ndf_mult_prev,
+                    ndf * ndf_mult,
+                    kernel_size=4,
+                    stride=2,
+                    padding=1,
+                    bias=False,
+                ),
                 nn.BatchNorm2d(ndf * ndf_mult),
-                nn.LeakyReLU(0.2, True)
+                nn.LeakyReLU(0.2, True),
             ]
 
         ndf_mult_prev = ndf_mult
-        ndf_mult = min(2 ** n_layers, 8)
+        ndf_mult = min(2**n_layers, 8)
 
         layers += [
-            nn.Conv2d(ndf * ndf_mult_prev, ndf * ndf_mult, kernel_size=4, stride=1, padding=1, bias=False),
+            nn.Conv2d(
+                ndf * ndf_mult_prev,
+                ndf * ndf_mult,
+                kernel_size=4,
+                stride=1,
+                padding=1,
+                bias=False,
+            ),
             nn.BatchNorm2d(ndf * ndf_mult),
-            nn.LeakyReLU(0.2, True)
+            nn.LeakyReLU(0.2, True),
         ]
 
         layers += [
-            nn.Conv2d(ndf * ndf_mult, 1, kernel_size=4, stride=1, padding=1)]  # output 1 channel prediction map
+            nn.Conv2d(ndf * ndf_mult, 1, kernel_size=4, stride=1, padding=1)
+        ]  # output 1 channel prediction map
         self.main = nn.Sequential(*layers)
 
         if model_path is not None:
-            chkpt = torch.load(model_path, map_location='cpu')
-            if 'params_d' in chkpt:
-                self.load_state_dict(torch.load(model_path, map_location='cpu')['params_d'])
-            elif 'params' in chkpt:
-                self.load_state_dict(torch.load(model_path, map_location='cpu')['params'])
+            chkpt = torch.load(model_path, map_location="cpu")
+            if "params_d" in chkpt:
+                self.load_state_dict(
+                    torch.load(model_path, map_location="cpu")["params_d"]
+                )
+            elif "params" in chkpt:
+                self.load_state_dict(
+                    torch.load(model_path, map_location="cpu")["params"]
+                )
             else:
-                raise ValueError(f'Wrong params!')
+                raise ValueError(f"Wrong params!")
 
     def forward(self, x):
         return self.main(x)

invokeai/backend/safety_checker.py

@@ -0,0 +1,82 @@
+'''
+SafetyChecker class - checks images against the StabilityAI NSFW filter
+and blurs images that contain potential NSFW content.
+'''
+import diffusers
+import numpy as np
+import torch
+import traceback
+from diffusers.pipelines.stable_diffusion.safety_checker import (
+    StableDiffusionSafetyChecker,
+)
+from pathlib import Path
+from PIL import Image, ImageFilter
+from transformers import AutoFeatureExtractor
+
+import invokeai.assets.web as web_assets
+from .globals import global_cache_dir
+from .util import CPU_DEVICE
+
+class SafetyChecker(object):
+    CAUTION_IMG = "caution.png"
+    
+    def __init__(self, device: torch.device):
+        path = Path(web_assets.__path__[0]) / self.CAUTION_IMG
+        caution = Image.open(path)
+        self.caution_img = caution.resize((caution.width // 2, caution.height // 2))
+        self.device = device
+        
+        try:
+            safety_model_id = "CompVis/stable-diffusion-safety-checker"
+            safety_model_path = global_cache_dir("hub")
+            self.safety_checker = StableDiffusionSafetyChecker.from_pretrained(
+                safety_model_id,
+                local_files_only=True,
+                cache_dir=safety_model_path,
+            )
+            self.safety_feature_extractor = AutoFeatureExtractor.from_pretrained(
+                safety_model_id,
+                local_files_only=True,
+                cache_dir=safety_model_path,
+            )
+        except Exception:
+            print(
+                "** An error was encountered while installing the safety checker:"
+            )
+            print(traceback.format_exc())
+
+    def check(self, image: Image.Image):
+        """
+        Check provided image against the StabilityAI safety checker and return
+
+        """
+
+        self.safety_checker.to(self.device)
+        features = self.safety_feature_extractor([image], return_tensors="pt")
+        features.to(self.device)
+        
+        # unfortunately checker requires the numpy version, so we have to convert back
+        x_image = np.array(image).astype(np.float32) / 255.0
+        x_image = x_image[None].transpose(0, 3, 1, 2)
+
+        diffusers.logging.set_verbosity_error()
+        checked_image, has_nsfw_concept = self.safety_checker(
+            images=x_image, clip_input=features.pixel_values
+        )
+        self.safety_checker.to(CPU_DEVICE) # offload
+        if has_nsfw_concept[0]:
+            print(
+                "** An image with potential non-safe content has been detected. A blurred image will be returned. **"
+            )
+            return self.blur(image)
+        else:
+            return image
+
+    def blur(self, input):
+        blurry = input.filter(filter=ImageFilter.GaussianBlur(radius=32))
+        try:
+            if caution := self.caution_img:
+                blurry.paste(caution, (0, 0), caution)
+        except FileNotFoundError:
+            pass
+        return blurry

invokeai/backend/stable_diffusion/__init__.py

@@ -0,0 +1,13 @@
+"""
+Initialization file for the invokeai.backend.stable_diffusion package
+"""
+from .concepts_lib import HuggingFaceConceptsLibrary
+from .diffusers_pipeline import (
+    ConditioningData,
+    PipelineIntermediateState,
+    StableDiffusionGeneratorPipeline,
+)
+from .diffusion import InvokeAIDiffuserComponent
+from .diffusion.cross_attention_map_saving import AttentionMapSaver
+from .diffusion.shared_invokeai_diffusion import PostprocessingSettings
+from .textual_inversion_manager import TextualInversionManager

invokeai/backend/stable_diffusion/concepts_lib.py

@@ -8,32 +8,50 @@ import os
 import re
 import traceback
 from typing import Callable
-from urllib import request, error as ul_error
-from huggingface_hub import HfFolder, hf_hub_url, ModelSearchArguments, ModelFilter, HfApi
-from ldm.invoke.globals import Globals
+from urllib import error as ul_error
+from urllib import request
+
+from huggingface_hub import (
+    HfApi,
+    HfFolder,
+    ModelFilter,
+    ModelSearchArguments,
+    hf_hub_url,
+)
+
+from invokeai.backend.globals import Globals
+
 
 class HuggingFaceConceptsLibrary(object):
     def __init__(self, root=None):
-        '''
+        """
         Initialize the Concepts object. May optionally pass a root directory.
-        '''
+        """
         self.root = root or Globals.root
         self.hf_api = HfApi()
         self.local_concepts = dict()
         self.concept_list = None
         self.concepts_loaded = dict()
-        self.triggers = dict()            # concept name to trigger phrase
-        self.concept_names = dict()       # trigger phrase to concept name
-        self.match_trigger = re.compile('(<[\w\- >]+>)') # trigger is slightly less restrictive than HF concept name
-        self.match_concept = re.compile('<([\w\-]+)>') # HF concept name can only contain A-Za-z0-9_-
+        self.triggers = dict()  # concept name to trigger phrase
+        self.concept_names = dict()  # trigger phrase to concept name
+        self.match_trigger = re.compile(
+            "(<[\w\- >]+>)"
+        )  # trigger is slightly less restrictive than HF concept name
+        self.match_concept = re.compile(
+            "<([\w\-]+)>"
+        )  # HF concept name can only contain A-Za-z0-9_-
 
-    def list_concepts(self)->list:
-        '''
+    def list_concepts(self) -> list:
+        """
         Return a list of all the concepts by name, without the 'sd-concepts-library' part.
         Also adds local concepts in invokeai/embeddings folder.
-        '''
-        local_concepts_now = self.get_local_concepts(os.path.join(self.root, 'embeddings'))
-        local_concepts_to_add = set(local_concepts_now).difference(set(self.local_concepts))
+        """
+        local_concepts_now = self.get_local_concepts(
+            os.path.join(self.root, "embeddings")
+        )
+        local_concepts_to_add = set(local_concepts_now).difference(
+            set(self.local_concepts)
+        )
         self.local_concepts.update(local_concepts_now)
 
         if self.concept_list is not None:
@@ -43,83 +61,96 @@ class HuggingFaceConceptsLibrary(object):
             return self.concept_list
         else:
             try:
-                models = self.hf_api.list_models(filter=ModelFilter(model_name='sd-concepts-library/'))
-                self.concept_list = [a.id.split('/')[1] for a in models]
+                models = self.hf_api.list_models(
+                    filter=ModelFilter(model_name="sd-concepts-library/")
+                )
+                self.concept_list = [a.id.split("/")[1] for a in models]
                 # when init, add all in dir. when not init, add only concepts added between init and now
                 self.concept_list.extend(list(local_concepts_to_add))
             except Exception as e:
-                print(f' ** WARNING: Hugging Face textual inversion concepts libraries could not be loaded. The error was {str(e)}.')
-                print(' ** You may load .bin and .pt file(s) manually using the --embedding_directory argument.')
+                print(
+                    f" ** WARNING: Hugging Face textual inversion concepts libraries could not be loaded. The error was {str(e)}."
+                )
+                print(
+                    " ** You may load .bin and .pt file(s) manually using the --embedding_directory argument."
+                )
             return self.concept_list
 
-    def get_concept_model_path(self, concept_name:str)->str:
-        '''
+    def get_concept_model_path(self, concept_name: str) -> str:
+        """
         Returns the path to the 'learned_embeds.bin' file in
         the named concept. Returns None if invalid or cannot
         be downloaded.
-        '''
+        """
         if not concept_name in self.list_concepts():
-            print(f'This concept is not a local embedding trigger, nor is it a HuggingFace concept. Generation will continue without the concept.')
+            print(
+                f"This concept is not a local embedding trigger, nor is it a HuggingFace concept. Generation will continue without the concept."
+            )
             return None
-        return self.get_concept_file(concept_name.lower(),'learned_embeds.bin')
+        return self.get_concept_file(concept_name.lower(), "learned_embeds.bin")
 
-    def concept_to_trigger(self, concept_name:str)->str:
-        '''
+    def concept_to_trigger(self, concept_name: str) -> str:
+        """
         Given a concept name returns its trigger by looking in the
         "token_identifier.txt" file.
-        '''
+        """
         if concept_name in self.triggers:
             return self.triggers[concept_name]
         elif self.concept_is_local(concept_name):
-            trigger = f'<{concept_name}>'
+            trigger = f"<{concept_name}>"
             self.triggers[concept_name] = trigger
             self.concept_names[trigger] = concept_name
             return trigger
 
-        file = self.get_concept_file(concept_name, 'token_identifier.txt', local_only=True)
+        file = self.get_concept_file(
+            concept_name, "token_identifier.txt", local_only=True
+        )
         if not file:
             return None
-        with open(file,'r') as f:
+        with open(file, "r") as f:
             trigger = f.readline()
             trigger = trigger.strip()
         self.triggers[concept_name] = trigger
         self.concept_names[trigger] = concept_name
         return trigger
 
-    def trigger_to_concept(self, trigger:str)->str:
-        '''
+    def trigger_to_concept(self, trigger: str) -> str:
+        """
         Given a trigger phrase, maps it to the concept library name.
         Only works if concept_to_trigger() has previously been called
         on this library. There needs to be a persistent database for
         this.
-        '''
-        concept = self.concept_names.get(trigger,None)
-        return f'<{concept}>' if concept else f'{trigger}'
+        """
+        concept = self.concept_names.get(trigger, None)
+        return f"<{concept}>" if concept else f"{trigger}"
 
-    def replace_triggers_with_concepts(self, prompt:str)->str:
-        '''
+    def replace_triggers_with_concepts(self, prompt: str) -> str:
+        """
         Given a prompt string that contains <trigger> tags, replace these
         tags with the concept name. The reason for this is so that the
         concept names get stored in the prompt metadata. There is no
         controlling of colliding triggers in the SD library, so it is
         better to store the concept name (unique) than the concept trigger
         (not necessarily unique!)
-        '''
+        """
         if not prompt:
             return prompt
         triggers = self.match_trigger.findall(prompt)
         if not triggers:
             return prompt
 
-        def do_replace(match)->str:
-            return self.trigger_to_concept(match.group(1)) or f'<{match.group(1)}>'
+        def do_replace(match) -> str:
+            return self.trigger_to_concept(match.group(1)) or f"<{match.group(1)}>"
+
         return self.match_trigger.sub(do_replace, prompt)
 
-    def replace_concepts_with_triggers(self,
-                                       prompt:str,
-                                       load_concepts_callback: Callable[[list], any],
-                                       excluded_tokens:list[str])->str:
-        '''
+    def replace_concepts_with_triggers(
+        self,
+        prompt: str,
+        load_concepts_callback: Callable[[list], any],
+        excluded_tokens: list[str],
+    ) -> str:
+        """
         Given a prompt string that contains `<concept_name>` tags, replace
         these tags with the appropriate trigger.
 
@@ -128,20 +159,30 @@ class HuggingFaceConceptsLibrary(object):
 
         `excluded_tokens` are any tokens that should not be replaced, typically because they
         are trigger tokens from a locally-loaded embedding.
-        '''
+        """
         concepts = self.match_concept.findall(prompt)
         if not concepts:
             return prompt
         load_concepts_callback(concepts)
 
-        def do_replace(match)->str:
-            if excluded_tokens and f'<{match.group(1)}>' in excluded_tokens:
-                return f'<{match.group(1)}>'
-            return self.concept_to_trigger(match.group(1)) or f'<{match.group(1)}>'
+        def do_replace(match) -> str:
+            if excluded_tokens and f"<{match.group(1)}>" in excluded_tokens:
+                return f"<{match.group(1)}>"
+            return self.concept_to_trigger(match.group(1)) or f"<{match.group(1)}>"
+
         return self.match_concept.sub(do_replace, prompt)
 
-    def get_concept_file(self, concept_name:str, file_name:str='learned_embeds.bin' , local_only:bool=False)->str:
-        if not (self.concept_is_downloaded(concept_name) or self.concept_is_local(concept_name) or local_only):
+    def get_concept_file(
+        self,
+        concept_name: str,
+        file_name: str = "learned_embeds.bin",
+        local_only: bool = False,
+    ) -> str:
+        if not (
+            self.concept_is_downloaded(concept_name)
+            or self.concept_is_local(concept_name)
+            or local_only
+        ):
             self.download_concept(concept_name)
 
         # get local path in invokeai/embeddings if local concept
@@ -153,19 +194,19 @@ class HuggingFaceConceptsLibrary(object):
             path = os.path.join(concept_path, file_name)
         return path if os.path.exists(path) else None
 
-    def concept_is_local(self, concept_name)->bool:
+    def concept_is_local(self, concept_name) -> bool:
         return concept_name in self.local_concepts
 
-    def concept_is_downloaded(self, concept_name)->bool:
+    def concept_is_downloaded(self, concept_name) -> bool:
         concept_directory = self._concept_path(concept_name)
         return os.path.exists(concept_directory)
 
-    def download_concept(self,concept_name)->bool:
+    def download_concept(self, concept_name) -> bool:
         repo_id = self._concept_id(concept_name)
         dest = self._concept_path(concept_name)
 
         access_token = HfFolder.get_token()
-        header = [("Authorization", f'Bearer {access_token}')] if access_token else []
+        header = [("Authorization", f"Bearer {access_token}")] if access_token else []
         opener = request.build_opener()
         opener.addheaders = header
         request.install_opener(opener)
@@ -174,45 +215,59 @@ class HuggingFaceConceptsLibrary(object):
         succeeded = True
 
         bytes = 0
+
         def tally_download_size(chunk, size, total):
             nonlocal bytes
-            if chunk==0:
+            if chunk == 0:
                 bytes += total
 
-        print(f'>> Downloading {repo_id}...',end='')
+        print(f">> Downloading {repo_id}...", end="")
         try:
-            for file in ('README.md','learned_embeds.bin','token_identifier.txt','type_of_concept.txt'):
+            for file in (
+                "README.md",
+                "learned_embeds.bin",
+                "token_identifier.txt",
+                "type_of_concept.txt",
+            ):
                 url = hf_hub_url(repo_id, file)
-                request.urlretrieve(url, os.path.join(dest,file),reporthook=tally_download_size)
+                request.urlretrieve(
+                    url, os.path.join(dest, file), reporthook=tally_download_size
+                )
         except ul_error.HTTPError as e:
-            if e.code==404:
-                print(f'This concept is not known to the Hugging Face library. Generation will continue without the concept.')
+            if e.code == 404:
+                print(
+                    f"This concept is not known to the Hugging Face library. Generation will continue without the concept."
+                )
             else:
-                print(f'Failed to download {concept_name}/{file} ({str(e)}. Generation will continue without the concept.)')
+                print(
+                    f"Failed to download {concept_name}/{file} ({str(e)}. Generation will continue without the concept.)"
+                )
             os.rmdir(dest)
             return False
         except ul_error.URLError as e:
-            print(f'ERROR: {str(e)}. This may reflect a network issue. Generation will continue without the concept.')
+            print(
+                f"ERROR: {str(e)}. This may reflect a network issue. Generation will continue without the concept."
+            )
             os.rmdir(dest)
             return False
-        print('...{:.2f}Kb'.format(bytes/1024))
+        print("...{:.2f}Kb".format(bytes / 1024))
         return succeeded
 
-    def _concept_id(self, concept_name:str)->str:
-        return f'sd-concepts-library/{concept_name}'
+    def _concept_id(self, concept_name: str) -> str:
+        return f"sd-concepts-library/{concept_name}"
 
-    def _concept_path(self, concept_name:str)->str:
-        return os.path.join(self.root,'models','sd-concepts-library',concept_name)
+    def _concept_path(self, concept_name: str) -> str:
+        return os.path.join(self.root, "models", "sd-concepts-library", concept_name)
 
-    def _concept_local_path(self, concept_name:str)->str:
+    def _concept_local_path(self, concept_name: str) -> str:
         filename = self.local_concepts[concept_name]
-        return os.path.join(self.root,'embeddings',filename)
+        return os.path.join(self.root, "embeddings", filename)
 
-    def get_local_concepts(self, loc_dir:str):
+    def get_local_concepts(self, loc_dir: str):
         locs_dic = dict()
         if os.path.isdir(loc_dir):
             for file in os.listdir(loc_dir):
                 f = os.path.splitext(file)
-                if f[1] == '.bin' or f[1] == '.pt':
+                if f[1] == ".bin" or f[1] == ".pt":
                     locs_dic[f[0]] = file
         return locs_dic

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -2,22 +2,29 @@ from __future__ import annotations
 
 import dataclasses
 import inspect
-import psutil
 import secrets
 from collections.abc import Sequence
 from dataclasses import dataclass, field
-from typing import List, Optional, Union, Callable, Type, TypeVar, Generic, Any
+from typing import Any, Callable, Generic, List, Optional, Type, TypeVar, Union
 
-import PIL.Image
 import einops
+import PIL.Image
+from accelerate.utils import set_seed
 import psutil
 import torch
 import torchvision.transforms as T
+from compel import EmbeddingsProvider
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
-from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipeline
-from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img import StableDiffusionImg2ImgPipeline
-from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import (
+    StableDiffusionPipeline,
+)
+from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img import (
+    StableDiffusionImg2ImgPipeline,
+)
+from diffusers.pipelines.stable_diffusion.safety_checker import (
+    StableDiffusionSafetyChecker,
+)
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.schedulers.scheduling_utils import SchedulerMixin, SchedulerOutput
 from diffusers.utils.import_utils import is_xformers_available
@@ -26,13 +33,16 @@ from torchvision.transforms.functional import resize as tv_resize
 from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 from typing_extensions import ParamSpec
 
-from ldm.invoke.globals import Globals
-from ldm.models.diffusion.shared_invokeai_diffusion import InvokeAIDiffuserComponent, PostprocessingSettings
-from ldm.modules.textual_inversion_manager import TextualInversionManager
-from ..devices import normalize_device, CPU_DEVICE
-from ..offloading import LazilyLoadedModelGroup, FullyLoadedModelGroup, ModelGroup
-from ...models.diffusion.cross_attention_map_saving import AttentionMapSaver
-from compel import EmbeddingsProvider
+from invokeai.backend.globals import Globals
+
+from ..util import CPU_DEVICE, normalize_device
+from .diffusion import (
+    AttentionMapSaver,
+    InvokeAIDiffuserComponent,
+    PostprocessingSettings,
+)
+from .offloading import FullyLoadedModelGroup, LazilyLoadedModelGroup, ModelGroup
+from .textual_inversion_manager import TextualInversionManager
 
 
 @dataclass
@@ -45,16 +55,6 @@ class PipelineIntermediateState:
     attention_map_saver: Optional[AttentionMapSaver] = None
 
 
-# copied from configs/stable-diffusion/v1-inference.yaml
-_default_personalization_config_params = dict(
-    placeholder_strings=["*"],
-    initializer_wods=["sculpture"],
-    per_image_tokens=False,
-    num_vectors_per_token=1,
-    progressive_words=False
-)
-
-
 @dataclass
 class AddsMaskLatents:
     """Add the channels required for inpainting model input.
@@ -64,29 +64,34 @@ class AddsMaskLatents:
 
     This class assumes the same mask and base image should apply to all items in the batch.
     """
+
     forward: Callable[[torch.Tensor, torch.Tensor, torch.Tensor], torch.Tensor]
     mask: torch.Tensor
     initial_image_latents: torch.Tensor
 
-    def __call__(self, latents: torch.Tensor, t: torch.Tensor, text_embeddings: torch.Tensor) -> torch.Tensor:
+    def __call__(
+        self, latents: torch.Tensor, t: torch.Tensor, text_embeddings: torch.Tensor
+    ) -> torch.Tensor:
         model_input = self.add_mask_channels(latents)
         return self.forward(model_input, t, text_embeddings)
 
     def add_mask_channels(self, latents):
         batch_size = latents.size(0)
         # duplicate mask and latents for each batch
-        mask = einops.repeat(self.mask, 'b c h w -> (repeat b) c h w', repeat=batch_size)
-        image_latents = einops.repeat(self.initial_image_latents, 'b c h w -> (repeat b) c h w', repeat=batch_size)
+        mask = einops.repeat(
+            self.mask, "b c h w -> (repeat b) c h w", repeat=batch_size
+        )
+        image_latents = einops.repeat(
+            self.initial_image_latents, "b c h w -> (repeat b) c h w", repeat=batch_size
+        )
         # add mask and image as additional channels
-        model_input, _ = einops.pack([latents, mask, image_latents], 'b * h w')
+        model_input, _ = einops.pack([latents, mask, image_latents], "b * h w")
         return model_input
 
 
 def are_like_tensors(a: torch.Tensor, b: object) -> bool:
-    return (
-        isinstance(b, torch.Tensor)
-        and (a.size() == b.size())
-    )
+    return isinstance(b, torch.Tensor) and (a.size() == b.size())
+
 
 @dataclass
 class AddsMaskGuidance:
@@ -96,7 +101,9 @@ class AddsMaskGuidance:
     noise: torch.Tensor
     _debug: Optional[Callable] = None
 
-    def __call__(self, step_output: BaseOutput | SchedulerOutput, t: torch.Tensor, conditioning) -> BaseOutput:
+    def __call__(
+        self, step_output: BaseOutput | SchedulerOutput, t: torch.Tensor, conditioning
+    ) -> BaseOutput:
         output_class = step_output.__class__  # We'll create a new one with masked data.
 
