chore: 4.2.9rc2 version bump

Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>

.github/workflows/build-container.yml

@@ -13,6 +13,12 @@ on:
     tags:
       - 'v*.*.*'
   workflow_dispatch:
+    inputs:
+      push-to-registry:
+        description: Push the built image to the container registry
+        required: false
+        type: boolean
+        default: false
 
 permissions:
   contents: write
@@ -50,16 +56,15 @@ jobs:
           df -h
 
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Docker meta
         id: meta
-        uses: docker/metadata-action@v4
+        uses: docker/metadata-action@v5
         with:
           github-token: ${{ secrets.GITHUB_TOKEN }}
           images: |
             ghcr.io/${{ github.repository }}
-            ${{ env.DOCKERHUB_REPOSITORY }}
           tags: |
             type=ref,event=branch
             type=ref,event=tag
@@ -72,49 +77,33 @@ jobs:
             suffix=-${{ matrix.gpu-driver }},onlatest=false
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: docker/setup-qemu-action@v3
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
         with:
           platforms: ${{ env.PLATFORMS }}
 
       - name: Login to GitHub Container Registry
         if: github.event_name != 'pull_request'
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           registry: ghcr.io
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
-      # - name: Login to Docker Hub
-      #   if: github.event_name != 'pull_request' && vars.DOCKERHUB_REPOSITORY != ''
-      #   uses: docker/login-action@v2
-      #   with:
-      #     username: ${{ secrets.DOCKERHUB_USERNAME }}
-      #     password: ${{ secrets.DOCKERHUB_TOKEN }}
-
       - name: Build container
         timeout-minutes: 40
         id: docker_build
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v6
         with:
           context: .
           file: docker/Dockerfile
           platforms: ${{ env.PLATFORMS }}
-          push: ${{ github.ref == 'refs/heads/main' || github.ref_type == 'tag' }}
+          push: ${{ github.ref == 'refs/heads/main' || github.ref_type == 'tag' || github.event.inputs.push-to-registry }}
           tags: ${{ steps.meta.outputs.tags }}
           labels: ${{ steps.meta.outputs.labels }}
           cache-from: |
             type=gha,scope=${{ github.ref_name }}-${{ matrix.gpu-driver }}
             type=gha,scope=main-${{ matrix.gpu-driver }}
           cache-to: type=gha,mode=max,scope=${{ github.ref_name }}-${{ matrix.gpu-driver }}
-
-      # - name: Docker Hub Description
-      #   if: github.ref == 'refs/heads/main' || github.ref == 'refs/tags/*' && vars.DOCKERHUB_REPOSITORY != ''
-      #   uses: peter-evans/dockerhub-description@v3
-      #   with:
-      #     username: ${{ secrets.DOCKERHUB_USERNAME }}
-      #     password: ${{ secrets.DOCKERHUB_TOKEN }}
-      #     repository: ${{ vars.DOCKERHUB_REPOSITORY }}
-      #     short-description: ${{ github.event.repository.description }}

docs/help/FAQ.md

@@ -196,6 +196,22 @@ tips to reduce the problem:
     === "12GB VRAM GPU"
 
         This should be sufficient to generate larger images up to about 1280x1280.
+		
+## Checkpoint Models Load Slowly or Use Too Much RAM
+
+The difference between diffusers models (a folder containing multiple
+subfolders) and checkpoint models (a file ending with .safetensors or
+.ckpt) is that InvokeAI is able to load diffusers models into memory
+incrementally, while checkpoint models must be loaded all at
+once. With very large models, or systems with limited RAM, you may
+experience slowdowns and other memory-related issues when loading
+checkpoint models.
+
+To solve this, go to the Model Manager tab (the cube), select the
+checkpoint model that's giving you trouble, and press the "Convert"
+button in the upper right of your browser window. This will conver the
+checkpoint into a diffusers model, after which loading should be
+faster and less memory-intensive.
 
 ## Memory Leak (Linux)
 

invokeai/app/api/routers/model_manager.py

@@ -3,8 +3,10 @@
 
 import io
 import pathlib
+import shutil
 import traceback
 from copy import deepcopy
+from enum import Enum
 from tempfile import TemporaryDirectory
 from typing import List, Optional, Type
 
@@ -17,6 +19,7 @@ from starlette.exceptions import HTTPException
 from typing_extensions import Annotated
 
 from invokeai.app.api.dependencies import ApiDependencies
+from invokeai.app.services.config import get_config
 from invokeai.app.services.model_images.model_images_common import ModelImageFileNotFoundException
 from invokeai.app.services.model_install.model_install_common import ModelInstallJob
 from invokeai.app.services.model_records import (
@@ -31,6 +34,7 @@ from invokeai.backend.model_manager.config import (
     ModelFormat,
     ModelType,
 )
+from invokeai.backend.model_manager.load.model_cache.model_cache_base import CacheStats
 from invokeai.backend.model_manager.metadata.fetch.huggingface import HuggingFaceMetadataFetch
 from invokeai.backend.model_manager.metadata.metadata_base import ModelMetadataWithFiles, UnknownMetadataException
 from invokeai.backend.model_manager.search import ModelSearch
@@ -50,6 +54,13 @@ class ModelsList(BaseModel):
     model_config = ConfigDict(use_enum_values=True)
 
 
+class CacheType(str, Enum):
+    """Cache type - one of vram or ram."""
+
+    RAM = "RAM"
+    VRAM = "VRAM"
+
+
 def add_cover_image_to_model_config(config: AnyModelConfig, dependencies: Type[ApiDependencies]) -> AnyModelConfig:
     """Add a cover image URL to a model configuration."""
     cover_image = dependencies.invoker.services.model_images.get_url(config.key)
@@ -797,3 +808,83 @@ async def get_starter_models() -> list[StarterModel]:
         model.dependencies = missing_deps
 
     return starter_models
+
+
+@model_manager_router.get(
+    "/model_cache",
+    operation_id="get_cache_size",
+    response_model=float,
+    summary="Get maximum size of model manager RAM or VRAM cache.",
+)
+async def get_cache_size(cache_type: CacheType = Query(description="The cache type", default=CacheType.RAM)) -> float:
+    """Return the current RAM or VRAM cache size setting (in GB)."""
+    cache = ApiDependencies.invoker.services.model_manager.load.ram_cache
+    value = 0.0
+    if cache_type == CacheType.RAM:
+        value = cache.max_cache_size
+    elif cache_type == CacheType.VRAM:
+        value = cache.max_vram_cache_size
+    return value
+
+
+@model_manager_router.put(
+    "/model_cache",
+    operation_id="set_cache_size",
+    response_model=float,
+    summary="Set maximum size of model manager RAM or VRAM cache, optionally writing new value out to invokeai.yaml config file.",
+)
+async def set_cache_size(
+    value: float = Query(description="The new value for the maximum cache size"),
+    cache_type: CacheType = Query(description="The cache type", default=CacheType.RAM),
+    persist: bool = Query(description="Write new value out to invokeai.yaml", default=False),
+) -> float:
+    """Set the current RAM or VRAM cache size setting (in GB). ."""
+    cache = ApiDependencies.invoker.services.model_manager.load.ram_cache
+    app_config = get_config()
+    # Record initial state.
+    vram_old = app_config.vram
+    ram_old = app_config.ram
+
+    # Prepare target state.
+    vram_new = vram_old
+    ram_new = ram_old
+    if cache_type == CacheType.RAM:
+        ram_new = value
+    elif cache_type == CacheType.VRAM:
+        vram_new = value
+    else:
+        raise ValueError(f"Unexpected {cache_type=}.")
+
+    config_path = app_config.config_file_path
+    new_config_path = config_path.with_suffix(".yaml.new")
+
+    try:
+        # Try to apply the target state.
+        cache.max_vram_cache_size = vram_new
+        cache.max_cache_size = ram_new
+        app_config.ram = ram_new
+        app_config.vram = vram_new
+        if persist:
+            app_config.write_file(new_config_path)
+            shutil.move(new_config_path, config_path)
+    except Exception as e:
+        # If there was a failure, restore the initial state.
+        cache.max_cache_size = ram_old
+        cache.max_vram_cache_size = vram_old
+        app_config.ram = ram_old
+        app_config.vram = vram_old
+
+        raise RuntimeError("Failed to update cache size") from e
+    return value
+
+
+@model_manager_router.get(
+    "/stats",
+    operation_id="get_stats",
+    response_model=Optional[CacheStats],
+    summary="Get model manager RAM cache performance statistics.",
+)
+async def get_stats() -> Optional[CacheStats]:
+    """Return performance statistics on the model manager's RAM cache. Will return null if no models have been loaded."""
+
+    return ApiDependencies.invoker.services.model_manager.load.ram_cache.stats

