Enable LoRAPatcher.apply_smart_lora_patches(...) throughout the stack.

Currently translated at 99.3% (1633 of 1643 strings)

Co-authored-by: Riccardo Giovanetti <riccardo.giovanetti@gmail.com>
Translate-URL: https://hosted.weblate.org/projects/invokeai/web-ui/it/
Translation: InvokeAI/Web UI

docker/Dockerfile

@@ -2,29 +2,42 @@
 
 ## Builder stage
 
-FROM library/ubuntu:23.04 AS builder
+FROM library/ubuntu:24.04 AS builder
 
 ARG DEBIAN_FRONTEND=noninteractive
 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
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt,sharing=locked \
     apt update && apt-get install -y \
-        git \
-        python3-venv \
-        python3-pip \
-        build-essential
+        build-essential \
+        git
 
-ENV INVOKEAI_SRC=/opt/invokeai
-ENV VIRTUAL_ENV=/opt/venv/invokeai
+# Install `uv` for package management
+COPY --from=ghcr.io/astral-sh/uv:0.5.5 /uv /uvx /bin/
 
+ENV VIRTUAL_ENV=/opt/venv
 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+ENV INVOKEAI_SRC=/opt/invokeai
+ENV PYTHON_VERSION=3.11
+ENV UV_COMPILE_BYTECODE=1
+ENV UV_LINK_MODE=copy
+
 ARG GPU_DRIVER=cuda
 ARG TARGETPLATFORM="linux/amd64"
 # unused but available
 ARG BUILDPLATFORM
 
-WORKDIR ${INVOKEAI_SRC}
+# Switch to the `ubuntu` user to work around dependency issues with uv-installed python
+RUN mkdir -p ${VIRTUAL_ENV} && \
+    mkdir -p ${INVOKEAI_SRC} && \
+    chmod -R a+w /opt
+USER ubuntu
 
+# Install python and create the venv
+RUN uv python install ${PYTHON_VERSION} && \
+    uv venv --relocatable --prompt "invoke" --python ${PYTHON_VERSION} ${VIRTUAL_ENV}
+
+WORKDIR ${INVOKEAI_SRC}
 COPY invokeai ./invokeai
 COPY pyproject.toml ./
 
@@ -32,25 +45,18 @@ COPY pyproject.toml ./
 # the local working copy can be bind-mounted into the image
 # at path defined by ${INVOKEAI_SRC}
 # NOTE: there are no pytorch builds for arm64 + cuda, only cpu
-# x86_64/CUDA is default
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m venv ${VIRTUAL_ENV} &&\
+# x86_64/CUDA is the default
+RUN --mount=type=cache,target=/home/ubuntu/.cache/uv,uid=1000,gid=1000 \
     if [ "$TARGETPLATFORM" = "linux/arm64" ] || [ "$GPU_DRIVER" = "cpu" ]; then \
         extra_index_url_arg="--extra-index-url https://download.pytorch.org/whl/cpu"; \
     elif [ "$GPU_DRIVER" = "rocm" ]; then \
         extra_index_url_arg="--extra-index-url https://download.pytorch.org/whl/rocm6.1"; \
     else \
         extra_index_url_arg="--extra-index-url https://download.pytorch.org/whl/cu124"; \
-    fi &&\
+    fi && \
+    uv pip install --python ${PYTHON_VERSION} $extra_index_url_arg -e "."
 
-    # xformers + triton fails to install on arm64
-    if [ "$GPU_DRIVER" = "cuda" ] && [ "$TARGETPLATFORM" = "linux/amd64" ]; then \
-        pip install $extra_index_url_arg -e ".[xformers]"; \
-    else \
-        pip install $extra_index_url_arg -e "."; \
-    fi
-
-# #### Build the Web UI ------------------------------------
+#### Build the Web UI ------------------------------------
 
 FROM node:20-slim AS web-builder
 ENV PNPM_HOME="/pnpm"
@@ -66,7 +72,7 @@ RUN npx vite build
 
 #### Runtime stage ---------------------------------------
 
-FROM library/ubuntu:23.04 AS runtime
+FROM library/ubuntu:24.04 AS runtime
 
 ARG DEBIAN_FRONTEND=noninteractive
 ENV PYTHONUNBUFFERED=1
@@ -83,17 +89,16 @@ RUN apt update && apt install -y --no-install-recommends \
         gosu \
         magic-wormhole \
         libglib2.0-0 \
-        libgl1-mesa-glx \
-        python3-venv \
-        python3-pip \
+        libgl1 \
+        libglx-mesa0 \
         build-essential \
         libopencv-dev \
         libstdc++-10-dev &&\
     apt-get clean && apt-get autoclean
 
-
 ENV INVOKEAI_SRC=/opt/invokeai
-ENV VIRTUAL_ENV=/opt/venv/invokeai
+ENV VIRTUAL_ENV=/opt/venv
+ENV PYTHON_VERSION=3.11
 ENV INVOKEAI_ROOT=/invokeai
 ENV INVOKEAI_HOST=0.0.0.0
 ENV INVOKEAI_PORT=9090
@@ -101,6 +106,14 @@ ENV PATH="$VIRTUAL_ENV/bin:$INVOKEAI_SRC:$PATH"
 ENV CONTAINER_UID=${CONTAINER_UID:-1000}
 ENV CONTAINER_GID=${CONTAINER_GID:-1000}
 
+# Install `uv` for package management
+# and install python for the ubuntu user (expected to exist on ubuntu >=24.x)
+# this is too tiny to optimize with multi-stage builds, but maybe we'll come back to it
+COPY --from=ghcr.io/astral-sh/uv:0.5.5 /uv /uvx /bin/
+USER ubuntu
+RUN uv python install ${PYTHON_VERSION}
+USER root
+
 # --link requires buldkit w/ dockerfile syntax 1.4
 COPY --link --from=builder ${INVOKEAI_SRC} ${INVOKEAI_SRC}
 COPY --link --from=builder ${VIRTUAL_ENV} ${VIRTUAL_ENV}
@@ -115,7 +128,7 @@ WORKDIR ${INVOKEAI_SRC}
 
 # build patchmatch
 RUN cd /usr/lib/$(uname -p)-linux-gnu/pkgconfig/ && ln -sf opencv4.pc opencv.pc
-RUN python3 -c "from patchmatch import patch_match"
+RUN python -c "from patchmatch import patch_match"
 
 RUN mkdir -p ${INVOKEAI_ROOT} && chown -R ${CONTAINER_UID}:${CONTAINER_GID} ${INVOKEAI_ROOT}
 

docker/docker-entrypoint.sh

@@ -16,6 +16,9 @@ set -e -o pipefail
 
 USER_ID=${CONTAINER_UID:-1000}
 USER=ubuntu
+# if the user does not exist, create it. It is expected to be present on ubuntu >=24.x
+_=$(id ${USER} 2>&1) || useradd -u ${USER_ID} ${USER}
+# ensure the UID is correct
 usermod -u ${USER_ID} ${USER} 1>/dev/null
 
 ### Set the $PUBLIC_KEY env var to enable SSH access.
@@ -36,6 +39,8 @@ fi
 mkdir -p "${INVOKEAI_ROOT}"
 chown --recursive ${USER} "${INVOKEAI_ROOT}" || true
 cd "${INVOKEAI_ROOT}"
+export HF_HOME=${HF_HOME:-$INVOKEAI_ROOT/.cache/huggingface}
+export MPLCONFIGDIR=${MPLCONFIGDIR:-$INVOKEAI_ROOT/.matplotlib}
 
 # Run the CMD as the Container User (not root).
 exec gosu ${USER} "$@"

invokeai/app/invocations/__init__.py

@@ -15,6 +15,11 @@ custom_nodes_readme_path = str(custom_nodes_path / "README.md")
 shutil.copy(Path(__file__).parent / "custom_nodes/init.py", custom_nodes_init_path)
 shutil.copy(Path(__file__).parent / "custom_nodes/README.md", custom_nodes_readme_path)
 
+# set the same permissions as the destination directory, in case our source is read-only,
+# so that the files are user-writable
+for p in custom_nodes_path.glob("**/*"):
+    p.chmod(custom_nodes_path.stat().st_mode)
+
 # Import custom nodes, see https://docs.python.org/3/library/importlib.html#importing-programmatically
 spec = spec_from_file_location("custom_nodes", custom_nodes_init_path)
 if spec is None or spec.loader is None:

invokeai/app/invocations/compel.py

@@ -82,10 +82,11 @@ class CompelInvocation(BaseInvocation):
             # apply all patches while the model is on the target device
             text_encoder_info.model_on_device() as (cached_weights, text_encoder),
             tokenizer_info as tokenizer,
-            LoRAPatcher.apply_lora_patches(
+            LoRAPatcher.apply_smart_lora_patches(
                 model=text_encoder,
                 patches=_lora_loader(),
                 prefix="lora_te_",
+                dtype=TorchDevice.choose_torch_dtype(),
                 cached_weights=cached_weights,
             ),
             # Apply CLIP Skip after LoRA to prevent LoRA application from failing on skipped layers.
@@ -179,10 +180,11 @@ class SDXLPromptInvocationBase:
             # apply all patches while the model is on the target device
             text_encoder_info.model_on_device() as (cached_weights, text_encoder),
             tokenizer_info as tokenizer,
-            LoRAPatcher.apply_lora_patches(
+            LoRAPatcher.apply_smart_lora_patches(
                 text_encoder,
                 patches=_lora_loader(),
                 prefix=lora_prefix,
+                dtype=TorchDevice.choose_torch_dtype(),
                 cached_weights=cached_weights,
             ),
             # Apply CLIP Skip after LoRA to prevent LoRA application from failing on skipped layers.

invokeai/app/invocations/denoise_latents.py

@@ -1003,10 +1003,11 @@ class DenoiseLatentsInvocation(BaseInvocation):
             ModelPatcher.apply_freeu(unet, self.unet.freeu_config),
             SeamlessExt.static_patch_model(unet, self.unet.seamless_axes),  # FIXME
             # Apply the LoRA after unet has been moved to its target device for faster patching.
-            LoRAPatcher.apply_lora_patches(
+            LoRAPatcher.apply_smart_lora_patches(
                 model=unet,
                 patches=_lora_loader(),
                 prefix="lora_unet_",
+                dtype=unet.dtype,
                 cached_weights=cached_weights,
             ),
         ):

invokeai/app/invocations/fields.py

@@ -56,6 +56,7 @@ class UIType(str, Enum, metaclass=MetaEnum):
     CLIPLEmbedModel = "CLIPLEmbedModelField"
     CLIPGEmbedModel = "CLIPGEmbedModelField"
     SpandrelImageToImageModel = "SpandrelImageToImageModelField"
+    StructuralLoRAModel = "StructuralLoRAModelField"
     # endregion
 
     # region Misc Field Types
@@ -143,6 +144,7 @@ class FieldDescriptions:
     controlnet_model = "ControlNet model to load"
     vae_model = "VAE model to load"
     lora_model = "LoRA model to load"
+    structural_lora_model = "Structural LoRA model to load"
     main_model = "Main model (UNet, VAE, CLIP) to load"
     flux_model = "Flux model (Transformer) to load"
     sd3_model = "SD3 model (MMDiTX) to load"

invokeai/app/invocations/flux_denoise.py

@@ -1,5 +1,5 @@
 from contextlib import ExitStack
-from typing import Callable, Iterator, Optional, Tuple
+from typing import Callable, Iterator, Optional, Tuple, Union
 
 import numpy as np
 import numpy.typing as npt
@@ -8,6 +8,8 @@ import torchvision.transforms as tv_transforms
 from torchvision.transforms.functional import resize as tv_resize
 from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
 
+from invokeai.backend.flux.modules.autoencoder import AutoEncoder
+
 from invokeai.app.invocations.baseinvocation import BaseInvocation, Classification, invocation
 from invokeai.app.invocations.fields import (
     DenoiseMaskField,
@@ -22,7 +24,7 @@ from invokeai.app.invocations.fields import (
 )
 from invokeai.app.invocations.flux_controlnet import FluxControlNetField
 from invokeai.app.invocations.ip_adapter import IPAdapterField
-from invokeai.app.invocations.model import TransformerField, VAEField
+from invokeai.app.invocations.model import TransformerField, VAEField, StructuralLoRAField, LoRAField
 from invokeai.app.invocations.primitives import LatentsOutput
 from invokeai.app.services.shared.invocation_context import InvocationContext
 from invokeai.backend.flux.controlnet.instantx_controlnet_flux import InstantXControlNetFlux
@@ -43,6 +45,8 @@ from invokeai.backend.flux.sampling_utils import (
     pack,
     unpack,
 )
+from invokeai.backend.flux.flux_tools_sampling_utils import prepare_control
+from invokeai.backend.flux.modules.conditioner import HFEncoder
 from invokeai.backend.flux.text_conditioning import FluxTextConditioning
 from invokeai.backend.lora.conversions.flux_lora_constants import FLUX_LORA_TRANSFORMER_PREFIX
 from invokeai.backend.lora.lora_model_raw import LoRAModelRaw
@@ -284,6 +288,16 @@ class FluxDenoiseInvocation(BaseInvocation, WithMetadata, WithBoard):
                 dtype=inference_dtype,
                 device=x.device,
             )
+            img_cond = None
+            if struct_lora := self.transformer.structural_lora:
+                # What should we do when we have multiple of these?
+                if not self.controlnet_vae:
+                    raise ValueError("controlnet_vae must be set when using a strutural lora")
+                ae_info = context.models.load(self.controlnet_vae.vae)
+                img = context.images.get_pil(struct_lora.img.image_name)
+                with ae_info as ae:
+                    assert isinstance(ae, AutoEncoder)
+                    img_cond = prepare_control(self.height, self.width, self.seed, ae, img)
 