         # The problem with taking SchedulerOutput instead of the model output is that we're less certain what's in it.
@@ -106,30 +113,41 @@ class AddsMaskGuidance:
         prev_sample = step_output[0]
         # Mask anything that has the same shape as prev_sample, return others as-is.
         return output_class(
-            {k: (self.apply_mask(v, self._t_for_field(k, t))
-                 if are_like_tensors(prev_sample, v) else v)
-            for k, v in step_output.items()}
+            {
+                k: (
+                    self.apply_mask(v, self._t_for_field(k, t))
+                    if are_like_tensors(prev_sample, v)
+                    else v
+                )
+                for k, v in step_output.items()
+            }
         )
 
-    def _t_for_field(self, field_name:str, t):
+    def _t_for_field(self, field_name: str, t):
         if field_name == "pred_original_sample":
             return torch.zeros_like(t, dtype=t.dtype)  # it represents t=0
         return t
 
     def apply_mask(self, latents: torch.Tensor, t) -> torch.Tensor:
         batch_size = latents.size(0)
-        mask = einops.repeat(self.mask, 'b c h w -> (repeat b) c h w', repeat=batch_size)
+        mask = einops.repeat(
+            self.mask, "b c h w -> (repeat b) c h w", repeat=batch_size
+        )
         if t.dim() == 0:
             # some schedulers expect t to be one-dimensional.
             # TODO: file diffusers bug about inconsistency?
-            t = einops.repeat(t, '-> batch', batch=batch_size)
+            t = einops.repeat(t, "-> batch", batch=batch_size)
         # Noise shouldn't be re-randomized between steps here. The multistep schedulers
         # get very confused about what is happening from step to step when we do that.
         mask_latents = self.scheduler.add_noise(self.mask_latents, self.noise, t)
         # TODO: Do we need to also apply scheduler.scale_model_input? Or is add_noise appropriately scaled already?
         # mask_latents = self.scheduler.scale_model_input(mask_latents, t)
-        mask_latents = einops.repeat(mask_latents, 'b c h w -> (repeat b) c h w', repeat=batch_size)
-        masked_input = torch.lerp(mask_latents.to(dtype=latents.dtype), latents, mask.to(dtype=latents.dtype))
+        mask_latents = einops.repeat(
+            mask_latents, "b c h w -> (repeat b) c h w", repeat=batch_size
+        )
+        masked_input = torch.lerp(
+            mask_latents.to(dtype=latents.dtype), latents, mask.to(dtype=latents.dtype)
+        )
         if self._debug:
             self._debug(masked_input, f"t={t} lerped")
         return masked_input
@@ -139,18 +157,22 @@ def trim_to_multiple_of(*args, multiple_of=8):
     return tuple((x - x % multiple_of) for x in args)
 
 
-def image_resized_to_grid_as_tensor(image: PIL.Image.Image, normalize: bool=True, multiple_of=8) -> torch.FloatTensor:
+def image_resized_to_grid_as_tensor(
+    image: PIL.Image.Image, normalize: bool = True, multiple_of=8
+) -> torch.FloatTensor:
     """
 
     :param image: input image
     :param normalize: scale the range to [-1, 1] instead of [0, 1]
     :param multiple_of: resize the input so both dimensions are a multiple of this
     """
-    w, h = trim_to_multiple_of(*image.size)
-    transformation = T.Compose([
-        T.Resize((h, w), T.InterpolationMode.LANCZOS),
-        T.ToTensor(),
-    ])
+    w, h = trim_to_multiple_of(*image.size, multiple_of=multiple_of)
+    transformation = T.Compose(
+        [
+            T.Resize((h, w), T.InterpolationMode.LANCZOS),
+            T.ToTensor(),
+        ]
+    )
     tensor = transformation(image)
     if normalize:
         tensor = tensor * 2.0 - 1.0
@@ -160,9 +182,11 @@ def image_resized_to_grid_as_tensor(image: PIL.Image.Image, normalize: bool=True
 def is_inpainting_model(unet: UNet2DConditionModel):
     return unet.conv_in.in_channels == 9
 
-CallbackType = TypeVar('CallbackType')
-ReturnType = TypeVar('ReturnType')
-ParamType = ParamSpec('ParamType')
+
+CallbackType = TypeVar("CallbackType")
+ReturnType = TypeVar("ReturnType")
+ParamType = ParamSpec("ParamType")
+
 
 @dataclass(frozen=True)
 class GeneratorToCallbackinator(Generic[ParamType, ReturnType, CallbackType]):
@@ -171,9 +195,12 @@ class GeneratorToCallbackinator(Generic[ParamType, ReturnType, CallbackType]):
     generator_method: Callable[ParamType, ReturnType]
     callback_arg_type: Type[CallbackType]
 
-    def __call__(self, *args: ParamType.args,
-                 callback:Callable[[CallbackType], Any]=None,
-                 **kwargs: ParamType.kwargs) -> ReturnType:
+    def __call__(
+        self,
+        *args: ParamType.args,
+        callback: Callable[[CallbackType], Any] = None,
+        **kwargs: ParamType.kwargs,
+    ) -> ReturnType:
         result = None
         for result in self.generator_method(*args, **kwargs):
             if callback is not None and isinstance(result, self.callback_arg_type):
@@ -218,6 +245,7 @@ class ConditioningData:
                 scheduler_args[name] = value
         return dataclasses.replace(self, scheduler_args=scheduler_args)
 
+
 @dataclass
 class InvokeAIStableDiffusionPipelineOutput(StableDiffusionPipelineOutput):
     r"""
@@ -253,10 +281,10 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
             [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
         unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
         scheduler ([`SchedulerMixin`]):
-            A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
             [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
         safety_checker ([`StableDiffusionSafetyChecker`]):
-            Classification module that estimates whether generated images could be considered offsensive or harmful.
+            Classification module that estimates whether generated images could be considered offensive or harmful.
             Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
         feature_extractor ([`CLIPFeatureExtractor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
@@ -275,10 +303,18 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         safety_checker: Optional[StableDiffusionSafetyChecker],
         feature_extractor: Optional[CLIPFeatureExtractor],
         requires_safety_checker: bool = False,
-        precision: str = 'float32',
+        precision: str = "float32",
     ):
-        super().__init__(vae, text_encoder, tokenizer, unet, scheduler,
-                         safety_checker, feature_extractor, requires_safety_checker)
+        super().__init__(
+            vae,
+            text_encoder,
+            tokenizer,
+            unet,
+            scheduler,
+            safety_checker,
+            feature_extractor,
+            requires_safety_checker,
+        )
 
         self.register_modules(
             vae=vae,
@@ -289,27 +325,34 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
-        self.invokeai_diffuser = InvokeAIDiffuserComponent(self.unet, self._unet_forward, is_running_diffusers=True)
-        use_full_precision = (precision == 'float32' or precision == 'autocast')
-        self.textual_inversion_manager = TextualInversionManager(tokenizer=self.tokenizer,
-                                                                 text_encoder=self.text_encoder,
-                                                                 full_precision=use_full_precision)
+        self.invokeai_diffuser = InvokeAIDiffuserComponent(
+            self.unet, self._unet_forward, is_running_diffusers=True
+        )
+        use_full_precision = precision == "float32" or precision == "autocast"
+        self.textual_inversion_manager = TextualInversionManager(
+            tokenizer=self.tokenizer,
+            text_encoder=self.text_encoder,
+            full_precision=use_full_precision,
+        )
         # InvokeAI's interface for text embeddings and whatnot
         self.embeddings_provider = EmbeddingsProvider(
             tokenizer=self.tokenizer,
             text_encoder=self.text_encoder,
-            textual_inversion_manager=self.textual_inversion_manager
+            textual_inversion_manager=self.textual_inversion_manager,
         )
 
         self._model_group = FullyLoadedModelGroup(self.unet.device)
         self._model_group.install(*self._submodels)
 
-
     def _adjust_memory_efficient_attention(self, latents: torch.Tensor):
         """
         if xformers is available, use it, otherwise use sliced attention.
         """
-        if torch.cuda.is_available() and is_xformers_available() and not Globals.disable_xformers:
+        if (
+            torch.cuda.is_available()
+            and is_xformers_available()
+            and not Globals.disable_xformers
+        ):
             self.enable_xformers_memory_efficient_attention()
         else:
             if torch.backends.mps.is_available():
@@ -318,25 +361,32 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
                 # fix is in https://github.com/kulinseth/pytorch/pull/222 but no idea when it will get merged to pytorch mainline.
                 pass
             else:
-                if self.device.type == 'cpu' or self.device.type == 'mps':
+                if self.device.type == "cpu" or self.device.type == "mps":
                     mem_free = psutil.virtual_memory().free
-                elif self.device.type == 'cuda':
+                elif self.device.type == "cuda":
                     mem_free, _ = torch.cuda.mem_get_info(normalize_device(self.device))
                 else:
                     raise ValueError(f"unrecognized device {self.device}")
                 # input tensor of [1, 4, h/8, w/8]
                 # output tensor of [16, (h/8 * w/8), (h/8 * w/8)]
-                bytes_per_element_needed_for_baddbmm_duplication = latents.element_size() + 4
-                max_size_required_for_baddbmm = \
-                    16 * \
-                    latents.size(dim=2) * latents.size(dim=3) * latents.size(dim=2) * latents.size(dim=3) * \
-                    bytes_per_element_needed_for_baddbmm_duplication
-                if max_size_required_for_baddbmm > (mem_free * 3.0 / 4.0): # 3.3 / 4.0 is from old Invoke code
-                    self.enable_attention_slicing(slice_size='max')
+                bytes_per_element_needed_for_baddbmm_duplication = (
+                    latents.element_size() + 4
+                )
+                max_size_required_for_baddbmm = (
+                    16
+                    * latents.size(dim=2)
+                    * latents.size(dim=3)
+                    * latents.size(dim=2)
+                    * latents.size(dim=3)
+                    * bytes_per_element_needed_for_baddbmm_duplication
+                )
+                if max_size_required_for_baddbmm > (
+                    mem_free * 3.0 / 4.0
+                ):  # 3.3 / 4.0 is from old Invoke code
+                    self.enable_attention_slicing(slice_size="max")
                 else:
                     self.disable_attention_slicing()
 
-
     def enable_offload_submodels(self, device: torch.device):
         """
         Offload each submodel when it's not in use.
@@ -377,11 +427,11 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         """
         Ready this pipeline's models.
 
-        i.e. pre-load them to the GPU if appropriate.
+        i.e. preload them to the GPU if appropriate.
         """
         self._model_group.ready()
 
-    def to(self, torch_device: Optional[Union[str, torch.device]] = None):
+    def to(self, torch_device: Optional[Union[str, torch.device]] = None, silence_dtype_warnings=False):
         # overridden method; types match the superclass.
         if torch_device is None:
             return self
@@ -398,12 +448,16 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         values = [getattr(self, name) for name in module_names.keys()]
         return [m for m in values if isinstance(m, torch.nn.Module)]
 
-    def image_from_embeddings(self, latents: torch.Tensor, num_inference_steps: int,
-                              conditioning_data: ConditioningData,
-                              *,
-                              noise: torch.Tensor,
-                              callback: Callable[[PipelineIntermediateState], None]=None,
-                              run_id=None) -> InvokeAIStableDiffusionPipelineOutput:
+    def image_from_embeddings(
+        self,
+        latents: torch.Tensor,
+        num_inference_steps: int,
+        conditioning_data: ConditioningData,
+        *,
+        noise: torch.Tensor,
+        callback: Callable[[PipelineIntermediateState], None] = None,
+        run_id=None,
+    ) -> InvokeAIStableDiffusionPipelineOutput:
         r"""
         Function invoked when calling the pipeline for generation.
 
@@ -417,71 +471,104 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         :param run_id:
         """
         result_latents, result_attention_map_saver = self.latents_from_embeddings(
-            latents, num_inference_steps,
+            latents,
+            num_inference_steps,
             conditioning_data,
             noise=noise,
             run_id=run_id,
-            callback=callback)
+            callback=callback,
+        )
         # https://discuss.huggingface.co/t/memory-usage-by-later-pipeline-stages/23699
         torch.cuda.empty_cache()
 
         with torch.inference_mode():
             image = self.decode_latents(result_latents)
-            output = InvokeAIStableDiffusionPipelineOutput(images=image, nsfw_content_detected=[], attention_map_saver=result_attention_map_saver)
+            output = InvokeAIStableDiffusionPipelineOutput(
+                images=image,
+                nsfw_content_detected=[],
+                attention_map_saver=result_attention_map_saver,
+            )
             return self.check_for_safety(output, dtype=conditioning_data.dtype)
 
-    def latents_from_embeddings(self, latents: torch.Tensor, num_inference_steps: int,
-                                conditioning_data: ConditioningData,
-                                *,
-                                noise: torch.Tensor,
-                                timesteps=None,
-                                additional_guidance: List[Callable] = None, run_id=None,
-                                callback: Callable[[PipelineIntermediateState], None] = None
-                                ) -> tuple[torch.Tensor, Optional[AttentionMapSaver]]:
+    def latents_from_embeddings(
+        self,
+        latents: torch.Tensor,
+        num_inference_steps: int,
+        conditioning_data: ConditioningData,
+        *,
+        noise: torch.Tensor,
+        timesteps=None,
+        additional_guidance: List[Callable] = None,
+        run_id=None,
+        callback: Callable[[PipelineIntermediateState], None] = None,
+    ) -> tuple[torch.Tensor, Optional[AttentionMapSaver]]:
         if timesteps is None:
-            self.scheduler.set_timesteps(num_inference_steps, device=self._model_group.device_for(self.unet))
+            self.scheduler.set_timesteps(
+                num_inference_steps, device=self._model_group.device_for(self.unet)
+            )
             timesteps = self.scheduler.timesteps
-        infer_latents_from_embeddings = GeneratorToCallbackinator(self.generate_latents_from_embeddings, PipelineIntermediateState)
+        infer_latents_from_embeddings = GeneratorToCallbackinator(
+            self.generate_latents_from_embeddings, PipelineIntermediateState
+        )
         result: PipelineIntermediateState = infer_latents_from_embeddings(
-            latents, timesteps, conditioning_data,
+            latents,
+            timesteps,
+            conditioning_data,
             noise=noise,
             additional_guidance=additional_guidance,
             run_id=run_id,
-            callback=callback)
+            callback=callback,
+        )
         return result.latents, result.attention_map_saver
 
-    def generate_latents_from_embeddings(self, latents: torch.Tensor, timesteps,
-                                         conditioning_data: ConditioningData,
-                                         *,
-                                         noise: torch.Tensor,
-                                         run_id: str = None,
-                                         additional_guidance: List[Callable] = None):
+    def generate_latents_from_embeddings(
+        self,
+        latents: torch.Tensor,
+        timesteps,
+        conditioning_data: ConditioningData,
+        *,
+        noise: torch.Tensor,
+        run_id: str = None,
+        additional_guidance: List[Callable] = None,
+    ):
         self._adjust_memory_efficient_attention(latents)
         if run_id is None:
             run_id = secrets.token_urlsafe(self.ID_LENGTH)
         if additional_guidance is None:
             additional_guidance = []
         extra_conditioning_info = conditioning_data.extra
-        with self.invokeai_diffuser.custom_attention_context(extra_conditioning_info=extra_conditioning_info,
-                                                             step_count=len(self.scheduler.timesteps)
-                                                             ):
-
-            yield PipelineIntermediateState(run_id=run_id, step=-1, timestep=self.scheduler.num_train_timesteps,
-                                            latents=latents)
+        with self.invokeai_diffuser.custom_attention_context(
+            extra_conditioning_info=extra_conditioning_info,
+            step_count=len(self.scheduler.timesteps),
+        ):
+            yield PipelineIntermediateState(
+                run_id=run_id,
+                step=-1,
+                timestep=self.scheduler.num_train_timesteps,
+                latents=latents,
+            )
 
             batch_size = latents.shape[0]
-            batched_t = torch.full((batch_size,), timesteps[0],
-                                   dtype=timesteps.dtype, device=self._model_group.device_for(self.unet))
+            batched_t = torch.full(
+                (batch_size,),
+                timesteps[0],
+                dtype=timesteps.dtype,
+                device=self._model_group.device_for(self.unet),
+            )
             latents = self.scheduler.add_noise(latents, noise, batched_t)
 
             attention_map_saver: Optional[AttentionMapSaver] = None
 
             for i, t in enumerate(self.progress_bar(timesteps)):
                 batched_t.fill_(t)
-                step_output = self.step(batched_t, latents, conditioning_data,
-                                        step_index=i,
-                                        total_step_count=len(timesteps),
-                                        additional_guidance=additional_guidance)
+                step_output = self.step(
+                    batched_t,
+                    latents,
+                    conditioning_data,
+                    step_index=i,
+                    total_step_count=len(timesteps),
+                    additional_guidance=additional_guidance,
+                )
                 latents = step_output.prev_sample
 
                 latents = self.invokeai_diffuser.do_latent_postprocessing(
@@ -489,28 +576,39 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
                     latents=latents,
                     sigma=batched_t,
                     step_index=i,
-                    total_step_count=len(timesteps)
+                    total_step_count=len(timesteps),
                 )
 
-                predicted_original = getattr(step_output, 'pred_original_sample', None)
+                predicted_original = getattr(step_output, "pred_original_sample", None)
 
                 # TODO resuscitate attention map saving
-                #if i == len(timesteps)-1 and extra_conditioning_info is not None:
+                # if i == len(timesteps)-1 and extra_conditioning_info is not None:
                 #    eos_token_index = extra_conditioning_info.tokens_count_including_eos_bos - 1
                 #    attention_map_token_ids = range(1, eos_token_index)
                 #    attention_map_saver = AttentionMapSaver(token_ids=attention_map_token_ids, latents_shape=latents.shape[-2:])
                 #    self.invokeai_diffuser.setup_attention_map_saving(attention_map_saver)
 