invokeai/app/invocations/denoise_latents.py

@@ -185,7 +185,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
     )
     denoise_mask: Optional[DenoiseMaskField] = InputField(
         default=None,
-        description=FieldDescriptions.mask,
+        description=FieldDescriptions.denoise_mask,
         input=Input.Connection,
         ui_order=8,
     )

invokeai/app/invocations/fields.py

@@ -181,7 +181,7 @@ class FieldDescriptions:
     )
     num_1 = "The first number"
     num_2 = "The second number"
-    mask = "The mask to use for the operation"
+    denoise_mask = "A mask of the region to apply the denoising process to."
     board = "The board to save the image to"
     image = "The image to process"
     tile_size = "Tile size"

invokeai/app/invocations/flux_denoise.py

@@ -0,0 +1,249 @@
+from typing import Callable, Optional
+
+import torch
+import torchvision.transforms as tv_transforms
+from torchvision.transforms.functional import resize as tv_resize
+
+from invokeai.app.invocations.baseinvocation import BaseInvocation, Classification, invocation
+from invokeai.app.invocations.fields import (
+    DenoiseMaskField,
+    FieldDescriptions,
+    FluxConditioningField,
+    Input,
+    InputField,
+    LatentsField,
+    WithBoard,
+    WithMetadata,
+)
+from invokeai.app.invocations.model import TransformerField
+from invokeai.app.invocations.primitives import LatentsOutput
+from invokeai.app.services.shared.invocation_context import InvocationContext
+from invokeai.backend.flux.denoise import denoise
+from invokeai.backend.flux.inpaint_extension import InpaintExtension
+from invokeai.backend.flux.model import Flux
+from invokeai.backend.flux.sampling_utils import (
+    clip_timestep_schedule,
+    generate_img_ids,
+    get_noise,
+    get_schedule,
+    pack,
+    unpack,
+)
+from invokeai.backend.stable_diffusion.diffusers_pipeline import PipelineIntermediateState
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import FLUXConditioningInfo
+from invokeai.backend.util.devices import TorchDevice
+
+
+@invocation(
+    "flux_denoise",
+    title="FLUX Denoise",
+    tags=["image", "flux"],
+    category="image",
+    version="1.0.0",
+    classification=Classification.Prototype,
+)
+class FluxDenoiseInvocation(BaseInvocation, WithMetadata, WithBoard):
+    """Run denoising process with a FLUX transformer model."""
+
+    # If latents is provided, this means we are doing image-to-image.
+    latents: Optional[LatentsField] = InputField(
+        default=None,
+        description=FieldDescriptions.latents,
+        input=Input.Connection,
+    )
+    # denoise_mask is used for image-to-image inpainting. Only the masked region is modified.
+    denoise_mask: Optional[DenoiseMaskField] = InputField(
+        default=None,
+        description=FieldDescriptions.denoise_mask,
+        input=Input.Connection,
+    )
+    denoising_start: float = InputField(
+        default=0.0,
+        ge=0,
+        le=1,
+        description=FieldDescriptions.denoising_start,
+    )
+    denoising_end: float = InputField(default=1.0, ge=0, le=1, description=FieldDescriptions.denoising_end)
+    transformer: TransformerField = InputField(
+        description=FieldDescriptions.flux_model,
+        input=Input.Connection,
+        title="Transformer",
+    )
+    positive_text_conditioning: FluxConditioningField = InputField(
+        description=FieldDescriptions.positive_cond, input=Input.Connection
+    )
+    width: int = InputField(default=1024, multiple_of=16, description="Width of the generated image.")
+    height: int = InputField(default=1024, multiple_of=16, description="Height of the generated image.")
+    num_steps: int = InputField(
+        default=4, description="Number of diffusion steps. Recommended values are schnell: 4, dev: 50."
+    )
+    guidance: float = InputField(
+        default=4.0,
+        description="The guidance strength. Higher values adhere more strictly to the prompt, and will produce less diverse images. FLUX dev only, ignored for schnell.",
+    )
+    seed: int = InputField(default=0, description="Randomness seed for reproducibility.")
+
+    @torch.no_grad()
+    def invoke(self, context: InvocationContext) -> LatentsOutput:
+        latents = self._run_diffusion(context)
+        latents = latents.detach().to("cpu")
+
+        name = context.tensors.save(tensor=latents)
+        return LatentsOutput.build(latents_name=name, latents=latents, seed=None)
+
+    def _run_diffusion(
+        self,
+        context: InvocationContext,
+    ):
+        inference_dtype = torch.bfloat16
+
+        # Load the conditioning data.
+        cond_data = context.conditioning.load(self.positive_text_conditioning.conditioning_name)
+        assert len(cond_data.conditionings) == 1
+        flux_conditioning = cond_data.conditionings[0]
+        assert isinstance(flux_conditioning, FLUXConditioningInfo)
+        flux_conditioning = flux_conditioning.to(dtype=inference_dtype)
+        t5_embeddings = flux_conditioning.t5_embeds
+        clip_embeddings = flux_conditioning.clip_embeds
+
+        # Load the input latents, if provided.
+        init_latents = context.tensors.load(self.latents.latents_name) if self.latents else None
+        if init_latents is not None:
+            init_latents = init_latents.to(device=TorchDevice.choose_torch_device(), dtype=inference_dtype)
+
+        # Prepare input noise.
+        noise = get_noise(
+            num_samples=1,
+            height=self.height,
+            width=self.width,
+            device=TorchDevice.choose_torch_device(),
+            dtype=inference_dtype,
+            seed=self.seed,
+        )
+
+        transformer_info = context.models.load(self.transformer.transformer)
+        is_schnell = "schnell" in transformer_info.config.config_path
+
+        # Calculate the timestep schedule.
+        image_seq_len = noise.shape[-1] * noise.shape[-2] // 4
+        timesteps = get_schedule(
+            num_steps=self.num_steps,
+            image_seq_len=image_seq_len,
+            shift=not is_schnell,
+        )
+
+        # Clip the timesteps schedule based on denoising_start and denoising_end.
+        timesteps = clip_timestep_schedule(timesteps, self.denoising_start, self.denoising_end)
+
+        # Prepare input latent image.
+        if init_latents is not None:
+            # If init_latents is provided, we are doing image-to-image.
+
+            if is_schnell:
+                context.logger.warning(
+                    "Running image-to-image with a FLUX schnell model. This is not recommended. The results are likely "
+                    "to be poor. Consider using a FLUX dev model instead."
+                )
+
+            # Noise the orig_latents by the appropriate amount for the first timestep.
+            t_0 = timesteps[0]
+            x = t_0 * noise + (1.0 - t_0) * init_latents
+        else:
+            # init_latents are not provided, so we are not doing image-to-image (i.e. we are starting from pure noise).
+            if self.denoising_start > 1e-5:
+                raise ValueError("denoising_start should be 0 when initial latents are not provided.")
+
+            x = noise
+
+        # If len(timesteps) == 1, then short-circuit. We are just noising the input latents, but not taking any
+        # denoising steps.
+        if len(timesteps) <= 1:
+            return x
+
+        inpaint_mask = self._prep_inpaint_mask(context, x)
+
+        b, _c, h, w = x.shape
+        img_ids = generate_img_ids(h=h, w=w, batch_size=b, device=x.device, dtype=x.dtype)
+
+        bs, t5_seq_len, _ = t5_embeddings.shape
+        txt_ids = torch.zeros(bs, t5_seq_len, 3, dtype=inference_dtype, device=TorchDevice.choose_torch_device())
+
+        # Pack all latent tensors.
+        init_latents = pack(init_latents) if init_latents is not None else None
+        inpaint_mask = pack(inpaint_mask) if inpaint_mask is not None else None
+        noise = pack(noise)
+        x = pack(x)
+
+        # Now that we have 'packed' the latent tensors, verify that we calculated the image_seq_len correctly.
+        assert image_seq_len == x.shape[1]
+
+        # Prepare inpaint extension.
+        inpaint_extension: InpaintExtension | None = None
+        if inpaint_mask is not None:
+            assert init_latents is not None
+            inpaint_extension = InpaintExtension(
+                init_latents=init_latents,
+                inpaint_mask=inpaint_mask,
+                noise=noise,
+            )
+
+        with transformer_info as transformer:
+            assert isinstance(transformer, Flux)
+
+            x = denoise(
+                model=transformer,
+                img=x,
+                img_ids=img_ids,
+                txt=t5_embeddings,
+                txt_ids=txt_ids,
+                vec=clip_embeddings,
+                timesteps=timesteps,
+                step_callback=self._build_step_callback(context),
+                guidance=self.guidance,
+                inpaint_extension=inpaint_extension,
+            )
+
+        x = unpack(x.float(), self.height, self.width)
+        return x
+
+    def _prep_inpaint_mask(self, context: InvocationContext, latents: torch.Tensor) -> torch.Tensor | None:
+        """Prepare the inpaint mask.
+
+        - Loads the mask
+        - Resizes if necessary
+        - Casts to same device/dtype as latents
+        - Expands mask to the same shape as latents so that they line up after 'packing'
+
+        Args:
+            context (InvocationContext): The invocation context, for loading the inpaint mask.
+            latents (torch.Tensor): A latent image tensor. In 'unpacked' format. Used to determine the target shape,
+                device, and dtype for the inpaint mask.
+
+        Returns:
+            torch.Tensor | None: Inpaint mask.
+        """
+        if self.denoise_mask is None:
+            return None
+
+        mask = context.tensors.load(self.denoise_mask.mask_name)
+
+        _, _, latent_height, latent_width = latents.shape
+        mask = tv_resize(
+            img=mask,
+            size=[latent_height, latent_width],
+            interpolation=tv_transforms.InterpolationMode.BILINEAR,
+            antialias=False,
+        )
+
+        mask = mask.to(device=latents.device, dtype=latents.dtype)
+
+        # Expand the inpaint mask to the same shape as `latents` so that when we 'pack' `mask` it lines up with
+        # `latents`.
+        return mask.expand_as(latents)
+
+    def _build_step_callback(self, context: InvocationContext) -> Callable[[PipelineIntermediateState], None]:
+        def step_callback(state: PipelineIntermediateState) -> None:
+            state.latents = unpack(state.latents.float(), self.height, self.width).squeeze()
+            context.util.flux_step_callback(state)
+
+        return step_callback