             # Load the transformer model.
             (cached_weights, transformer) = exit_stack.enter_context(transformer_info.model_on_device())
@@ -296,10 +310,11 @@ class FluxDenoiseInvocation(BaseInvocation, WithMetadata, WithBoard):
             if config.format in [ModelFormat.Checkpoint]:
                 # The model is non-quantized, so we can apply the LoRA weights directly into the model.
                 exit_stack.enter_context(
-                    LoRAPatcher.apply_lora_patches(
+                    LoRAPatcher.apply_smart_lora_patches(
                         model=transformer,
                         patches=self._lora_iterator(context),
                         prefix=FLUX_LORA_TRANSFORMER_PREFIX,
+                        dtype=inference_dtype,
                         cached_weights=cached_weights,
                     )
                 )
@@ -311,7 +326,7 @@ class FluxDenoiseInvocation(BaseInvocation, WithMetadata, WithBoard):
                 # The model is quantized, so apply the LoRA weights as sidecar layers. This results in slower inference,
                 # than directly patching the weights, but is agnostic to the quantization format.
                 exit_stack.enter_context(
-                    LoRAPatcher.apply_lora_sidecar_patches(
+                    LoRAPatcher.apply_lora_wrapper_patches(
                         model=transformer,
                         patches=self._lora_iterator(context),
                         prefix=FLUX_LORA_TRANSFORMER_PREFIX,
@@ -345,6 +360,7 @@ class FluxDenoiseInvocation(BaseInvocation, WithMetadata, WithBoard):
                 controlnet_extensions=controlnet_extensions,
                 pos_ip_adapter_extensions=pos_ip_adapter_extensions,
                 neg_ip_adapter_extensions=neg_ip_adapter_extensions,
+                img_cond=img_cond
             )
 
         x = unpack(x.float(), self.height, self.width)
@@ -682,7 +698,10 @@ class FluxDenoiseInvocation(BaseInvocation, WithMetadata, WithBoard):
         return pos_ip_adapter_extensions, neg_ip_adapter_extensions
 
     def _lora_iterator(self, context: InvocationContext) -> Iterator[Tuple[LoRAModelRaw, float]]:
-        for lora in self.transformer.loras:
+        loras: list[Union[LoRAField, StructuralLoRAField]] = [*self.transformer.loras]
+        if self.transformer.structural_lora:
+            loras.append(self.transformer.structural_lora)
+        for lora in loras:
             lora_info = context.models.load(lora.lora)
             assert isinstance(lora_info.model, LoRAModelRaw)
             yield (lora_info.model, lora.weight)

invokeai/app/invocations/flux_model_loader.py

@@ -81,8 +81,8 @@ class FluxModelLoaderInvocation(BaseInvocation):
         assert isinstance(transformer_config, CheckpointConfigBase)
 
         return FluxModelLoaderOutput(
-            transformer=TransformerField(transformer=transformer, loras=[]),
-            clip=CLIPField(tokenizer=tokenizer, text_encoder=clip_encoder, loras=[], skipped_layers=0),
+            transformer=TransformerField(transformer=transformer, loras=[], structural_loras=[]),
+            clip=CLIPField(tokenizer=tokenizer, text_encoder=clip_encoder, loras=[], structural_loras=[], skipped_layers=0),
             t5_encoder=T5EncoderField(tokenizer=tokenizer2, text_encoder=t5_encoder),
             vae=VAEField(vae=vae),
             max_seq_len=max_seq_lengths[transformer_config.config_path],

invokeai/app/invocations/flux_structural_lora_loader.py

@@ -0,0 +1,70 @@
+from typing import Optional, Literal
+
+from invokeai.app.invocations.baseinvocation import (
+    BaseInvocation,
+    BaseInvocationOutput,
+    Classification,
+    invocation,
+    invocation_output,
+)
+from invokeai.app.invocations.fields import FieldDescriptions, Input, InputField, OutputField, UIType, ImageField
+from invokeai.app.invocations.model import VAEField, StructuralLoRAField, ModelIdentifierField, TransformerField
+from invokeai.app.services.shared.invocation_context import InvocationContext
+
+
+@invocation_output("flux_structural_lora_loader_output")
+class FluxStructuralLoRALoaderOutput(BaseInvocationOutput):
+    """Flux Structural LoRA Loader Output"""
+
+    transformer: Optional[TransformerField] = OutputField(
+        default=None, description=FieldDescriptions.transformer, title="FLUX Transformer"
+    )
+
+
+@invocation(
+    "flux_structural_lora_loader",
+    title="Flux Structural LoRA",
+    tags=["lora", "model", "flux"],
+    category="model",
+    version="1.1.0",
+    classification=Classification.Prototype,
+)
+class FluxStructuralLoRALoaderInvocation(BaseInvocation):
+    """Apply a LoRA model to a FLUX transformer and/or text encoder."""
+
+    lora: ModelIdentifierField = InputField(
+        description=FieldDescriptions.structural_lora_model, title="Structural LoRA", ui_type=UIType.StructuralLoRAModel
+    )
+    transformer: TransformerField | None = InputField(
+        default=None,
+        description=FieldDescriptions.transformer,
+        input=Input.Connection,
+        title="FLUX Transformer",
+    )
+    image: ImageField = InputField(
+        description="The image to encode.",
+    )
+    weight: float = InputField(default=0.75, description=FieldDescriptions.lora_weight)
+
+    def invoke(self, context: InvocationContext) -> FluxStructuralLoRALoaderOutput:
+        lora_key = self.lora.key
+
+        if not context.models.exists(lora_key):
+            raise ValueError(f"Unknown lora: {lora_key}!")
+
+        # Check for existing LoRAs with the same key.
+        if self.transformer and self.transformer.structural_lora and self.transformer.structural_lora.lora.key == lora_key:
+            raise ValueError(f'Structural LoRA "{lora_key}" already applied to transformer.')
+
+        output = FluxStructuralLoRALoaderOutput()
+
+        # Attach LoRA layers to the models.
+        if self.transformer is not None:
+            output.transformer = self.transformer.model_copy(deep=True)
+            output.transformer.structural_lora = StructuralLoRAField(
+                lora=self.lora,
+                img=self.image,
+                weight=self.weight,
+            )
+
+        return output

invokeai/app/invocations/flux_text_encoder.py

@@ -22,6 +22,7 @@ from invokeai.backend.lora.lora_model_raw import LoRAModelRaw
 from invokeai.backend.lora.lora_patcher import LoRAPatcher
 from invokeai.backend.model_manager.config import ModelFormat
 from invokeai.backend.stable_diffusion.diffusion.conditioning_data import ConditioningFieldData, FLUXConditioningInfo
+from invokeai.backend.util.devices import TorchDevice
 
 
 @invocation(
@@ -111,10 +112,11 @@ class FluxTextEncoderInvocation(BaseInvocation):
             if clip_text_encoder_config.format in [ModelFormat.Diffusers]:
                 # The model is non-quantized, so we can apply the LoRA weights directly into the model.
                 exit_stack.enter_context(
-                    LoRAPatcher.apply_lora_patches(
+                    LoRAPatcher.apply_smart_lora_patches(
                         model=clip_text_encoder,
                         patches=self._clip_lora_iterator(context),
                         prefix=FLUX_LORA_CLIP_PREFIX,
+                        dtype=TorchDevice.choose_torch_dtype(),
                         cached_weights=cached_weights,
                     )
                 )

invokeai/app/invocations/model.py

@@ -1,5 +1,5 @@
 import copy
-from typing import List, Optional
+from typing import List, Optional, Literal
 
 from pydantic import BaseModel, Field
 
@@ -10,7 +10,7 @@ from invokeai.app.invocations.baseinvocation import (
     invocation,
     invocation_output,
 )
-from invokeai.app.invocations.fields import FieldDescriptions, Input, InputField, OutputField, UIType
+from invokeai.app.invocations.fields import FieldDescriptions, Input, InputField, OutputField, UIType, ImageField
 from invokeai.app.services.shared.invocation_context import InvocationContext
 from invokeai.app.shared.models import FreeUConfig
 from invokeai.backend.model_manager.config import (
@@ -65,11 +65,6 @@ class CLIPField(BaseModel):
     loras: List[LoRAField] = Field(description="LoRAs to apply on model loading")
 
 
-class TransformerField(BaseModel):
-    transformer: ModelIdentifierField = Field(description="Info to load Transformer submodel")
-    loras: List[LoRAField] = Field(description="LoRAs to apply on model loading")
-
-
 class T5EncoderField(BaseModel):
     tokenizer: ModelIdentifierField = Field(description="Info to load tokenizer submodel")
     text_encoder: ModelIdentifierField = Field(description="Info to load text_encoder submodel")
@@ -79,6 +74,13 @@ class VAEField(BaseModel):
     vae: ModelIdentifierField = Field(description="Info to load vae submodel")
     seamless_axes: List[str] = Field(default_factory=list, description='Axes("x" and "y") to which apply seamless')
 
+class StructuralLoRAField(LoRAField):
+    img: ImageField = Field(description="Image to use in structural conditioning")
+
+class TransformerField(BaseModel):
+    transformer: ModelIdentifierField = Field(description="Info to load Transformer submodel")
+    loras: List[LoRAField] = Field(description="LoRAs to apply on model loading")
+    structural_lora: Optional[StructuralLoRAField] = Field(description="Structural LoRAs to apply on model loading", default=None)
 
 @invocation_output("unet_output")
 class UNetOutput(BaseInvocationOutput):

invokeai/app/invocations/sd3_text_encoder.py

@@ -21,6 +21,7 @@ from invokeai.backend.lora.lora_model_raw import LoRAModelRaw
 from invokeai.backend.lora.lora_patcher import LoRAPatcher
 from invokeai.backend.model_manager.config import ModelFormat
 from invokeai.backend.stable_diffusion.diffusion.conditioning_data import ConditioningFieldData, SD3ConditioningInfo
+from invokeai.backend.util.devices import TorchDevice
 
 # The SD3 T5 Max Sequence Length set based on the default in diffusers.
 SD3_T5_MAX_SEQ_LEN = 256
@@ -150,10 +151,11 @@ class Sd3TextEncoderInvocation(BaseInvocation):
             if clip_text_encoder_config.format in [ModelFormat.Diffusers]:
                 # The model is non-quantized, so we can apply the LoRA weights directly into the model.
                 exit_stack.enter_context(
-                    LoRAPatcher.apply_lora_patches(
+                    LoRAPatcher.apply_smart_lora_patches(
                         model=clip_text_encoder,
                         patches=self._clip_lora_iterator(context, clip_model),
                         prefix=FLUX_LORA_CLIP_PREFIX,
+                        dtype=TorchDevice.choose_torch_dtype(),
                         cached_weights=cached_weights,
                     )
                 )

invokeai/app/invocations/tiled_multi_diffusion_denoise_latents.py

@@ -207,7 +207,9 @@ class TiledMultiDiffusionDenoiseLatents(BaseInvocation):
         with (
             ExitStack() as exit_stack,
             unet_info as unet,
-            LoRAPatcher.apply_lora_patches(model=unet, patches=_lora_loader(), prefix="lora_unet_"),
+            LoRAPatcher.apply_smart_lora_patches(
+                model=unet, patches=_lora_loader(), prefix="lora_unet_", dtype=unet.dtype
+            ),
         ):
             assert isinstance(unet, UNet2DConditionModel)
             latents = latents.to(device=unet.device, dtype=unet.dtype)

invokeai/app/services/config/config_default.py

@@ -4,6 +4,7 @@
 from __future__ import annotations
 
 import copy
+import filecmp
 import locale
 import os
 import re
@@ -525,9 +526,35 @@ def get_config() -> InvokeAIAppConfig:
     ]
     example_config.write_file(config.config_file_path.with_suffix(".example.yaml"), as_example=True)
 
-    # Copy all legacy configs - We know `__path__[0]` is correct here
+    # Copy all legacy configs only if needed
+    # We know `__path__[0]` is correct here
     configs_src = Path(model_configs.__path__[0])  # pyright: ignore [reportUnknownMemberType, reportUnknownArgumentType, reportAttributeAccessIssue]
-    shutil.copytree(configs_src, config.legacy_conf_path, dirs_exist_ok=True)
+    dest_path = config.legacy_conf_path
+
+    # Create destination (we don't need to check for existence)
+    dest_path.mkdir(parents=True, exist_ok=True)
+
+    # Compare directories recursively
+    comparison = filecmp.dircmp(configs_src, dest_path)
+    need_copy = any(
+        [
+            comparison.left_only,  # Files exist only in source
+            comparison.diff_files,  # Files that differ
+            comparison.common_funny,  # Files that couldn't be compared
+        ]
+    )
+
+    if need_copy:
+        # Get permissions from destination directory
+        dest_mode = dest_path.stat().st_mode
+
+        # Copy directory tree
+        shutil.copytree(configs_src, dest_path, dirs_exist_ok=True)
+
+        # Set permissions on copied files to match destination directory
+        dest_path.chmod(dest_mode)
+        for p in dest_path.glob("**/*"):
+            p.chmod(dest_mode)
 
     if config.config_file_path.exists():
         config_from_file = load_and_migrate_config(config.config_file_path)

invokeai/app/services/session_processor/session_processor_default.py

@@ -379,7 +379,7 @@ class DefaultSessionProcessor(SessionProcessorBase):
 
     async def _on_queue_item_status_changed(self, event: FastAPIEvent[QueueItemStatusChangedEvent]) -> None:
         # Make sure the cancel event is for the currently processing queue item
-        if self._queue_item and self._queue_item.item_id is not event[1].item_id:
+        if self._queue_item and self._queue_item.item_id != event[1].item_id:
             return
         if self._queue_item and event[1].status in ["completed", "failed", "canceled"]:
             # When the queue item is canceled via HTTP, the queue item status is set to `"canceled"` and this event is

invokeai/app/services/shared/sqlite_migrator/migrations/migration_11.py

@@ -35,7 +35,7 @@ class Migration11Callback:
 
     def _remove_convert_cache(self) -> None:
         """Rename models/.cache to models/.convert_cache."""
-        self._logger.info("Removing .cache directory. Converted models will now be cached in .convert_cache.")
+        self._logger.info("Removing models/.cache directory. Converted models will now be cached in .convert_cache.")
         legacy_convert_path = self._app_config.root_path / "models" / ".cache"
         shutil.rmtree(legacy_convert_path, ignore_errors=True)
 

invokeai/backend/assets/model_base_conf_files/controlnet_sd15/config.json

@@ -1,42 +0,0 @@
-{
-  "_class_name": "ControlNetModel",
-  "_diffusers_version": "0.16.0.dev0",
-  "_name_or_path": "controlnet_v1_1/control_v11p_sd15_canny",
-  "act_fn": "silu",
-  "attention_head_dim": 8,
-  "block_out_channels": [
-    320,
-    640,
-    1280,
-    1280
-  ],
-  "class_embed_type": null,
-  "conditioning_embedding_out_channels": [
-    16,
-    32,
-    96,
-    256
-  ],
-  "controlnet_conditioning_channel_order": "rgb",
-  "cross_attention_dim": 768,
-  "down_block_types": [
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D",
-    "DownBlock2D"
-  ],
-  "downsample_padding": 1,
-  "flip_sin_to_cos": true,
-  "freq_shift": 0,
-  "in_channels": 4,
-  "layers_per_block": 2,
-  "mid_block_scale_factor": 1,
-  "norm_eps": 1e-05,
-  "norm_num_groups": 32,
-  "num_class_embeds": null,
-  "only_cross_attention": false,
-  "projection_class_embeddings_input_dim": null,
-  "resnet_time_scale_shift": "default",
-  "upcast_attention": false,
-  "use_linear_projection": false
-}

invokeai/backend/assets/model_base_conf_files/controlnet_sdxl/config.json

@@ -1,56 +0,0 @@
-{
-  "_class_name": "ControlNetModel",
-  "_diffusers_version": "0.19.3",
-  "act_fn": "silu",
-  "addition_embed_type": "text_time",
-  "addition_embed_type_num_heads": 64,
-  "addition_time_embed_dim": 256,
-  "attention_head_dim": [
-    5,
-    10,
-    20
-  ],
-  "block_out_channels": [
-    320,
-    640,
-    1280
-  ],
-  "class_embed_type": null,
-  "conditioning_channels": 3,
-  "conditioning_embedding_out_channels": [
-    16,
-    32,
-    96,
-    256
-  ],
-  "controlnet_conditioning_channel_order": "rgb",
-  "cross_attention_dim": 2048,
-  "down_block_types": [
-    "DownBlock2D",
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D"
-  ],
-  "downsample_padding": 1,
-  "encoder_hid_dim": null,
-  "encoder_hid_dim_type": null,
-  "flip_sin_to_cos": true,
-  "freq_shift": 0,
-  "global_pool_conditions": false,
-  "in_channels": 4,
-  "layers_per_block": 2,
-  "mid_block_scale_factor": 1,
-  "norm_eps": 1e-05,
-  "norm_num_groups": 32,
-  "num_attention_heads": null,
-  "num_class_embeds": null,
-  "only_cross_attention": false,
-  "projection_class_embeddings_input_dim": 2816,
-  "resnet_time_scale_shift": "default",
-  "transformer_layers_per_block": [
-    1,
-    2,
-    10
-  ],
-  "upcast_attention": null,
-  "use_linear_projection": true
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-epsilon/feature_extractor/preprocessor_config.json

@@ -1,20 +0,0 @@
-{
-  "crop_size": 224,
-  "do_center_crop": true,
-  "do_convert_rgb": true,
-  "do_normalize": true,
-  "do_resize": true,
-  "feature_extractor_type": "CLIPFeatureExtractor",
-  "image_mean": [
-    0.48145466,
-    0.4578275,
-    0.40821073
-  ],
-  "image_std": [
-    0.26862954,
-    0.26130258,
-    0.27577711
-  ],
-  "resample": 3,
-  "size": 224
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-epsilon/model_index.json