-                yield PipelineIntermediateState(run_id=run_id, step=i, timestep=int(t), latents=latents,
-                                                predicted_original=predicted_original, attention_map_saver=attention_map_saver)
+                yield PipelineIntermediateState(
+                    run_id=run_id,
+                    step=i,
+                    timestep=int(t),
+                    latents=latents,
+                    predicted_original=predicted_original,
+                    attention_map_saver=attention_map_saver,
+                )
 
             return latents, attention_map_saver
 
     @torch.inference_mode()
-    def step(self, t: torch.Tensor, latents: torch.Tensor,
-             conditioning_data: ConditioningData,
-             step_index:int, total_step_count:int,
-             additional_guidance: List[Callable] = None):
+    def step(
+        self,
+        t: torch.Tensor,
+        latents: torch.Tensor,
+        conditioning_data: ConditioningData,
+        step_index: int,
+        total_step_count: int,
+        additional_guidance: List[Callable] = None,
+    ):
         # invokeai_diffuser has batched timesteps, but diffusers schedulers expect a single value
         timestep = t[0]
 
@@ -523,16 +621,19 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
 
         # predict the noise residual
         noise_pred = self.invokeai_diffuser.do_diffusion_step(
-            latent_model_input, t,
-            conditioning_data.unconditioned_embeddings, conditioning_data.text_embeddings,
+            latent_model_input,
+            t,
+            conditioning_data.unconditioned_embeddings,
+            conditioning_data.text_embeddings,
             conditioning_data.guidance_scale,
             step_index=step_index,
             total_step_count=total_step_count,
         )
 
         # compute the previous noisy sample x_t -> x_t-1
-        step_output = self.scheduler.step(noise_pred, timestep, latents,
-                                          **conditioning_data.scheduler_args)
+        step_output = self.scheduler.step(
+            noise_pred, timestep, latents, **conditioning_data.scheduler_args
+        )
 
         # TODO: this additional_guidance extension point feels redundant with InvokeAIDiffusionComponent.
         #    But the way things are now, scheduler runs _after_ that, so there was
@@ -542,7 +643,13 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
 
         return step_output
 
-    def _unet_forward(self, latents, t, text_embeddings, cross_attention_kwargs: Optional[dict[str,Any]] = None):
+    def _unet_forward(
+        self,
+        latents,
+        t,
+        text_embeddings,
+        cross_attention_kwargs: Optional[dict[str, Any]] = None,
+    ):
         """predict the noise residual"""
         if is_inpainting_model(self.unet) and latents.size(1) == 4:
             # Pad out normal non-inpainting inputs for an inpainting model.
@@ -551,67 +658,105 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
             #     use of AddsMaskLatents.
             latents = AddsMaskLatents(
                 self._unet_forward,
-                mask=torch.ones_like(latents[:1, :1], device=latents.device, dtype=latents.dtype),
-                initial_image_latents=torch.zeros_like(latents[:1], device=latents.device, dtype=latents.dtype)
+                mask=torch.ones_like(
+                    latents[:1, :1], device=latents.device, dtype=latents.dtype
+                ),
+                initial_image_latents=torch.zeros_like(
+                    latents[:1], device=latents.device, dtype=latents.dtype
+                ),
             ).add_mask_channels(latents)
 
         # First three args should be positional, not keywords, so torch hooks can see them.
-        return self.unet(latents, t, text_embeddings,
-                         cross_attention_kwargs=cross_attention_kwargs).sample
+        return self.unet(
+            latents, t, text_embeddings, cross_attention_kwargs=cross_attention_kwargs
+        ).sample
 
-    def img2img_from_embeddings(self,
-                                init_image: Union[torch.FloatTensor, PIL.Image.Image],
-                                strength: float,
-                                num_inference_steps: int,
-                                conditioning_data: ConditioningData,
-                                *, callback: Callable[[PipelineIntermediateState], None] = None,
-                                run_id=None,
-                                noise_func=None
-                                ) -> InvokeAIStableDiffusionPipelineOutput:
+    def img2img_from_embeddings(
+        self,
+        init_image: Union[torch.FloatTensor, PIL.Image.Image],
+        strength: float,
+        num_inference_steps: int,
+        conditioning_data: ConditioningData,
+        *,
+        callback: Callable[[PipelineIntermediateState], None] = None,
+        run_id=None,
+        noise_func=None,
+        seed=None,
+    ) -> InvokeAIStableDiffusionPipelineOutput:
         if isinstance(init_image, PIL.Image.Image):
-            init_image = image_resized_to_grid_as_tensor(init_image.convert('RGB'))
+            init_image = image_resized_to_grid_as_tensor(init_image.convert("RGB"))
 
         if init_image.dim() == 3:
-            init_image = einops.rearrange(init_image, 'c h w -> 1 c h w')
+            init_image = einops.rearrange(init_image, "c h w -> 1 c h w")
 
         # 6. Prepare latent variables
         initial_latents = self.non_noised_latents_from_image(
-            init_image, device=self._model_group.device_for(self.unet),
-            dtype=self.unet.dtype)
+            init_image,
+            device=self._model_group.device_for(self.unet),
+            dtype=self.unet.dtype,
+        )
+        if seed is not None:
+            set_seed(seed)
         noise = noise_func(initial_latents)
 
-        return self.img2img_from_latents_and_embeddings(initial_latents, num_inference_steps,
-                                                        conditioning_data,
-                                                        strength,
-                                                        noise, run_id, callback)
+        return self.img2img_from_latents_and_embeddings(
+            initial_latents,
+            num_inference_steps,
+            conditioning_data,
+            strength,
+            noise,
+            run_id,
+            callback,
+        )
 
-    def img2img_from_latents_and_embeddings(self, initial_latents, num_inference_steps,
-                                            conditioning_data: ConditioningData,
-                                            strength,
-                                            noise: torch.Tensor, run_id=None, callback=None
-                                            ) -> InvokeAIStableDiffusionPipelineOutput:
-        timesteps, _ = self.get_img2img_timesteps(num_inference_steps, strength,
-                                                  device=self._model_group.device_for(self.unet))
+    def img2img_from_latents_and_embeddings(
+        self,
+        initial_latents,
+        num_inference_steps,
+        conditioning_data: ConditioningData,
+        strength,
+        noise: torch.Tensor,
+        run_id=None,
+        callback=None,
+    ) -> InvokeAIStableDiffusionPipelineOutput:
+        timesteps, _ = self.get_img2img_timesteps(
+            num_inference_steps,
+            strength,
+            device=self._model_group.device_for(self.unet),
+        )
         result_latents, result_attention_maps = self.latents_from_embeddings(
-            initial_latents, num_inference_steps, conditioning_data,
+            latents=initial_latents if strength < 1.0 else torch.zeros_like(
+                initial_latents, device=initial_latents.device, dtype=initial_latents.dtype
+            ),
+            num_inference_steps=num_inference_steps,
+            conditioning_data=conditioning_data,
             timesteps=timesteps,
             noise=noise,
             run_id=run_id,
-            callback=callback)
+            callback=callback,
+        )
 
         # https://discuss.huggingface.co/t/memory-usage-by-later-pipeline-stages/23699
         torch.cuda.empty_cache()
 
         with torch.inference_mode():
             image = self.decode_latents(result_latents)
-            output = InvokeAIStableDiffusionPipelineOutput(images=image, nsfw_content_detected=[], attention_map_saver=result_attention_maps)
+            output = InvokeAIStableDiffusionPipelineOutput(
+                images=image,
+                nsfw_content_detected=[],
+                attention_map_saver=result_attention_maps,
+            )
             return self.check_for_safety(output, dtype=conditioning_data.dtype)
 
-    def get_img2img_timesteps(self, num_inference_steps: int, strength: float, device) -> (torch.Tensor, int):
+    def get_img2img_timesteps(
+        self, num_inference_steps: int, strength: float, device
+    ) -> (torch.Tensor, int):
         img2img_pipeline = StableDiffusionImg2ImgPipeline(**self.components)
         assert img2img_pipeline.scheduler is self.scheduler
         img2img_pipeline.scheduler.set_timesteps(num_inference_steps, device=device)
-        timesteps, adjusted_steps = img2img_pipeline.get_timesteps(num_inference_steps, strength, device=device)
+        timesteps, adjusted_steps = img2img_pipeline.get_timesteps(
+            num_inference_steps, strength, device=device
+        )
         # Workaround for low strength resulting in zero timesteps.
         # TODO: submit upstream fix for zero-step img2img
         if timesteps.numel() == 0:
@@ -620,21 +765,23 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         return timesteps, adjusted_steps
 
     def inpaint_from_embeddings(
-            self,
-            init_image: torch.FloatTensor,
-            mask: torch.FloatTensor,
-            strength: float,
-            num_inference_steps: int,
-            conditioning_data: ConditioningData,
-            *, callback: Callable[[PipelineIntermediateState], None] = None,
-            run_id=None,
-            noise_func=None,
-            ) -> InvokeAIStableDiffusionPipelineOutput:
+        self,
+        init_image: torch.FloatTensor,
+        mask: torch.FloatTensor,
+        strength: float,
+        num_inference_steps: int,
+        conditioning_data: ConditioningData,
+        *,
+        callback: Callable[[PipelineIntermediateState], None] = None,
+        run_id=None,
+        noise_func=None,
+        seed=None,
+    ) -> InvokeAIStableDiffusionPipelineOutput:
         device = self._model_group.device_for(self.unet)
         latents_dtype = self.unet.dtype
 
         if isinstance(init_image, PIL.Image.Image):
-            init_image = image_resized_to_grid_as_tensor(init_image.convert('RGB'))
+            init_image = image_resized_to_grid_as_tensor(init_image.convert("RGB"))
 
         init_image = init_image.to(device=device, dtype=latents_dtype)
         mask = mask.to(device=device, dtype=latents_dtype)
@@ -642,18 +789,25 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         if init_image.dim() == 3:
             init_image = init_image.unsqueeze(0)
 
-        timesteps, _ = self.get_img2img_timesteps(num_inference_steps, strength, device=device)
+        timesteps, _ = self.get_img2img_timesteps(
+            num_inference_steps, strength, device=device
+        )
 
         # 6. Prepare latent variables
         # can't quite use upstream StableDiffusionImg2ImgPipeline.prepare_latents
         # because we have our own noise function
-        init_image_latents = self.non_noised_latents_from_image(init_image, device=device, dtype=latents_dtype)
+        init_image_latents = self.non_noised_latents_from_image(
+            init_image, device=device, dtype=latents_dtype
+        )
+        if seed is not None:
+            set_seed(seed)
         noise = noise_func(init_image_latents)
 
         if mask.dim() == 3:
             mask = mask.unsqueeze(0)
-        latent_mask = tv_resize(mask, init_image_latents.shape[-2:], T.InterpolationMode.BILINEAR) \
-            .to(device=device, dtype=latents_dtype)
+        latent_mask = tv_resize(
+            mask, init_image_latents.shape[-2:], T.InterpolationMode.BILINEAR
+        ).to(device=device, dtype=latents_dtype)
 
         guidance: List[Callable] = []
 
@@ -661,20 +815,32 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
             # You'd think the inpainting model wouldn't be paying attention to the area it is going to repaint
             # (that's why there's a mask!) but it seems to really want that blanked out.
             masked_init_image = init_image * torch.where(mask < 0.5, 1, 0)
-            masked_latents = self.non_noised_latents_from_image(masked_init_image, device=device, dtype=latents_dtype)
+            masked_latents = self.non_noised_latents_from_image(
+                masked_init_image, device=device, dtype=latents_dtype
+            )
 
             # TODO: we should probably pass this in so we don't have to try/finally around setting it.
-            self.invokeai_diffuser.model_forward_callback = \
-                AddsMaskLatents(self._unet_forward, latent_mask, masked_latents)
+            self.invokeai_diffuser.model_forward_callback = AddsMaskLatents(
+                self._unet_forward, latent_mask, masked_latents
+            )
         else:
-            guidance.append(AddsMaskGuidance(latent_mask, init_image_latents, self.scheduler, noise))
+            guidance.append(
+                AddsMaskGuidance(latent_mask, init_image_latents, self.scheduler, noise)
+            )
 
         try:
             result_latents, result_attention_maps = self.latents_from_embeddings(
-                init_image_latents, num_inference_steps,
-                conditioning_data, noise=noise, timesteps=timesteps,
+                latents=init_image_latents if strength < 1.0 else torch.zeros_like(
+                    init_image_latents, device=init_image_latents.device, dtype=init_image_latents.dtype
+                ),
+                num_inference_steps=num_inference_steps,
+                conditioning_data=conditioning_data,
+                noise=noise,
+                timesteps=timesteps,
                 additional_guidance=guidance,
-                run_id=run_id, callback=callback)
+                run_id=run_id,
+                callback=callback,
+            )
         finally:
             self.invokeai_diffuser.model_forward_callback = self._unet_forward
 
@@ -683,13 +849,17 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
 
         with torch.inference_mode():
             image = self.decode_latents(result_latents)
-            output = InvokeAIStableDiffusionPipelineOutput(images=image, nsfw_content_detected=[], attention_map_saver=result_attention_maps)
+            output = InvokeAIStableDiffusionPipelineOutput(
+                images=image,
+                nsfw_content_detected=[],
+                attention_map_saver=result_attention_maps,
+            )
             return self.check_for_safety(output, dtype=conditioning_data.dtype)
 
     def non_noised_latents_from_image(self, init_image, *, device: torch.device, dtype):
         init_image = init_image.to(device=device, dtype=dtype)
         with torch.inference_mode():
-            if device.type == 'mps':
+            if device.type == "mps":
                 # workaround for torch MPS bug that has been fixed in https://github.com/kulinseth/pytorch/pull/222
                 # TODO remove this workaround once kulinseth#222 is merged to pytorch mainline
                 self.vae.to(CPU_DEVICE)
@@ -697,8 +867,10 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
             else:
                 self._model_group.load(self.vae)
             init_latent_dist = self.vae.encode(init_image).latent_dist
-            init_latents = init_latent_dist.sample().to(dtype=dtype)  # FIXME: uses torch.randn. make reproducible!
-            if device.type == 'mps':
+            init_latents = init_latent_dist.sample().to(
+                dtype=dtype
+            )  # FIXME: uses torch.randn. make reproducible!
+            if device.type == "mps":
                 self.vae.to(device)
                 init_latents = init_latents.to(device)
 
@@ -707,14 +879,18 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
 
     def check_for_safety(self, output, dtype):
         with torch.inference_mode():
-            screened_images, has_nsfw_concept = self.run_safety_checker(output.images, dtype=dtype)
+            screened_images, has_nsfw_concept = self.run_safety_checker(
+                output.images, dtype=dtype
+            )
         screened_attention_map_saver = None
         if has_nsfw_concept is None or not has_nsfw_concept:
             screened_attention_map_saver = output.attention_map_saver
-        return InvokeAIStableDiffusionPipelineOutput(screened_images,
-                                                     has_nsfw_concept,
-                                                     # block the attention maps if NSFW content is detected
-                                                     attention_map_saver=screened_attention_map_saver)
+        return InvokeAIStableDiffusionPipelineOutput(
+            screened_images,
+            has_nsfw_concept,
+            # block the attention maps if NSFW content is detected
+            attention_map_saver=screened_attention_map_saver,
+        )
 
     def run_safety_checker(self, image, device=None, dtype=None):
         # overriding to use the model group for device info instead of requiring the caller to know.
@@ -723,28 +899,17 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         return super().run_safety_checker(image, device, dtype)
 
     @torch.inference_mode()
-    def get_learned_conditioning(self, c: List[List[str]], *, return_tokens=True, fragment_weights=None):
+    def get_learned_conditioning(
+        self, c: List[List[str]], *, return_tokens=True, fragment_weights=None
+    ):
         """
-        Compatibility function for ldm.models.diffusion.ddpm.LatentDiffusion.
+        Compatibility function for invokeai.models.diffusion.ddpm.LatentDiffusion.
         """
         return self.embeddings_provider.get_embeddings_for_weighted_prompt_fragments(
             text_batch=c,
             fragment_weights_batch=fragment_weights,
             should_return_tokens=return_tokens,
-            device=self._model_group.device_for(self.unet))
-
-    @property
-    def cond_stage_model(self):
-        return self.embeddings_provider
-
-    @torch.inference_mode()
-    def _tokenize(self, prompt: Union[str, List[str]]):
-        return self.tokenizer(
-            prompt,
-            padding="max_length",
-            max_length=self.tokenizer.model_max_length,
-            truncation=True,
-            return_tensors="pt",
+            device=self._model_group.device_for(self.unet),
         )
 
     @property
@@ -758,8 +923,10 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         return super().decode_latents(latents)
 
     def debug_latents(self, latents, msg):
+        from invokeai.backend.image_util import debug_image
         with torch.inference_mode():
-            from ldm.util import debug_image
             decoded = self.numpy_to_pil(self.decode_latents(latents))
-            for i, img in enumerate(decoded):
-                debug_image(img, f"latents {msg} {i+1}/{len(decoded)}", debug_status=True)
+        for i, img in enumerate(decoded):
+            debug_image(
+                img, f"latents {msg} {i+1}/{len(decoded)}", debug_status=True
+            )

invokeai/backend/stable_diffusion/diffusion/__init__.py

@@ -0,0 +1,6 @@
+"""
+Initialization file for invokeai.models.diffusion
+"""
+from .cross_attention_control import InvokeAICrossAttentionMixin
+from .cross_attention_map_saving import AttentionMapSaver
+from .shared_invokeai_diffusion import InvokeAIDiffuserComponent, PostprocessingSettings

invokeai/backend/stable_diffusion/diffusion/cross_attention_control.py

@@ -1,21 +1,20 @@
-
 # adapted from bloc97's CrossAttentionControl colab
 # https://github.com/bloc97/CrossAttentionControl
 
 
 import enum
 import math
-from typing import Optional, Callable
+from typing import Callable, Optional
 
+import diffusers
 import psutil
 import torch
-import diffusers
+from compel.cross_attention_control import Arguments
+from diffusers.models.cross_attention import AttnProcessor
+from diffusers.models.unet_2d_condition import UNet2DConditionModel
 from torch import nn
 
-from compel.cross_attention_control import Arguments
-from diffusers.models.unet_2d_condition import UNet2DConditionModel
-from diffusers.models.cross_attention import AttnProcessor
-from ldm.invoke.devices import torch_dtype
+from ...util import torch_dtype
 
 
 class CrossAttentionType(enum.Enum):
@@ -24,13 +23,12 @@ class CrossAttentionType(enum.Enum):
 
 
 class Context:
-
     cross_attention_mask: Optional[torch.Tensor]
     cross_attention_index_map: Optional[torch.Tensor]
 
     class Action(enum.Enum):
         NONE = 0
-        SAVE = 1,
+        SAVE = (1,)
         APPLY = 2
 
     def __init__(self, arguments: Arguments, step_count: int):
@@ -53,11 +51,13 @@ class Context:
         self.clear_requests(cleanup=True)
 
     def register_cross_attention_modules(self, model):
-        for name,module in get_cross_attention_modules(model, CrossAttentionType.SELF):
+        for name, module in get_cross_attention_modules(model, CrossAttentionType.SELF):
             if name in self.self_cross_attention_module_identifiers:
                 assert False, f"name {name} cannot appear more than once"
             self.self_cross_attention_module_identifiers.append(name)
-        for name,module in get_cross_attention_modules(model, CrossAttentionType.TOKENS):
+        for name, module in get_cross_attention_modules(
+            model, CrossAttentionType.TOKENS
+        ):
             if name in self.tokens_cross_attention_module_identifiers:
                 assert False, f"name {name} cannot appear more than once"
             self.tokens_cross_attention_module_identifiers.append(name)
@@ -68,7 +68,9 @@ class Context:
         else:
             self.tokens_cross_attention_action = Context.Action.SAVE
 
-    def request_apply_saved_attention_maps(self, cross_attention_type: CrossAttentionType):
+    def request_apply_saved_attention_maps(
+        self, cross_attention_type: CrossAttentionType
+    ):
         if cross_attention_type == CrossAttentionType.SELF:
             self.self_cross_attention_action = Context.Action.APPLY
         else:
@@ -91,8 +93,9 @@ class Context:
             return self.tokens_cross_attention_action == Context.Action.APPLY
         return False
 