invokeai/app/invocations/flux_text_to_image.py

@@ -1,169 +0,0 @@
-import torch
-from einops import rearrange
-from PIL import Image
-
-from invokeai.app.invocations.baseinvocation import BaseInvocation, Classification, invocation
-from invokeai.app.invocations.fields import (
-    FieldDescriptions,
-    FluxConditioningField,
-    Input,
-    InputField,
-    WithBoard,
-    WithMetadata,
-)
-from invokeai.app.invocations.model import TransformerField, VAEField
-from invokeai.app.invocations.primitives import ImageOutput
-from invokeai.app.services.session_processor.session_processor_common import CanceledException
-from invokeai.app.services.shared.invocation_context import InvocationContext
-from invokeai.backend.flux.model import Flux
-from invokeai.backend.flux.modules.autoencoder import AutoEncoder
-from invokeai.backend.flux.sampling import denoise, get_noise, get_schedule, prepare_latent_img_patches, unpack
-from invokeai.backend.stable_diffusion.diffusion.conditioning_data import FLUXConditioningInfo
-from invokeai.backend.util.devices import TorchDevice
-
-
-@invocation(
-    "flux_text_to_image",
-    title="FLUX Text to Image",
-    tags=["image", "flux"],
-    category="image",
-    version="1.0.0",
-    classification=Classification.Prototype,
-)
-class FluxTextToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
-    """Text-to-image generation using a FLUX model."""
-
-    transformer: TransformerField = InputField(
-        description=FieldDescriptions.flux_model,
-        input=Input.Connection,
-        title="Transformer",
-    )
-    vae: VAEField = InputField(
-        description=FieldDescriptions.vae,
-        input=Input.Connection,
-    )
-    positive_text_conditioning: FluxConditioningField = InputField(
-        description=FieldDescriptions.positive_cond, input=Input.Connection
-    )
-    width: int = InputField(default=1024, multiple_of=16, description="Width of the generated image.")
-    height: int = InputField(default=1024, multiple_of=16, description="Height of the generated image.")
-    num_steps: int = InputField(
-        default=4, description="Number of diffusion steps. Recommend values are schnell: 4, dev: 50."
-    )
-    guidance: float = InputField(
-        default=4.0,
-        description="The guidance strength. Higher values adhere more strictly to the prompt, and will produce less diverse images. FLUX dev only, ignored for schnell.",
-    )
-    seed: int = InputField(default=0, description="Randomness seed for reproducibility.")
-
-    @torch.no_grad()
-    def invoke(self, context: InvocationContext) -> ImageOutput:
-        latents = self._run_diffusion(context)
-        image = self._run_vae_decoding(context, latents)
-        image_dto = context.images.save(image=image)
-        return ImageOutput.build(image_dto)
-
-    def _run_diffusion(
-        self,
-        context: InvocationContext,
-    ):
-        inference_dtype = torch.bfloat16
-
-        # Load the conditioning data.
-        cond_data = context.conditioning.load(self.positive_text_conditioning.conditioning_name)
-        assert len(cond_data.conditionings) == 1
-        flux_conditioning = cond_data.conditionings[0]
-        assert isinstance(flux_conditioning, FLUXConditioningInfo)
-        flux_conditioning = flux_conditioning.to(dtype=inference_dtype)
-        t5_embeddings = flux_conditioning.t5_embeds
-        clip_embeddings = flux_conditioning.clip_embeds
-
-        transformer_info = context.models.load(self.transformer.transformer)
-
-        # Prepare input noise.
-        x = get_noise(
-            num_samples=1,
-            height=self.height,
-            width=self.width,
-            device=TorchDevice.choose_torch_device(),
-            dtype=inference_dtype,
-            seed=self.seed,
-        )
-
-        x, img_ids = prepare_latent_img_patches(x)
-
-        is_schnell = "schnell" in transformer_info.config.config_path
-
-        timesteps = get_schedule(
-            num_steps=self.num_steps,
-            image_seq_len=x.shape[1],
-            shift=not is_schnell,
-        )
-
-        bs, t5_seq_len, _ = t5_embeddings.shape
-        txt_ids = torch.zeros(bs, t5_seq_len, 3, dtype=inference_dtype, device=TorchDevice.choose_torch_device())
-
-        with transformer_info as transformer:
-            assert isinstance(transformer, Flux)
-
-            def step_callback() -> None:
-                if context.util.is_canceled():
-                    raise CanceledException
-
-                # TODO: Make this look like the image before re-enabling
-                # latent_image = unpack(img.float(), self.height, self.width)
-                # latent_image = latent_image.squeeze()  # Remove unnecessary dimensions
-                # flattened_tensor = latent_image.reshape(-1)  # Flatten to shape [48*128*128]
-
-                # # Create a new tensor of the required shape [255, 255, 3]
-                # latent_image = flattened_tensor[: 255 * 255 * 3].reshape(255, 255, 3)  # Reshape to RGB format
-
-                # # Convert to a NumPy array and then to a PIL Image
-                # image = Image.fromarray(latent_image.cpu().numpy().astype(np.uint8))
-
-                # (width, height) = image.size
-                # width *= 8
-                # height *= 8
-
-                # dataURL = image_to_dataURL(image, image_format="JPEG")
-
-                # # TODO: move this whole function to invocation context to properly reference these variables
-                # context._services.events.emit_invocation_denoise_progress(
-                #     context._data.queue_item,
-                #     context._data.invocation,
-                #     state,
-                #     ProgressImage(dataURL=dataURL, width=width, height=height),
-                # )
-
-            x = denoise(
-                model=transformer,
-                img=x,
-                img_ids=img_ids,
-                txt=t5_embeddings,
-                txt_ids=txt_ids,
-                vec=clip_embeddings,
-                timesteps=timesteps,
-                step_callback=step_callback,
-                guidance=self.guidance,
-            )
-
-        x = unpack(x.float(), self.height, self.width)
-
-        return x
-
-    def _run_vae_decoding(
-        self,
-        context: InvocationContext,
-        latents: torch.Tensor,
-    ) -> Image.Image:
-        vae_info = context.models.load(self.vae.vae)
-        with vae_info as vae:
-            assert isinstance(vae, AutoEncoder)
-            latents = latents.to(dtype=TorchDevice.choose_torch_dtype())
-            img = vae.decode(latents)
-
-        img = img.clamp(-1, 1)
-        img = rearrange(img[0], "c h w -> h w c")
-        img_pil = Image.fromarray((127.5 * (img + 1.0)).byte().cpu().numpy())
-
-        return img_pil