@@ -1,32 +0,0 @@
-{
-  "_class_name": "StableDiffusionPipeline",
-  "_diffusers_version": "0.6.0",
-  "feature_extractor": [
-    "transformers",
-    "CLIPImageProcessor"
-  ],
-  "safety_checker": [
-    "stable_diffusion",
-    "StableDiffusionSafetyChecker"
-  ],
-  "scheduler": [
-    "diffusers",
-    "PNDMScheduler"
-  ],
-  "text_encoder": [
-    "transformers",
-    "CLIPTextModel"
-  ],
-  "tokenizer": [
-    "transformers",
-    "CLIPTokenizer"
-  ],
-  "unet": [
-    "diffusers",
-    "UNet2DConditionModel"
-  ],
-  "vae": [
-    "diffusers",
-    "AutoencoderKL"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-epsilon/safety_checker/config.json

@@ -1,175 +0,0 @@
-{
-  "_commit_hash": "4bb648a606ef040e7685bde262611766a5fdd67b",
-  "_name_or_path": "CompVis/stable-diffusion-safety-checker",
-  "architectures": [
-    "StableDiffusionSafetyChecker"
-  ],
-  "initializer_factor": 1.0,
-  "logit_scale_init_value": 2.6592,
-  "model_type": "clip",
-  "projection_dim": 768,
-  "text_config": {
-    "_name_or_path": "",
-    "add_cross_attention": false,
-    "architectures": null,
-    "attention_dropout": 0.0,
-    "bad_words_ids": null,
-    "bos_token_id": 0,
-    "chunk_size_feed_forward": 0,
-    "cross_attention_hidden_size": null,
-    "decoder_start_token_id": null,
-    "diversity_penalty": 0.0,
-    "do_sample": false,
-    "dropout": 0.0,
-    "early_stopping": false,
-    "encoder_no_repeat_ngram_size": 0,
-    "eos_token_id": 2,
-    "exponential_decay_length_penalty": null,
-    "finetuning_task": null,
-    "forced_bos_token_id": null,
-    "forced_eos_token_id": null,
-    "hidden_act": "quick_gelu",
-    "hidden_size": 768,
-    "id2label": {
-      "0": "LABEL_0",
-      "1": "LABEL_1"
-    },
-    "initializer_factor": 1.0,
-    "initializer_range": 0.02,
-    "intermediate_size": 3072,
-    "is_decoder": false,
-    "is_encoder_decoder": false,
-    "label2id": {
-      "LABEL_0": 0,
-      "LABEL_1": 1
-    },
-    "layer_norm_eps": 1e-05,
-    "length_penalty": 1.0,
-    "max_length": 20,
-    "max_position_embeddings": 77,
-    "min_length": 0,
-    "model_type": "clip_text_model",
-    "no_repeat_ngram_size": 0,
-    "num_attention_heads": 12,
-    "num_beam_groups": 1,
-    "num_beams": 1,
-    "num_hidden_layers": 12,
-    "num_return_sequences": 1,
-    "output_attentions": false,
-    "output_hidden_states": false,
-    "output_scores": false,
-    "pad_token_id": 1,
-    "prefix": null,
-    "problem_type": null,
-    "pruned_heads": {},
-    "remove_invalid_values": false,
-    "repetition_penalty": 1.0,
-    "return_dict": true,
-    "return_dict_in_generate": false,
-    "sep_token_id": null,
-    "task_specific_params": null,
-    "temperature": 1.0,
-    "tf_legacy_loss": false,
-    "tie_encoder_decoder": false,
-    "tie_word_embeddings": true,
-    "tokenizer_class": null,
-    "top_k": 50,
-    "top_p": 1.0,
-    "torch_dtype": null,
-    "torchscript": false,
-    "transformers_version": "4.22.0.dev0",
-    "typical_p": 1.0,
-    "use_bfloat16": false,
-    "vocab_size": 49408
-  },
-  "text_config_dict": {
-    "hidden_size": 768,
-    "intermediate_size": 3072,
-    "num_attention_heads": 12,
-    "num_hidden_layers": 12
-  },
-  "torch_dtype": "float32",
-  "transformers_version": null,
-  "vision_config": {
-    "_name_or_path": "",
-    "add_cross_attention": false,
-    "architectures": null,
-    "attention_dropout": 0.0,
-    "bad_words_ids": null,
-    "bos_token_id": null,
-    "chunk_size_feed_forward": 0,
-    "cross_attention_hidden_size": null,
-    "decoder_start_token_id": null,
-    "diversity_penalty": 0.0,
-    "do_sample": false,
-    "dropout": 0.0,
-    "early_stopping": false,
-    "encoder_no_repeat_ngram_size": 0,
-    "eos_token_id": null,
-    "exponential_decay_length_penalty": null,
-    "finetuning_task": null,
-    "forced_bos_token_id": null,
-    "forced_eos_token_id": null,
-    "hidden_act": "quick_gelu",
-    "hidden_size": 1024,
-    "id2label": {
-      "0": "LABEL_0",
-      "1": "LABEL_1"
-    },
-    "image_size": 224,
-    "initializer_factor": 1.0,
-    "initializer_range": 0.02,
-    "intermediate_size": 4096,
-    "is_decoder": false,
-    "is_encoder_decoder": false,
-    "label2id": {
-      "LABEL_0": 0,
-      "LABEL_1": 1
-    },
-    "layer_norm_eps": 1e-05,
-    "length_penalty": 1.0,
-    "max_length": 20,
-    "min_length": 0,
-    "model_type": "clip_vision_model",
-    "no_repeat_ngram_size": 0,
-    "num_attention_heads": 16,
-    "num_beam_groups": 1,
-    "num_beams": 1,
-    "num_channels": 3,
-    "num_hidden_layers": 24,
-    "num_return_sequences": 1,
-    "output_attentions": false,
-    "output_hidden_states": false,
-    "output_scores": false,
-    "pad_token_id": null,
-    "patch_size": 14,
-    "prefix": null,
-    "problem_type": null,
-    "pruned_heads": {},
-    "remove_invalid_values": false,
-    "repetition_penalty": 1.0,
-    "return_dict": true,
-    "return_dict_in_generate": false,
-    "sep_token_id": null,
-    "task_specific_params": null,
-    "temperature": 1.0,
-    "tf_legacy_loss": false,
-    "tie_encoder_decoder": false,
-    "tie_word_embeddings": true,
-    "tokenizer_class": null,
-    "top_k": 50,
-    "top_p": 1.0,
-    "torch_dtype": null,
-    "torchscript": false,
-    "transformers_version": "4.22.0.dev0",
-    "typical_p": 1.0,
-    "use_bfloat16": false
-  },
-  "vision_config_dict": {
-    "hidden_size": 1024,
-    "intermediate_size": 4096,
-    "num_attention_heads": 16,
-    "num_hidden_layers": 24,
-    "patch_size": 14
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-epsilon/scheduler/scheduler_config.json

@@ -1,13 +0,0 @@
-{
-  "_class_name": "PNDMScheduler",
-  "_diffusers_version": "0.6.0",
-  "beta_end": 0.012,
-  "beta_schedule": "scaled_linear",
-  "beta_start": 0.00085,
-  "num_train_timesteps": 1000,
-  "set_alpha_to_one": false,
-  "skip_prk_steps": true,
-  "steps_offset": 1,
-  "trained_betas": null,
-  "clip_sample": false
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-epsilon/text_encoder/config.json

@@ -1,25 +0,0 @@
-{
-  "_name_or_path": "openai/clip-vit-large-patch14",
-  "architectures": [
-    "CLIPTextModel"
-  ],
-  "attention_dropout": 0.0,
-  "bos_token_id": 0,
-  "dropout": 0.0,
-  "eos_token_id": 2,
-  "hidden_act": "quick_gelu",
-  "hidden_size": 768,
-  "initializer_factor": 1.0,
-  "initializer_range": 0.02,
-  "intermediate_size": 3072,
-  "layer_norm_eps": 1e-05,
-  "max_position_embeddings": 77,
-  "model_type": "clip_text_model",
-  "num_attention_heads": 12,
-  "num_hidden_layers": 12,
-  "pad_token_id": 1,
-  "projection_dim": 768,
-  "torch_dtype": "float32",
-  "transformers_version": "4.22.0.dev0",
-  "vocab_size": 49408
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-epsilon/tokenizer/special_tokens_map.json

@@ -1,24 +0,0 @@
-{
-  "bos_token": {
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "eos_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "pad_token": "<|endoftext|>",
-  "unk_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-epsilon/tokenizer/tokenizer_config.json

@@ -1,34 +0,0 @@
-{
-  "add_prefix_space": false,
-  "bos_token": {
-    "__type": "AddedToken",
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "do_lower_case": true,
-  "eos_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "errors": "replace",
-  "model_max_length": 77,
-  "name_or_path": "openai/clip-vit-large-patch14",
-  "pad_token": "<|endoftext|>",
-  "special_tokens_map_file": "./special_tokens_map.json",
-  "tokenizer_class": "CLIPTokenizer",
-  "unk_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-epsilon/unet/config.json

@@ -1,36 +0,0 @@
-{
-  "_class_name": "UNet2DConditionModel",
-  "_diffusers_version": "0.6.0",
-  "act_fn": "silu",
-  "attention_head_dim": 8,
-  "block_out_channels": [
-    320,
-    640,
-    1280,
-    1280
-  ],
-  "center_input_sample": false,
-  "cross_attention_dim": 768,
-  "down_block_types": [
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D",
-    "DownBlock2D"
-  ],
-  "downsample_padding": 1,
-  "flip_sin_to_cos": true,
-  "freq_shift": 0,
-  "in_channels": 4,
-  "layers_per_block": 2,
-  "mid_block_scale_factor": 1,
-  "norm_eps": 1e-05,
-  "norm_num_groups": 32,
-  "out_channels": 4,
-  "sample_size": 64,
-  "up_block_types": [
-    "UpBlock2D",
-    "CrossAttnUpBlock2D",
-    "CrossAttnUpBlock2D",
-    "CrossAttnUpBlock2D"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-epsilon/vae/config.json

@@ -1,29 +0,0 @@
-{
-  "_class_name": "AutoencoderKL",
-  "_diffusers_version": "0.6.0",
-  "act_fn": "silu",
-  "block_out_channels": [
-    128,
-    256,
-    512,
-    512
-  ],
-  "down_block_types": [
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D"
-  ],
-  "in_channels": 3,
-  "latent_channels": 4,
-  "layers_per_block": 2,
-  "norm_num_groups": 32,
-  "out_channels": 3,
-  "sample_size": 512,
-  "up_block_types": [
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-v_prediction/feature_extractor/preprocessor_config.json

@@ -1,28 +0,0 @@
-{
-  "crop_size": {
-    "height": 224,
-    "width": 224
-  },
-  "do_center_crop": true,
-  "do_convert_rgb": true,
-  "do_normalize": true,
-  "do_rescale": true,
-  "do_resize": true,
-  "feature_extractor_type": "CLIPFeatureExtractor",
-  "image_mean": [
-    0.48145466,
-    0.4578275,
-    0.40821073
-  ],
-  "image_processor_type": "CLIPFeatureExtractor",
-  "image_std": [
-    0.26862954,
-    0.26130258,
-    0.27577711
-  ],
-  "resample": 3,
-  "rescale_factor": 0.00392156862745098,
-  "size": {
-    "shortest_edge": 224
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-v_prediction/model_index.json

@@ -1,33 +0,0 @@
-{
-  "_class_name": "StableDiffusionPipeline",
-  "_diffusers_version": "0.18.0.dev0",
-  "feature_extractor": [
-    "transformers",
-    "CLIPFeatureExtractor"
-  ],
-  "requires_safety_checker": true,
-  "safety_checker": [
-    "stable_diffusion",
-    "StableDiffusionSafetyChecker"
-  ],
-  "scheduler": [
-    "diffusers",
-    "DPMSolverMultistepScheduler"
-  ],
-  "text_encoder": [
-    "transformers",
-    "CLIPTextModel"
-  ],
-  "tokenizer": [
-    "transformers",
-    "CLIPTokenizer"
-  ],
-  "unet": [
-    "diffusers",
-    "UNet2DConditionModel"
-  ],
-  "vae": [
-    "diffusers",
-    "AutoencoderKL"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-v_prediction/safety_checker/config.json

@@ -1,168 +0,0 @@
-{
-  "_commit_hash": "cb41f3a270d63d454d385fc2e4f571c487c253c5",
-  "_name_or_path": "CompVis/stable-diffusion-safety-checker",
-  "architectures": [
-    "StableDiffusionSafetyChecker"
-  ],
-  "initializer_factor": 1.0,
-  "logit_scale_init_value": 2.6592,
-  "model_type": "clip",
-  "projection_dim": 768,
-  "text_config": {
-    "_name_or_path": "",
-    "add_cross_attention": false,
-    "architectures": null,
-    "attention_dropout": 0.0,
-    "bad_words_ids": null,
-    "begin_suppress_tokens": null,
-    "bos_token_id": 0,
-    "chunk_size_feed_forward": 0,
-    "cross_attention_hidden_size": null,
-    "decoder_start_token_id": null,
-    "diversity_penalty": 0.0,
-    "do_sample": false,
-    "dropout": 0.0,
-    "early_stopping": false,
-    "encoder_no_repeat_ngram_size": 0,
-    "eos_token_id": 2,
-    "exponential_decay_length_penalty": null,
-    "finetuning_task": null,
-    "forced_bos_token_id": null,
-    "forced_eos_token_id": null,
-    "hidden_act": "quick_gelu",
-    "hidden_size": 768,
-    "id2label": {
-      "0": "LABEL_0",
-      "1": "LABEL_1"
-    },
-    "initializer_factor": 1.0,
-    "initializer_range": 0.02,
-    "intermediate_size": 3072,
-    "is_decoder": false,
-    "is_encoder_decoder": false,
-    "label2id": {
-      "LABEL_0": 0,
-      "LABEL_1": 1
-    },
-    "layer_norm_eps": 1e-05,
-    "length_penalty": 1.0,
-    "max_length": 20,
-    "max_position_embeddings": 77,
-    "min_length": 0,
-    "model_type": "clip_text_model",
-    "no_repeat_ngram_size": 0,
-    "num_attention_heads": 12,
-    "num_beam_groups": 1,
-    "num_beams": 1,
-    "num_hidden_layers": 12,
-    "num_return_sequences": 1,
-    "output_attentions": false,
-    "output_hidden_states": false,
-    "output_scores": false,
-    "pad_token_id": 1,
-    "prefix": null,
-    "problem_type": null,
-    "projection_dim": 512,
-    "pruned_heads": {},
-    "remove_invalid_values": false,
-    "repetition_penalty": 1.0,
-    "return_dict": true,
-    "return_dict_in_generate": false,
-    "sep_token_id": null,
-    "suppress_tokens": null,
-    "task_specific_params": null,
-    "temperature": 1.0,
-    "tf_legacy_loss": false,
-    "tie_encoder_decoder": false,
-    "tie_word_embeddings": true,
-    "tokenizer_class": null,
-    "top_k": 50,
-    "top_p": 1.0,
-    "torch_dtype": null,
-    "torchscript": false,
-    "transformers_version": "4.30.2",
-    "typical_p": 1.0,
-    "use_bfloat16": false,
-    "vocab_size": 49408
-  },
-  "torch_dtype": "float16",
-  "transformers_version": null,
-  "vision_config": {
-    "_name_or_path": "",
-    "add_cross_attention": false,
-    "architectures": null,
-    "attention_dropout": 0.0,
-    "bad_words_ids": null,
-    "begin_suppress_tokens": null,
-    "bos_token_id": null,
-    "chunk_size_feed_forward": 0,
-    "cross_attention_hidden_size": null,
-    "decoder_start_token_id": null,
-    "diversity_penalty": 0.0,
-    "do_sample": false,
-    "dropout": 0.0,
-    "early_stopping": false,
-    "encoder_no_repeat_ngram_size": 0,
-    "eos_token_id": null,
-    "exponential_decay_length_penalty": null,
-    "finetuning_task": null,
-    "forced_bos_token_id": null,
-    "forced_eos_token_id": null,
-    "hidden_act": "quick_gelu",
-    "hidden_size": 1024,
-    "id2label": {
-      "0": "LABEL_0",
-      "1": "LABEL_1"
-    },
-    "image_size": 224,
-    "initializer_factor": 1.0,
-    "initializer_range": 0.02,
-    "intermediate_size": 4096,
-    "is_decoder": false,
-    "is_encoder_decoder": false,
-    "label2id": {
-      "LABEL_0": 0,
-      "LABEL_1": 1
-    },
-    "layer_norm_eps": 1e-05,
-    "length_penalty": 1.0,
-    "max_length": 20,
-    "min_length": 0,
-    "model_type": "clip_vision_model",
-    "no_repeat_ngram_size": 0,
-    "num_attention_heads": 16,
-    "num_beam_groups": 1,
-    "num_beams": 1,
-    "num_channels": 3,
-    "num_hidden_layers": 24,
-    "num_return_sequences": 1,
-    "output_attentions": false,
-    "output_hidden_states": false,
-    "output_scores": false,
-    "pad_token_id": null,
-    "patch_size": 14,
-    "prefix": null,
-    "problem_type": null,
-    "projection_dim": 512,
-    "pruned_heads": {},
-    "remove_invalid_values": false,
-    "repetition_penalty": 1.0,
-    "return_dict": true,
-    "return_dict_in_generate": false,
-    "sep_token_id": null,
-    "suppress_tokens": null,
-    "task_specific_params": null,
-    "temperature": 1.0,
-    "tf_legacy_loss": false,
-    "tie_encoder_decoder": false,
-    "tie_word_embeddings": true,
-    "tokenizer_class": null,
-    "top_k": 50,
-    "top_p": 1.0,
-    "torch_dtype": null,
-    "torchscript": false,
-    "transformers_version": "4.30.2",
-    "typical_p": 1.0,
-    "use_bfloat16": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-v_prediction/scheduler/scheduler_config.json