-    def get_active_cross_attention_control_types_for_step(self, percent_through:float=None)\
-            -> list[CrossAttentionType]:
+    def get_active_cross_attention_control_types_for_step(
+        self, percent_through: float = None
+    ) -> list[CrossAttentionType]:
         """
         Should cross-attention control be applied on the given step?
         :param percent_through: How far through the step sequence are we (0.0=pure noise, 1.0=completely denoised image). Expected range 0.0..<1.0.
@@ -103,50 +106,73 @@ class Context:
 
         opts = self.arguments.edit_options
         to_control = []
-        if opts['s_start'] <= percent_through < opts['s_end']:
+        if opts["s_start"] <= percent_through < opts["s_end"]:
             to_control.append(CrossAttentionType.SELF)
-        if opts['t_start'] <= percent_through < opts['t_end']:
+        if opts["t_start"] <= percent_through < opts["t_end"]:
             to_control.append(CrossAttentionType.TOKENS)
         return to_control
 
-    def save_slice(self, identifier: str, slice: torch.Tensor, dim: Optional[int], offset: int,
-                   slice_size: Optional[int]):
+    def save_slice(
+        self,
+        identifier: str,
+        slice: torch.Tensor,
+        dim: Optional[int],
+        offset: int,
+        slice_size: Optional[int],
+    ):
         if identifier not in self.saved_cross_attention_maps:
             self.saved_cross_attention_maps[identifier] = {
-                'dim': dim,
-                'slice_size': slice_size,
-                'slices': {offset or 0: slice}
+                "dim": dim,
+                "slice_size": slice_size,
+                "slices": {offset or 0: slice},
             }
         else:
-            self.saved_cross_attention_maps[identifier]['slices'][offset or 0] = slice
+            self.saved_cross_attention_maps[identifier]["slices"][offset or 0] = slice
 
-    def get_slice(self, identifier: str, requested_dim: Optional[int], requested_offset: int, slice_size: int):
+    def get_slice(
+        self,
+        identifier: str,
+        requested_dim: Optional[int],
+        requested_offset: int,
+        slice_size: int,
+    ):
         saved_attention_dict = self.saved_cross_attention_maps[identifier]
         if requested_dim is None:
-            if saved_attention_dict['dim'] is not None:
-                raise RuntimeError(f"dim mismatch: expected dim=None, have {saved_attention_dict['dim']}")
-            return saved_attention_dict['slices'][0]
-
-        if saved_attention_dict['dim'] == requested_dim:
-            if slice_size != saved_attention_dict['slice_size']:
+            if saved_attention_dict["dim"] is not None:
                 raise RuntimeError(
-                    f"slice_size mismatch: expected slice_size={slice_size}, have {saved_attention_dict['slice_size']}")
-            return saved_attention_dict['slices'][requested_offset]
+                    f"dim mismatch: expected dim=None, have {saved_attention_dict['dim']}"
+                )
+            return saved_attention_dict["slices"][0]
 
-        if saved_attention_dict['dim'] is None:
-            whole_saved_attention = saved_attention_dict['slices'][0]
+        if saved_attention_dict["dim"] == requested_dim:
+            if slice_size != saved_attention_dict["slice_size"]:
+                raise RuntimeError(
+                    f"slice_size mismatch: expected slice_size={slice_size}, have {saved_attention_dict['slice_size']}"
+                )
+            return saved_attention_dict["slices"][requested_offset]
+
+        if saved_attention_dict["dim"] is None:
+            whole_saved_attention = saved_attention_dict["slices"][0]
             if requested_dim == 0:
-                return whole_saved_attention[requested_offset:requested_offset + slice_size]
+                return whole_saved_attention[
+                    requested_offset : requested_offset + slice_size
+                ]
             elif requested_dim == 1:
-                return whole_saved_attention[:, requested_offset:requested_offset + slice_size]
+                return whole_saved_attention[
+                    :, requested_offset : requested_offset + slice_size
+                ]
 
-        raise RuntimeError(f"Cannot convert dim {saved_attention_dict['dim']} to requested dim {requested_dim}")
+        raise RuntimeError(
+            f"Cannot convert dim {saved_attention_dict['dim']} to requested dim {requested_dim}"
+        )
 
-    def get_slicing_strategy(self, identifier: str) -> tuple[Optional[int], Optional[int]]:
+    def get_slicing_strategy(
+        self, identifier: str
+    ) -> tuple[Optional[int], Optional[int]]:
         saved_attention = self.saved_cross_attention_maps.get(identifier, None)
         if saved_attention is None:
             return None, None
-        return saved_attention['dim'], saved_attention['slice_size']
+        return saved_attention["dim"], saved_attention["slice_size"]
 
     def clear_requests(self, cleanup=True):
         self.tokens_cross_attention_action = Context.Action.NONE
@@ -156,9 +182,8 @@ class Context:
 
     def offload_saved_attention_slices_to_cpu(self):
         for key, map_dict in self.saved_cross_attention_maps.items():
-            for offset, slice in map_dict['slices'].items():
-                map_dict[offset] = slice.to('cpu')
-
+            for offset, slice in map_dict["slices"].items():
+                map_dict[offset] = slice.to("cpu")
 
 
 class InvokeAICrossAttentionMixin:
@@ -167,14 +192,20 @@ class InvokeAICrossAttentionMixin:
     through both to an attention_slice_wrangler and a slicing_strategy_getter for custom attention map wrangling
     and dymamic slicing strategy selection.
     """
+
     def __init__(self):
         self.mem_total_gb = psutil.virtual_memory().total // (1 << 30)
         self.attention_slice_wrangler = None
         self.slicing_strategy_getter = None
         self.attention_slice_calculated_callback = None
 
-    def set_attention_slice_wrangler(self, wrangler: Optional[Callable[[nn.Module, torch.Tensor, int, int, int], torch.Tensor]]):
-        '''
+    def set_attention_slice_wrangler(
+        self,
+        wrangler: Optional[
+            Callable[[nn.Module, torch.Tensor, int, int, int], torch.Tensor]
+        ],
+    ):
+        """
         Set custom attention calculator to be called when attention is calculated
         :param wrangler: Callback, with args (module, suggested_attention_slice, dim, offset, slice_size),
         which returns either the suggested_attention_slice or an adjusted equivalent.
@@ -185,20 +216,30 @@ class InvokeAICrossAttentionMixin:
 
         Pass None to use the default attention calculation.
         :return:
-        '''
+        """
         self.attention_slice_wrangler = wrangler
 
-    def set_slicing_strategy_getter(self, getter: Optional[Callable[[nn.Module], tuple[int,int]]]):
+    def set_slicing_strategy_getter(
+        self, getter: Optional[Callable[[nn.Module], tuple[int, int]]]
+    ):
         self.slicing_strategy_getter = getter
 
-    def set_attention_slice_calculated_callback(self, callback: Optional[Callable[[torch.Tensor], None]]):
+    def set_attention_slice_calculated_callback(
+        self, callback: Optional[Callable[[torch.Tensor], None]]
+    ):
         self.attention_slice_calculated_callback = callback
 
     def einsum_lowest_level(self, query, key, value, dim, offset, slice_size):
         # calculate attention scores
-        #attention_scores = torch.einsum('b i d, b j d -> b i j', q, k)
+        # attention_scores = torch.einsum('b i d, b j d -> b i j', q, k)
         attention_scores = torch.baddbmm(
-            torch.empty(query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device),
+            torch.empty(
+                query.shape[0],
+                query.shape[1],
+                key.shape[1],
+                dtype=query.dtype,
+                device=query.device,
+            ),
             query,
             key.transpose(-1, -2),
             beta=0,
@@ -206,35 +247,49 @@ class InvokeAICrossAttentionMixin:
         )
 
         # calculate attention slice by taking the best scores for each latent pixel
-        default_attention_slice = attention_scores.softmax(dim=-1, dtype=attention_scores.dtype)
+        default_attention_slice = attention_scores.softmax(
+            dim=-1, dtype=attention_scores.dtype
+        )
         attention_slice_wrangler = self.attention_slice_wrangler
         if attention_slice_wrangler is not None:
-            attention_slice = attention_slice_wrangler(self, default_attention_slice, dim, offset, slice_size)
+            attention_slice = attention_slice_wrangler(
+                self, default_attention_slice, dim, offset, slice_size
+            )
         else:
             attention_slice = default_attention_slice
 
         if self.attention_slice_calculated_callback is not None:
-            self.attention_slice_calculated_callback(attention_slice, dim, offset, slice_size)
+            self.attention_slice_calculated_callback(
+                attention_slice, dim, offset, slice_size
+            )
 
         hidden_states = torch.bmm(attention_slice, value)
         return hidden_states
 
     def einsum_op_slice_dim0(self, q, k, v, slice_size):
-        r = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
+        r = torch.zeros(
+            q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype
+        )
         for i in range(0, q.shape[0], slice_size):
             end = i + slice_size
-            r[i:end] = self.einsum_lowest_level(q[i:end], k[i:end], v[i:end], dim=0, offset=i, slice_size=slice_size)
+            r[i:end] = self.einsum_lowest_level(
+                q[i:end], k[i:end], v[i:end], dim=0, offset=i, slice_size=slice_size
+            )
         return r
 
     def einsum_op_slice_dim1(self, q, k, v, slice_size):
-        r = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
+        r = torch.zeros(
+            q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype
+        )
         for i in range(0, q.shape[1], slice_size):
             end = i + slice_size
-            r[:, i:end] = self.einsum_lowest_level(q[:, i:end], k, v, dim=1, offset=i, slice_size=slice_size)
+            r[:, i:end] = self.einsum_lowest_level(
+                q[:, i:end], k, v, dim=1, offset=i, slice_size=slice_size
+            )
         return r
 
     def einsum_op_mps_v1(self, q, k, v):
-        if q.shape[1] <= 4096: # (512x512) max q.shape[1]: 4096
+        if q.shape[1] <= 4096:  # (512x512) max q.shape[1]: 4096
             return self.einsum_lowest_level(q, k, v, None, None, None)
         else:
             slice_size = math.floor(2**30 / (q.shape[0] * q.shape[1]))
@@ -272,13 +327,12 @@ class InvokeAICrossAttentionMixin:
         # Divide factor of safety as there's copying and fragmentation
         return self.einsum_op_tensor_mem(q, k, v, mem_free_total / 3.3 / (1 << 20))
 
-
     def get_invokeai_attention_mem_efficient(self, q, k, v):
-        if q.device.type == 'cuda':
-            #print("in get_attention_mem_efficient with q shape", q.shape, ", k shape", k.shape, ", free memory is", get_mem_free_total(q.device))
+        if q.device.type == "cuda":
+            # print("in get_attention_mem_efficient with q shape", q.shape, ", k shape", k.shape, ", free memory is", get_mem_free_total(q.device))
             return self.einsum_op_cuda(q, k, v)
 
-        if q.device.type == 'mps' or q.device.type == 'cpu':
+        if q.device.type == "mps" or q.device.type == "cpu":
             if self.mem_total_gb >= 32:
                 return self.einsum_op_mps_v1(q, k, v)
             return self.einsum_op_mps_v2(q, k, v)
@@ -288,8 +342,11 @@ class InvokeAICrossAttentionMixin:
         return self.einsum_op_tensor_mem(q, k, v, 32)
 
 
-
-def restore_default_cross_attention(model, is_running_diffusers: bool, restore_attention_processor: Optional[AttnProcessor]=None):
+def restore_default_cross_attention(
+    model,
+    is_running_diffusers: bool,
+    restore_attention_processor: Optional[AttnProcessor] = None,
+):
     if is_running_diffusers:
         unet = model
         unet.set_attn_processor(restore_attention_processor or CrossAttnProcessor())
@@ -297,7 +354,7 @@ def restore_default_cross_attention(model, is_running_diffusers: bool, restore_a
         remove_attention_function(model)
 
 
-def override_cross_attention(model, context: Context, is_running_diffusers = False):
+def override_cross_attention(model, context: Context, is_running_diffusers=False):
     """
     Inject attention parameters and functions into the passed in model to enable cross attention editing.
 
@@ -316,7 +373,7 @@ def override_cross_attention(model, context: Context, is_running_diffusers = Fal
     indices = torch.arange(max_length, dtype=torch.long)
     for name, a0, a1, b0, b1 in context.arguments.edit_opcodes:
         if b0 < max_length:
-            if name == "equal":# or (name == "replace" and a1 - a0 == b1 - b0):
+            if name == "equal":  # or (name == "replace" and a1 - a0 == b1 - b0):
                 # these tokens have not been edited
                 indices[b0:b1] = indices_target[a0:a1]
                 mask[b0:b1] = 1
@@ -332,7 +389,14 @@ def override_cross_attention(model, context: Context, is_running_diffusers = Fal
         else:
             # try to re-use an existing slice size
             default_slice_size = 4
-            slice_size = next((p.slice_size for p in old_attn_processors.values() if type(p) is SlicedAttnProcessor), default_slice_size)
+            slice_size = next(
+                (
+                    p.slice_size
+                    for p in old_attn_processors.values()
+                    if type(p) is SlicedAttnProcessor
+                ),
+                default_slice_size,
+            )
             unet.set_attn_processor(SlicedSwapCrossAttnProcesser(slice_size=slice_size))
         return old_attn_processors
     else:
@@ -341,65 +405,96 @@ def override_cross_attention(model, context: Context, is_running_diffusers = Fal
         return None
 
 
+def get_cross_attention_modules(
+    model, which: CrossAttentionType
+) -> list[tuple[str, InvokeAICrossAttentionMixin]]:
+    from ldm.modules.attention import CrossAttention  # avoid circular import
 
-
-def get_cross_attention_modules(model, which: CrossAttentionType) -> list[tuple[str, InvokeAICrossAttentionMixin]]:
-    from ldm.modules.attention import CrossAttention # avoid circular import
-    cross_attention_class: type = InvokeAIDiffusersCrossAttention if isinstance(model,UNet2DConditionModel) else CrossAttention
+    cross_attention_class: type = (
+        InvokeAIDiffusersCrossAttention
+        if isinstance(model, UNet2DConditionModel)
+        else CrossAttention
+    )
     which_attn = "attn1" if which is CrossAttentionType.SELF else "attn2"
-    attention_module_tuples = [(name,module) for name, module in model.named_modules() if
-                isinstance(module, cross_attention_class) and which_attn in name]
+    attention_module_tuples = [
+        (name, module)
+        for name, module in model.named_modules()
+        if isinstance(module, cross_attention_class) and which_attn in name
+    ]
     cross_attention_modules_in_model_count = len(attention_module_tuples)
     expected_count = 16
     if cross_attention_modules_in_model_count != expected_count:
         # non-fatal error but .swap() won't work.
-        print(f"Error! CrossAttentionControl found an unexpected number of {cross_attention_class} modules in the model " +
-              f"(expected {expected_count}, found {cross_attention_modules_in_model_count}). Either monkey-patching failed " +
-              f"or some assumption has changed about the structure of the model itself. Please fix the monkey-patching, " +
-              f"and/or update the {expected_count} above to an appropriate number, and/or find and inform someone who knows " +
-              f"what it means. This error is non-fatal, but it is likely that .swap() and attention map display will not " +
-              f"work properly until it is fixed.")
+        print(
+            f"Error! CrossAttentionControl found an unexpected number of {cross_attention_class} modules in the model "
+            + f"(expected {expected_count}, found {cross_attention_modules_in_model_count}). Either monkey-patching failed "
+            + f"or some assumption has changed about the structure of the model itself. Please fix the monkey-patching, "
+            + f"and/or update the {expected_count} above to an appropriate number, and/or find and inform someone who knows "
+            + f"what it means. This error is non-fatal, but it is likely that .swap() and attention map display will not "
+            + f"work properly until it is fixed."
+        )
     return attention_module_tuples
 
 
 def inject_attention_function(unet, context: Context):
     # ORIGINAL SOURCE CODE: https://github.com/huggingface/diffusers/blob/91ddd2a25b848df0fa1262d4f1cd98c7ccb87750/src/diffusers/models/attention.py#L276
 
-    def attention_slice_wrangler(module, suggested_attention_slice:torch.Tensor, dim, offset, slice_size):
-
-        #memory_usage = suggested_attention_slice.element_size() * suggested_attention_slice.nelement()
+    def attention_slice_wrangler(
+        module, suggested_attention_slice: torch.Tensor, dim, offset, slice_size
+    ):
+        # memory_usage = suggested_attention_slice.element_size() * suggested_attention_slice.nelement()
 
         attention_slice = suggested_attention_slice
 
         if context.get_should_save_maps(module.identifier):
-            #print(module.identifier, "saving suggested_attention_slice of shape",
+            # print(module.identifier, "saving suggested_attention_slice of shape",
             #      suggested_attention_slice.shape, "dim", dim, "offset", offset)
-            slice_to_save = attention_slice.to('cpu') if dim is not None else attention_slice
-            context.save_slice(module.identifier, slice_to_save, dim=dim, offset=offset, slice_size=slice_size)
+            slice_to_save = (
+                attention_slice.to("cpu") if dim is not None else attention_slice
+            )
+            context.save_slice(
+                module.identifier,
+                slice_to_save,
+                dim=dim,
+                offset=offset,
+                slice_size=slice_size,
+            )
         elif context.get_should_apply_saved_maps(module.identifier):
-            #print(module.identifier, "applying saved attention slice for dim", dim, "offset", offset)
-            saved_attention_slice = context.get_slice(module.identifier, dim, offset, slice_size)
+            # print(module.identifier, "applying saved attention slice for dim", dim, "offset", offset)
+            saved_attention_slice = context.get_slice(
+                module.identifier, dim, offset, slice_size
+            )
 
             # slice may have been offloaded to CPU
-            saved_attention_slice = saved_attention_slice.to(suggested_attention_slice.device)
+            saved_attention_slice = saved_attention_slice.to(
+                suggested_attention_slice.device
+            )
 
             if context.is_tokens_cross_attention(module.identifier):
                 index_map = context.cross_attention_index_map
-                remapped_saved_attention_slice = torch.index_select(saved_attention_slice, -1, index_map)
+                remapped_saved_attention_slice = torch.index_select(
+                    saved_attention_slice, -1, index_map
+                )
                 this_attention_slice = suggested_attention_slice
 
-                mask = context.cross_attention_mask.to(torch_dtype(suggested_attention_slice.device))
+                mask = context.cross_attention_mask.to(
+                    torch_dtype(suggested_attention_slice.device)
+                )
                 saved_mask = mask
                 this_mask = 1 - mask
-                attention_slice = remapped_saved_attention_slice * saved_mask + \
-                                  this_attention_slice * this_mask
+                attention_slice = (
+                    remapped_saved_attention_slice * saved_mask
+                    + this_attention_slice * this_mask
+                )
             else:
                 # just use everything
                 attention_slice = saved_attention_slice
 
         return attention_slice
 
-    cross_attention_modules = get_cross_attention_modules(unet, CrossAttentionType.TOKENS) + get_cross_attention_modules(unet, CrossAttentionType.SELF)
+    cross_attention_modules = get_cross_attention_modules(
+        unet, CrossAttentionType.TOKENS
+    ) + get_cross_attention_modules(unet, CrossAttentionType.SELF)
     for identifier, module in cross_attention_modules:
         module.identifier = identifier
         try:
@@ -408,56 +503,61 @@ def inject_attention_function(unet, context: Context):
                 lambda module: context.get_slicing_strategy(identifier)
             )
         except AttributeError as e:
-            if is_attribute_error_about(e, 'set_attention_slice_wrangler'):
-                print(f"TODO: implement set_attention_slice_wrangler for {type(module)}")  # TODO
+            if is_attribute_error_about(e, "set_attention_slice_wrangler"):
+                print(
+                    f"TODO: implement set_attention_slice_wrangler for {type(module)}"
+                )  # TODO
             else:
                 raise
 
 
 def remove_attention_function(unet):
-    cross_attention_modules = get_cross_attention_modules(unet, CrossAttentionType.TOKENS) + get_cross_attention_modules(unet, CrossAttentionType.SELF)
+    cross_attention_modules = get_cross_attention_modules(
+        unet, CrossAttentionType.TOKENS
+    ) + get_cross_attention_modules(unet, CrossAttentionType.SELF)
     for identifier, module in cross_attention_modules:
         try:
             # clear wrangler callback
             module.set_attention_slice_wrangler(None)
             module.set_slicing_strategy_getter(None)
         except AttributeError as e:
-            if is_attribute_error_about(e, 'set_attention_slice_wrangler'):
-                print(f"TODO: implement set_attention_slice_wrangler for {type(module)}")
+            if is_attribute_error_about(e, "set_attention_slice_wrangler"):
+                print(
+                    f"TODO: implement set_attention_slice_wrangler for {type(module)}"
+                )
             else:
                 raise
 
 
 def is_attribute_error_about(error: AttributeError, attribute: str):
-    if hasattr(error, 'name'):  # Python 3.10
+    if hasattr(error, "name"):  # Python 3.10
         return error.name == attribute
     else:  # Python 3.9
         return attribute in str(error)
 