invokeai/app/invocations/flux_vae_decode.py

@@ -0,0 +1,60 @@
+import torch
+from einops import rearrange
+from PIL import Image
+
+from invokeai.app.invocations.baseinvocation import BaseInvocation, invocation
+from invokeai.app.invocations.fields import (
+    FieldDescriptions,
+    Input,
+    InputField,
+    LatentsField,
+    WithBoard,
+    WithMetadata,
+)
+from invokeai.app.invocations.model import VAEField
+from invokeai.app.invocations.primitives import ImageOutput
+from invokeai.app.services.shared.invocation_context import InvocationContext
+from invokeai.backend.flux.modules.autoencoder import AutoEncoder
+from invokeai.backend.model_manager.load.load_base import LoadedModel
+from invokeai.backend.util.devices import TorchDevice
+
+
+@invocation(
+    "flux_vae_decode",
+    title="FLUX Latents to Image",
+    tags=["latents", "image", "vae", "l2i", "flux"],
+    category="latents",
+    version="1.0.0",
+)
+class FluxVaeDecodeInvocation(BaseInvocation, WithMetadata, WithBoard):
+    """Generates an image from latents."""
+
+    latents: LatentsField = InputField(
+        description=FieldDescriptions.latents,
+        input=Input.Connection,
+    )
+    vae: VAEField = InputField(
+        description=FieldDescriptions.vae,
+        input=Input.Connection,
+    )
+
+    def _vae_decode(self, vae_info: LoadedModel, latents: torch.Tensor) -> Image.Image:
+        with vae_info as vae:
+            assert isinstance(vae, AutoEncoder)
+            latents = latents.to(device=TorchDevice.choose_torch_device(), dtype=TorchDevice.choose_torch_dtype())
+            img = vae.decode(latents)
+
+        img = img.clamp(-1, 1)
+        img = rearrange(img[0], "c h w -> h w c")  # noqa: F821
+        img_pil = Image.fromarray((127.5 * (img + 1.0)).byte().cpu().numpy())
+        return img_pil
+
+    @torch.no_grad()
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        latents = context.tensors.load(self.latents.latents_name)
+        vae_info = context.models.load(self.vae.vae)
+        image = self._vae_decode(vae_info=vae_info, latents=latents)
+
+        TorchDevice.empty_cache()
+        image_dto = context.images.save(image=image)
+        return ImageOutput.build(image_dto)

invokeai/app/invocations/flux_vae_encode.py

@@ -0,0 +1,67 @@
+import einops
+import torch
+
+from invokeai.app.invocations.baseinvocation import BaseInvocation, invocation
+from invokeai.app.invocations.fields import (
+    FieldDescriptions,
+    ImageField,
+    Input,
+    InputField,
+)
+from invokeai.app.invocations.model import VAEField
+from invokeai.app.invocations.primitives import LatentsOutput
+from invokeai.app.services.shared.invocation_context import InvocationContext
+from invokeai.backend.flux.modules.autoencoder import AutoEncoder
+from invokeai.backend.model_manager import LoadedModel
+from invokeai.backend.stable_diffusion.diffusers_pipeline import image_resized_to_grid_as_tensor
+from invokeai.backend.util.devices import TorchDevice
+
+
+@invocation(
+    "flux_vae_encode",
+    title="FLUX Image to Latents",
+    tags=["latents", "image", "vae", "i2l", "flux"],
+    category="latents",
+    version="1.0.0",
+)
+class FluxVaeEncodeInvocation(BaseInvocation):
+    """Encodes an image into latents."""
+
+    image: ImageField = InputField(
+        description="The image to encode.",
+    )
+    vae: VAEField = InputField(
+        description=FieldDescriptions.vae,
+        input=Input.Connection,
+    )
+
+    @staticmethod
+    def vae_encode(vae_info: LoadedModel, image_tensor: torch.Tensor) -> torch.Tensor:
+        # TODO(ryand): Expose seed parameter at the invocation level.
+        # TODO(ryand): Write a util function for generating random tensors that is consistent across devices / dtypes.
+        # There's a starting point in get_noise(...), but it needs to be extracted and generalized. This function
+        # should be used for VAE encode sampling.
+        generator = torch.Generator(device=TorchDevice.choose_torch_device()).manual_seed(0)
+        with vae_info as vae:
+            assert isinstance(vae, AutoEncoder)
+            image_tensor = image_tensor.to(
+                device=TorchDevice.choose_torch_device(), dtype=TorchDevice.choose_torch_dtype()
+            )
+            latents = vae.encode(image_tensor, sample=True, generator=generator)
+            return latents
+
+    @torch.no_grad()
+    def invoke(self, context: InvocationContext) -> LatentsOutput:
+        image = context.images.get_pil(self.image.image_name)
+
+        vae_info = context.models.load(self.vae.vae)
+
+        image_tensor = image_resized_to_grid_as_tensor(image.convert("RGB"))
+        if image_tensor.dim() == 3:
+            image_tensor = einops.rearrange(image_tensor, "c h w -> 1 c h w")
+
+        latents = self.vae_encode(vae_info=vae_info, image_tensor=image_tensor)
+
+        latents = latents.to("cpu")
+        name = context.tensors.save(tensor=latents)
+        return LatentsOutput.build(latents_name=name, latents=latents, seed=None)

invokeai/app/invocations/mask.py

@@ -126,7 +126,7 @@ class ImageMaskToTensorInvocation(BaseInvocation, WithMetadata):
     title="Tensor Mask to Image",
     tags=["mask"],
     category="mask",
-    version="1.0.0",
+    version="1.1.0",
 )
 class MaskTensorToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
     """Convert a mask tensor to an image."""
@@ -135,6 +135,11 @@ class MaskTensorToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
 
     def invoke(self, context: InvocationContext) -> ImageOutput:
         mask = context.tensors.load(self.mask.tensor_name)
+
+        # Squeeze the channel dimension if it exists.
+        if mask.dim() == 3:
+            mask = mask.squeeze(0)
+
         # Ensure that the mask is binary.
         if mask.dtype != torch.bool:
             mask = mask > 0.5

invokeai/app/services/model_install/model_install_common.py

@@ -103,7 +103,7 @@ class HFModelSource(StringLikeSource):
         if self.variant:
             base += f":{self.variant or ''}"
         if self.subfolder:
-            base += f":{self.subfolder}"
+            base += f"::{self.subfolder.as_posix()}"
         return base
 
 

invokeai/app/services/shared/invocation_context.py

@@ -14,7 +14,7 @@ from invokeai.app.services.image_records.image_records_common import ImageCatego
 from invokeai.app.services.images.images_common import ImageDTO
 from invokeai.app.services.invocation_services import InvocationServices
 from invokeai.app.services.model_records.model_records_base import UnknownModelException
-from invokeai.app.util.step_callback import stable_diffusion_step_callback
+from invokeai.app.util.step_callback import flux_step_callback, stable_diffusion_step_callback
 from invokeai.backend.model_manager.config import (
     AnyModel,
     AnyModelConfig,
@@ -557,6 +557,24 @@ class UtilInterface(InvocationContextInterface):
             is_canceled=self.is_canceled,
         )
 