@@ -1,26 +0,0 @@
-{
-  "_class_name": "DPMSolverMultistepScheduler",
-  "_diffusers_version": "0.18.0.dev0",
-  "algorithm_type": "dpmsolver++",
-  "beta_end": 0.012,
-  "beta_schedule": "scaled_linear",
-  "beta_start": 0.00085,
-  "clip_sample": false,
-  "clip_sample_range": 1.0,
-  "dynamic_thresholding_ratio": 0.995,
-  "lambda_min_clipped": -Infinity,
-  "lower_order_final": true,
-  "num_train_timesteps": 1000,
-  "prediction_type": "v_prediction",
-  "rescale_betas_zero_snr": false,
-  "sample_max_value": 1.0,
-  "set_alpha_to_one": false,
-  "solver_order": 2,
-  "solver_type": "midpoint",
-  "steps_offset": 1,
-  "thresholding": false,
-  "timestep_spacing": "leading",
-  "trained_betas": null,
-  "use_karras_sigmas": false,
-  "variance_type": null
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-v_prediction/text_encoder/config.json

@@ -1,25 +0,0 @@
-{
-  "_name_or_path": "openai/clip-vit-large-patch14",
-  "architectures": [
-    "CLIPTextModel"
-  ],
-  "attention_dropout": 0.0,
-  "bos_token_id": 0,
-  "dropout": 0.0,
-  "eos_token_id": 2,
-  "hidden_act": "quick_gelu",
-  "hidden_size": 768,
-  "initializer_factor": 1.0,
-  "initializer_range": 0.02,
-  "intermediate_size": 3072,
-  "layer_norm_eps": 1e-05,
-  "max_position_embeddings": 77,
-  "model_type": "clip_text_model",
-  "num_attention_heads": 12,
-  "num_hidden_layers": 12,
-  "pad_token_id": 1,
-  "projection_dim": 768,
-  "torch_dtype": "float16",
-  "transformers_version": "4.30.2",
-  "vocab_size": 49408
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-v_prediction/tokenizer/special_tokens_map.json

@@ -1,24 +0,0 @@
-{
-  "bos_token": {
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "eos_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "pad_token": "<|endoftext|>",
-  "unk_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-v_prediction/tokenizer/tokenizer_config.json

@@ -1,33 +0,0 @@
-{
-  "add_prefix_space": false,
-  "bos_token": {
-    "__type": "AddedToken",
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "clean_up_tokenization_spaces": true,
-  "do_lower_case": true,
-  "eos_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "errors": "replace",
-  "model_max_length": 77,
-  "pad_token": "<|endoftext|>",
-  "tokenizer_class": "CLIPTokenizer",
-  "unk_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-v_prediction/unet/config.json

@@ -1,62 +0,0 @@
-{
-  "_class_name": "UNet2DConditionModel",
-  "_diffusers_version": "0.18.0.dev0",
-  "act_fn": "silu",
-  "addition_embed_type": null,
-  "addition_embed_type_num_heads": 64,
-  "attention_head_dim": 8,
-  "block_out_channels": [
-    320,
-    640,
-    1280,
-    1280
-  ],
-  "center_input_sample": false,
-  "class_embed_type": null,
-  "class_embeddings_concat": false,
-  "conv_in_kernel": 3,
-  "conv_out_kernel": 3,
-  "cross_attention_dim": 768,
-  "cross_attention_norm": null,
-  "down_block_types": [
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D",
-    "DownBlock2D"
-  ],
-  "downsample_padding": 1,
-  "dual_cross_attention": false,
-  "encoder_hid_dim": null,
-  "encoder_hid_dim_type": null,
-  "flip_sin_to_cos": true,
-  "freq_shift": 0,
-  "in_channels": 4,
-  "layers_per_block": 2,
-  "mid_block_only_cross_attention": null,
-  "mid_block_scale_factor": 1,
-  "mid_block_type": "UNetMidBlock2DCrossAttn",
-  "norm_eps": 1e-05,
-  "norm_num_groups": 32,
-  "num_attention_heads": null,
-  "num_class_embeds": null,
-  "only_cross_attention": false,
-  "out_channels": 4,
-  "projection_class_embeddings_input_dim": null,
-  "resnet_out_scale_factor": 1.0,
-  "resnet_skip_time_act": false,
-  "resnet_time_scale_shift": "default",
-  "sample_size": 96,
-  "time_cond_proj_dim": null,
-  "time_embedding_act_fn": null,
-  "time_embedding_dim": null,
-  "time_embedding_type": "positional",
-  "timestep_post_act": null,
-  "up_block_types": [
-    "UpBlock2D",
-    "CrossAttnUpBlock2D",
-    "CrossAttnUpBlock2D",
-    "CrossAttnUpBlock2D"
-  ],
-  "upcast_attention": null,
-  "use_linear_projection": false
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-1.5-v_prediction/vae/config.json

@@ -1,30 +0,0 @@
-{
-  "_class_name": "AutoencoderKL",
-  "_diffusers_version": "0.18.0.dev0",
-  "act_fn": "silu",
-  "block_out_channels": [
-    128,
-    256,
-    512,
-    512
-  ],
-  "down_block_types": [
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D"
-  ],
-  "in_channels": 3,
-  "latent_channels": 4,
-  "layers_per_block": 2,
-  "norm_num_groups": 32,
-  "out_channels": 3,
-  "sample_size": 768,
-  "scaling_factor": 0.18215,
-  "up_block_types": [
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-2.0-v_prediction/feature_extractor/preprocessor_config.json

@@ -1,20 +0,0 @@
-{
-  "crop_size": 224,
-  "do_center_crop": true,
-  "do_convert_rgb": true,
-  "do_normalize": true,
-  "do_resize": true,
-  "feature_extractor_type": "CLIPFeatureExtractor",
-  "image_mean": [
-    0.48145466,
-    0.4578275,
-    0.40821073
-  ],
-  "image_std": [
-    0.26862954,
-    0.26130258,
-    0.27577711
-  ],
-  "resample": 3,
-  "size": 224
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-2.0-v_prediction/model_index.json

@@ -1,33 +0,0 @@
-{
-  "_class_name": "StableDiffusionPipeline",
-  "_diffusers_version": "0.8.0",
-  "feature_extractor": [
-    "transformers",
-    "CLIPImageProcessor"
-  ],
-  "requires_safety_checker": false,
-  "safety_checker": [
-    null,
-    null
-  ],
-  "scheduler": [
-    "diffusers",
-    "DDIMScheduler"
-  ],
-  "text_encoder": [
-    "transformers",
-    "CLIPTextModel"
-  ],
-  "tokenizer": [
-    "transformers",
-    "CLIPTokenizer"
-  ],
-  "unet": [
-    "diffusers",
-    "UNet2DConditionModel"
-  ],
-  "vae": [
-    "diffusers",
-    "AutoencoderKL"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-2.0-v_prediction/scheduler/scheduler_config.json

@@ -1,14 +0,0 @@
-{
-  "_class_name": "DDIMScheduler",
-  "_diffusers_version": "0.8.0",
-  "beta_end": 0.012,
-  "beta_schedule": "scaled_linear",
-  "beta_start": 0.00085,
-  "clip_sample": false,
-  "num_train_timesteps": 1000,
-  "prediction_type": "v_prediction",
-  "set_alpha_to_one": false,
-  "skip_prk_steps": true,
-  "steps_offset": 1,
-  "trained_betas": null
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-2.0-v_prediction/text_encoder/config.json

@@ -1,25 +0,0 @@
-{
-  "_name_or_path": "hf-models/stable-diffusion-v2-768x768/text_encoder",
-  "architectures": [
-    "CLIPTextModel"
-  ],
-  "attention_dropout": 0.0,
-  "bos_token_id": 0,
-  "dropout": 0.0,
-  "eos_token_id": 2,
-  "hidden_act": "gelu",
-  "hidden_size": 1024,
-  "initializer_factor": 1.0,
-  "initializer_range": 0.02,
-  "intermediate_size": 4096,
-  "layer_norm_eps": 1e-05,
-  "max_position_embeddings": 77,
-  "model_type": "clip_text_model",
-  "num_attention_heads": 16,
-  "num_hidden_layers": 23,
-  "pad_token_id": 1,
-  "projection_dim": 512,
-  "torch_dtype": "float32",
-  "transformers_version": "4.25.0.dev0",
-  "vocab_size": 49408
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-2.0-v_prediction/tokenizer/special_tokens_map.json

@@ -1,24 +0,0 @@
-{
-  "bos_token": {
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "eos_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "pad_token": "!",
-  "unk_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-2.0-v_prediction/tokenizer/tokenizer_config.json

@@ -1,34 +0,0 @@
-{
-  "add_prefix_space": false,
-  "bos_token": {
-    "__type": "AddedToken",
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "do_lower_case": true,
-  "eos_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "errors": "replace",
-  "model_max_length": 77,
-  "name_or_path": "hf-models/stable-diffusion-v2-768x768/tokenizer",
-  "pad_token": "<|endoftext|>",
-  "special_tokens_map_file": "./special_tokens_map.json",
-  "tokenizer_class": "CLIPTokenizer",
-  "unk_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-2.0-v_prediction/unet/config.json

@@ -1,46 +0,0 @@
-{
-  "_class_name": "UNet2DConditionModel",
-  "_diffusers_version": "0.10.0.dev0",
-  "act_fn": "silu",
-  "attention_head_dim": [
-    5,
-    10,
-    20,
-    20
-  ],
-  "block_out_channels": [
-    320,
-    640,
-    1280,
-    1280
-  ],
-  "center_input_sample": false,
-  "cross_attention_dim": 1024,
-  "down_block_types": [
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D",
-    "DownBlock2D"
-  ],
-  "downsample_padding": 1,
-  "dual_cross_attention": false,
-  "flip_sin_to_cos": true,
-  "freq_shift": 0,
-  "in_channels": 4,
-  "layers_per_block": 2,
-  "mid_block_scale_factor": 1,
-  "norm_eps": 1e-05,
-  "norm_num_groups": 32,
-  "num_class_embeds": null,
-  "only_cross_attention": false,
-  "out_channels": 4,
-  "sample_size": 96,
-  "up_block_types": [
-    "UpBlock2D",
-    "CrossAttnUpBlock2D",
-    "CrossAttnUpBlock2D",
-    "CrossAttnUpBlock2D"
-  ],
-  "use_linear_projection": true,
-  "upcast_attention": true
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-2.0-v_prediction/vae/config.json

@@ -1,30 +0,0 @@
-{
-  "_class_name": "AutoencoderKL",
-  "_diffusers_version": "0.8.0",
-  "_name_or_path": "hf-models/stable-diffusion-v2-768x768/vae",
-  "act_fn": "silu",
-  "block_out_channels": [
-    128,
-    256,
-    512,
-    512
-  ],
-  "down_block_types": [
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D"
-  ],
-  "in_channels": 3,
-  "latent_channels": 4,
-  "layers_per_block": 2,
-  "norm_num_groups": 32,
-  "out_channels": 3,
-  "sample_size": 768,
-  "up_block_types": [
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/model_index.json

@@ -1,34 +0,0 @@
-{
-  "_class_name": "StableDiffusionXLPipeline",
-  "_diffusers_version": "0.19.0.dev0",
-  "force_zeros_for_empty_prompt": true,
-  "add_watermarker": null,
-  "scheduler": [
-    "diffusers",
-    "EulerDiscreteScheduler"
-  ],
-  "text_encoder": [
-    "transformers",
-    "CLIPTextModel"
-  ],
-  "text_encoder_2": [
-    "transformers",
-    "CLIPTextModelWithProjection"
-  ],
-  "tokenizer": [
-    "transformers",
-    "CLIPTokenizer"
-  ],
-  "tokenizer_2": [
-    "transformers",
-    "CLIPTokenizer"
-  ],
-  "unet": [
-    "diffusers",
-    "UNet2DConditionModel"
-  ],
-  "vae": [
-    "diffusers",
-    "AutoencoderKL"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/scheduler/scheduler_config.json

@@ -1,18 +0,0 @@
-{
-  "_class_name": "EulerDiscreteScheduler",
-  "_diffusers_version": "0.19.0.dev0",
-  "beta_end": 0.012,
-  "beta_schedule": "scaled_linear",
-  "beta_start": 0.00085,
-  "clip_sample": false,
-  "interpolation_type": "linear",
-  "num_train_timesteps": 1000,
-  "prediction_type": "epsilon",
-  "sample_max_value": 1.0,
-  "set_alpha_to_one": false,
-  "skip_prk_steps": true,
-  "steps_offset": 1,
-  "timestep_spacing": "leading",
-  "trained_betas": null,
-  "use_karras_sigmas": false
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/text_encoder/config.json

@@ -1,24 +0,0 @@
-{
-  "architectures": [
-    "CLIPTextModel"
-  ],
-  "attention_dropout": 0.0,
-  "bos_token_id": 0,
-  "dropout": 0.0,
-  "eos_token_id": 2,
-  "hidden_act": "quick_gelu",
-  "hidden_size": 768,
-  "initializer_factor": 1.0,
-  "initializer_range": 0.02,
-  "intermediate_size": 3072,
-  "layer_norm_eps": 1e-05,
-  "max_position_embeddings": 77,
-  "model_type": "clip_text_model",
-  "num_attention_heads": 12,
-  "num_hidden_layers": 12,
-  "pad_token_id": 1,
-  "projection_dim": 768,
-  "torch_dtype": "float16",
-  "transformers_version": "4.32.0.dev0",
-  "vocab_size": 49408
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/text_encoder_2/config.json

@@ -1,24 +0,0 @@
-{
-  "architectures": [
-    "CLIPTextModelWithProjection"
-  ],
-  "attention_dropout": 0.0,
-  "bos_token_id": 0,
-  "dropout": 0.0,
-  "eos_token_id": 2,
-  "hidden_act": "gelu",
-  "hidden_size": 1280,
-  "initializer_factor": 1.0,
-  "initializer_range": 0.02,
-  "intermediate_size": 5120,
-  "layer_norm_eps": 1e-05,
-  "max_position_embeddings": 77,
-  "model_type": "clip_text_model",
-  "num_attention_heads": 20,
-  "num_hidden_layers": 32,
-  "pad_token_id": 1,
-  "projection_dim": 1280,
-  "torch_dtype": "float16",
-  "transformers_version": "4.32.0.dev0",
-  "vocab_size": 49408
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/tokenizer/special_tokens_map.json