 
-
 def get_mem_free_total(device):
-    #only on cuda
+    # only on cuda
     if not torch.cuda.is_available():
         return None
     stats = torch.cuda.memory_stats(device)
-    mem_active = stats['active_bytes.all.current']
-    mem_reserved = stats['reserved_bytes.all.current']
+    mem_active = stats["active_bytes.all.current"]
+    mem_reserved = stats["reserved_bytes.all.current"]
     mem_free_cuda, _ = torch.cuda.mem_get_info(device)
     mem_free_torch = mem_reserved - mem_active
     mem_free_total = mem_free_cuda + mem_free_torch
     return mem_free_total
 
 
-
-class InvokeAIDiffusersCrossAttention(diffusers.models.attention.CrossAttention, InvokeAICrossAttentionMixin):
-
+class InvokeAIDiffusersCrossAttention(
+    diffusers.models.attention.CrossAttention, InvokeAICrossAttentionMixin
+):
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
         InvokeAICrossAttentionMixin.__init__(self)
 
     def _attention(self, query, key, value, attention_mask=None):
-        #default_result = super()._attention(query,  key, value)
+        # default_result = super()._attention(query,  key, value)
         if attention_mask is not None:
             print(f"{type(self).__name__} ignoring passed-in attention_mask")
         attention_result = self.get_invokeai_attention_mem_efficient(query, key, value)
@@ -466,9 +566,6 @@ class InvokeAIDiffusersCrossAttention(diffusers.models.attention.CrossAttention,
         return hidden_states
 
 
-
-
-
 ## 🧨diffusers implementation follows
 
 
@@ -501,25 +598,30 @@ class CrossAttnProcessor:
         return hidden_states
 
 """
-from dataclasses import field, dataclass
+from dataclasses import dataclass, field
 
 import torch
-
-from diffusers.models.cross_attention import CrossAttention, CrossAttnProcessor, SlicedAttnProcessor
+from diffusers.models.cross_attention import (
+    CrossAttention,
+    CrossAttnProcessor,
+    SlicedAttnProcessor,
+)
 
 
 @dataclass
 class SwapCrossAttnContext:
     modified_text_embeddings: torch.Tensor
-    index_map: torch.Tensor # maps from original prompt token indices to the equivalent tokens in the modified prompt
-    mask: torch.Tensor # in the target space of the index_map
+    index_map: torch.Tensor  # maps from original prompt token indices to the equivalent tokens in the modified prompt
+    mask: torch.Tensor  # in the target space of the index_map
     cross_attention_types_to_do: list[CrossAttentionType] = field(default_factory=list)
 
-    def __int__(self,
-                cac_types_to_do: [CrossAttentionType],
-                modified_text_embeddings: torch.Tensor,
-                index_map: torch.Tensor,
-                mask: torch.Tensor):
+    def __int__(
+        self,
+        cac_types_to_do: [CrossAttentionType],
+        modified_text_embeddings: torch.Tensor,
+        index_map: torch.Tensor,
+        mask: torch.Tensor,
+    ):
         self.cross_attention_types_to_do = cac_types_to_do
         self.modified_text_embeddings = modified_text_embeddings
         self.index_map = index_map
@@ -529,9 +631,9 @@ class SwapCrossAttnContext:
         return attn_type in self.cross_attention_types_to_do
 
     @classmethod
-    def make_mask_and_index_map(cls, edit_opcodes: list[tuple[str, int, int, int, int]], max_length: int) \
-            -> tuple[torch.Tensor, torch.Tensor]:
-
+    def make_mask_and_index_map(
+        cls, edit_opcodes: list[tuple[str, int, int, int, int]], max_length: int
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         # mask=1 means use original prompt attention, mask=0 means use modified prompt attention
         mask = torch.zeros(max_length)
         indices_target = torch.arange(max_length, dtype=torch.long)
@@ -547,28 +649,42 @@ class SwapCrossAttnContext:
 
 
 class SlicedSwapCrossAttnProcesser(SlicedAttnProcessor):
-
     # TODO: dynamically pick slice size based on memory conditions
 
-    def __call__(self, attn: CrossAttention, hidden_states, encoder_hidden_states=None, attention_mask=None,
-                 # kwargs
-                 swap_cross_attn_context: SwapCrossAttnContext=None):
-
-        attention_type = CrossAttentionType.SELF if encoder_hidden_states is None else CrossAttentionType.TOKENS
+    def __call__(
+        self,
+        attn: CrossAttention,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        # kwargs
+        swap_cross_attn_context: SwapCrossAttnContext = None,
+    ):
+        attention_type = (
+            CrossAttentionType.SELF
+            if encoder_hidden_states is None
+            else CrossAttentionType.TOKENS
+        )
 
         # if cross-attention control is not in play, just call through to the base implementation.
-        if attention_type is CrossAttentionType.SELF or \
-            swap_cross_attn_context is None or \
-            not swap_cross_attn_context.wants_cross_attention_control(attention_type):
-            #print(f"SwapCrossAttnContext for {attention_type} not active - passing request to superclass")
-            return super().__call__(attn, hidden_states, encoder_hidden_states, attention_mask)
-        #else:
+        if (
+            attention_type is CrossAttentionType.SELF
+            or swap_cross_attn_context is None
+            or not swap_cross_attn_context.wants_cross_attention_control(attention_type)
+        ):
+            # print(f"SwapCrossAttnContext for {attention_type} not active - passing request to superclass")
+            return super().__call__(
+                attn, hidden_states, encoder_hidden_states, attention_mask
+            )
+        # else:
         #    print(f"SwapCrossAttnContext for {attention_type} active")
 
         batch_size, sequence_length, _ = hidden_states.shape
         attention_mask = attn.prepare_attention_mask(
-            attention_mask=attention_mask, target_length=sequence_length,
-            batch_size=batch_size)
+            attention_mask=attention_mask,
+            target_length=sequence_length,
+            batch_size=batch_size,
+        )
 
         query = attn.to_q(hidden_states)
         dim = query.shape[-1]
@@ -589,41 +705,51 @@ class SlicedSwapCrossAttnProcesser(SlicedAttnProcessor):
         # compute slices and prepare output tensor
         batch_size_attention = query.shape[0]
         hidden_states = torch.zeros(
-            (batch_size_attention, sequence_length, dim // attn.heads), device=query.device, dtype=query.dtype
+            (batch_size_attention, sequence_length, dim // attn.heads),
+            device=query.device,
+            dtype=query.dtype,
         )
 
         # do slices
-        for i in range(max(1,hidden_states.shape[0] // self.slice_size)):
+        for i in range(max(1, hidden_states.shape[0] // self.slice_size)):
             start_idx = i * self.slice_size
             end_idx = (i + 1) * self.slice_size
 
             query_slice = query[start_idx:end_idx]
             original_key_slice = original_text_key[start_idx:end_idx]
             modified_key_slice = modified_text_key[start_idx:end_idx]
-            attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None
+            attn_mask_slice = (
+                attention_mask[start_idx:end_idx]
+                if attention_mask is not None
+                else None
+            )
 
-            original_attn_slice = attn.get_attention_scores(query_slice, original_key_slice, attn_mask_slice)
-            modified_attn_slice = attn.get_attention_scores(query_slice, modified_key_slice, attn_mask_slice)
+            original_attn_slice = attn.get_attention_scores(
+                query_slice, original_key_slice, attn_mask_slice
+            )
+            modified_attn_slice = attn.get_attention_scores(
+                query_slice, modified_key_slice, attn_mask_slice
+            )
 
             # because the prompt modifications may result in token sequences shifted forwards or backwards,
             # the original attention probabilities must be remapped to account for token index changes in the
             # modified prompt
-            remapped_original_attn_slice = torch.index_select(original_attn_slice, -1,
-                                                              swap_cross_attn_context.index_map)
+            remapped_original_attn_slice = torch.index_select(
+                original_attn_slice, -1, swap_cross_attn_context.index_map
+            )
 
             # only some tokens taken from the original attention probabilities. this is controlled by the mask.
             mask = swap_cross_attn_context.mask
             inverse_mask = 1 - mask
-            attn_slice = \
-                remapped_original_attn_slice * mask + \
-                modified_attn_slice * inverse_mask
+            attn_slice = (
+                remapped_original_attn_slice * mask + modified_attn_slice * inverse_mask
+            )
 
             del remapped_original_attn_slice, modified_attn_slice
 
             attn_slice = torch.bmm(attn_slice, modified_value[start_idx:end_idx])
             hidden_states[start_idx:end_idx] = attn_slice
 
-
         # done
         hidden_states = attn.batch_to_head_dim(hidden_states)
 
@@ -636,7 +762,7 @@ class SlicedSwapCrossAttnProcesser(SlicedAttnProcessor):
 
 
 class SwapCrossAttnProcessor(SlicedSwapCrossAttnProcesser):
-
     def __init__(self):
-        super(SwapCrossAttnProcessor, self).__init__(slice_size=int(1e9)) # massive slice size = don't slice
-
+        super(SwapCrossAttnProcessor, self).__init__(
+            slice_size=int(1e9)
+        )  # massive slice size = don't slice

invokeai/backend/stable_diffusion/diffusion/cross_attention_map_saving.py

@@ -2,17 +2,17 @@ import math
 
 import PIL
 import torch
-from torchvision.transforms.functional import resize as tv_resize, InterpolationMode
+from torchvision.transforms.functional import InterpolationMode
+from torchvision.transforms.functional import resize as tv_resize
 
-from ldm.models.diffusion.cross_attention_control import get_cross_attention_modules, CrossAttentionType
+from .cross_attention_control import CrossAttentionType, get_cross_attention_modules
 
 
-class AttentionMapSaver():
-
+class AttentionMapSaver:
     def __init__(self, token_ids: range, latents_shape: torch.Size):
         self.token_ids = token_ids
         self.latents_shape = latents_shape
-        #self.collated_maps = #torch.zeros([len(token_ids), latents_shape[0], latents_shape[1]])
+        # self.collated_maps = #torch.zeros([len(token_ids), latents_shape[0], latents_shape[1]])
         self.collated_maps = {}
 
     def clear_maps(self):
@@ -25,7 +25,7 @@ class AttentionMapSaver():
         :param key: Storage key. If a map already exists for this key it will be summed with the incoming data. In this case the maps sizes (H and W) should match.
         :return: None
         """
-        key_and_size = f'{key}_{maps.shape[1]}'
+        key_and_size = f"{key}_{maps.shape[1]}"
 
         # extract desired tokens
         maps = maps[:, :, self.token_ids]
@@ -35,12 +35,12 @@ class AttentionMapSaver():
 
         # store
         if key_and_size not in self.collated_maps:
-            self.collated_maps[key_and_size] = torch.zeros_like(maps, device='cpu')
+            self.collated_maps[key_and_size] = torch.zeros_like(maps, device="cpu")
         self.collated_maps[key_and_size] += maps.cpu()
 
     def write_maps_to_disk(self, path: str):
         pil_image = self.get_stacked_maps_image()
-        pil_image.save(path, 'PNG')
+        pil_image.save(path, "PNG")
 
     def get_stacked_maps_image(self) -> PIL.Image:
         """
@@ -57,39 +57,50 @@ class AttentionMapSaver():
         merged = None
 
         for key, maps in self.collated_maps.items():
-
             # maps has shape [(H*W), N] for N tokens
             # but we want [N, H, W]
-            this_scale_factor = math.sqrt(maps.shape[0] / (latents_width * latents_height))
+            this_scale_factor = math.sqrt(
+                maps.shape[0] / (latents_width * latents_height)
+            )
             this_maps_height = int(float(latents_height) * this_scale_factor)
             this_maps_width = int(float(latents_width) * this_scale_factor)
             # and we need to do some dimension juggling
-            maps = torch.reshape(torch.swapdims(maps, 0, 1), [num_tokens, this_maps_height, this_maps_width])
+            maps = torch.reshape(
+                torch.swapdims(maps, 0, 1),
+                [num_tokens, this_maps_height, this_maps_width],
+            )
 
             # scale to output size if necessary
             if this_scale_factor != 1:
-                maps = tv_resize(maps, [latents_height, latents_width], InterpolationMode.BICUBIC)
+                maps = tv_resize(
+                    maps, [latents_height, latents_width], InterpolationMode.BICUBIC
+                )
 
             # normalize
             maps_min = torch.min(maps)
             maps_range = torch.max(maps) - maps_min
-            #print(f"map {key} size {[this_maps_width, this_maps_height]} range {[maps_min, maps_min + maps_range]}")
+            # print(f"map {key} size {[this_maps_width, this_maps_height]} range {[maps_min, maps_min + maps_range]}")
             maps_normalized = (maps - maps_min) / maps_range
             # expand to (-0.1, 1.1) and clamp
             maps_normalized_expanded = maps_normalized * 1.1 - 0.05
-            maps_normalized_expanded_clamped = torch.clamp(maps_normalized_expanded, 0, 1)
+            maps_normalized_expanded_clamped = torch.clamp(
+                maps_normalized_expanded, 0, 1
+            )
 
             # merge together, producing a vertical stack
-            maps_stacked = torch.reshape(maps_normalized_expanded_clamped, [num_tokens * latents_height, latents_width])
+            maps_stacked = torch.reshape(
+                maps_normalized_expanded_clamped,
+                [num_tokens * latents_height, latents_width],
+            )
 
             if merged is None:
                 merged = maps_stacked
             else:
                 # screen blend
-                merged = 1 - (1 - maps_stacked)*(1 - merged)
+                merged = 1 - (1 - maps_stacked) * (1 - merged)
 
         if merged is None:
             return None
 
-        merged_bytes = merged.mul(0xff).byte()
-        return PIL.Image.fromarray(merged_bytes.numpy(), mode='L')
+        merged_bytes = merged.mul(0xFF).byte()
+        return PIL.Image.fromarray(merged_bytes.numpy(), mode="L")

invokeai/backend/stable_diffusion/diffusion/shared_invokeai_diffusion.py

@@ -1,25 +1,36 @@
 from contextlib import contextmanager
 from dataclasses import dataclass
 from math import ceil
-from typing import Callable, Optional, Union, Any, Dict
+from typing import Any, Callable, Dict, Optional, Union
 
 import numpy as np
 import torch
 from diffusers.models.cross_attention import AttnProcessor
 from typing_extensions import TypeAlias
 
-from ldm.invoke.globals import Globals
-from ldm.models.diffusion.cross_attention_control import Arguments, \
-    restore_default_cross_attention, override_cross_attention, Context, get_cross_attention_modules, \
-    CrossAttentionType, SwapCrossAttnContext
-from ldm.models.diffusion.cross_attention_map_saving import AttentionMapSaver
+from invokeai.backend.globals import Globals
+
+from .cross_attention_control import (
+    Arguments,
+    Context,
+    CrossAttentionType,
+    SwapCrossAttnContext,
+    get_cross_attention_modules,
+    override_cross_attention,
+    restore_default_cross_attention,
+)
+from .cross_attention_map_saving import AttentionMapSaver
 
 ModelForwardCallback: TypeAlias = Union[
     # x, t, conditioning, Optional[cross-attention kwargs]
-    Callable[[torch.Tensor, torch.Tensor, torch.Tensor, Optional[dict[str, Any]]], torch.Tensor],
-    Callable[[torch.Tensor, torch.Tensor, torch.Tensor], torch.Tensor]
+    Callable[
+        [torch.Tensor, torch.Tensor, torch.Tensor, Optional[dict[str, Any]]],
+        torch.Tensor,
+    ],
+    Callable[[torch.Tensor, torch.Tensor, torch.Tensor], torch.Tensor],
 ]
 
+
 @dataclass(frozen=True)
 class PostprocessingSettings:
     threshold: float
@@ -29,20 +40,20 @@ class PostprocessingSettings:
 
 
 class InvokeAIDiffuserComponent:
-    '''
+    """
     The aim of this component is to provide a single place for code that can be applied identically to
     all InvokeAI diffusion procedures.
 
     At the moment it includes the following features:
     * Cross attention control ("prompt2prompt")
     * Hybrid conditioning (used for inpainting)
-    '''
+    """
+
     debug_thresholding = False
     sequential_guidance = False
 
     @dataclass
     class ExtraConditioningInfo:
-
         tokens_count_including_eos_bos: int
         cross_attention_control_args: Optional[Arguments] = None
 
@@ -50,10 +61,12 @@ class InvokeAIDiffuserComponent:
         def wants_cross_attention_control(self):
             return self.cross_attention_control_args is not None
 
-
-    def __init__(self, model, model_forward_callback: ModelForwardCallback,
-                 is_running_diffusers: bool=False,
-                 ):
+    def __init__(
+        self,
+        model,
+        model_forward_callback: ModelForwardCallback,
+        is_running_diffusers: bool = False,
+    ):
         """
         :param model: the unet model to pass through to cross attention control
         :param model_forward_callback: a lambda with arguments (x, sigma, conditioning_to_apply). will be called repeatedly. most likely, this should simply call model.forward(x, sigma, conditioning)
@@ -66,23 +79,29 @@ class InvokeAIDiffuserComponent:
         self.sequential_guidance = Globals.sequential_guidance
 
     @contextmanager
-    def custom_attention_context(self,
-                                 extra_conditioning_info: Optional[ExtraConditioningInfo],
-                                 step_count: int):
-        do_swap = extra_conditioning_info is not None and extra_conditioning_info.wants_cross_attention_control
+    def custom_attention_context(
+        self, extra_conditioning_info: Optional[ExtraConditioningInfo], step_count: int
+    ):
+        do_swap = (
+            extra_conditioning_info is not None
+            and extra_conditioning_info.wants_cross_attention_control
+        )
         old_attn_processor = None
         if do_swap:
-            old_attn_processor = self.override_cross_attention(extra_conditioning_info,
-                                                               step_count=step_count)
+            old_attn_processor = self.override_cross_attention(
+                extra_conditioning_info, step_count=step_count
+            )
         try:
             yield None
         finally:
             if old_attn_processor is not None:
                 self.restore_default_cross_attention(old_attn_processor)
             # TODO resuscitate attention map saving
-            #self.remove_attention_map_saving()
+            # self.remove_attention_map_saving()
 
-    def override_cross_attention(self, conditioning: ExtraConditioningInfo, step_count: int) -> Dict[str, AttnProcessor]:
+    def override_cross_attention(
+        self, conditioning: ExtraConditioningInfo, step_count: int
+    ) -> Dict[str, AttnProcessor]:
         """
         setup cross attention .swap control. for diffusers this replaces the attention processor, so
         the previous attention processor is returned so that the caller can restore it later.
@@ -90,18 +109,24 @@ class InvokeAIDiffuserComponent:
         self.conditioning = conditioning
         self.cross_attention_control_context = Context(
             arguments=self.conditioning.cross_attention_control_args,
-            step_count=step_count
+            step_count=step_count,
+        )
+        return override_cross_attention(
+            self.model,
+            self.cross_attention_control_context,
+            is_running_diffusers=self.is_running_diffusers,
         )
-        return override_cross_attention(self.model,
-                                        self.cross_attention_control_context,
-                                        is_running_diffusers=self.is_running_diffusers)
 
-    def restore_default_cross_attention(self, restore_attention_processor: Optional['AttnProcessor']=None):
+    def restore_default_cross_attention(
+        self, restore_attention_processor: Optional["AttnProcessor"] = None
+    ):
         self.conditioning = None
         self.cross_attention_control_context = None
-        restore_default_cross_attention(self.model,
-                                        is_running_diffusers=self.is_running_diffusers,
-                                        restore_attention_processor=restore_attention_processor)
+        restore_default_cross_attention(
+            self.model,
+            is_running_diffusers=self.is_running_diffusers,
+            restore_attention_processor=restore_attention_processor,
+        )
 
     def setup_attention_map_saving(self, saver: AttentionMapSaver):
         def callback(slice, dim, offset, slice_size, key):
@@ -110,26 +135,40 @@ class InvokeAIDiffuserComponent:
                 return
             saver.add_attention_maps(slice, key)
 