+    def flux_step_callback(self, intermediate_state: PipelineIntermediateState) -> None:
+        """
+        The step callback emits a progress event with the current step, the total number of
+        steps, a preview image, and some other internal metadata.
+
+        This should be called after each denoising step.
+
+        Args:
+            intermediate_state: The intermediate state of the diffusion pipeline.
+        """
+
+        flux_step_callback(
+            context_data=self._data,
+            intermediate_state=intermediate_state,
+            events=self._services.events,
+            is_canceled=self.is_canceled,
+        )
+
 
 class InvocationContext:
     """Provides access to various services and data for the current invocation.

invokeai/app/services/workflow_records/default_workflows/FLUX Image to Image.json

@@ -0,0 +1,407 @@
+{
+  "name": "FLUX Image to Image",
+  "author": "InvokeAI",
+  "description": "A simple image-to-image workflow using a FLUX dev model. ",
+  "version": "1.0.4",
+  "contact": "",
+  "tags": "image2image, flux, image-to-image",
+  "notes": "Prerequisite model downloads: T5 Encoder, CLIP-L Encoder, and FLUX VAE. Quantized and un-quantized versions can be found in the starter models tab within your Model Manager. We recommend using FLUX dev models for image-to-image workflows. The image-to-image performance with FLUX schnell models is poor.",
+  "exposedFields": [
+    {
+      "nodeId": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "fieldName": "model"
+    },
+    {
+      "nodeId": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "fieldName": "t5_encoder_model"
+    },
+    {
+      "nodeId": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "fieldName": "clip_embed_model"
+    },
+    {
+      "nodeId": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "fieldName": "vae_model"
+    },
+    {
+      "nodeId": "ace0258f-67d7-4eee-a218-6fff27065214",
+      "fieldName": "denoising_start"
+    },
+    {
+      "nodeId": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
+      "fieldName": "prompt"
+    },
+    {
+      "nodeId": "ace0258f-67d7-4eee-a218-6fff27065214",
+      "fieldName": "num_steps"
+    }
+  ],
+  "meta": {
+    "version": "3.0.0",
+    "category": "default"
+  },
+  "nodes": [
+    {
+      "id": "2981a67c-480f-4237-9384-26b68dbf912b",
+      "type": "invocation",
+      "data": {
+        "id": "2981a67c-480f-4237-9384-26b68dbf912b",
+        "type": "flux_vae_encode",
+        "version": "1.0.0",
+        "label": "",
+        "notes": "",
+        "isOpen": true,
+        "isIntermediate": true,
+        "useCache": true,
+        "inputs": {
+          "image": {
+            "name": "image",
+            "label": "",
+            "value": {
+              "image_name": "8a5c62aa-9335-45d2-9c71-89af9fc1f8d4.png"
+            }
+          },
+          "vae": {
+            "name": "vae",
+            "label": ""
+          }
+        }
+      },
+      "position": {
+        "x": 732.7680166609682,
+        "y": -24.37398171806909
+      }
+    },
+    {
+      "id": "ace0258f-67d7-4eee-a218-6fff27065214",
+      "type": "invocation",
+      "data": {
+        "id": "ace0258f-67d7-4eee-a218-6fff27065214",
+        "type": "flux_denoise",
+        "version": "1.0.0",
+        "label": "",
+        "notes": "",
+        "isOpen": true,
+        "isIntermediate": true,
+        "useCache": true,
+        "inputs": {
+          "board": {
+            "name": "board",
+            "label": ""
+          },
+          "metadata": {
+            "name": "metadata",
+            "label": ""
+          },
+          "latents": {
+            "name": "latents",
+            "label": ""
+          },
+          "denoise_mask": {
+            "name": "denoise_mask",
+            "label": ""
+          },
+          "denoising_start": {
+            "name": "denoising_start",
+            "label": "",
+            "value": 0.04
+          },
+          "denoising_end": {
+            "name": "denoising_end",
+            "label": "",
+            "value": 1
+          },
+          "transformer": {
+            "name": "transformer",
+            "label": ""
+          },
+          "positive_text_conditioning": {
+            "name": "positive_text_conditioning",
+            "label": ""
+          },
+          "width": {
+            "name": "width",
+            "label": "",
+            "value": 1024
+          },
+          "height": {
+            "name": "height",
+            "label": "",
+            "value": 1024
+          },
+          "num_steps": {
+            "name": "num_steps",
+            "label": "Steps (Recommend 30 for Dev, 4 for Schnell)",
+            "value": 30
+          },
+          "guidance": {
+            "name": "guidance",
+            "label": "",
+            "value": 4
+          },
+          "seed": {
+            "name": "seed",
+            "label": "",
+            "value": 0
+          }
+        }
+      },
+      "position": {
+        "x": 1182.8836633018684,
+        "y": -251.38882958913183
+      }
+    },
+    {
+      "id": "7e5172eb-48c1-44db-a770-8fd83e1435d1",
+      "type": "invocation",
+      "data": {
+        "id": "7e5172eb-48c1-44db-a770-8fd83e1435d1",
+        "type": "flux_vae_decode",
+        "version": "1.0.0",
+        "label": "",
+        "notes": "",
+        "isOpen": true,
+        "isIntermediate": false,
+        "useCache": true,
+        "inputs": {
+          "board": {
+            "name": "board",
+            "label": ""
+          },
+          "metadata": {
+            "name": "metadata",
+            "label": ""
+          },
+          "latents": {
+            "name": "latents",
+            "label": ""
+          },
+          "vae": {
+            "name": "vae",
+            "label": ""
+          }
+        }
+      },
+      "position": {
+        "x": 1575.5797431839133,
+        "y": -209.00150975507415
+      }
+    },
+    {
+      "id": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "type": "invocation",
+      "data": {
+        "id": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+        "type": "flux_model_loader",
+        "version": "1.0.4",
+        "label": "",
+        "notes": "",
+        "isOpen": true,
+        "isIntermediate": true,
+        "useCache": false,
+        "inputs": {
+          "model": {
+            "name": "model",
+            "label": "Model (dev variant recommended for Image-to-Image)"
+          },
+          "t5_encoder_model": {
+            "name": "t5_encoder_model",
+            "label": ""
+          },
+          "clip_embed_model": {
+            "name": "clip_embed_model",
+            "label": "",
+            "value": {