@@ -1,24 +0,0 @@
-{
-  "bos_token": {
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "eos_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "pad_token": "<|endoftext|>",
-  "unk_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/tokenizer/tokenizer_config.json

@@ -1,33 +0,0 @@
-{
-  "add_prefix_space": false,
-  "bos_token": {
-    "__type": "AddedToken",
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "clean_up_tokenization_spaces": true,
-  "do_lower_case": true,
-  "eos_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "errors": "replace",
-  "model_max_length": 77,
-  "pad_token": "<|endoftext|>",
-  "tokenizer_class": "CLIPTokenizer",
-  "unk_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/tokenizer_2/special_tokens_map.json

@@ -1,24 +0,0 @@
-{
-  "bos_token": {
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "eos_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "pad_token": "!",
-  "unk_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/tokenizer_2/tokenizer_config.json

@@ -1,33 +0,0 @@
-{
-  "add_prefix_space": false,
-  "bos_token": {
-    "__type": "AddedToken",
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "clean_up_tokenization_spaces": true,
-  "do_lower_case": true,
-  "eos_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "errors": "replace",
-  "model_max_length": 77,
-  "pad_token": "!",
-  "tokenizer_class": "CLIPTokenizer",
-  "unk_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/unet/config.json

@@ -1,69 +0,0 @@
-{
-  "_class_name": "UNet2DConditionModel",
-  "_diffusers_version": "0.19.0.dev0",
-  "act_fn": "silu",
-  "addition_embed_type": "text_time",
-  "addition_embed_type_num_heads": 64,
-  "addition_time_embed_dim": 256,
-  "attention_head_dim": [
-    5,
-    10,
-    20
-  ],
-  "block_out_channels": [
-    320,
-    640,
-    1280
-  ],
-  "center_input_sample": false,
-  "class_embed_type": null,
-  "class_embeddings_concat": false,
-  "conv_in_kernel": 3,
-  "conv_out_kernel": 3,
-  "cross_attention_dim": 2048,
-  "cross_attention_norm": null,
-  "down_block_types": [
-    "DownBlock2D",
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D"
-  ],
-  "downsample_padding": 1,
-  "dual_cross_attention": false,
-  "encoder_hid_dim": null,
-  "encoder_hid_dim_type": null,
-  "flip_sin_to_cos": true,
-  "freq_shift": 0,
-  "in_channels": 4,
-  "layers_per_block": 2,
-  "mid_block_only_cross_attention": null,
-  "mid_block_scale_factor": 1,
-  "mid_block_type": "UNetMidBlock2DCrossAttn",
-  "norm_eps": 1e-05,
-  "norm_num_groups": 32,
-  "num_attention_heads": null,
-  "num_class_embeds": null,
-  "only_cross_attention": false,
-  "out_channels": 4,
-  "projection_class_embeddings_input_dim": 2816,
-  "resnet_out_scale_factor": 1.0,
-  "resnet_skip_time_act": false,
-  "resnet_time_scale_shift": "default",
-  "sample_size": 128,
-  "time_cond_proj_dim": null,
-  "time_embedding_act_fn": null,
-  "time_embedding_dim": null,
-  "time_embedding_type": "positional",
-  "timestep_post_act": null,
-  "transformer_layers_per_block": [
-    1,
-    2,
-    10
-  ],
-  "up_block_types": [
-    "CrossAttnUpBlock2D",
-    "CrossAttnUpBlock2D",
-    "UpBlock2D"
-  ],
-  "upcast_attention": null,
-  "use_linear_projection": true
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/vae/config.json

@@ -1,32 +0,0 @@
-{
-  "_class_name": "AutoencoderKL",
-  "_diffusers_version": "0.20.0.dev0",
-  "_name_or_path": "../sdxl-vae/",
-  "act_fn": "silu",
-  "block_out_channels": [
-    128,
-    256,
-    512,
-    512
-  ],
-  "down_block_types": [
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D"
-  ],
-  "force_upcast": true,
-  "in_channels": 3,
-  "latent_channels": 4,
-  "layers_per_block": 2,
-  "norm_num_groups": 32,
-  "out_channels": 3,
-  "sample_size": 1024,
-  "scaling_factor": 0.13025,
-  "up_block_types": [
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/vae_1_0/config.json

@@ -1,31 +0,0 @@
-{
-  "_class_name": "AutoencoderKL",
-  "_diffusers_version": "0.19.0.dev0",
-  "act_fn": "silu",
-  "block_out_channels": [
-    128,
-    256,
-    512,
-    512
-  ],
-  "down_block_types": [
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D"
-  ],
-  "force_upcast": true,
-  "in_channels": 3,
-  "latent_channels": 4,
-  "layers_per_block": 2,
-  "norm_num_groups": 32,
-  "out_channels": 3,
-  "sample_size": 1024,
-  "scaling_factor": 0.13025,
-  "up_block_types": [
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/vae_decoder/config.json

@@ -1,31 +0,0 @@
-{
-  "_class_name": "AutoencoderKL",
-  "_diffusers_version": "0.19.0.dev0",
-  "act_fn": "silu",
-  "block_out_channels": [
-    128,
-    256,
-    512,
-    512
-  ],
-  "down_block_types": [
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D"
-  ],
-  "force_upcast": true,
-  "in_channels": 3,
-  "latent_channels": 4,
-  "layers_per_block": 2,
-  "norm_num_groups": 32,
-  "out_channels": 3,
-  "sample_size": 1024,
-  "scaling_factor": 0.13025,
-  "up_block_types": [
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-base-1.0/vae_encoder/config.json

@@ -1,31 +0,0 @@
-{
-  "_class_name": "AutoencoderKL",
-  "_diffusers_version": "0.19.0.dev0",
-  "act_fn": "silu",
-  "block_out_channels": [
-    128,
-    256,
-    512,
-    512
-  ],
-  "down_block_types": [
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D"
-  ],
-  "force_upcast": true,
-  "in_channels": 3,
-  "latent_channels": 4,
-  "layers_per_block": 2,
-  "norm_num_groups": 32,
-  "out_channels": 3,
-  "sample_size": 1024,
-  "scaling_factor": 0.13025,
-  "up_block_types": [
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-refiner-1.0/model_index.json

@@ -1,35 +0,0 @@
-{
-  "_class_name": "StableDiffusionXLImg2ImgPipeline",
-  "_diffusers_version": "0.19.0.dev0",
-  "force_zeros_for_empty_prompt": false,
-  "add_watermarker": null,
-  "requires_aesthetics_score": true,
-  "scheduler": [
-    "diffusers",
-    "EulerDiscreteScheduler"
-  ],
-  "text_encoder": [
-    null,
-    null
-  ],
-  "text_encoder_2": [
-    "transformers",
-    "CLIPTextModelWithProjection"
-  ],
-  "tokenizer": [
-    null,
-    null
-  ],
-  "tokenizer_2": [
-    "transformers",
-    "CLIPTokenizer"
-  ],
-  "unet": [
-    "diffusers",
-    "UNet2DConditionModel"
-  ],
-  "vae": [
-    "diffusers",
-    "AutoencoderKL"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-refiner-1.0/scheduler/scheduler_config.json

@@ -1,18 +0,0 @@
-{
-  "_class_name": "EulerDiscreteScheduler",
-  "_diffusers_version": "0.19.0.dev0",
-  "beta_end": 0.012,
-  "beta_schedule": "scaled_linear",
-  "beta_start": 0.00085,
-  "clip_sample": false,
-  "interpolation_type": "linear",
-  "num_train_timesteps": 1000,
-  "prediction_type": "epsilon",
-  "sample_max_value": 1.0,
-  "set_alpha_to_one": false,
-  "skip_prk_steps": true,
-  "steps_offset": 1,
-  "timestep_spacing": "leading",
-  "trained_betas": null,
-  "use_karras_sigmas": false
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-refiner-1.0/text_encoder_2/config.json

@@ -1,24 +0,0 @@
-{
-  "architectures": [
-    "CLIPTextModelWithProjection"
-  ],
-  "attention_dropout": 0.0,
-  "bos_token_id": 0,
-  "dropout": 0.0,
-  "eos_token_id": 2,
-  "hidden_act": "gelu",
-  "hidden_size": 1280,
-  "initializer_factor": 1.0,
-  "initializer_range": 0.02,
-  "intermediate_size": 5120,
-  "layer_norm_eps": 1e-05,
-  "max_position_embeddings": 77,
-  "model_type": "clip_text_model",
-  "num_attention_heads": 20,
-  "num_hidden_layers": 32,
-  "pad_token_id": 1,
-  "projection_dim": 1280,
-  "torch_dtype": "float16",
-  "transformers_version": "4.32.0.dev0",
-  "vocab_size": 49408
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-refiner-1.0/tokenizer_2/special_tokens_map.json

@@ -1,24 +0,0 @@
-{
-  "bos_token": {
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "eos_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "pad_token": "!",
-  "unk_token": {
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-refiner-1.0/tokenizer_2/tokenizer_config.json

@@ -1,33 +0,0 @@
-{
-  "add_prefix_space": false,
-  "bos_token": {
-    "__type": "AddedToken",
-    "content": "<|startoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "clean_up_tokenization_spaces": true,
-  "do_lower_case": true,
-  "eos_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  },
-  "errors": "replace",
-  "model_max_length": 77,
-  "pad_token": "!",
-  "tokenizer_class": "CLIPTokenizer",
-  "unk_token": {
-    "__type": "AddedToken",
-    "content": "<|endoftext|>",
-    "lstrip": false,
-    "normalized": true,
-    "rstrip": false,
-    "single_word": false
-  }
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-refiner-1.0/unet/config.json

@@ -1,69 +0,0 @@
-{
-  "_class_name": "UNet2DConditionModel",
-  "_diffusers_version": "0.19.0.dev0",
-  "act_fn": "silu",
-  "addition_embed_type": "text_time",
-  "addition_embed_type_num_heads": 64,
-  "addition_time_embed_dim": 256,
-  "attention_head_dim": [
-    6,
-    12,
-    24,
-    24
-  ],
-  "block_out_channels": [
-    384,
-    768,
-    1536,
-    1536
-  ],
-  "center_input_sample": false,
-  "class_embed_type": null,
-  "class_embeddings_concat": false,
-  "conv_in_kernel": 3,
-  "conv_out_kernel": 3,
-  "cross_attention_dim": 1280,
-  "cross_attention_norm": null,
-  "down_block_types": [
-    "DownBlock2D",
-    "CrossAttnDownBlock2D",
-    "CrossAttnDownBlock2D",
-    "DownBlock2D"
-  ],
-  "downsample_padding": 1,
-  "dual_cross_attention": false,
-  "encoder_hid_dim": null,
-  "encoder_hid_dim_type": null,
-  "flip_sin_to_cos": true,
-  "freq_shift": 0,
-  "in_channels": 4,
-  "layers_per_block": 2,
-  "mid_block_only_cross_attention": null,
-  "mid_block_scale_factor": 1,
-  "mid_block_type": "UNetMidBlock2DCrossAttn",
-  "norm_eps": 1e-05,
-  "norm_num_groups": 32,
-  "num_attention_heads": null,
-  "num_class_embeds": null,
-  "only_cross_attention": false,
-  "out_channels": 4,
-  "projection_class_embeddings_input_dim": 2560,
-  "resnet_out_scale_factor": 1.0,
-  "resnet_skip_time_act": false,
-  "resnet_time_scale_shift": "default",
-  "sample_size": 128,
-  "time_cond_proj_dim": null,
-  "time_embedding_act_fn": null,
-  "time_embedding_dim": null,
-  "time_embedding_type": "positional",
-  "timestep_post_act": null,
-  "transformer_layers_per_block": 4,
-  "up_block_types": [
-    "UpBlock2D",
-    "CrossAttnUpBlock2D",
-    "CrossAttnUpBlock2D",
-    "UpBlock2D"
-  ],
-  "upcast_attention": null,
-  "use_linear_projection": true
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-refiner-1.0/vae/config.json

@@ -1,32 +0,0 @@
-{
-  "_class_name": "AutoencoderKL",
-  "_diffusers_version": "0.20.0.dev0",
-  "_name_or_path": "../sdxl-vae/",
-  "act_fn": "silu",
-  "block_out_channels": [
-    128,
-    256,
-    512,
-    512
-  ],
-  "down_block_types": [
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D"
-  ],
-  "force_upcast": true,
-  "in_channels": 3,
-  "latent_channels": 4,
-  "layers_per_block": 2,
-  "norm_num_groups": 32,
-  "out_channels": 3,
-  "sample_size": 1024,
-  "scaling_factor": 0.13025,
-  "up_block_types": [
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D"
-  ]
-}

invokeai/backend/assets/model_base_conf_files/stable-diffusion-xl-refiner-1.0/vae_1_0/config.json

@@ -1,31 +0,0 @@
-{
-  "_class_name": "AutoencoderKL",
-  "_diffusers_version": "0.19.0.dev0",
-  "act_fn": "silu",
-  "block_out_channels": [
-    128,
-    256,
-    512,
-    512
-  ],
-  "down_block_types": [
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D"
-  ],
-  "force_upcast": true,
-  "in_channels": 3,
-  "latent_channels": 4,
-  "layers_per_block": 2,
-  "norm_num_groups": 32,
-  "out_channels": 3,
-  "sample_size": 1024,
-  "scaling_factor": 0.13025,
-  "up_block_types": [
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D"
-  ]
-}

invokeai/backend/flux/denoise.py

@@ -30,6 +30,8 @@ def denoise(
     controlnet_extensions: list[XLabsControlNetExtension | InstantXControlNetExtension],
     pos_ip_adapter_extensions: list[XLabsIPAdapterExtension],
     neg_ip_adapter_extensions: list[XLabsIPAdapterExtension],
+    # extra img tokens
+    img_cond: torch.Tensor | None = None,
 ):
     # step 0 is the initial state
     total_steps = len(timesteps) - 1
@@ -69,9 +71,9 @@ def denoise(
         # controlnet_residuals datastructure is efficient in that it likely contains multiple references to the same
         # tensors. Calculating the sum materializes each tensor into its own instance.
         merged_controlnet_residuals = sum_controlnet_flux_outputs(controlnet_residuals)
-
+        pred_img = torch.cat((img, img_cond), dim=-1) if img_cond is not None else img
         pred = model(
-            img=img,
+            img=pred_img,
             img_ids=img_ids,
             txt=pos_regional_prompting_extension.regional_text_conditioning.t5_embeddings,
             txt_ids=pos_regional_prompting_extension.regional_text_conditioning.t5_txt_ids,

invokeai/backend/flux/flux_tools_sampling_utils.py

@@ -0,0 +1,27 @@
+import torch
+import numpy as np
+from PIL import Image
+from einops import rearrange
+
+from invokeai.backend.flux.modules.autoencoder import AutoEncoder
+
+def prepare_control(
+    height: int,
+    width: int,
+    seed: int,
+    ae: AutoEncoder,
+    cond_image: Image.Image,
+) -> torch.Tensor:
+    # load and encode the conditioning image
+    img_cond = cond_image.convert("RGB")
+    img_cond = img_cond.resize((width, height), Image.Resampling.LANCZOS)
+    img_cond = np.array(img_cond)
+    img_cond = torch.from_numpy(img_cond).float()
+    img_cond = rearrange(img_cond, "h w c -> 1 c h w")
+    ae_dtype = next(iter(ae.parameters())).dtype
+    ae_device = next(iter(ae.parameters())).device
+    img_cond = img_cond.to(device=ae_device, dtype=ae_dtype)
+    generator = torch.Generator(device=ae_device).manual_seed(seed)
+    img_cond = ae.encode(img_cond, sample=True, generator=generator)
+    img_cond = rearrange(img_cond, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+    return img_cond