-        tokens_cross_attention_modules = get_cross_attention_modules(self.model, CrossAttentionType.TOKENS)
+        tokens_cross_attention_modules = get_cross_attention_modules(
+            self.model, CrossAttentionType.TOKENS
+        )
         for identifier, module in tokens_cross_attention_modules:
-            key = ('down' if identifier.startswith('down') else
-                   'up' if identifier.startswith('up') else
-                   'mid')
+            key = (
+                "down"
+                if identifier.startswith("down")
+                else "up"
+                if identifier.startswith("up")
+                else "mid"
+            )
             module.set_attention_slice_calculated_callback(
-                lambda slice, dim, offset, slice_size, key=key: callback(slice, dim, offset, slice_size, key))
+                lambda slice, dim, offset, slice_size, key=key: callback(
+                    slice, dim, offset, slice_size, key
+                )
+            )
 
     def remove_attention_map_saving(self):
-        tokens_cross_attention_modules = get_cross_attention_modules(self.model, CrossAttentionType.TOKENS)
+        tokens_cross_attention_modules = get_cross_attention_modules(
+            self.model, CrossAttentionType.TOKENS
+        )
         for _, module in tokens_cross_attention_modules:
             module.set_attention_slice_calculated_callback(None)
 
-    def do_diffusion_step(self, x: torch.Tensor, sigma: torch.Tensor,
-                                unconditioning: Union[torch.Tensor,dict],
-                                conditioning: Union[torch.Tensor,dict],
-                                unconditional_guidance_scale: float,
-                                step_index: Optional[int]=None,
-                                total_step_count: Optional[int]=None,
-                          ):
+    def do_diffusion_step(
+        self,
+        x: torch.Tensor,
+        sigma: torch.Tensor,
+        unconditioning: Union[torch.Tensor, dict],
+        conditioning: Union[torch.Tensor, dict],
+        unconditional_guidance_scale: float,
+        step_index: Optional[int] = None,
+        total_step_count: Optional[int] = None,
+    ):
         """
         :param x: current latents
         :param sigma: aka t, passed to the internal model to control how much denoising will occur
@@ -140,33 +179,55 @@ class InvokeAIDiffuserComponent:
         :return: the new latents after applying the model to x using unscaled unconditioning and CFG-scaled conditioning.
         """
 
-
         cross_attention_control_types_to_do = []
         context: Context = self.cross_attention_control_context
         if self.cross_attention_control_context is not None:
-            percent_through = self.calculate_percent_through(sigma, step_index, total_step_count)
-            cross_attention_control_types_to_do = context.get_active_cross_attention_control_types_for_step(percent_through)
+            percent_through = self.calculate_percent_through(
+                sigma, step_index, total_step_count
+            )
+            cross_attention_control_types_to_do = (
+                context.get_active_cross_attention_control_types_for_step(
+                    percent_through
+                )
+            )
 
-        wants_cross_attention_control = (len(cross_attention_control_types_to_do) > 0)
+        wants_cross_attention_control = len(cross_attention_control_types_to_do) > 0
         wants_hybrid_conditioning = isinstance(conditioning, dict)
 
         if wants_hybrid_conditioning:
-            unconditioned_next_x, conditioned_next_x = self._apply_hybrid_conditioning(x, sigma, unconditioning,
-                                                                                       conditioning)
+            unconditioned_next_x, conditioned_next_x = self._apply_hybrid_conditioning(
+                x, sigma, unconditioning, conditioning
+            )
         elif wants_cross_attention_control:
-            unconditioned_next_x, conditioned_next_x = self._apply_cross_attention_controlled_conditioning(x, sigma,
-                                                                                                           unconditioning,
-                                                                                                           conditioning,
-                                                                                                           cross_attention_control_types_to_do)
+            (
+                unconditioned_next_x,
+                conditioned_next_x,
+            ) = self._apply_cross_attention_controlled_conditioning(
+                x,
+                sigma,
+                unconditioning,
+                conditioning,
+                cross_attention_control_types_to_do,
+            )
         elif self.sequential_guidance:
-            unconditioned_next_x, conditioned_next_x = self._apply_standard_conditioning_sequentially(
-                x, sigma, unconditioning, conditioning)
+            (
+                unconditioned_next_x,
+                conditioned_next_x,
+            ) = self._apply_standard_conditioning_sequentially(
+                x, sigma, unconditioning, conditioning
+            )
 
         else:
-            unconditioned_next_x, conditioned_next_x = self._apply_standard_conditioning(
-                x, sigma, unconditioning, conditioning)
+            (
+                unconditioned_next_x,
+                conditioned_next_x,
+            ) = self._apply_standard_conditioning(
+                x, sigma, unconditioning, conditioning
+            )
 
-        combined_next_x = self._combine(unconditioned_next_x, conditioned_next_x, unconditional_guidance_scale)
+        combined_next_x = self._combine(
+            unconditioned_next_x, conditioned_next_x, unconditional_guidance_scale
+        )
 
         return combined_next_x
 
@@ -176,24 +237,33 @@ class InvokeAIDiffuserComponent:
         latents: torch.Tensor,
         sigma,
         step_index,
-        total_step_count
+        total_step_count,
     ) -> torch.Tensor:
         if postprocessing_settings is not None:
-            percent_through = self.calculate_percent_through(sigma, step_index, total_step_count)
-            latents = self.apply_threshold(postprocessing_settings, latents, percent_through)
-            latents = self.apply_symmetry(postprocessing_settings, latents, percent_through)
+            percent_through = self.calculate_percent_through(
+                sigma, step_index, total_step_count
+            )
+            latents = self.apply_threshold(
+                postprocessing_settings, latents, percent_through
+            )
+            latents = self.apply_symmetry(
+                postprocessing_settings, latents, percent_through
+            )
         return latents
 
     def calculate_percent_through(self, sigma, step_index, total_step_count):
         if step_index is not None and total_step_count is not None:
             # 🧨diffusers codepath
-            percent_through = step_index / total_step_count  # will never reach 1.0 - this is deliberate
+            percent_through = (
+                step_index / total_step_count
+            )  # will never reach 1.0 - this is deliberate
         else:
             # legacy compvis codepath
             # TODO remove when compvis codepath support is dropped
             if step_index is None and sigma is None:
                 raise ValueError(
-                    f"Either step_index or sigma is required when doing cross attention control, but both are None.")
+                    f"Either step_index or sigma is required when doing cross attention control, but both are None."
+                )
             percent_through = self.estimate_percent_through(step_index, sigma)
         return percent_through
 
@@ -204,24 +274,30 @@ class InvokeAIDiffuserComponent:
         x_twice = torch.cat([x] * 2)
         sigma_twice = torch.cat([sigma] * 2)
         both_conditionings = torch.cat([unconditioning, conditioning])
-        both_results = self.model_forward_callback(x_twice, sigma_twice, both_conditionings)
+        both_results = self.model_forward_callback(
+            x_twice, sigma_twice, both_conditionings
+        )
         unconditioned_next_x, conditioned_next_x = both_results.chunk(2)
-        if conditioned_next_x.device.type == 'mps':
+        if conditioned_next_x.device.type == "mps":
             # prevent a result filled with zeros. seems to be a torch bug.
             conditioned_next_x = conditioned_next_x.clone()
         return unconditioned_next_x, conditioned_next_x
 
-
-    def _apply_standard_conditioning_sequentially(self, x: torch.Tensor, sigma, unconditioning: torch.Tensor, conditioning: torch.Tensor):
+    def _apply_standard_conditioning_sequentially(
+        self,
+        x: torch.Tensor,
+        sigma,
+        unconditioning: torch.Tensor,
+        conditioning: torch.Tensor,
+    ):
         # low-memory sequential path
         unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning)
         conditioned_next_x = self.model_forward_callback(x, sigma, conditioning)
-        if conditioned_next_x.device.type == 'mps':
+        if conditioned_next_x.device.type == "mps":
             # prevent a result filled with zeros. seems to be a torch bug.
             conditioned_next_x = conditioned_next_x.clone()
         return unconditioned_next_x, conditioned_next_x
 
-
     def _apply_hybrid_conditioning(self, x, sigma, unconditioning, conditioning):
         assert isinstance(conditioning, dict)
         assert isinstance(unconditioning, dict)
@@ -236,48 +312,80 @@ class InvokeAIDiffuserComponent:
                 ]
             else:
                 both_conditionings[k] = torch.cat([unconditioning[k], conditioning[k]])
-        unconditioned_next_x, conditioned_next_x = self.model_forward_callback(x_twice, sigma_twice, both_conditionings).chunk(2)
+        unconditioned_next_x, conditioned_next_x = self.model_forward_callback(
+            x_twice, sigma_twice, both_conditionings
+        ).chunk(2)
         return unconditioned_next_x, conditioned_next_x
 
-
-    def _apply_cross_attention_controlled_conditioning(self,
-                                                     x: torch.Tensor,
-                                                     sigma,
-                                                     unconditioning,
-                                                     conditioning,
-                                                     cross_attention_control_types_to_do):
+    def _apply_cross_attention_controlled_conditioning(
+        self,
+        x: torch.Tensor,
+        sigma,
+        unconditioning,
+        conditioning,
+        cross_attention_control_types_to_do,
+    ):
         if self.is_running_diffusers:
-            return self._apply_cross_attention_controlled_conditioning__diffusers(x, sigma, unconditioning,
-                                                                                  conditioning,
-                                                                                  cross_attention_control_types_to_do)
+            return self._apply_cross_attention_controlled_conditioning__diffusers(
+                x,
+                sigma,
+                unconditioning,
+                conditioning,
+                cross_attention_control_types_to_do,
+            )
         else:
-            return self._apply_cross_attention_controlled_conditioning__compvis(x, sigma, unconditioning, conditioning,
-                                                                                cross_attention_control_types_to_do)
+            return self._apply_cross_attention_controlled_conditioning__compvis(
+                x,
+                sigma,
+                unconditioning,
+                conditioning,
+                cross_attention_control_types_to_do,
+            )
 
-    def _apply_cross_attention_controlled_conditioning__diffusers(self,
-                                                                 x: torch.Tensor,
-                                                                 sigma,
-                                                                 unconditioning,
-                                                                 conditioning,
-                                                                 cross_attention_control_types_to_do):
+    def _apply_cross_attention_controlled_conditioning__diffusers(
+        self,
+        x: torch.Tensor,
+        sigma,
+        unconditioning,
+        conditioning,
+        cross_attention_control_types_to_do,
+    ):
         context: Context = self.cross_attention_control_context
 
-        cross_attn_processor_context = SwapCrossAttnContext(modified_text_embeddings=context.arguments.edited_conditioning,
-                                                            index_map=context.cross_attention_index_map,
-                                                            mask=context.cross_attention_mask,
-                                                            cross_attention_types_to_do=[])
+        cross_attn_processor_context = SwapCrossAttnContext(
+            modified_text_embeddings=context.arguments.edited_conditioning,
+            index_map=context.cross_attention_index_map,
+            mask=context.cross_attention_mask,
+            cross_attention_types_to_do=[],
+        )
         # no cross attention for unconditioning (negative prompt)
-        unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning,
-                                                           {"swap_cross_attn_context": cross_attn_processor_context})
+        unconditioned_next_x = self.model_forward_callback(
+            x,
+            sigma,
+            unconditioning,
+            {"swap_cross_attn_context": cross_attn_processor_context},
+        )
 
         # do requested cross attention types for conditioning (positive prompt)
-        cross_attn_processor_context.cross_attention_types_to_do = cross_attention_control_types_to_do
-        conditioned_next_x = self.model_forward_callback(x, sigma, conditioning,
-                                                         {"swap_cross_attn_context": cross_attn_processor_context})
+        cross_attn_processor_context.cross_attention_types_to_do = (
+            cross_attention_control_types_to_do
+        )
+        conditioned_next_x = self.model_forward_callback(
+            x,
+            sigma,
+            conditioning,
+            {"swap_cross_attn_context": cross_attn_processor_context},
+        )
         return unconditioned_next_x, conditioned_next_x
 
-
-    def _apply_cross_attention_controlled_conditioning__compvis(self, x:torch.Tensor, sigma, unconditioning, conditioning, cross_attention_control_types_to_do):
+    def _apply_cross_attention_controlled_conditioning__compvis(
+        self,
+        x: torch.Tensor,
+        sigma,
+        unconditioning,
+        conditioning,
+        cross_attention_control_types_to_do,
+    ):
         # print('pct', percent_through, ': doing cross attention control on', cross_attention_control_types_to_do)
         # slower non-batched path (20% slower on mac MPS)
         # We are only interested in using attention maps for conditioned_next_x, but batching them with generation of
@@ -287,24 +395,28 @@ class InvokeAIDiffuserComponent:
         # representing batched uncond + cond, but then when it comes to applying the saved attention, the
         # wrangler gets an attention tensor which only has shape[0]=8, representing just self.edited_conditionings.)
         # todo: give CrossAttentionControl's `wrangler` function more info so it can work with a batched call as well.
-        context:Context = self.cross_attention_control_context
+        context: Context = self.cross_attention_control_context
 
         try:
             unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning)
 
             # process x using the original prompt, saving the attention maps
-            #print("saving attention maps for", cross_attention_control_types_to_do)
+            # print("saving attention maps for", cross_attention_control_types_to_do)
             for ca_type in cross_attention_control_types_to_do:
                 context.request_save_attention_maps(ca_type)
             _ = self.model_forward_callback(x, sigma, conditioning)
             context.clear_requests(cleanup=False)
 
             # process x again, using the saved attention maps to control where self.edited_conditioning will be applied
-            #print("applying saved attention maps for", cross_attention_control_types_to_do)
+            # print("applying saved attention maps for", cross_attention_control_types_to_do)
             for ca_type in cross_attention_control_types_to_do:
                 context.request_apply_saved_attention_maps(ca_type)
-            edited_conditioning = self.conditioning.cross_attention_control_args.edited_conditioning
-            conditioned_next_x = self.model_forward_callback(x, sigma, edited_conditioning)
+            edited_conditioning = (
+                self.conditioning.cross_attention_control_args.edited_conditioning
+            )
+            conditioned_next_x = self.model_forward_callback(
+                x, sigma, edited_conditioning
+            )
             context.clear_requests(cleanup=True)
 
         except:
@@ -323,17 +435,21 @@ class InvokeAIDiffuserComponent:
         self,
         postprocessing_settings: PostprocessingSettings,
         latents: torch.Tensor,
-        percent_through: float
+        percent_through: float,
     ) -> torch.Tensor:
-
-        if postprocessing_settings.threshold is None or postprocessing_settings.threshold == 0.0:
+        if (
+            postprocessing_settings.threshold is None
+            or postprocessing_settings.threshold == 0.0
+        ):
             return latents
 
         threshold = postprocessing_settings.threshold
         warmup = postprocessing_settings.warmup
 
         if percent_through < warmup:
-            current_threshold = threshold + threshold * 5 * (1 - (percent_through / warmup))
+            current_threshold = threshold + threshold * 5 * (
+                1 - (percent_through / warmup)
+            )
         else:
             current_threshold = threshold
 
@@ -347,10 +463,14 @@ class InvokeAIDiffuserComponent:
 
         if self.debug_thresholding:
             std, mean = [i.item() for i in torch.std_mean(latents)]
-            outside = torch.count_nonzero((latents < -current_threshold) | (latents > current_threshold))
-            print(f"\nThreshold: %={percent_through} threshold={current_threshold:.3f} (of {threshold:.3f})\n"
-                  f"  | min, mean, max = {minval:.3f}, {mean:.3f}, {maxval:.3f}\tstd={std}\n"
-                  f"  | {outside / latents.numel() * 100:.2f}% values outside threshold")
+            outside = torch.count_nonzero(
+                (latents < -current_threshold) | (latents > current_threshold)
+            )
+            print(
+                f"\nThreshold: %={percent_through} threshold={current_threshold:.3f} (of {threshold:.3f})\n"
+                f"  | min, mean, max = {minval:.3f}, {mean:.3f}, {maxval:.3f}\tstd={std}\n"
+                f"  | {outside / latents.numel() * 100:.2f}% values outside threshold"
+            )
 
         if maxval < current_threshold and minval > -current_threshold:
             return latents
@@ -363,17 +483,23 @@ class InvokeAIDiffuserComponent:
             latents = torch.clone(latents)
             maxval = np.clip(maxval * scale, 1, current_threshold)
             num_altered += torch.count_nonzero(latents > maxval)
-            latents[latents > maxval] = torch.rand_like(latents[latents > maxval]) * maxval
+            latents[latents > maxval] = (
+                torch.rand_like(latents[latents > maxval]) * maxval
+            )
 
         if minval < -current_threshold:
             latents = torch.clone(latents)
             minval = np.clip(minval * scale, -current_threshold, -1)
             num_altered += torch.count_nonzero(latents < minval)
-            latents[latents < minval] = torch.rand_like(latents[latents < minval]) * minval
+            latents[latents < minval] = (
+                torch.rand_like(latents[latents < minval]) * minval
+            )
 
         if self.debug_thresholding:
-            print(f"  | min,     , max = {minval:.3f},        , {maxval:.3f}\t(scaled by {scale})\n"
-                  f"  | {num_altered / latents.numel() * 100:.2f}% values altered")
+            print(
+                f"  | min,     , max = {minval:.3f},        , {maxval:.3f}\t(scaled by {scale})\n"
+                f"  | {num_altered / latents.numel() * 100:.2f}% values altered"
+            )
 
         return latents
 
@@ -381,9 +507,8 @@ class InvokeAIDiffuserComponent:
         self,
         postprocessing_settings: PostprocessingSettings,
         latents: torch.Tensor,
-        percent_through: float
+        percent_through: float,
     ) -> torch.Tensor:
-
         # Reset our last percent through if this is our first step.
         if percent_through == 0.0:
             self.last_percent_through = 0.0
@@ -393,36 +518,52 @@ class InvokeAIDiffuserComponent:
 
         # Check for out of bounds
         h_symmetry_time_pct = postprocessing_settings.h_symmetry_time_pct
-        if (h_symmetry_time_pct is not None and (h_symmetry_time_pct <= 0.0 or h_symmetry_time_pct > 1.0)):
+        if h_symmetry_time_pct is not None and (
+            h_symmetry_time_pct <= 0.0 or h_symmetry_time_pct > 1.0
+        ):
             h_symmetry_time_pct = None
 
         v_symmetry_time_pct = postprocessing_settings.v_symmetry_time_pct
-        if (v_symmetry_time_pct is not None and (v_symmetry_time_pct <= 0.0 or v_symmetry_time_pct > 1.0)):
+        if v_symmetry_time_pct is not None and (
+            v_symmetry_time_pct <= 0.0 or v_symmetry_time_pct > 1.0
+        ):
             v_symmetry_time_pct = None
 
         dev = latents.device.type
 
-        latents.to(device='cpu')
+        latents.to(device="cpu")
 
         if (
-            h_symmetry_time_pct != None and
-            self.last_percent_through < h_symmetry_time_pct and
-            percent_through >= h_symmetry_time_pct
+            h_symmetry_time_pct != None
+            and self.last_percent_through < h_symmetry_time_pct
+            and percent_through >= h_symmetry_time_pct
         ):
             # Horizontal symmetry occurs on the 3rd dimension of the latent
             width = latents.shape[3]
             x_flipped = torch.flip(latents, dims=[3])
-            latents = torch.cat([latents[:, :, :, 0:int(width/2)], x_flipped[:, :, :, int(width/2):int(width)]], dim=3)
+            latents = torch.cat(
+                [
+                    latents[:, :, :, 0 : int(width / 2)],
+                    x_flipped[:, :, :, int(width / 2) : int(width)],
+                ],
+                dim=3,
+            )
 
         if (
-            v_symmetry_time_pct != None and
-            self.last_percent_through < v_symmetry_time_pct and
-            percent_through >= v_symmetry_time_pct
+            v_symmetry_time_pct != None
+            and self.last_percent_through < v_symmetry_time_pct
+            and percent_through >= v_symmetry_time_pct
         ):
             # Vertical symmetry occurs on the 2nd dimension of the latent
             height = latents.shape[2]
             y_flipped = torch.flip(latents, dims=[2])
-            latents = torch.cat([latents[:, :, 0:int(height / 2)], y_flipped[:, :, int(height / 2):int(height)]], dim=2)
+            latents = torch.cat(
+                [
+                    latents[:, :, 0 : int(height / 2)],
+                    y_flipped[:, :, int(height / 2) : int(height)],
+                ],
+                dim=2,
+            )
 
         self.last_percent_through = percent_through
         return latents.to(device=dev)
@@ -430,7 +571,9 @@ class InvokeAIDiffuserComponent:
     def estimate_percent_through(self, step_index, sigma):
         if step_index is not None and self.cross_attention_control_context is not None:
             # percent_through will never reach 1.0 (but this is intended)
-            return float(step_index) / float(self.cross_attention_control_context.step_count)
+            return float(step_index) / float(
+                self.cross_attention_control_context.step_count
+            )
         # find the best possible index of the current sigma in the sigma sequence
         smaller_sigmas = torch.nonzero(self.model.sigmas <= sigma)
         sigma_index = smaller_sigmas[-1].item() if smaller_sigmas.shape[0] > 0 else 0
@@ -439,33 +582,38 @@ class InvokeAIDiffuserComponent:
         return 1.0 - float(sigma_index + 1) / float(self.model.sigmas.shape[0])
         # print('estimated percent_through', percent_through, 'from sigma', sigma.item())
 