+              "key": "fa23a584-b623-415d-832a-21b5098ff1a1",
+              "hash": "blake3:17c19f0ef941c3b7609a9c94a659ca5364de0be364a91d4179f0e39ba17c3b70",
+              "name": "clip-vit-large-patch14",
+              "base": "any",
+              "type": "clip_embed"
+            }
+          },
+          "vae_model": {
+            "name": "vae_model",
+            "label": "",
+            "value": {
+              "key": "74fc82ba-c0a8-479d-a890-2126f82da758",
+              "hash": "blake3:ce21cb76364aa6e2421311cf4a4b5eb052a76c4f1cd207b50703d8978198a068",
+              "name": "FLUX.1-schnell_ae",
+              "base": "flux",
+              "type": "vae"
+            }
+          }
+        }
+      },
+      "position": {
+        "x": 328.1809894659957,
+        "y": -90.2241133566946
+      }
+    },
+    {
+      "id": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
+      "type": "invocation",
+      "data": {
+        "id": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
+        "type": "flux_text_encoder",
+        "version": "1.0.0",
+        "label": "",
+        "notes": "",
+        "isOpen": true,
+        "isIntermediate": true,
+        "useCache": true,
+        "inputs": {
+          "clip": {
+            "name": "clip",
+            "label": ""
+          },
+          "t5_encoder": {
+            "name": "t5_encoder",
+            "label": ""
+          },
+          "t5_max_seq_len": {
+            "name": "t5_max_seq_len",
+            "label": "T5 Max Seq Len",
+            "value": 256
+          },
+          "prompt": {
+            "name": "prompt",
+            "label": "",
+            "value": "a cat wearing a birthday hat"
+          }
+        }
+      },
+      "position": {
+        "x": 745.8823365057267,
+        "y": -299.60249175851914
+      }
+    },
+    {
+      "id": "4754c534-a5f3-4ad0-9382-7887985e668c",
+      "type": "invocation",
+      "data": {
+        "id": "4754c534-a5f3-4ad0-9382-7887985e668c",
+        "type": "rand_int",
+        "version": "1.0.1",
+        "label": "",
+        "notes": "",
+        "isOpen": true,
+        "isIntermediate": true,
+        "useCache": false,
+        "inputs": {
+          "low": {
+            "name": "low",
+            "label": "",
+            "value": 0
+          },
+          "high": {
+            "name": "high",
+            "label": "",
+            "value": 2147483647
+          }
+        }
+      },
+      "position": {
+        "x": 725.834098928012,
+        "y": 496.2710031089931
+      }
+    }
+  ],
+  "edges": [
+    {
+      "id": "reactflow__edge-2981a67c-480f-4237-9384-26b68dbf912bheight-ace0258f-67d7-4eee-a218-6fff27065214height",
+      "type": "default",
+      "source": "2981a67c-480f-4237-9384-26b68dbf912b",
+      "target": "ace0258f-67d7-4eee-a218-6fff27065214",
+      "sourceHandle": "height",
+      "targetHandle": "height"
+    },
+    {
+      "id": "reactflow__edge-2981a67c-480f-4237-9384-26b68dbf912bwidth-ace0258f-67d7-4eee-a218-6fff27065214width",
+      "type": "default",
+      "source": "2981a67c-480f-4237-9384-26b68dbf912b",
+      "target": "ace0258f-67d7-4eee-a218-6fff27065214",
+      "sourceHandle": "width",
+      "targetHandle": "width"
+    },
+    {
+      "id": "reactflow__edge-2981a67c-480f-4237-9384-26b68dbf912blatents-ace0258f-67d7-4eee-a218-6fff27065214latents",
+      "type": "default",
+      "source": "2981a67c-480f-4237-9384-26b68dbf912b",
+      "target": "ace0258f-67d7-4eee-a218-6fff27065214",
+      "sourceHandle": "latents",
+      "targetHandle": "latents"
+    },
+    {
+      "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90vae-2981a67c-480f-4237-9384-26b68dbf912bvae",
+      "type": "default",
+      "source": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "target": "2981a67c-480f-4237-9384-26b68dbf912b",
+      "sourceHandle": "vae",
+      "targetHandle": "vae"
+    },
+    {
+      "id": "reactflow__edge-ace0258f-67d7-4eee-a218-6fff27065214latents-7e5172eb-48c1-44db-a770-8fd83e1435d1latents",
+      "type": "default",
+      "source": "ace0258f-67d7-4eee-a218-6fff27065214",
+      "target": "7e5172eb-48c1-44db-a770-8fd83e1435d1",
+      "sourceHandle": "latents",
+      "targetHandle": "latents"
+    },
+    {
+      "id": "reactflow__edge-4754c534-a5f3-4ad0-9382-7887985e668cvalue-ace0258f-67d7-4eee-a218-6fff27065214seed",
+      "type": "default",
+      "source": "4754c534-a5f3-4ad0-9382-7887985e668c",
+      "target": "ace0258f-67d7-4eee-a218-6fff27065214",
+      "sourceHandle": "value",
+      "targetHandle": "seed"
+    },
+    {
+      "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90transformer-ace0258f-67d7-4eee-a218-6fff27065214transformer",
+      "type": "default",
+      "source": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "target": "ace0258f-67d7-4eee-a218-6fff27065214",
+      "sourceHandle": "transformer",
+      "targetHandle": "transformer"
+    },
+    {
+      "id": "reactflow__edge-01f674f8-b3d1-4df1-acac-6cb8e0bfb63cconditioning-ace0258f-67d7-4eee-a218-6fff27065214positive_text_conditioning",
+      "type": "default",
+      "source": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
+      "target": "ace0258f-67d7-4eee-a218-6fff27065214",
+      "sourceHandle": "conditioning",
+      "targetHandle": "positive_text_conditioning"
+    },
+    {
+      "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90vae-7e5172eb-48c1-44db-a770-8fd83e1435d1vae",
+      "type": "default",
+      "source": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "target": "7e5172eb-48c1-44db-a770-8fd83e1435d1",
+      "sourceHandle": "vae",
+      "targetHandle": "vae"
+    },
+    {
+      "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90max_seq_len-01f674f8-b3d1-4df1-acac-6cb8e0bfb63ct5_max_seq_len",
+      "type": "default",
+      "source": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "target": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
+      "sourceHandle": "max_seq_len",
+      "targetHandle": "t5_max_seq_len"
+    },
+    {
+      "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90t5_encoder-01f674f8-b3d1-4df1-acac-6cb8e0bfb63ct5_encoder",
+      "type": "default",
+      "source": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "target": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
+      "sourceHandle": "t5_encoder",
+      "targetHandle": "t5_encoder"
+    },
+    {
+      "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90clip-01f674f8-b3d1-4df1-acac-6cb8e0bfb63cclip",
+      "type": "default",
+      "source": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "target": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
+      "sourceHandle": "clip",
+      "targetHandle": "clip"
+    }
+  ]
+}