invokeai/backend/flux/math.py

@@ -32,4 +32,4 @@ def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor) -> tuple[Tensor, Tenso
     xk_ = xk.view(*xk.shape[:-1], -1, 1, 2)
     xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
     xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
-    return xq_out.view(*xq.shape), xk_out.view(*xk.shape)
+    return xq_out.view(*xq.shape).type_as(xq), xk_out.view(*xk.shape).type_as(xk)

invokeai/backend/flux/model.py

@@ -4,6 +4,7 @@ from dataclasses import dataclass
 
 import torch
 from torch import Tensor, nn
+from typing import Optional
 
 from invokeai.backend.flux.custom_block_processor import (
     CustomDoubleStreamBlockProcessor,
@@ -35,6 +36,7 @@ class FluxParams:
     theta: int
     qkv_bias: bool
     guidance_embed: bool
+    out_channels: Optional[int] = None
 
 
 class Flux(nn.Module):
@@ -47,7 +49,7 @@ class Flux(nn.Module):
 
         self.params = params
         self.in_channels = params.in_channels
-        self.out_channels = self.in_channels
+        self.out_channels = params.out_channels or self.in_channels
         if params.hidden_size % params.num_heads != 0:
             raise ValueError(f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}")
         pe_dim = params.hidden_size // params.num_heads

invokeai/backend/flux/modules/image_embedders.py

@@ -0,0 +1,50 @@
+import os
+import cv2
+import numpy as np
+import torch
+
+from einops import rearrange, repeat
+from PIL import Image
+from safetensors.torch import load_file as load_sft
+from torch import nn
+from transformers import AutoModelForDepthEstimation, AutoProcessor, SiglipImageProcessor, SiglipVisionModel
+
+class DepthImageEncoder:
+    depth_model_name = "LiheYoung/depth-anything-large-hf"
+    def __init__(self, device):
+        self.device = device
+        self.depth_model = AutoModelForDepthEstimation.from_pretrained(self.depth_model_name).to(device)
+        self.processor = AutoProcessor.from_pretrained(self.depth_model_name)
+    def __call__(self, img: torch.Tensor) -> torch.Tensor:
+        hw = img.shape[-2:]
+        img = torch.clamp(img, -1.0, 1.0)
+        img_byte = ((img + 1.0) * 127.5).byte()
+        img = self.processor(img_byte, return_tensors="pt")["pixel_values"]
+        depth = self.depth_model(img.to(self.device)).predicted_depth
+        depth = repeat(depth, "b h w -> b 3 h w")
+        depth = torch.nn.functional.interpolate(depth, hw, mode="bicubic", antialias=True)
+        depth = depth / 127.5 - 1.0
+        return depth
+
+class CannyImageEncoder:
+    def __init__(
+        self,
+        device,
+        min_t: int = 50,
+        max_t: int = 200,
+    ):
+        self.device = device
+        self.min_t = min_t
+        self.max_t = max_t
+    def __call__(self, img: torch.Tensor) -> torch.Tensor:
+        assert img.shape[0] == 1, "Only batch size 1 is supported"
+        img = rearrange(img[0], "c h w -> h w c")
+        img = torch.clamp(img, -1.0, 1.0)
+        img_np = ((img + 1.0) * 127.5).numpy().astype(np.uint8)
+        # Apply Canny edge detection
+        canny = cv2.Canny(img_np, self.min_t, self.max_t)
+        # Convert back to torch tensor and reshape
+        canny = torch.from_numpy(canny).float() / 127.5 - 1.0
+        canny = rearrange(canny, "h w -> 1 1 h w")
+        canny = repeat(canny, "b 1 ... -> b 3 ...")
+        return canny.to(self.device)

invokeai/backend/lora/conversions/flux_control_lora_utils.py

@@ -0,0 +1,65 @@
+import re
+import torch
+
+from typing import Any, Dict
+from invokeai.backend.lora.layers.any_lora_layer import AnyLoRALayer
+from invokeai.backend.lora.layers.utils import any_lora_layer_from_state_dict
+from invokeai.backend.lora.lora_model_raw import LoRAModelRaw
+from invokeai.backend.lora.conversions.flux_lora_constants import FLUX_LORA_TRANSFORMER_PREFIX
+from invokeai.backend.lora.layers.lora_layer import LoRALayer
+from invokeai.backend.lora.layers.set_parameter_layer import SetParameterLayer
+
+
+# A regex pattern that matches all of the keys in the Flux Dev/Canny LoRA format.
+# Example keys:
+#   guidance_in.in_layer.lora_B.bias
+#   single_blocks.0.linear1.lora_A.weight
+#   double_blocks.0.img_attn.norm.key_norm.scale
+FLUX_STRUCTURAL_TRANSFORMER_KEY_REGEX = r"(final_layer|vector_in|txt_in|time_in|img_in|guidance_in|\w+_blocks)(\.(\d+))?\.(lora_(A|B)|(in|out)_layer|adaLN_modulation|img_attn|img_mlp|img_mod|txt_attn|txt_mlp|txt_mod|linear|linear1|linear2|modulation|norm)\.?(.*)"
+
+def is_state_dict_likely_flux_control(state_dict: Dict[str, Any]) -> bool:
+    """Checks if the provided state dict is likely in the FLUX Control LoRA format.
+
+    This is intended to be a high-precision detector, but it is not guaranteed to have perfect precision. (A
+    perfect-precision detector would require checking all keys against a whitelist and verifying tensor shapes.)
+    """
+    return all(
+        re.match(FLUX_STRUCTURAL_TRANSFORMER_KEY_REGEX, k) or re.match(FLUX_STRUCTURAL_TRANSFORMER_KEY_REGEX, k)
+        for k in state_dict.keys()
+    )
+
+def lora_model_from_flux_control_state_dict(state_dict: Dict[str, torch.Tensor]) -> LoRAModelRaw:
+    # converted_state_dict = _convert_lora_bfl_control(state_dict=state_dict)
+    # Group keys by layer.
+    grouped_state_dict: dict[str, dict[str, torch.Tensor]] = {}
+    for key, value in state_dict.items():
+        key_props = key.split(".")
+        # Got it loading using lora_down and lora_up but it didn't seem to match this lora's structure
+        # Leaving this in since it doesn't hurt anything and may be better
+        layer_prop_size = -2 if any(prop in key for prop in ["lora_B", "lora_A"]) else -1
+        layer_name = ".".join(key_props[:layer_prop_size])
+        param_name = ".".join(key_props[layer_prop_size:])
+        if layer_name not in grouped_state_dict:
+            grouped_state_dict[layer_name] = {}
+        grouped_state_dict[layer_name][param_name] = value
+
+    # Create LoRA layers.
+    layers: dict[str, AnyLoRALayer] = {}
+    for layer_key, layer_state_dict in grouped_state_dict.items():
+        # Convert to a full layer diff
+        prefixed_key = f"{FLUX_LORA_TRANSFORMER_PREFIX}{layer_key}"
+        if all(k in layer_state_dict for k in ["lora_A.weight", "lora_B.bias", "lora_B.weight"]):
+            layers[prefixed_key] = LoRALayer(
+                layer_state_dict["lora_B.weight"],
+                None,
+                layer_state_dict["lora_A.weight"],
+                None,
+                layer_state_dict["lora_B.bias"]
+            )
+        elif "scale" in layer_state_dict:
+            layers[prefixed_key] = SetParameterLayer("scale", layer_state_dict["scale"])
+        else:
+            raise AssertionError(f"{layer_key} not expected")
+    # Create and return the LoRAModelRaw.
+    return LoRAModelRaw(layers=layers)
+

invokeai/backend/lora/layers/any_lora_layer.py

@@ -7,5 +7,6 @@ from invokeai.backend.lora.layers.loha_layer import LoHALayer
 from invokeai.backend.lora.layers.lokr_layer import LoKRLayer
 from invokeai.backend.lora.layers.lora_layer import LoRALayer
 from invokeai.backend.lora.layers.norm_layer import NormLayer
+from invokeai.backend.lora.layers.set_parameter_layer import SetParameterLayer
 
-AnyLoRALayer = Union[LoRALayer, LoHALayer, LoKRLayer, FullLayer, IA3Layer, NormLayer, ConcatenatedLoRALayer]
+AnyLoRALayer = Union[LoRALayer, LoHALayer, LoKRLayer, FullLayer, IA3Layer, NormLayer, ConcatenatedLoRALayer, SetParameterLayer]

invokeai/backend/lora/layers/reshape_weight_layer.py

@@ -0,0 +1,34 @@
+from typing import Dict, Optional
+
+import torch
+
+from invokeai.backend.lora.layers.lora_layer_base import LoRALayerBase
+from invokeai.backend.util.calc_tensor_size import calc_tensor_size
+
+
+class ReshapeWeightLayer(LoRALayerBase):
+    # TODO: Just everything in this class 
+    def __init__(self, weight: Optional[torch.Tensor], bias: Optional[torch.Tensor], scale: Optional[torch.Tensor]):
+        super().__init__(alpha=None, bias=bias)
+        self.weight = torch.nn.Parameter(weight) if weight is not None else None
+        self.bias = torch.nn.Parameter(bias) if bias is not None else None
+        self.manual_scale = scale
+
+    def scale(self):
+        return self.manual_scale.float() if self.manual_scale is not None else super().scale()
+
+    def rank(self) -> int | None:
+        return None
+
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+        return orig_weight
+
+    def to(self, device: torch.device | None = None, dtype: torch.dtype | None = None):
+        super().to(device=device, dtype=dtype)
+        if self.weight is not None:
+            self.weight = self.weight.to(device=device, dtype=dtype)
+        if self.manual_scale is not None:
+            self.manual_scale = self.manual_scale.to(device=device, dtype=dtype)
+
+    def calc_size(self) -> int:
+        return super().calc_size() + calc_tensor_size(self.manual_scale)

invokeai/backend/lora/layers/set_parameter_layer.py

@@ -0,0 +1,29 @@
+from typing import Dict, Optional
+
+import torch
+
+from invokeai.backend.lora.layers.lora_layer_base import LoRALayerBase
+from invokeai.backend.util.calc_tensor_size import calc_tensor_size
+
+
+class SetParameterLayer(LoRALayerBase):
+    def __init__(self, param_name: str, weight: torch.Tensor):
+        super().__init__(None, None)
+        self.weight = weight
+        self.param_name = param_name
+
+    def rank(self) -> int | None:
+        return None
+
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+        return self.weight - orig_weight
+    
+    def get_parameters(self, orig_module: torch.nn.Module) -> Dict[str, torch.Tensor]:
+        return {self.param_name: self.get_weight(orig_module.get_parameter(self.param_name))}
+
+    def to(self, device: torch.device | None = None, dtype: torch.dtype | None = None):
+        super().to(device=device, dtype=dtype)
+        self.weight = self.weight.to(device=device, dtype=dtype)
+
+    def calc_size(self) -> int:
+        return super().calc_size() + calc_tensor_size(self.weight)

invokeai/backend/lora/layers/utils.py

@@ -9,6 +9,7 @@ from invokeai.backend.lora.layers.loha_layer import LoHALayer
 from invokeai.backend.lora.layers.lokr_layer import LoKRLayer
 from invokeai.backend.lora.layers.lora_layer import LoRALayer
 from invokeai.backend.lora.layers.norm_layer import NormLayer
+from invokeai.backend.lora.layers.set_parameter_layer import SetParameterLayer
 
 
 def any_lora_layer_from_state_dict(state_dict: Dict[str, torch.Tensor]) -> AnyLoRALayer:

invokeai/backend/lora/lora_layer_wrappers.py

@@ -0,0 +1,133 @@
+import torch
+
+from invokeai.backend.lora.layers.any_lora_layer import AnyLoRALayer
+from invokeai.backend.lora.layers.concatenated_lora_layer import ConcatenatedLoRALayer
+from invokeai.backend.lora.layers.lora_layer import LoRALayer
+
+
+class LoRASidecarWrapper(torch.nn.Module):
+    def __init__(self, orig_module: torch.nn.Module, lora_layers: list[AnyLoRALayer], lora_weights: list[float]):
+        super().__init__()
+        self._orig_module = orig_module
+        self._lora_layers = lora_layers
+        self._lora_weights = lora_weights
+
+    @property
+    def orig_module(self) -> torch.nn.Module:
+        return self._orig_module
+
+    def add_lora_layer(self, lora_layer: AnyLoRALayer, lora_weight: float):
+        self._lora_layers.append(lora_layer)
+        self._lora_weights.append(lora_weight)
+
+    @torch.no_grad()
+    def _get_lora_patched_parameters(
+        self, orig_params: dict[str, torch.Tensor], lora_layers: list[AnyLoRALayer], lora_weights: list[float]
+    ) -> dict[str, torch.Tensor]:
+        params: dict[str, torch.Tensor] = {}
+        for lora_layer, lora_weight in zip(lora_layers, lora_weights, strict=True):
+            layer_params = lora_layer.get_parameters(self._orig_module)
+            for param_name, param_weight in layer_params.items():
+                if orig_params[param_name].shape != param_weight.shape:
+                    param_weight = param_weight.reshape(orig_params[param_name].shape)
+
+                if param_name not in params:
+                    params[param_name] = param_weight * (lora_layer.scale() * lora_weight)
+                else:
+                    params[param_name] += param_weight * (lora_layer.scale() * lora_weight)
+
+        return params
+
+
+class LoRALinearWrapper(LoRASidecarWrapper):
+    def _lora_linear_forward(self, input: torch.Tensor, lora_layer: LoRALayer, lora_weight: float) -> torch.Tensor:
+        """An optimized implementation of the residual calculation for a Linear LoRALayer."""
+        x = torch.nn.functional.linear(input, lora_layer.down)
+        if lora_layer.mid is not None:
+            x = torch.nn.functional.linear(x, lora_layer.mid)
+        x = torch.nn.functional.linear(x, lora_layer.up, bias=lora_layer.bias)
+        x *= lora_weight * lora_layer.scale()
+        return x
+
+    def _concatenated_lora_forward(
+        self, input: torch.Tensor, concatenated_lora_layer: ConcatenatedLoRALayer, lora_weight: float
+    ) -> torch.Tensor:
+        """An optimized implementation of the residual calculation for a Linear ConcatenatedLoRALayer."""
+        x_chunks: list[torch.Tensor] = []
+        for lora_layer in concatenated_lora_layer.lora_layers:
+            x_chunk = torch.nn.functional.linear(input, lora_layer.down)
+            if lora_layer.mid is not None:
+                x_chunk = torch.nn.functional.linear(x_chunk, lora_layer.mid)
+            x_chunk = torch.nn.functional.linear(x_chunk, lora_layer.up, bias=lora_layer.bias)
+            x_chunk *= lora_weight * lora_layer.scale()
+            x_chunks.append(x_chunk)
+
+        # TODO(ryand): Generalize to support concat_axis != 0.
+        assert concatenated_lora_layer.concat_axis == 0
+        x = torch.cat(x_chunks, dim=-1)
+        return x
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        # Split the LoRA layers into those that have optimized implementations and those that don't.
+        optimized_layer_types = (LoRALayer, ConcatenatedLoRALayer)
+        optimized_layers = [
+            (layer, weight)
+            for layer, weight in zip(self._lora_layers, self._lora_weights, strict=True)
+            if isinstance(layer, optimized_layer_types)
+        ]
+        non_optimized_layers = [
+            (layer, weight)
+            for layer, weight in zip(self._lora_layers, self._lora_weights, strict=True)
+            if not isinstance(layer, optimized_layer_types)
+        ]
+
+        # First, calculate the residual for LoRA layers for which there is an optimized implementation.
+        residual = None
+        for lora_layer, lora_weight in optimized_layers:
+            if isinstance(lora_layer, LoRALayer):
+                added_residual = self._lora_linear_forward(input, lora_layer, lora_weight)
+            elif isinstance(lora_layer, ConcatenatedLoRALayer):
+                added_residual = self._concatenated_lora_forward(input, lora_layer, lora_weight)
+            else:
+                raise ValueError(f"Unsupported LoRA layer type: {type(lora_layer)}")
+
+            if residual is None:
+                residual = added_residual
+            else:
+                residual += added_residual
+
+        # Next, calculate the residuals for the LoRA layers for which there is no optimized implementation.
+        if non_optimized_layers:
+            unoptimized_layers, unoptimized_weights = zip(*non_optimized_layers, strict=True)
+            params = self._get_lora_patched_parameters(
+                orig_params={"weight": self._orig_module.weight, "bias": self._orig_module.bias},
+                lora_layers=unoptimized_layers,
+                lora_weights=unoptimized_weights,
+            )
+            added_residual = torch.nn.functional.linear(input, params["weight"], params.get("bias", None))
+            if residual is None:
+                residual = added_residual
+            else:
+                residual += added_residual
+
+        return self.orig_module(input) + residual
+
+
+class LoRAConv1dWrapper(LoRASidecarWrapper):
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        params = self._get_lora_patched_parameters(
+            orig_params={"weight": self._orig_module.weight, "bias": self._orig_module.bias},
+            lora_layers=self._lora_layers,
+            lora_weights=self._lora_weights,
+        )
+        return self.orig_module(input) + torch.nn.functional.conv1d(input, params["weight"], params.get("bias", None))
+
+
+class LoRAConv2dWrapper(LoRASidecarWrapper):
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        params = self._get_lora_patched_parameters(
+            orig_params={"weight": self._orig_module.weight, "bias": self._orig_module.bias},
+            lora_layers=self._lora_layers,
+            lora_weights=self._lora_weights,
+        )
+        return self.orig_module(input) + torch.nn.functional.conv2d(input, params["weight"], params.get("bias", None))