-
     # todo: make this work
     @classmethod
-    def apply_conjunction(cls, x, t, forward_func, uc, c_or_weighted_c_list, global_guidance_scale):
+    def apply_conjunction(
+        cls, x, t, forward_func, uc, c_or_weighted_c_list, global_guidance_scale
+    ):
         x_in = torch.cat([x] * 2)
-        t_in = torch.cat([t] * 2) # aka sigmas
+        t_in = torch.cat([t] * 2)  # aka sigmas
 
         deltas = None
         uncond_latents = None
-        weighted_cond_list = c_or_weighted_c_list if type(c_or_weighted_c_list) is list else [(c_or_weighted_c_list, 1)]
+        weighted_cond_list = (
+            c_or_weighted_c_list
+            if type(c_or_weighted_c_list) is list
+            else [(c_or_weighted_c_list, 1)]
+        )
 
         # below is fugly omg
         num_actual_conditionings = len(c_or_weighted_c_list)
-        conditionings = [uc] + [c for c,weight in weighted_cond_list]
-        weights = [1] + [weight for c,weight in weighted_cond_list]
-        chunk_count = ceil(len(conditionings)/2)
+        conditionings = [uc] + [c for c, weight in weighted_cond_list]
+        weights = [1] + [weight for c, weight in weighted_cond_list]
+        chunk_count = ceil(len(conditionings) / 2)
         deltas = None
         for chunk_index in range(chunk_count):
-            offset = chunk_index*2
-            chunk_size = min(2, len(conditionings)-offset)
+            offset = chunk_index * 2
+            chunk_size = min(2, len(conditionings) - offset)
 
             if chunk_size == 1:
                 c_in = conditionings[offset]
                 latents_a = forward_func(x_in[:-1], t_in[:-1], c_in)
                 latents_b = None
             else:
-                c_in = torch.cat(conditionings[offset:offset+2])
+                c_in = torch.cat(conditionings[offset : offset + 2])
                 latents_a, latents_b = forward_func(x_in, t_in, c_in).chunk(2)
 
             # first chunk is guaranteed to be 2 entries: uncond_latents + first conditioining
@@ -478,11 +626,15 @@ class InvokeAIDiffuserComponent:
                     deltas = torch.cat((deltas, latents_b - uncond_latents))
 
         # merge the weighted deltas together into a single merged delta
-        per_delta_weights = torch.tensor(weights[1:], dtype=deltas.dtype, device=deltas.device)
+        per_delta_weights = torch.tensor(
+            weights[1:], dtype=deltas.dtype, device=deltas.device
+        )
         normalize = False
         if normalize:
             per_delta_weights /= torch.sum(per_delta_weights)
-        reshaped_weights = per_delta_weights.reshape(per_delta_weights.shape + (1, 1, 1))
+        reshaped_weights = per_delta_weights.reshape(
+            per_delta_weights.shape + (1, 1, 1)
+        )
         deltas_merged = torch.sum(deltas * reshaped_weights, dim=0, keepdim=True)
 
         # old_return_value = super().forward(x, sigma, uncond, cond, cond_scale)

invokeai/backend/stable_diffusion/image_degradation/bsrgan.py

@@ -10,20 +10,19 @@
 # --------------------------------------------
 """
 
-import numpy as np
-import cv2
-import torch
-
-from functools import partial
 import random
-from scipy import ndimage
+from functools import partial
+
+import albumentations
+import cv2
+import ldm.modules.image_degradation.utils_image as util
+import numpy as np
 import scipy
 import scipy.stats as ss
+import torch
+from scipy import ndimage
 from scipy.interpolate import interp2d
 from scipy.linalg import orth
-import albumentations
-
-import ldm.modules.image_degradation.utils_image as util
 
 
 def modcrop_np(img, sf):
@@ -54,9 +53,7 @@ def analytic_kernel(k):
     # Loop over the small kernel to fill the big one
     for r in range(k_size):
         for c in range(k_size):
-            big_k[2 * r : 2 * r + k_size, 2 * c : 2 * c + k_size] += (
-                k[r, c] * k
-            )
+            big_k[2 * r : 2 * r + k_size, 2 * c : 2 * c + k_size] += k[r, c] * k
     # Crop the edges of the big kernel to ignore very small values and increase run time of SR
     crop = k_size // 2
     cropped_big_k = big_k[crop:-crop, crop:-crop]
@@ -77,9 +74,7 @@ def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
     """
 
     v = np.dot(
-        np.array(
-            [[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]
-        ),
+        np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]),
         np.array([1.0, 0.0]),
     )
     V = np.array([[v[0], v[1]], [v[1], -v[0]]])
@@ -139,13 +134,11 @@ def blur(x, k):
     """
     n, c = x.shape[:2]
     p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
-    x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
+    x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode="replicate")
     k = k.repeat(1, c, 1, 1)
     k = k.view(-1, 1, k.shape[2], k.shape[3])
     x = x.view(1, -1, x.shape[2], x.shape[3])
-    x = torch.nn.functional.conv2d(
-        x, k, bias=None, stride=1, padding=0, groups=n * c
-    )
+    x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
     x = x.view(n, c, x.shape[2], x.shape[3])
 
     return x
@@ -172,16 +165,12 @@ def gen_kernel(
 
     # Set COV matrix using Lambdas and Theta
     LAMBDA = np.diag([lambda_1, lambda_2])
-    Q = np.array(
-        [[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]
-    )
+    Q = np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]])
     SIGMA = Q @ LAMBDA @ Q.T
     INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
 
     # Set expectation position (shifting kernel for aligned image)
-    MU = k_size // 2 - 0.5 * (
-        scale_factor - 1
-    )  # - 0.5 * (scale_factor - k_size % 2)
+    MU = k_size // 2 - 0.5 * (scale_factor - 1)  # - 0.5 * (scale_factor - k_size % 2)
     MU = MU[None, None, :, None]
 
     # Create meshgrid for Gaussian
@@ -206,9 +195,7 @@ def fspecial_gaussian(hsize, sigma):
     hsize = [hsize, hsize]
     siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
     std = sigma
-    [x, y] = np.meshgrid(
-        np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1)
-    )
+    [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
     arg = -(x * x + y * y) / (2 * std * std)
     h = np.exp(arg)
     h[h < scipy.finfo(float).eps * h.max()] = 0
@@ -232,9 +219,9 @@ def fspecial(filter_type, *args, **kwargs):
     python code from:
     https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
     """
-    if filter_type == 'gaussian':
+    if filter_type == "gaussian":
         return fspecial_gaussian(*args, **kwargs)
-    if filter_type == 'laplacian':
+    if filter_type == "laplacian":
         return fspecial_laplacian(*args, **kwargs)
 
 
@@ -275,7 +262,7 @@ def srmd_degradation(x, k, sf=3):
         }
     """
     x = ndimage.filters.convolve(
-        x, np.expand_dims(k, axis=2), mode='wrap'
+        x, np.expand_dims(k, axis=2), mode="wrap"
     )  # 'nearest' | 'mirror'
     x = bicubic_degradation(x, sf=sf)
     return x
@@ -299,7 +286,7 @@ def dpsr_degradation(x, k, sf=3):
         }
     """
     x = bicubic_degradation(x, sf=sf)
-    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode="wrap")
     return x
 
 
@@ -312,7 +299,7 @@ def classical_degradation(x, k, sf=3):
     Return:
         downsampled LR image
     """
-    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode="wrap")
     # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
     st = 0
     return x[st::sf, st::sf, ...]
@@ -336,7 +323,7 @@ def add_sharpening(img, weight=0.5, radius=50, threshold=10):
     blur = cv2.GaussianBlur(img, (radius, radius), 0)
     residual = img - blur
     mask = np.abs(residual) * 255 > threshold
-    mask = mask.astype('float32')
+    mask = mask.astype("float32")
     soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
 
     K = img + weight * residual
@@ -357,12 +344,8 @@ def add_blur(img, sf=4):
             l2=l2,
         )
     else:
-        k = fspecial(
-            'gaussian', 2 * random.randint(2, 11) + 3, wd * random.random()
-        )
-    img = ndimage.filters.convolve(
-        img, np.expand_dims(k, axis=2), mode='mirror'
-    )
+        k = fspecial("gaussian", 2 * random.randint(2, 11) + 3, wd * random.random())
+    img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode="mirror")
 
     return img
 
@@ -430,9 +413,9 @@ def add_speckle_noise(img, noise_level1=2, noise_level2=25):
     img = np.clip(img, 0.0, 1.0)
     rnum = random.random()
     if rnum > 0.6:
-        img += img * np.random.normal(
-            0, noise_level / 255.0, img.shape
-        ).astype(np.float32)
+        img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(
+            np.float32
+        )
     elif rnum < 0.4:
         img += img * np.random.normal(
             0, noise_level / 255.0, (*img.shape[:2], 1)
@@ -458,8 +441,7 @@ def add_Poisson_noise(img):
         img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
         img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.0
         noise_gray = (
-            np.random.poisson(img_gray * vals).astype(np.float32) / vals
-            - img_gray
+            np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
         )
         img += noise_gray[:, :, np.newaxis]
     img = np.clip(img, 0.0, 1.0)
@@ -470,7 +452,7 @@ def add_JPEG_noise(img):
     quality_factor = random.randint(30, 95)
     img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
     result, encimg = cv2.imencode(
-        '.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor]
+        ".jpg", img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor]
     )
     img = cv2.imdecode(encimg, 1)
     img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
@@ -513,7 +495,7 @@ def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
     h, w = img.shape[:2]
 
     if h < lq_patchsize * sf or w < lq_patchsize * sf:
-        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+        raise ValueError(f"img size ({h1}X{w1}) is too small!")
 
     hq = img.copy()
 
@@ -538,7 +520,6 @@ def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
         )
 
     for i in shuffle_order:
-
         if i == 0:
             img = add_blur(img, sf=sf)
 
@@ -556,13 +537,11 @@ def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
                     interpolation=random.choice([1, 2, 3]),
                 )
             else:
-                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k = fspecial("gaussian", 25, random.uniform(0.1, 0.6 * sf))
                 k_shifted = shift_pixel(k, sf)
-                k_shifted = (
-                    k_shifted / k_shifted.sum()
-                )  # blur with shifted kernel
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
                 img = ndimage.filters.convolve(
-                    img, np.expand_dims(k_shifted, axis=2), mode='mirror'
+                    img, np.expand_dims(k_shifted, axis=2), mode="mirror"
                 )
                 img = img[0::sf, 0::sf, ...]  # nearest downsampling
             img = np.clip(img, 0.0, 1.0)
@@ -644,7 +623,6 @@ def degradation_bsrgan_variant(image, sf=4, isp_model=None):
         )
 
     for i in shuffle_order:
-
         if i == 0:
             image = add_blur(image, sf=sf)
 
@@ -665,13 +643,11 @@ def degradation_bsrgan_variant(image, sf=4, isp_model=None):
                     interpolation=random.choice([1, 2, 3]),
                 )
             else:
-                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k = fspecial("gaussian", 25, random.uniform(0.1, 0.6 * sf))
                 k_shifted = shift_pixel(k, sf)
-                k_shifted = (
-                    k_shifted / k_shifted.sum()
-                )  # blur with shifted kernel
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
                 image = ndimage.filters.convolve(
-                    image, np.expand_dims(k_shifted, axis=2), mode='mirror'
+                    image, np.expand_dims(k_shifted, axis=2), mode="mirror"
                 )
                 image = image[0::sf, 0::sf, ...]  # nearest downsampling
             image = np.clip(image, 0.0, 1.0)
@@ -703,7 +679,7 @@ def degradation_bsrgan_variant(image, sf=4, isp_model=None):
     # add final JPEG compression noise
     image = add_JPEG_noise(image)
     image = util.single2uint(image)
-    example = {'image': image}
+    example = {"image": image}
     return example
 
 
@@ -735,7 +711,7 @@ def degradation_bsrgan_plus(
     h, w = img.shape[:2]
 
     if h < lq_patchsize * sf or w < lq_patchsize * sf:
-        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+        raise ValueError(f"img size ({h1}X{w1}) is too small!")
 
     if use_sharp:
         img = add_sharpening(img)
@@ -746,12 +722,8 @@ def degradation_bsrgan_plus(
     else:
         shuffle_order = list(range(13))
         # local shuffle for noise, JPEG is always the last one
-        shuffle_order[2:6] = random.sample(
-            shuffle_order[2:6], len(range(2, 6))
-        )
-        shuffle_order[9:13] = random.sample(
-            shuffle_order[9:13], len(range(9, 13))
-        )
+        shuffle_order[2:6] = random.sample(shuffle_order[2:6], len(range(2, 6)))
+        shuffle_order[9:13] = random.sample(shuffle_order[9:13], len(range(9, 13)))
 
     poisson_prob, speckle_prob, isp_prob = 0.1, 0.1, 0.1
 
@@ -791,7 +763,7 @@ def degradation_bsrgan_plus(
                 with torch.no_grad():
                     img, hq = isp_model.forward(img.copy(), hq)
         else:
-            print('check the shuffle!')
+            print("check the shuffle!")
 
     # resize to desired size
     img = cv2.resize(
@@ -809,15 +781,15 @@ def degradation_bsrgan_plus(
     return img, hq
 
 
-if __name__ == '__main__':
-    print('hey')
-    img = util.imread_uint('utils/test.png', 3)
+if __name__ == "__main__":
+    print("hey")
+    img = util.imread_uint("utils/test.png", 3)
     print(img)
     img = util.uint2single(img)
     print(img)
     img = img[:448, :448]
     h = img.shape[0] // 4
-    print('resizing to', h)
+    print("resizing to", h)
     sf = 4
     deg_fn = partial(degradation_bsrgan_variant, sf=sf)
     for i in range(20):
@@ -826,9 +798,9 @@ if __name__ == '__main__':
         print(img_lq)
         img_lq_bicubic = albumentations.SmallestMaxSize(
             max_size=h, interpolation=cv2.INTER_CUBIC
-        )(image=img)['image']
+        )(image=img)["image"]
         print(img_lq.shape)
-        print('bicubic', img_lq_bicubic.shape)
+        print("bicubic", img_lq_bicubic.shape)
         print(img_hq.shape)
         lq_nearest = cv2.resize(
             util.single2uint(img_lq),
@@ -843,4 +815,4 @@ if __name__ == '__main__':
         img_concat = np.concatenate(
             [lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1
         )
-        util.imsave(img_concat, str(i) + '.png')
+        util.imsave(img_concat, str(i) + ".png")

invokeai/backend/stable_diffusion/image_degradation/bsrgan_light.py

@@ -1,18 +1,17 @@
 # -*- coding: utf-8 -*-
-import numpy as np
-import cv2
-import torch
-
-from functools import partial
 import random
-from scipy import ndimage
+from functools import partial
+
+import albumentations
+import cv2
+import ldm.modules.image_degradation.utils_image as util
+import numpy as np
 import scipy
 import scipy.stats as ss
+import torch
+from scipy import ndimage
 from scipy.interpolate import interp2d
 from scipy.linalg import orth
-import albumentations
-
-import ldm.modules.image_degradation.utils_image as util
 
 """
 # --------------------------------------------
@@ -54,9 +53,7 @@ def analytic_kernel(k):
     # Loop over the small kernel to fill the big one
     for r in range(k_size):
         for c in range(k_size):
-            big_k[2 * r : 2 * r + k_size, 2 * c : 2 * c + k_size] += (
-                k[r, c] * k
-            )
+            big_k[2 * r : 2 * r + k_size, 2 * c : 2 * c + k_size] += k[r, c] * k
     # Crop the edges of the big kernel to ignore very small values and increase run time of SR
     crop = k_size // 2
     cropped_big_k = big_k[crop:-crop, crop:-crop]
@@ -77,9 +74,7 @@ def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
     """
 
     v = np.dot(
-        np.array(
-            [[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]
-        ),
+        np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]),
         np.array([1.0, 0.0]),
     )
     V = np.array([[v[0], v[1]], [v[1], -v[0]]])
@@ -139,13 +134,11 @@ def blur(x, k):
     """
     n, c = x.shape[:2]
     p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
-    x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
+    x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode="replicate")
     k = k.repeat(1, c, 1, 1)
     k = k.view(-1, 1, k.shape[2], k.shape[3])
     x = x.view(1, -1, x.shape[2], x.shape[3])
-    x = torch.nn.functional.conv2d(
-        x, k, bias=None, stride=1, padding=0, groups=n * c
-    )
+    x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
     x = x.view(n, c, x.shape[2], x.shape[3])
 
     return x
@@ -172,16 +165,12 @@ def gen_kernel(
 
     # Set COV matrix using Lambdas and Theta
     LAMBDA = np.diag([lambda_1, lambda_2])
-    Q = np.array(
-        [[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]
-    )
+    Q = np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]])
     SIGMA = Q @ LAMBDA @ Q.T
     INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
 
     # Set expectation position (shifting kernel for aligned image)
-    MU = k_size // 2 - 0.5 * (
-        scale_factor - 1
-    )  # - 0.5 * (scale_factor - k_size % 2)
+    MU = k_size // 2 - 0.5 * (scale_factor - 1)  # - 0.5 * (scale_factor - k_size % 2)
     MU = MU[None, None, :, None]
 
     # Create meshgrid for Gaussian
@@ -206,9 +195,7 @@ def fspecial_gaussian(hsize, sigma):
     hsize = [hsize, hsize]
     siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
     std = sigma
-    [x, y] = np.meshgrid(
-        np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1)
-    )
+    [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
     arg = -(x * x + y * y) / (2 * std * std)
     h = np.exp(arg)
     h[h < scipy.finfo(float).eps * h.max()] = 0
@@ -232,9 +219,9 @@ def fspecial(filter_type, *args, **kwargs):
     python code from:
     https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
     """
-    if filter_type == 'gaussian':
+    if filter_type == "gaussian":
         return fspecial_gaussian(*args, **kwargs)
-    if filter_type == 'laplacian':
+    if filter_type == "laplacian":
         return fspecial_laplacian(*args, **kwargs)
 
 
@@ -275,7 +262,7 @@ def srmd_degradation(x, k, sf=3):
         }
     """
     x = ndimage.filters.convolve(
-        x, np.expand_dims(k, axis=2), mode='wrap'
+        x, np.expand_dims(k, axis=2), mode="wrap"
     )  # 'nearest' | 'mirror'
     x = bicubic_degradation(x, sf=sf)
     return x
@@ -299,7 +286,7 @@ def dpsr_degradation(x, k, sf=3):
         }
     """
     x = bicubic_degradation(x, sf=sf)
-    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode="wrap")
     return x
 
 
@@ -312,7 +299,7 @@ def classical_degradation(x, k, sf=3):
     Return:
         downsampled LR image
     """
-    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode="wrap")
     # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
     st = 0
     return x[st::sf, st::sf, ...]
@@ -336,7 +323,7 @@ def add_sharpening(img, weight=0.5, radius=50, threshold=10):
     blur = cv2.GaussianBlur(img, (radius, radius), 0)
     residual = img - blur
     mask = np.abs(residual) * 255 > threshold
-    mask = mask.astype('float32')
+    mask = mask.astype("float32")
     soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
 