invokeai/app/services/workflow_records/default_workflows/Flux Text to Image.json

@@ -1,7 +1,7 @@
 {
   "name": "FLUX Text to Image",
   "author": "InvokeAI",
-  "description": "A simple text-to-image workflow using FLUX dev or schnell models. Prerequisite model downloads: T5 Encoder, CLIP-L Encoder, and FLUX VAE. Quantized and un-quantized versions can be found in the starter models tab within your Model Manager. We recommend 4 steps for FLUX schnell models and 30 steps for FLUX dev models.",
+  "description": "A simple text-to-image workflow using FLUX dev or schnell models.",
   "version": "1.0.4",
   "contact": "",
   "tags": "text2image, flux",
@@ -11,17 +11,25 @@
       "nodeId": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
       "fieldName": "model"
     },
+    {
+      "nodeId": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "fieldName": "t5_encoder_model"
+    },
+    {
+      "nodeId": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "fieldName": "clip_embed_model"
+    },
+    {
+      "nodeId": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "fieldName": "vae_model"
+    },
     {
       "nodeId": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
       "fieldName": "prompt"
     },
     {
-      "nodeId": "159bdf1b-79e7-4174-b86e-d40e646964c8",
+      "nodeId": "4fe24f07-f906-4f55-ab2c-9beee56ef5bd",
       "fieldName": "num_steps"
-    },
-    {
-      "nodeId": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
-      "fieldName": "t5_encoder_model"
     }
   ],
   "meta": {
@@ -29,6 +37,121 @@
     "category": "default"
   },
   "nodes": [
+    {
+      "id": "4fe24f07-f906-4f55-ab2c-9beee56ef5bd",
+      "type": "invocation",
+      "data": {
+        "id": "4fe24f07-f906-4f55-ab2c-9beee56ef5bd",
+        "type": "flux_denoise",
+        "version": "1.0.0",
+        "label": "",
+        "notes": "",
+        "isOpen": true,
+        "isIntermediate": true,
+        "useCache": true,
+        "inputs": {
+          "board": {
+            "name": "board",
+            "label": ""
+          },
+          "metadata": {
+            "name": "metadata",
+            "label": ""
+          },
+          "latents": {
+            "name": "latents",
+            "label": ""
+          },
+          "denoise_mask": {
+            "name": "denoise_mask",
+            "label": ""
+          },
+          "denoising_start": {
+            "name": "denoising_start",
+            "label": "",
+            "value": 0
+          },
+          "denoising_end": {
+            "name": "denoising_end",
+            "label": "",
+            "value": 1
+          },
+          "transformer": {
+            "name": "transformer",
+            "label": ""
+          },
+          "positive_text_conditioning": {
+            "name": "positive_text_conditioning",
+            "label": ""
+          },
+          "width": {
+            "name": "width",
+            "label": "",
+            "value": 1024
+          },
+          "height": {
+            "name": "height",
+            "label": "",
+            "value": 1024
+          },
+          "num_steps": {
+            "name": "num_steps",
+            "label": "Steps (Recommend 30 for Dev, 4 for Schnell)",
+            "value": 30
+          },
+          "guidance": {
+            "name": "guidance",
+            "label": "",
+            "value": 4
+          },
+          "seed": {
+            "name": "seed",
+            "label": "",
+            "value": 0
+          }
+        }
+      },
+      "position": {
+        "x": 1186.1868226120378,
+        "y": -214.9459927686657
+      }
+    },
+    {
+      "id": "7e5172eb-48c1-44db-a770-8fd83e1435d1",
+      "type": "invocation",
+      "data": {
+        "id": "7e5172eb-48c1-44db-a770-8fd83e1435d1",
+        "type": "flux_vae_decode",
+        "version": "1.0.0",
+        "label": "",
+        "notes": "",
+        "isOpen": true,
+        "isIntermediate": false,
+        "useCache": true,
+        "inputs": {
+          "board": {
+            "name": "board",
+            "label": ""
+          },
+          "metadata": {
+            "name": "metadata",
+            "label": ""
+          },
+          "latents": {
+            "name": "latents",
+            "label": ""
+          },
+          "vae": {
+            "name": "vae",
+            "label": ""
+          }
+        }
+      },
+      "position": {
+        "x": 1575.5797431839133,
+        "y": -209.00150975507415
+      }
+    },
     {
       "id": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
       "type": "invocation",
@@ -99,8 +222,8 @@
         }
       },
       "position": {
-        "x": 824.1970602278849,
-        "y": 146.98251001061735
+        "x": 778.4899149328337,
+        "y": -100.36469216659502
       }
     },
     {
@@ -129,77 +252,52 @@
         }
       },
       "position": {
-        "x": 822.9899179655476,
-        "y": 360.9657214885052
-      }
-    },
-    {
-      "id": "159bdf1b-79e7-4174-b86e-d40e646964c8",
-      "type": "invocation",
-      "data": {
-        "id": "159bdf1b-79e7-4174-b86e-d40e646964c8",
-        "type": "flux_text_to_image",
-        "version": "1.0.0",
-        "label": "",
-        "notes": "",
-        "isOpen": true,
-        "isIntermediate": false,
-        "useCache": true,
-        "inputs": {
-          "board": {
-            "name": "board",
-            "label": ""
-          },
-          "metadata": {
-            "name": "metadata",
-            "label": ""
-          },
-          "transformer": {
-            "name": "transformer",
-            "label": ""
-          },
-          "vae": {
-            "name": "vae",
-            "label": ""
-          },
-          "positive_text_conditioning": {
-            "name": "positive_text_conditioning",
-            "label": ""
-          },
-          "width": {
-            "name": "width",
-            "label": "",
-            "value": 1024
-          },
-          "height": {
-            "name": "height",
-            "label": "",
-            "value": 1024
-          },
-          "num_steps": {
-            "name": "num_steps",
-            "label": "Steps (Recommend 30 for Dev, 4 for Schnell)",
-            "value": 30
-          },
-          "guidance": {
-            "name": "guidance",
-            "label": "",
-            "value": 4
-          },
-          "seed": {
-            "name": "seed",
-            "label": "",
-            "value": 0
-          }
-        }
-      },
-      "position": {
-        "x": 1216.3900791301849,
-        "y": 5.500841807102248
+        "x": 800.9667463219505,
+        "y": 285.8297267547506
       }
     }
   ],
   "edges": [
+    {
+      "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90transformer-4fe24f07-f906-4f55-ab2c-9beee56ef5bdtransformer",
+      "type": "default",
+      "source": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "target": "4fe24f07-f906-4f55-ab2c-9beee56ef5bd",
+      "sourceHandle": "transformer",
+      "targetHandle": "transformer"
+    },
+    {
+      "id": "reactflow__edge-01f674f8-b3d1-4df1-acac-6cb8e0bfb63cconditioning-4fe24f07-f906-4f55-ab2c-9beee56ef5bdpositive_text_conditioning",
+      "type": "default",
+      "source": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
+      "target": "4fe24f07-f906-4f55-ab2c-9beee56ef5bd",
+      "sourceHandle": "conditioning",
+      "targetHandle": "positive_text_conditioning"
+    },
+    {
+      "id": "reactflow__edge-4754c534-a5f3-4ad0-9382-7887985e668cvalue-4fe24f07-f906-4f55-ab2c-9beee56ef5bdseed",
+      "type": "default",
+      "source": "4754c534-a5f3-4ad0-9382-7887985e668c",
+      "target": "4fe24f07-f906-4f55-ab2c-9beee56ef5bd",
+      "sourceHandle": "value",
+      "targetHandle": "seed"
+    },
+    {
+      "id": "reactflow__edge-4fe24f07-f906-4f55-ab2c-9beee56ef5bdlatents-7e5172eb-48c1-44db-a770-8fd83e1435d1latents",
+      "type": "default",
+      "source": "4fe24f07-f906-4f55-ab2c-9beee56ef5bd",
+      "target": "7e5172eb-48c1-44db-a770-8fd83e1435d1",
+      "sourceHandle": "latents",
+      "targetHandle": "latents"
+    },
+    {
+      "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90vae-7e5172eb-48c1-44db-a770-8fd83e1435d1vae",
+      "type": "default",
+      "source": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
+      "target": "7e5172eb-48c1-44db-a770-8fd83e1435d1",
+      "sourceHandle": "vae",
+      "targetHandle": "vae"
+    },
     {
       "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90max_seq_len-01f674f8-b3d1-4df1-acac-6cb8e0bfb63ct5_max_seq_len",
       "type": "default",
@@ -208,14 +306,6 @@
       "sourceHandle": "max_seq_len",
       "targetHandle": "t5_max_seq_len"
     },
-    {
-      "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90vae-159bdf1b-79e7-4174-b86e-d40e646964c8vae",
-      "type": "default",
-      "source": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
-      "target": "159bdf1b-79e7-4174-b86e-d40e646964c8",
-      "sourceHandle": "vae",
-      "targetHandle": "vae"
-    },
     {
       "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90t5_encoder-01f674f8-b3d1-4df1-acac-6cb8e0bfb63ct5_encoder",
       "type": "default",
@@ -231,30 +321,6 @@
       "target": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
       "sourceHandle": "clip",
       "targetHandle": "clip"
-    },
-    {
-      "id": "reactflow__edge-f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90transformer-159bdf1b-79e7-4174-b86e-d40e646964c8transformer",
-      "type": "default",
-      "source": "f8d9d7c8-9ed7-4bd7-9e42-ab0e89bfac90",
-      "target": "159bdf1b-79e7-4174-b86e-d40e646964c8",
-      "sourceHandle": "transformer",
-      "targetHandle": "transformer"
-    },
-    {
-      "id": "reactflow__edge-01f674f8-b3d1-4df1-acac-6cb8e0bfb63cconditioning-159bdf1b-79e7-4174-b86e-d40e646964c8positive_text_conditioning",
-      "type": "default",
-      "source": "01f674f8-b3d1-4df1-acac-6cb8e0bfb63c",
-      "target": "159bdf1b-79e7-4174-b86e-d40e646964c8",
-      "sourceHandle": "conditioning",
-      "targetHandle": "positive_text_conditioning"
-    },
-    {
-      "id": "reactflow__edge-4754c534-a5f3-4ad0-9382-7887985e668cvalue-159bdf1b-79e7-4174-b86e-d40e646964c8seed",
-      "type": "default",
-      "source": "4754c534-a5f3-4ad0-9382-7887985e668c",
-      "target": "159bdf1b-79e7-4174-b86e-d40e646964c8",
-      "sourceHandle": "value",
-      "targetHandle": "seed"
     }
   ]
 }

invokeai/app/util/step_callback.py

@@ -38,6 +38,25 @@ SD1_5_LATENT_RGB_FACTORS = [
     [-0.1307, -0.1874, -0.7445],  # L4
 ]
 