invokeai/backend/lora/lora_patcher.py

@@ -4,19 +4,126 @@ from typing import Dict, Iterable, Optional, Tuple
 import torch
 
 from invokeai.backend.lora.layers.any_lora_layer import AnyLoRALayer
-from invokeai.backend.lora.layers.concatenated_lora_layer import ConcatenatedLoRALayer
-from invokeai.backend.lora.layers.lora_layer import LoRALayer
-from invokeai.backend.lora.lora_model_raw import LoRAModelRaw
-from invokeai.backend.lora.sidecar_layers.concatenated_lora.concatenated_lora_linear_sidecar_layer import (
-    ConcatenatedLoRALinearSidecarLayer,
+from invokeai.backend.lora.lora_layer_wrappers import (
+    LoRAConv1dWrapper,
+    LoRAConv2dWrapper,
+    LoRALinearWrapper,
+    LoRASidecarWrapper,
 )
-from invokeai.backend.lora.sidecar_layers.lora.lora_linear_sidecar_layer import LoRALinearSidecarLayer
-from invokeai.backend.lora.sidecar_layers.lora_sidecar_module import LoRASidecarModule
+from invokeai.backend.lora.lora_model_raw import LoRAModelRaw
 from invokeai.backend.util.devices import TorchDevice
 from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
 
 
 class LoRAPatcher:
+    @staticmethod
+    @torch.no_grad()
+    @contextmanager
+    def apply_smart_lora_patches(
+        model: torch.nn.Module,
+        patches: Iterable[Tuple[LoRAModelRaw, float]],
+        prefix: str,
+        dtype: torch.dtype,
+        cached_weights: Optional[Dict[str, torch.Tensor]] = None,
+    ):
+        """Apply 'smart' LoRA patching that chooses whether to use direct patching or a sidecar wrapper for each module."""
+
+        # original_weights are stored for unpatching layers that are directly patched.
+        original_weights = OriginalWeightsStorage(cached_weights)
+        # original_modules are stored for unpatching layers that are wrapped in a LoRASidecarWrapper.
+        original_modules: dict[str, torch.nn.Module] = {}
+        try:
+            for patch, patch_weight in patches:
+                LoRAPatcher._apply_smart_lora_patch(
+                    model=model,
+                    prefix=prefix,
+                    patch=patch,
+                    patch_weight=patch_weight,
+                    original_weights=original_weights,
+                    original_modules=original_modules,
+                    dtype=dtype,
+                )
+
+            yield
+        finally:
+            # Restore directly patched layers.
+            for param_key, weight in original_weights.get_changed_weights():
+                model.get_parameter(param_key).copy_(weight)
+
+            # Restore LoRASidecarWrapper modules.
+            # Note: This logic assumes no nested modules in original_modules.
+            for module_key, orig_module in original_modules.items():
+                module_parent_key, module_name = LoRAPatcher._split_parent_key(module_key)
+                parent_module = model.get_submodule(module_parent_key)
+                LoRAPatcher._set_submodule(parent_module, module_name, orig_module)
+
+    @staticmethod
+    @torch.no_grad()
+    def _apply_smart_lora_patch(
+        model: torch.nn.Module,
+        prefix: str,
+        patch: LoRAModelRaw,
+        patch_weight: float,
+        original_weights: OriginalWeightsStorage,
+        original_modules: dict[str, torch.nn.Module],
+        dtype: torch.dtype,
+    ):
+        """Apply a single LoRA patch to a model using the 'smart' patching strategy that chooses whether to use direct
+        patching or a sidecar wrapper for each module.
+        """
+        if patch_weight == 0:
+            return
+
+        # If the layer keys contain a dot, then they are not flattened, and can be directly used to access model
+        # submodules. If the layer keys do not contain a dot, then they are flattened, meaning that all '.' have been
+        # replaced with '_'. Non-flattened keys are preferred, because they allow submodules to be accessed directly
+        # without searching, but some legacy code still uses flattened keys.
+        layer_keys_are_flattened = "." not in next(iter(patch.layers.keys()))
+
+        prefix_len = len(prefix)
+
+        for layer_key, layer in patch.layers.items():
+            if not layer_key.startswith(prefix):
+                continue
+
+            module_key, module = LoRAPatcher._get_submodule(
+                model, layer_key[prefix_len:], layer_key_is_flattened=layer_keys_are_flattened
+            )
+
+            # Decide whether to use direct patching or a sidecar wrapper.
+            # Direct patching is preferred, because it results in better runtime speed.
+            # Reasons to use sidecar patching:
+            # - The module is already wrapped in a LoRASidecarWrapper.
+            # - The module is quantized.
+            # - The module is on the CPU (and we don't want to store a second full copy of the original weights on the
+            #   CPU, since this would double the RAM usage)
+            # NOTE: For now, we don't check if the layer is quantized here. We assume that this is checked in the caller
+            # and that the caller will use the 'apply_lora_wrapper_patches' method if the layer is quantized.
+            # TODO(ryand): Handle the case where we are running without a GPU. Should we set a config flag that allows
+            # forcing full patching even on the CPU?
+            if isinstance(module, LoRASidecarWrapper) or LoRAPatcher._is_any_part_of_layer_on_cpu(module):
+                LoRAPatcher._apply_lora_layer_wrapper_patch(
+                    model=model,
+                    module_to_patch=module,
+                    module_to_patch_key=module_key,
+                    patch=layer,
+                    patch_weight=patch_weight,
+                    original_modules=original_modules,
+                    dtype=dtype,
+                )
+            else:
+                LoRAPatcher._apply_lora_layer_patch(
+                    module_to_patch=module,
+                    module_to_patch_key=module_key,
+                    patch=layer,
+                    patch_weight=patch_weight,
+                    original_weights=original_weights,
+                )
+
+    @staticmethod
+    def _is_any_part_of_layer_on_cpu(layer: torch.nn.Module) -> bool:
+        return any(p.device.type == "cpu" for p in layer.parameters())
+
     @staticmethod
     @torch.no_grad()
     @contextmanager
@@ -40,7 +147,7 @@ class LoRAPatcher:
         original_weights = OriginalWeightsStorage(cached_weights)
         try:
             for patch, patch_weight in patches:
-                LoRAPatcher.apply_lora_patch(
+                LoRAPatcher._apply_lora_patch(
                     model=model,
                     prefix=prefix,
                     patch=patch,
@@ -52,11 +159,12 @@ class LoRAPatcher:
             yield
         finally:
             for param_key, weight in original_weights.get_changed_weights():
-                model.get_parameter(param_key).copy_(weight)
+                cur_param = model.get_parameter(param_key)
+                cur_param.data = weight.to(dtype=cur_param.dtype, device=cur_param.device, copy=True)
 
     @staticmethod
     @torch.no_grad()
-    def apply_lora_patch(
+    def _apply_lora_patch(
         model: torch.nn.Module,
         prefix: str,
         patch: LoRAModelRaw,
@@ -91,48 +199,84 @@ class LoRAPatcher:
                 model, layer_key[prefix_len:], layer_key_is_flattened=layer_keys_are_flattened
             )
 
-            # All of the LoRA weight calculations will be done on the same device as the module weight.
-            # (Performance will be best if this is a CUDA device.)
-            device = module.weight.device
-            dtype = module.weight.dtype
+            LoRAPatcher._apply_lora_layer_patch(
+                module_to_patch=module,
+                module_to_patch_key=module_key,
+                patch=layer,
+                patch_weight=patch_weight,
+                original_weights=original_weights,
+            )
 
-            layer_scale = layer.scale()
+    @staticmethod
+    @torch.no_grad()
+    def _apply_lora_layer_patch(
+        module_to_patch: torch.nn.Module,
+        module_to_patch_key: str,
+        patch: AnyLoRALayer,
+        patch_weight: float,
+        original_weights: OriginalWeightsStorage,
+    ):
+        # All of the LoRA weight calculations will be done on the same device as the module weight.
+        # (Performance will be best if this is a CUDA device.)
+        first_param = next(module_to_patch.parameters())
+        device = first_param.device
+        dtype = first_param.dtype
 
-            # We intentionally move to the target device first, then cast. Experimentally, this was found to
-            # be significantly faster for 16-bit CPU tensors being moved to a CUDA device than doing the
-            # same thing in a single call to '.to(...)'.
-            layer.to(device=device)
-            layer.to(dtype=torch.float32)
+        layer_scale = patch.scale()
 
-            # TODO(ryand): Using torch.autocast(...) over explicit casting may offer a speed benefit on CUDA
-            # devices here. Experimentally, it was found to be very slow on CPU. More investigation needed.
-            for param_name, lora_param_weight in layer.get_parameters(module).items():
-                param_key = module_key + "." + param_name
-                module_param = module.get_parameter(param_name)
+        # We intentionally move to the target device first, then cast. Experimentally, this was found to
+        # be significantly faster for 16-bit CPU tensors being moved to a CUDA device than doing the
+        # same thing in a single call to '.to(...)'.
+        patch.to(device=device)
+        patch.to(dtype=torch.float32)
 
-                # Save original weight
-                original_weights.save(param_key, module_param)
+        # TODO(ryand): Using torch.autocast(...) over explicit casting may offer a speed benefit on CUDA
+        # devices here. Experimentally, it was found to be very slow on CPU. More investigation needed.
+        for param_name, lora_param_weight in patch.get_parameters(module_to_patch).items():
+            param_key = module_to_patch_key + "." + param_name
+            module_param = module_to_patch.get_parameter(param_name)
 
-                if module_param.shape != lora_param_weight.shape:
+            # Save original weight
+            original_weights.save(param_key, module_param)
+
+            if module_param.shape != lora_param_weight.shape:
+                if module_param.nelement() == lora_param_weight.nelement():
                     lora_param_weight = lora_param_weight.reshape(module_param.shape)
+                else:
+                    # This condition was added to handle layers in FLUX control LoRAs.
+                    # TODO(ryand): Move the weight update into the LoRA layer so that the LoRAPatcher doesn't need
+                    # to worry about this?
+                    expanded_weight = torch.zeros_like(
+                        lora_param_weight, dtype=module_param.dtype, device=module_param.device
+                    )
+                    slices = tuple(slice(0, dim) for dim in module_param.shape)
+                    expanded_weight[slices] = module_param
+                    setattr(
+                        module,
+                        param_name,
+                        torch.nn.Parameter(expanded_weight, requires_grad=module_param.requires_grad),
+                    )
+                    module_param = expanded_weight
 
-                lora_param_weight *= patch_weight * layer_scale
-                module_param += lora_param_weight.to(dtype=dtype)
+            lora_param_weight *= patch_weight * layer_scale
+            module_param += lora_param_weight.to(dtype=dtype)
 
-            layer.to(device=TorchDevice.CPU_DEVICE)
+        patch.to(device=TorchDevice.CPU_DEVICE)
 
     @staticmethod
     @torch.no_grad()
     @contextmanager
-    def apply_lora_sidecar_patches(
+    def apply_lora_wrapper_patches(
         model: torch.nn.Module,
         patches: Iterable[Tuple[LoRAModelRaw, float]],
         prefix: str,
         dtype: torch.dtype,
     ):
-        """Apply one or more LoRA sidecar patches to a model within a context manager. Sidecar patches incur some
-        overhead compared to normal LoRA patching, but they allow for LoRA layers to applied to base layers in any
-        quantization format.
+        """Apply one or more LoRA wrapper patches to a model within a context manager. Wrapper patches incur some
+        runtime overhead compared to normal LoRA patching, but they enable:
+        - LoRA layers to be applied to quantized models
+        - LoRA layers to be applied to CPU layers without needing to store a full copy of the original weights (i.e.
+          avoid doubling the memory requirements).
 
         Args:
             model (torch.nn.Module): The model to patch.
@@ -140,14 +284,11 @@ class LoRAPatcher:
                 associated weights. An iterator is used so that the LoRA patches do not need to be loaded into memory
                 all at once.
             prefix (str): The keys in the patches will be filtered to only include weights with this prefix.
-            dtype (torch.dtype): The compute dtype of the sidecar layers. This cannot easily be inferred from the model,
-                since the sidecar layers are typically applied on top of quantized layers whose weight dtype is
-                different from their compute dtype.
         """
         original_modules: dict[str, torch.nn.Module] = {}
         try:
             for patch, patch_weight in patches:
-                LoRAPatcher._apply_lora_sidecar_patch(
+                LoRAPatcher._apply_lora_wrapper_patch(
                     model=model,
                     prefix=prefix,
                     patch=patch,
@@ -165,7 +306,7 @@ class LoRAPatcher:
                 LoRAPatcher._set_submodule(parent_module, module_name, orig_module)
 
     @staticmethod
-    def _apply_lora_sidecar_patch(
+    def _apply_lora_wrapper_patch(
         model: torch.nn.Module,
         patch: LoRAModelRaw,
         patch_weight: float,
@@ -173,7 +314,7 @@ class LoRAPatcher:
         original_modules: dict[str, torch.nn.Module],
         dtype: torch.dtype,
     ):
-        """Apply a single LoRA sidecar patch to a model."""
+        """Apply a single LoRA wrapper patch to a model."""
 
         if patch_weight == 0:
             return
@@ -194,28 +335,47 @@ class LoRAPatcher:
                 model, layer_key[prefix_len:], layer_key_is_flattened=layer_keys_are_flattened
             )
 
-            # Initialize the LoRA sidecar layer.
-            lora_sidecar_layer = LoRAPatcher._initialize_lora_sidecar_layer(module, layer, patch_weight)
+            LoRAPatcher._apply_lora_layer_wrapper_patch(
+                model=model,
+                module_to_patch=module,
+                module_to_patch_key=module_key,
+                patch=layer,
+                patch_weight=patch_weight,
+                original_modules=original_modules,
+                dtype=dtype,
+            )
 
-            # Replace the original module with a LoRASidecarModule if it has not already been done.
-            if module_key in original_modules:
-                # The module has already been patched with a LoRASidecarModule. Append to it.
-                assert isinstance(module, LoRASidecarModule)
-                lora_sidecar_module = module
-            else:
-                # The module has not yet been patched with a LoRASidecarModule. Create one.
-                lora_sidecar_module = LoRASidecarModule(module, [])
-                original_modules[module_key] = module
-                module_parent_key, module_name = LoRAPatcher._split_parent_key(module_key)
-                module_parent = model.get_submodule(module_parent_key)
-                LoRAPatcher._set_submodule(module_parent, module_name, lora_sidecar_module)
+    @staticmethod
+    @torch.no_grad()
+    def _apply_lora_layer_wrapper_patch(
+        model: torch.nn.Module,
+        module_to_patch: torch.nn.Module,
+        module_to_patch_key: str,
+        patch: AnyLoRALayer,
+        patch_weight: float,
+        original_modules: dict[str, torch.nn.Module],
+        dtype: torch.dtype,
+    ):
+        """Apply a single LoRA wrapper patch to a model."""
 