     K = img + weight * residual
@@ -361,12 +348,8 @@ def add_blur(img, sf=4):
             l2=l2,
         )
     else:
-        k = fspecial(
-            'gaussian', random.randint(2, 4) + 3, wd * random.random()
-        )
-    img = ndimage.filters.convolve(
-        img, np.expand_dims(k, axis=2), mode='mirror'
-    )
+        k = fspecial("gaussian", random.randint(2, 4) + 3, wd * random.random())
+    img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode="mirror")
 
     return img
 
@@ -434,9 +417,9 @@ def add_speckle_noise(img, noise_level1=2, noise_level2=25):
     img = np.clip(img, 0.0, 1.0)
     rnum = random.random()
     if rnum > 0.6:
-        img += img * np.random.normal(
-            0, noise_level / 255.0, img.shape
-        ).astype(np.float32)
+        img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(
+            np.float32
+        )
     elif rnum < 0.4:
         img += img * np.random.normal(
             0, noise_level / 255.0, (*img.shape[:2], 1)
@@ -462,8 +445,7 @@ def add_Poisson_noise(img):
         img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
         img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.0
         noise_gray = (
-            np.random.poisson(img_gray * vals).astype(np.float32) / vals
-            - img_gray
+            np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
         )
         img += noise_gray[:, :, np.newaxis]
     img = np.clip(img, 0.0, 1.0)
@@ -474,7 +456,7 @@ def add_JPEG_noise(img):
     quality_factor = random.randint(80, 95)
     img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
     result, encimg = cv2.imencode(
-        '.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor]
+        ".jpg", img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor]
     )
     img = cv2.imdecode(encimg, 1)
     img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
@@ -517,7 +499,7 @@ def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
     h, w = img.shape[:2]
 
     if h < lq_patchsize * sf or w < lq_patchsize * sf:
-        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+        raise ValueError(f"img size ({h1}X{w1}) is too small!")
 
     hq = img.copy()
 
@@ -542,7 +524,6 @@ def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
         )
 
     for i in shuffle_order:
-
         if i == 0:
             img = add_blur(img, sf=sf)
 
@@ -560,13 +541,11 @@ def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
                     interpolation=random.choice([1, 2, 3]),
                 )
             else:
-                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k = fspecial("gaussian", 25, random.uniform(0.1, 0.6 * sf))
                 k_shifted = shift_pixel(k, sf)
-                k_shifted = (
-                    k_shifted / k_shifted.sum()
-                )  # blur with shifted kernel
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
                 img = ndimage.filters.convolve(
-                    img, np.expand_dims(k_shifted, axis=2), mode='mirror'
+                    img, np.expand_dims(k_shifted, axis=2), mode="mirror"
                 )
                 img = img[0::sf, 0::sf, ...]  # nearest downsampling
             img = np.clip(img, 0.0, 1.0)
@@ -648,7 +627,6 @@ def degradation_bsrgan_variant(image, sf=4, isp_model=None):
         )
 
     for i in shuffle_order:
-
         if i == 0:
             image = add_blur(image, sf=sf)
 
@@ -672,13 +650,11 @@ def degradation_bsrgan_variant(image, sf=4, isp_model=None):
                     interpolation=random.choice([1, 2, 3]),
                 )
             else:
-                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k = fspecial("gaussian", 25, random.uniform(0.1, 0.6 * sf))
                 k_shifted = shift_pixel(k, sf)
-                k_shifted = (
-                    k_shifted / k_shifted.sum()
-                )  # blur with shifted kernel
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
                 image = ndimage.filters.convolve(
-                    image, np.expand_dims(k_shifted, axis=2), mode='mirror'
+                    image, np.expand_dims(k_shifted, axis=2), mode="mirror"
                 )
                 image = image[0::sf, 0::sf, ...]  # nearest downsampling
 
@@ -711,29 +687,29 @@ def degradation_bsrgan_variant(image, sf=4, isp_model=None):
     # add final JPEG compression noise
     image = add_JPEG_noise(image)
     image = util.single2uint(image)
-    example = {'image': image}
+    example = {"image": image}
     return example
 
 
-if __name__ == '__main__':
-    print('hey')
-    img = util.imread_uint('utils/test.png', 3)
+if __name__ == "__main__":
+    print("hey")
+    img = util.imread_uint("utils/test.png", 3)
     img = img[:448, :448]
     h = img.shape[0] // 4
-    print('resizing to', h)
+    print("resizing to", h)
     sf = 4
     deg_fn = partial(degradation_bsrgan_variant, sf=sf)
     for i in range(20):
         print(i)
         img_hq = img
-        img_lq = deg_fn(img)['image']
+        img_lq = deg_fn(img)["image"]
         img_hq, img_lq = util.uint2single(img_hq), util.uint2single(img_lq)
         print(img_lq)
         img_lq_bicubic = albumentations.SmallestMaxSize(
             max_size=h, interpolation=cv2.INTER_CUBIC
-        )(image=img_hq)['image']
+        )(image=img_hq)["image"]
         print(img_lq.shape)
-        print('bicubic', img_lq_bicubic.shape)
+        print("bicubic", img_lq_bicubic.shape)
         print(img_hq.shape)
         lq_nearest = cv2.resize(
             util.single2uint(img_lq),
@@ -748,4 +724,4 @@ if __name__ == '__main__':
         img_concat = np.concatenate(
             [lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1
         )
-        util.imsave(img_concat, str(i) + '.png')
+        util.imsave(img_concat, str(i) + ".png")

invokeai/backend/stable_diffusion/image_degradation/utils_image.py

@@ -1,16 +1,17 @@
-import os
 import math
+import os
 import random
+from datetime import datetime
+
+import cv2
 import numpy as np
 import torch
-import cv2
 from torchvision.utils import make_grid
-from datetime import datetime
 
 # import matplotlib.pyplot as plt   # TODO: check with Dominik, also bsrgan.py vs bsrgan_light.py
 
 
-os.environ['KMP_DUPLICATE_LIB_OK'] = 'TRUE'
+os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
 
 
 """
@@ -25,17 +26,17 @@ os.environ['KMP_DUPLICATE_LIB_OK'] = 'TRUE'
 
 
 IMG_EXTENSIONS = [
-    '.jpg',
-    '.JPG',
-    '.jpeg',
-    '.JPEG',
-    '.png',
-    '.PNG',
-    '.ppm',
-    '.PPM',
-    '.bmp',
-    '.BMP',
-    '.tif',
+    ".jpg",
+    ".JPG",
+    ".jpeg",
+    ".JPEG",
+    ".png",
+    ".PNG",
+    ".ppm",
+    ".PPM",
+    ".bmp",
+    ".BMP",
+    ".tif",
 ]
 
 
@@ -44,12 +45,12 @@ def is_image_file(filename):
 
 
 def get_timestamp():
-    return datetime.now().strftime('%y%m%d-%H%M%S')
+    return datetime.now().strftime("%y%m%d-%H%M%S")
 
 
 def imshow(x, title=None, cbar=False, figsize=None):
     plt.figure(figsize=figsize)
-    plt.imshow(np.squeeze(x), interpolation='nearest', cmap='gray')
+    plt.imshow(np.squeeze(x), interpolation="nearest", cmap="gray")
     if title:
         plt.title(title)
     if cbar:
@@ -57,9 +58,9 @@ def imshow(x, title=None, cbar=False, figsize=None):
     plt.show()
 
 
-def surf(Z, cmap='rainbow', figsize=None):
+def surf(Z, cmap="rainbow", figsize=None):
     plt.figure(figsize=figsize)
-    ax3 = plt.axes(projection='3d')
+    ax3 = plt.axes(projection="3d")
 
     w, h = Z.shape[:2]
     xx = np.arange(0, w, 1)
@@ -85,14 +86,14 @@ def get_image_paths(dataroot):
 
 
 def _get_paths_from_images(path):
-    assert os.path.isdir(path), '{:s} is not a valid directory'.format(path)
+    assert os.path.isdir(path), "{:s} is not a valid directory".format(path)
     images = []
     for dirpath, _, fnames in sorted(os.walk(path)):
         for fname in sorted(fnames):
             if is_image_file(fname):
                 img_path = os.path.join(dirpath, fname)
                 images.append(img_path)
-    assert images, '{:s} has no valid image file'.format(path)
+    assert images, "{:s} has no valid image file".format(path)
     return images
 
 
@@ -133,7 +134,7 @@ def imssave(imgs, img_path):
             img = img[:, :, [2, 1, 0]]
         new_path = os.path.join(
             os.path.dirname(img_path),
-            img_name + str('_s{:04d}'.format(i)) + '.png',
+            img_name + str("_s{:04d}".format(i)) + ".png",
         )
         cv2.imwrite(new_path, img)
 
@@ -162,9 +163,7 @@ def split_imageset(
         # img_name, ext = os.path.splitext(os.path.basename(img_path))
         img = imread_uint(img_path, n_channels=n_channels)
         patches = patches_from_image(img, p_size, p_overlap, p_max)
-        imssave(
-            patches, os.path.join(taget_dataroot, os.path.basename(img_path))
-        )
+        imssave(patches, os.path.join(taget_dataroot, os.path.basename(img_path)))
         # if original_dataroot == taget_dataroot:
         # del img_path
 
@@ -191,8 +190,8 @@ def mkdirs(paths):
 
 def mkdir_and_rename(path):
     if os.path.exists(path):
-        new_name = path + '_archived_' + get_timestamp()
-        print('Path already exists. Rename it to [{:s}]'.format(new_name))
+        new_name = path + "_archived_" + get_timestamp()
+        print("Path already exists. Rename it to [{:s}]".format(new_name))
         os.replace(path, new_name)
     os.makedirs(path)
 
@@ -273,22 +272,18 @@ def read_img(path):
 
 
 def uint2single(img):
-
     return np.float32(img / 255.0)
 
 
 def single2uint(img):
-
     return np.uint8((img.clip(0, 1) * 255.0).round())
 
 
 def uint162single(img):
-
     return np.float32(img / 65535.0)
 
 
 def single2uint16(img):
-
     return np.uint16((img.clip(0, 1) * 65535.0).round())
 
 
@@ -315,10 +310,7 @@ def uint2tensor3(img):
     if img.ndim == 2:
         img = np.expand_dims(img, axis=2)
     return (
-        torch.from_numpy(np.ascontiguousarray(img))
-        .permute(2, 0, 1)
-        .float()
-        .div(255.0)
+        torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.0)
     )
 
 
@@ -379,18 +371,11 @@ def single2tensor5(img):
 
 
 def single32tensor5(img):
-    return (
-        torch.from_numpy(np.ascontiguousarray(img))
-        .float()
-        .unsqueeze(0)
-        .unsqueeze(0)
-    )
+    return torch.from_numpy(np.ascontiguousarray(img)).float().unsqueeze(0).unsqueeze(0)
 
 
 def single42tensor4(img):
-    return (
-        torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float()
-    )
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float()
 
 
 # from skimage.io import imread, imsave
@@ -403,15 +388,11 @@ def tensor2img(tensor, out_type=np.uint8, min_max=(0, 1)):
     tensor = (
         tensor.squeeze().float().cpu().clamp_(*min_max)
     )  # squeeze first, then clamp
-    tensor = (tensor - min_max[0]) / (
-        min_max[1] - min_max[0]
-    )  # to range [0,1]
+    tensor = (tensor - min_max[0]) / (min_max[1] - min_max[0])  # to range [0,1]
     n_dim = tensor.dim()
     if n_dim == 4:
         n_img = len(tensor)
-        img_np = make_grid(
-            tensor, nrow=int(math.sqrt(n_img)), normalize=False
-        ).numpy()
+        img_np = make_grid(tensor, nrow=int(math.sqrt(n_img)), normalize=False).numpy()
         img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0))  # HWC, BGR
     elif n_dim == 3:
         img_np = tensor.numpy()
@@ -420,7 +401,7 @@ def tensor2img(tensor, out_type=np.uint8, min_max=(0, 1)):
         img_np = tensor.numpy()
     else:
         raise TypeError(
-            'Only support 4D, 3D and 2D tensor. But received with dimension: {:d}'.format(
+            "Only support 4D, 3D and 2D tensor. But received with dimension: {:d}".format(
                 n_dim
             )
         )
@@ -564,7 +545,7 @@ def modcrop(img_in, scale):
         H_r, W_r = H % scale, W % scale
         img = img[: H - H_r, : W - W_r, :]
     else:
-        raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim))
+        raise ValueError("Wrong img ndim: [{:d}].".format(img.ndim))
     return img
 
 
@@ -675,13 +656,13 @@ def bgr2ycbcr(img, only_y=True):
 
 def channel_convert(in_c, tar_type, img_list):
     # conversion among BGR, gray and y
-    if in_c == 3 and tar_type == 'gray':  # BGR to gray
+    if in_c == 3 and tar_type == "gray":  # BGR to gray
         gray_list = [cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) for img in img_list]
         return [np.expand_dims(img, axis=2) for img in gray_list]
-    elif in_c == 3 and tar_type == 'y':  # BGR to y
+    elif in_c == 3 and tar_type == "y":  # BGR to y
         y_list = [bgr2ycbcr(img, only_y=True) for img in img_list]
         return [np.expand_dims(img, axis=2) for img in y_list]
-    elif in_c == 1 and tar_type == 'RGB':  # gray/y to BGR
+    elif in_c == 1 and tar_type == "RGB":  # gray/y to BGR
         return [cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) for img in img_list]
     else:
         return img_list
@@ -702,7 +683,7 @@ def calculate_psnr(img1, img2, border=0):
     # img1 = img1.squeeze()
     # img2 = img2.squeeze()
     if not img1.shape == img2.shape:
-        raise ValueError('Input images must have the same dimensions.')
+        raise ValueError("Input images must have the same dimensions.")
     h, w = img1.shape[:2]
     img1 = img1[border : h - border, border : w - border]
     img2 = img2[border : h - border, border : w - border]
@@ -711,7 +692,7 @@ def calculate_psnr(img1, img2, border=0):
     img2 = img2.astype(np.float64)
     mse = np.mean((img1 - img2) ** 2)
     if mse == 0:
-        return float('inf')
+        return float("inf")
     return 20 * math.log10(255.0 / math.sqrt(mse))
 
 
@@ -726,7 +707,7 @@ def calculate_ssim(img1, img2, border=0):
     # img1 = img1.squeeze()
     # img2 = img2.squeeze()
     if not img1.shape == img2.shape:
-        raise ValueError('Input images must have the same dimensions.')
+        raise ValueError("Input images must have the same dimensions.")
     h, w = img1.shape[:2]
     img1 = img1[border : h - border, border : w - border]
     img2 = img2[border : h - border, border : w - border]
@@ -742,7 +723,7 @@ def calculate_ssim(img1, img2, border=0):
         elif img1.shape[2] == 1:
             return ssim(np.squeeze(img1), np.squeeze(img2))
     else:
-        raise ValueError('Wrong input image dimensions.')
+        raise ValueError("Wrong input image dimensions.")
 
 
 def ssim(img1, img2):
@@ -861,7 +842,7 @@ def imresize(img, scale, antialiasing=True):
         math.ceil(in_W * scale),
     )
     kernel_width = 4
-    kernel = 'cubic'
+    kernel = "cubic"
 
     # Return the desired dimension order for performing the resize.  The
     # strategy is to perform the resize first along the dimension with the
@@ -896,9 +877,7 @@ def imresize(img, scale, antialiasing=True):
         idx = int(indices_H[i][0])
         for j in range(out_C):
             out_1[j, i, :] = (
-                img_aug[j, idx : idx + kernel_width, :]
-                .transpose(0, 1)
-                .mv(weights_H[i])
+                img_aug[j, idx : idx + kernel_width, :].transpose(0, 1).mv(weights_H[i])
             )
 
     # process W dimension
@@ -921,9 +900,7 @@ def imresize(img, scale, antialiasing=True):
     for i in range(out_W):
         idx = int(indices_W[i][0])
         for j in range(out_C):
-            out_2[j, :, i] = out_1_aug[j, :, idx : idx + kernel_width].mv(
-                weights_W[i]
-            )
+            out_2[j, :, i] = out_1_aug[j, :, idx : idx + kernel_width].mv(weights_W[i])
     if need_squeeze:
         out_2.squeeze_()
     return out_2
@@ -948,7 +925,7 @@ def imresize_np(img, scale, antialiasing=True):
         math.ceil(in_W * scale),
     )
     kernel_width = 4
-    kernel = 'cubic'
+    kernel = "cubic"
 
     # Return the desired dimension order for performing the resize.  The
     # strategy is to perform the resize first along the dimension with the
@@ -983,9 +960,7 @@ def imresize_np(img, scale, antialiasing=True):
         idx = int(indices_H[i][0])
         for j in range(out_C):
             out_1[i, :, j] = (
-                img_aug[idx : idx + kernel_width, :, j]
-                .transpose(0, 1)
-                .mv(weights_H[i])
+                img_aug[idx : idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i])
             )
 
     # process W dimension
@@ -1008,17 +983,15 @@ def imresize_np(img, scale, antialiasing=True):
     for i in range(out_W):
         idx = int(indices_W[i][0])
         for j in range(out_C):
-            out_2[:, i, j] = out_1_aug[:, idx : idx + kernel_width, j].mv(
-                weights_W[i]
-            )
+            out_2[:, i, j] = out_1_aug[:, idx : idx + kernel_width, j].mv(weights_W[i])
     if need_squeeze:
         out_2.squeeze_()
 
     return out_2.numpy()
 
 
-if __name__ == '__main__':
-    print('---')
+if __name__ == "__main__":
+    print("---")
 #    img = imread_uint('test.bmp', 3)
 #    img = uint2single(img)
 #    img_bicubic = imresize_np(img, 1/4)

invokeai/backend/stable_diffusion/offloading.py

@@ -12,6 +12,7 @@ from torch.utils.hooks import RemovableHandle
 
 OFFLOAD_DEVICE = torch.device("cpu")
 
+
 class _NoModel:
     """Symbol that indicates no model is loaded.
 
@@ -28,6 +29,7 @@ class _NoModel:
     def __repr__(self):
         return "<NO MODEL>"
 
+
 NO_MODEL = _NoModel()
 
 
@@ -93,8 +95,10 @@ class ModelGroup(metaclass=ABCMeta):
         pass
 
     def __repr__(self) -> str:
-        return f"<{self.__class__.__name__} object at {id(self):x}: " \
-               f"device={self.execution_device} >"
+        return (
+            f"<{self.__class__.__name__} object at {id(self):x}: "
+            f"device={self.execution_device} >"
+        )
 
 
 class LazilyLoadedModelGroup(ModelGroup):
@@ -138,8 +142,11 @@ class LazilyLoadedModelGroup(ModelGroup):
     def _pre_hook(self, module: torch.nn.Module, forward_input):
         self.load(module)
         if len(forward_input) == 0:
-            warnings.warn(f"Hook for {module.__class__.__name__} got no input. "
-                          f"Inputs must be positional, not keywords.", stacklevel=3)
+            warnings.warn(
+                f"Hook for {module.__class__.__name__} got no input. "
+                f"Inputs must be positional, not keywords.",
+                stacklevel=3,
+            )
         return send_to_device(forward_input, self.execution_device)
 
     def load(self, module):
@@ -154,7 +161,9 @@ class LazilyLoadedModelGroup(ModelGroup):
         self.clear_current_model()
 
     def _load(self, module: torch.nn.Module) -> torch.nn.Module:
-        assert self.is_empty(), f"A model is already loaded: {self._current_model_ref()}"
+        assert (
+            self.is_empty()
+        ), f"A model is already loaded: {self._current_model_ref()}"
         module = module.to(self.execution_device)
         self.set_current_model(module)
         return module
@@ -183,8 +192,12 @@ class LazilyLoadedModelGroup(ModelGroup):
 
     def device_for(self, model):
         if model not in self:
-            raise KeyError(f"This does not manage this model {type(model).__name__}", model)
-        return self.execution_device  # this implementation only dispatches to one device
+            raise KeyError(
+                f"This does not manage this model {type(model).__name__}", model
+            )
+        return (
+            self.execution_device
+        )  # this implementation only dispatches to one device
 
     def ready(self):
         pass  # always ready to load on-demand
@@ -193,8 +206,10 @@ class LazilyLoadedModelGroup(ModelGroup):
         return model in self._hooks
 
     def __repr__(self) -> str:
-        return f"<{self.__class__.__name__} object at {id(self):x}: " \
-               f"current_model={type(self._current_model_ref()).__name__} >"
+        return (
+            f"<{self.__class__.__name__} object at {id(self):x}: "
+            f"current_model={type(self._current_model_ref()).__name__} >"
+        )
 
 
 class FullyLoadedModelGroup(ModelGroup):
@@ -203,6 +218,7 @@ class FullyLoadedModelGroup(ModelGroup):
 
     :py:meth:`.ready` loads _all_ the models to the execution device at once.
     """
+
     _models: weakref.WeakSet
 
     def __init__(self, execution_device: torch.device):
@@ -240,8 +256,12 @@ class FullyLoadedModelGroup(ModelGroup):
 
     def device_for(self, model):
         if model not in self:
-            raise KeyError("This does not manage this model f{type(model).__name__}", model)
-        return self.execution_device  # this implementation only dispatches to one device
+            raise KeyError(
+                "This does not manage this model f{type(model).__name__}", model
+            )
+        return (
+            self.execution_device
+        )  # this implementation only dispatches to one device
 
     def __contains__(self, model):
         return model in self._models