-            # Move the LoRA sidecar layer to the same device/dtype as the orig module.
-            # TODO(ryand): Experiment with moving to the device first, then casting. This could be faster.
-            lora_sidecar_layer.to(device=lora_sidecar_module.orig_module.weight.device, dtype=dtype)
+        # Replace the original module with a LoRASidecarWrapper if it has not already been done.
+        if not isinstance(module_to_patch, LoRASidecarWrapper):
+            lora_wrapper_layer = LoRAPatcher._initialize_lora_wrapper_layer(module_to_patch)
+            original_modules[module_to_patch_key] = module_to_patch
+            module_parent_key, module_name = LoRAPatcher._split_parent_key(module_to_patch_key)
+            module_parent = model.get_submodule(module_parent_key)
+            LoRAPatcher._set_submodule(module_parent, module_name, lora_wrapper_layer)
+            orig_module = module_to_patch
+        else:
+            assert module_to_patch_key in original_modules
+            lora_wrapper_layer = module_to_patch
+            orig_module = module_to_patch.orig_module
 
-            # Add the LoRA sidecar layer to the LoRASidecarModule.
-            lora_sidecar_module.add_lora_layer(lora_sidecar_layer)
+        # Move the LoRA layer to the same device/dtype as the orig module.
+        patch.to(device=orig_module.weight.device, dtype=dtype)
+
+        # Add the LoRA wrapper layer to the LoRASidecarWrapper.
+        lora_wrapper_layer.add_lora_layer(patch, patch_weight)
 
     @staticmethod
     def _split_parent_key(module_key: str) -> tuple[str, str]:
@@ -236,17 +396,13 @@ class LoRAPatcher:
             raise ValueError(f"Invalid module key: {module_key}")
 
     @staticmethod
-    def _initialize_lora_sidecar_layer(orig_layer: torch.nn.Module, lora_layer: AnyLoRALayer, patch_weight: float):
-        # TODO(ryand): Add support for more original layer types and LoRA layer types.
-        if isinstance(orig_layer, torch.nn.Linear) or (
-            isinstance(orig_layer, LoRASidecarModule) and isinstance(orig_layer.orig_module, torch.nn.Linear)
-        ):
-            if isinstance(lora_layer, LoRALayer):
-                return LoRALinearSidecarLayer(lora_layer=lora_layer, weight=patch_weight)
-            elif isinstance(lora_layer, ConcatenatedLoRALayer):
-                return ConcatenatedLoRALinearSidecarLayer(concatenated_lora_layer=lora_layer, weight=patch_weight)
-            else:
-                raise ValueError(f"Unsupported Linear LoRA layer type: {type(lora_layer)}")
+    def _initialize_lora_wrapper_layer(orig_layer: torch.nn.Module):
+        if isinstance(orig_layer, torch.nn.Linear):
+            return LoRALinearWrapper(orig_layer, [], [])
+        elif isinstance(orig_layer, torch.nn.Conv1d):
+            return LoRAConv1dWrapper(orig_layer, [], [])
+        elif isinstance(orig_layer, torch.nn.Conv2d):
+            return LoRAConv2dWrapper(orig_layer, [], [])
         else:
             raise ValueError(f"Unsupported layer type: {type(orig_layer)}")
 

invokeai/backend/lora/sidecar_layers/concatenated_lora/concatenated_lora_linear_sidecar_layer.py

@@ -1,34 +0,0 @@
-import torch
-
-from invokeai.backend.lora.layers.concatenated_lora_layer import ConcatenatedLoRALayer
-
-
-class ConcatenatedLoRALinearSidecarLayer(torch.nn.Module):
-    def __init__(
-        self,
-        concatenated_lora_layer: ConcatenatedLoRALayer,
-        weight: float,
-    ):
-        super().__init__()
-
-        self._concatenated_lora_layer = concatenated_lora_layer
-        self._weight = weight
-
-    def forward(self, input: torch.Tensor) -> torch.Tensor:
-        x_chunks: list[torch.Tensor] = []
-        for lora_layer in self._concatenated_lora_layer.lora_layers:
-            x_chunk = torch.nn.functional.linear(input, lora_layer.down)
-            if lora_layer.mid is not None:
-                x_chunk = torch.nn.functional.linear(x_chunk, lora_layer.mid)
-            x_chunk = torch.nn.functional.linear(x_chunk, lora_layer.up, bias=lora_layer.bias)
-            x_chunk *= self._weight * lora_layer.scale()
-            x_chunks.append(x_chunk)
-
-        # TODO(ryand): Generalize to support concat_axis != 0.
-        assert self._concatenated_lora_layer.concat_axis == 0
-        x = torch.cat(x_chunks, dim=-1)
-        return x
-
-    def to(self, device: torch.device | None = None, dtype: torch.dtype | None = None):
-        self._concatenated_lora_layer.to(device=device, dtype=dtype)
-        return self

invokeai/backend/lora/sidecar_layers/lora/lora_linear_sidecar_layer.py

@@ -1,27 +0,0 @@
-import torch
-
-from invokeai.backend.lora.layers.lora_layer import LoRALayer
-
-
-class LoRALinearSidecarLayer(torch.nn.Module):
-    def __init__(
-        self,
-        lora_layer: LoRALayer,
-        weight: float,
-    ):
-        super().__init__()
-
-        self._lora_layer = lora_layer
-        self._weight = weight
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        x = torch.nn.functional.linear(x, self._lora_layer.down)
-        if self._lora_layer.mid is not None:
-            x = torch.nn.functional.linear(x, self._lora_layer.mid)
-        x = torch.nn.functional.linear(x, self._lora_layer.up, bias=self._lora_layer.bias)
-        x *= self._weight * self._lora_layer.scale()
-        return x
-
-    def to(self, device: torch.device | None = None, dtype: torch.dtype | None = None):
-        self._lora_layer.to(device=device, dtype=dtype)
-        return self

invokeai/backend/lora/sidecar_layers/lora_sidecar_module.py

@@ -1,24 +0,0 @@
-import torch
-
-
-class LoRASidecarModule(torch.nn.Module):
-    """A LoRA sidecar module that wraps an original module and adds LoRA layers to it."""
-
-    def __init__(self, orig_module: torch.nn.Module, lora_layers: list[torch.nn.Module]):
-        super().__init__()
-        self.orig_module = orig_module
-        self._lora_layers = lora_layers
-
-    def add_lora_layer(self, lora_layer: torch.nn.Module):
-        self._lora_layers.append(lora_layer)
-
-    def forward(self, input: torch.Tensor) -> torch.Tensor:
-        x = self.orig_module(input)
-        for lora_layer in self._lora_layers:
-            x += lora_layer(input)
-        return x
-
-    def to(self, device: torch.device | None = None, dtype: torch.dtype | None = None):
-        self._orig_module.to(device=device, dtype=dtype)
-        for lora_layer in self._lora_layers:
-            lora_layer.to(device=device, dtype=dtype)

invokeai/backend/model_manager/config.py

@@ -67,6 +67,7 @@ class ModelType(str, Enum):
     Main = "main"
     VAE = "vae"
     LoRA = "lora"
+    StructuralLoRa = "structural_lora"
     ControlNet = "controlnet"  # used by model_probe
     TextualInversion = "embedding"
     IPAdapter = "ip_adapter"
@@ -273,6 +274,18 @@ class LoRALyCORISConfig(LoRAConfigBase):
         return Tag(f"{ModelType.LoRA.value}.{ModelFormat.LyCORIS.value}")
 
 
+class StructuralLoRALyCORISConfig(ModelConfigBase):
+    """Model config for Structural LoRA/Lycoris models."""
+
+    type: Literal[ModelType.StructuralLoRa] = ModelType.StructuralLoRa
+    trigger_phrases: Optional[set[str]] = Field(description="Set of trigger phrases for this model", default=None)
+    format: Literal[ModelFormat.LyCORIS] = ModelFormat.LyCORIS
+
+    @staticmethod
+    def get_tag() -> Tag:
+        return Tag(f"{ModelType.StructuralLoRa.value}.{ModelFormat.LyCORIS.value}")
+
+
 class LoRADiffusersConfig(LoRAConfigBase):
     """Model config for LoRA/Diffusers models."""
 
@@ -535,6 +548,7 @@ AnyModelConfig = Annotated[
         Annotated[ControlNetDiffusersConfig, ControlNetDiffusersConfig.get_tag()],
         Annotated[ControlNetCheckpointConfig, ControlNetCheckpointConfig.get_tag()],
         Annotated[LoRALyCORISConfig, LoRALyCORISConfig.get_tag()],
+        Annotated[StructuralLoRALyCORISConfig, StructuralLoRALyCORISConfig.get_tag()],
         Annotated[LoRADiffusersConfig, LoRADiffusersConfig.get_tag()],
         Annotated[T5EncoderConfig, T5EncoderConfig.get_tag()],
         Annotated[T5EncoderBnbQuantizedLlmInt8bConfig, T5EncoderBnbQuantizedLlmInt8bConfig.get_tag()],

invokeai/backend/model_manager/load/model_loaders/controlnet.py

@@ -1,12 +1,10 @@
 # Copyright (c) 2024, Lincoln D. Stein and the InvokeAI Development Team
 """Class for ControlNet model loading in InvokeAI."""
 
-from pathlib import Path
 from typing import Optional
 
 from diffusers import ControlNetModel
 
-import invokeai.backend.assets.model_base_conf_files as conf_file_cache
 from invokeai.backend.model_manager import (
     AnyModel,
     AnyModelConfig,
@@ -48,20 +46,9 @@ class ControlNetLoader(GenericDiffusersLoader):
         config: AnyModelConfig,
         submodel_type: Optional[SubModelType] = None,
     ) -> AnyModel:
-        config_dirs = {
-            BaseModelType.StableDiffusion1: "controlnet_sd15",
-            BaseModelType.StableDiffusionXL: "controlnet_sdxl",
-        }
-        try:
-            config_dir = config_dirs[config.base]
-        except KeyError:
-            raise Exception(f"No configuration template known for controlnet model with base={config.base}")
-
         if isinstance(config, ControlNetCheckpointConfig):
             return ControlNetModel.from_single_file(
                 config.path,
-                config=Path(conf_file_cache.__path__[0], config_dir).as_posix(),
-                local_files_only=True,
                 torch_dtype=self._torch_dtype,
             )
         else:

invokeai/backend/model_manager/load/model_loaders/lora.py

@@ -13,8 +13,9 @@ from invokeai.backend.lora.conversions.flux_diffusers_lora_conversion_utils impo
     lora_model_from_flux_diffusers_state_dict,
 )
 from invokeai.backend.lora.conversions.flux_kohya_lora_conversion_utils import (
-    lora_model_from_flux_kohya_state_dict,
+    is_state_dict_likely_in_flux_kohya_format, lora_model_from_flux_kohya_state_dict,
 )
+from invokeai.backend.lora.conversions.flux_control_lora_utils import is_state_dict_likely_flux_control, lora_model_from_flux_control_state_dict
 from invokeai.backend.lora.conversions.sd_lora_conversion_utils import lora_model_from_sd_state_dict
 from invokeai.backend.lora.conversions.sdxl_lora_conversion_utils import convert_sdxl_keys_to_diffusers_format
 from invokeai.backend.model_manager import (
@@ -32,6 +33,7 @@ from invokeai.backend.model_manager.load.model_loader_registry import ModelLoade
 
 @ModelLoaderRegistry.register(base=BaseModelType.Any, type=ModelType.LoRA, format=ModelFormat.Diffusers)
 @ModelLoaderRegistry.register(base=BaseModelType.Any, type=ModelType.LoRA, format=ModelFormat.LyCORIS)
+@ModelLoaderRegistry.register(base=BaseModelType.Flux, type=ModelType.StructuralLoRa, format=ModelFormat.LyCORIS)
 class LoRALoader(ModelLoader):
     """Class to load LoRA models."""
 
@@ -75,7 +77,10 @@ class LoRALoader(ModelLoader):
                 # https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora_flux.py#L1194
                 model = lora_model_from_flux_diffusers_state_dict(state_dict=state_dict, alpha=None)
             elif config.format == ModelFormat.LyCORIS:
-                model = lora_model_from_flux_kohya_state_dict(state_dict=state_dict)
+                if is_state_dict_likely_in_flux_kohya_format(state_dict=state_dict):
+                    model = lora_model_from_flux_kohya_state_dict(state_dict=state_dict)
+                elif is_state_dict_likely_flux_control(state_dict=state_dict):
+                    model = lora_model_from_flux_control_state_dict(state_dict=state_dict)
             else:
                 raise ValueError(f"LoRA model is in unsupported FLUX format: {config.format}")
         elif self._model_base in [BaseModelType.StableDiffusion1, BaseModelType.StableDiffusion2]:

invokeai/backend/model_manager/load/model_loaders/stable_diffusion.py

@@ -11,7 +11,6 @@ from diffusers import (
     StableDiffusionXLPipeline,
 )
 
-import invokeai.backend.assets.model_base_conf_files as conf_file_cache
 from invokeai.backend.model_manager import (
     AnyModel,
     AnyModelConfig,
@@ -19,7 +18,6 @@ from invokeai.backend.model_manager import (
     ModelFormat,
     ModelType,
     ModelVariantType,
-    SchedulerPredictionType,
     SubModelType,
 )
 from invokeai.backend.model_manager.config import (
@@ -108,33 +106,11 @@ class StableDiffusionDiffusersModel(GenericDiffusersLoader):
                 ModelVariantType.Normal: StableDiffusionXLPipeline,
             },
         }
-        config_dirs = {
-            BaseModelType.StableDiffusion1: {
-                SchedulerPredictionType.Epsilon: "stable-diffusion-1.5-epsilon",
-                SchedulerPredictionType.VPrediction: "stable-diffusion-1.5-v_prediction",
-            },
-            BaseModelType.StableDiffusion2: {
-                SchedulerPredictionType.VPrediction: "stable-diffusion-2.0-v_prediction",
-            },
-            BaseModelType.StableDiffusionXL: {
-                SchedulerPredictionType.Epsilon: "stable-diffusion-xl-base-1.0",
-            },
-            BaseModelType.StableDiffusionXLRefiner: {
-                SchedulerPredictionType.Epsilon: "stable-diffusion-xl-refiner-1.0",
-            },
-        }
-
         assert isinstance(config, MainCheckpointConfig)
         try:
             load_class = load_classes[config.base][config.variant]
         except KeyError as e:
             raise Exception(f"No diffusers pipeline known for base={config.base}, variant={config.variant}") from e
-        try:
-            config_dir = config_dirs[config.base][config.prediction_type]
-        except KeyError as e:
-            raise Exception(
-                f"No configuration template known for base={config.base}, prediction_type={config.prediction_type}"
-            ) from e
 
         # Without SilenceWarnings we get log messages like this:
         # site-packages/huggingface_hub/file_download.py:1132: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
@@ -144,17 +120,8 @@ class StableDiffusionDiffusersModel(GenericDiffusersLoader):
         # Some weights of the model checkpoint were not used when initializing CLIPTextModelWithProjection:
         # ['text_model.embeddings.position_ids']
 
-        original_config_file = self._app_config.legacy_conf_path / config.config_path
-
         with SilenceWarnings():
-            pipeline = load_class.from_single_file(
-                config.path,
-                config=Path(conf_file_cache.__path__[0], config_dir).as_posix(),
-                original_config=original_config_file,
-                torch_dtype=self._torch_dtype,
-                local_files_only=True,
-                kwargs={"load_safety_checker": False},
-            )
+            pipeline = load_class.from_single_file(config.path, torch_dtype=self._torch_dtype)
 
         if not submodel_type:
             return pipeline