+FLUX_LATENT_RGB_FACTORS = [
+    [-0.0412, 0.0149, 0.0521],
+    [0.0056, 0.0291, 0.0768],
+    [0.0342, -0.0681, -0.0427],
+    [-0.0258, 0.0092, 0.0463],
+    [0.0863, 0.0784, 0.0547],
+    [-0.0017, 0.0402, 0.0158],
+    [0.0501, 0.1058, 0.1152],
+    [-0.0209, -0.0218, -0.0329],
+    [-0.0314, 0.0083, 0.0896],
+    [0.0851, 0.0665, -0.0472],
+    [-0.0534, 0.0238, -0.0024],
+    [0.0452, -0.0026, 0.0048],
+    [0.0892, 0.0831, 0.0881],
+    [-0.1117, -0.0304, -0.0789],
+    [0.0027, -0.0479, -0.0043],
+    [-0.1146, -0.0827, -0.0598],
+]
+
 
 def sample_to_lowres_estimated_image(
     samples: torch.Tensor, latent_rgb_factors: torch.Tensor, smooth_matrix: Optional[torch.Tensor] = None
@@ -94,3 +113,32 @@ def stable_diffusion_step_callback(
         intermediate_state,
         ProgressImage(dataURL=dataURL, width=width, height=height),
     )
+
+
+def flux_step_callback(
+    context_data: "InvocationContextData",
+    intermediate_state: PipelineIntermediateState,
+    events: "EventServiceBase",
+    is_canceled: Callable[[], bool],
+) -> None:
+    if is_canceled():
+        raise CanceledException
+    sample = intermediate_state.latents
+    latent_rgb_factors = torch.tensor(FLUX_LATENT_RGB_FACTORS, dtype=sample.dtype, device=sample.device)
+    latent_image_perm = sample.permute(1, 2, 0).to(dtype=sample.dtype, device=sample.device)
+    latent_image = latent_image_perm @ latent_rgb_factors
+    latents_ubyte = (
+        ((latent_image + 1) / 2).clamp(0, 1).mul(0xFF)  # change scale from -1..1 to 0..1  # to 0..255
+    ).to(device="cpu", dtype=torch.uint8)
+    image = Image.fromarray(latents_ubyte.cpu().numpy())
+    (width, height) = image.size
+    width *= 8
+    height *= 8
+    dataURL = image_to_dataURL(image, image_format="JPEG")
+
+    events.emit_invocation_denoise_progress(
+        context_data.queue_item,
+        context_data.invocation,
+        intermediate_state,
+        ProgressImage(dataURL=dataURL, width=width, height=height),
+    )

invokeai/backend/flux/denoise.py

@@ -0,0 +1,56 @@
+from typing import Callable
+
+import torch
+from tqdm import tqdm
+
+from invokeai.backend.flux.inpaint_extension import InpaintExtension
+from invokeai.backend.flux.model import Flux
+from invokeai.backend.stable_diffusion.diffusers_pipeline import PipelineIntermediateState
+
+
+def denoise(
+    model: Flux,
+    # model input
+    img: torch.Tensor,
+    img_ids: torch.Tensor,
+    txt: torch.Tensor,
+    txt_ids: torch.Tensor,
+    vec: torch.Tensor,
+    # sampling parameters
+    timesteps: list[float],
+    step_callback: Callable[[PipelineIntermediateState], None],
+    guidance: float,
+    inpaint_extension: InpaintExtension | None,
+):
+    step = 0
+    # guidance_vec is ignored for schnell.
+    guidance_vec = torch.full((img.shape[0],), guidance, device=img.device, dtype=img.dtype)
+    for t_curr, t_prev in tqdm(list(zip(timesteps[:-1], timesteps[1:], strict=True))):
+        t_vec = torch.full((img.shape[0],), t_curr, dtype=img.dtype, device=img.device)
+        pred = model(
+            img=img,
+            img_ids=img_ids,
+            txt=txt,
+            txt_ids=txt_ids,
+            y=vec,
+            timesteps=t_vec,
+            guidance=guidance_vec,
+        )
+        preview_img = img - t_curr * pred
+        img = img + (t_prev - t_curr) * pred
+
+        if inpaint_extension is not None:
+            img = inpaint_extension.merge_intermediate_latents_with_init_latents(img, t_prev)
+
+        step_callback(
+            PipelineIntermediateState(
+                step=step,
+                order=1,
+                total_steps=len(timesteps),
+                timestep=int(t_curr),
+                latents=preview_img,
+            ),
+        )
+        step += 1
+
+    return img

invokeai/backend/flux/inpaint_extension.py

@@ -0,0 +1,35 @@
+import torch
+
+
+class InpaintExtension:
+    """A class for managing inpainting with FLUX."""
+
+    def __init__(self, init_latents: torch.Tensor, inpaint_mask: torch.Tensor, noise: torch.Tensor):
+        """Initialize InpaintExtension.
+
+        Args:
+            init_latents (torch.Tensor): The initial latents (i.e. un-noised at timestep 0). In 'packed' format.
+            inpaint_mask (torch.Tensor): A mask specifying which elements to inpaint. Range [0, 1]. Values of 1 will be
+                re-generated. Values of 0 will remain unchanged. Values between 0 and 1 can be used to blend the
+                inpainted region with the background. In 'packed' format.
+            noise (torch.Tensor): The noise tensor used to noise the init_latents. In 'packed' format.
+        """
+        assert init_latents.shape == inpaint_mask.shape == noise.shape
+        self._init_latents = init_latents
+        self._inpaint_mask = inpaint_mask
+        self._noise = noise
+
+    def merge_intermediate_latents_with_init_latents(
+        self, intermediate_latents: torch.Tensor, timestep: float
+    ) -> torch.Tensor:
+        """Merge the intermediate latents with the initial latents for the current timestep using the inpaint mask. I.e.
+        update the intermediate latents to keep the regions that are not being inpainted on the correct noise
+        trajectory.
+
+        This function should be called after each denoising step.
+        """
+        # Noise the init latents for the current timestep.
+        noised_init_latents = self._noise * timestep + (1.0 - timestep) * self._init_latents
+
+        # Merge the intermediate latents with the noised_init_latents using the inpaint_mask.
+        return intermediate_latents * self._inpaint_mask + noised_init_latents * (1.0 - self._inpaint_mask)

invokeai/backend/flux/modules/autoencoder.py

@@ -258,16 +258,17 @@ class Decoder(nn.Module):
 
 
 class DiagonalGaussian(nn.Module):
-    def __init__(self, sample: bool = True, chunk_dim: int = 1):
+    def __init__(self, chunk_dim: int = 1):
         super().__init__()
-        self.sample = sample
         self.chunk_dim = chunk_dim
 
-    def forward(self, z: Tensor) -> Tensor:
+    def forward(self, z: Tensor, sample: bool = True, generator: torch.Generator | None = None) -> Tensor:
         mean, logvar = torch.chunk(z, 2, dim=self.chunk_dim)
-        if self.sample:
+        if sample:
             std = torch.exp(0.5 * logvar)
-            return mean + std * torch.randn_like(mean)
+            # Unfortunately, torch.randn_like(...) does not accept a generator argument at the time of writing, so we
+            # have to use torch.randn(...) instead.
+            return mean + std * torch.randn(size=mean.size(), generator=generator, dtype=mean.dtype, device=mean.device)
         else:
             return mean
 
@@ -297,8 +298,21 @@ class AutoEncoder(nn.Module):
         self.scale_factor = params.scale_factor
         self.shift_factor = params.shift_factor
 
-    def encode(self, x: Tensor) -> Tensor:
-        z = self.reg(self.encoder(x))
+    def encode(self, x: Tensor, sample: bool = True, generator: torch.Generator | None = None) -> Tensor:
+        """Run VAE encoding on input tensor x.
+
+        Args:
+            x (Tensor): Input image tensor. Shape: (batch_size, in_channels, height, width).
+            sample (bool, optional): If True, sample from the encoded distribution, else, return the distribution mean.
+                Defaults to True.
+            generator (torch.Generator | None, optional): Optional random number generator for reproducibility.
+                Defaults to None.
+
+        Returns:
+            Tensor: Encoded latent tensor. Shape: (batch_size, z_channels, latent_height, latent_width).
+        """
+
+        z = self.reg(self.encoder(x), sample=sample, generator=generator)
         z = self.scale_factor * (z - self.shift_factor)
         return z
 

invokeai/backend/flux/sampling.py

@@ -1,167 +0,0 @@
-# Initially pulled from https://github.com/black-forest-labs/flux
-
-import math
-from typing import Callable
-
-import torch
-from einops import rearrange, repeat
-from torch import Tensor
-from tqdm import tqdm
-
-from invokeai.backend.flux.model import Flux
-from invokeai.backend.flux.modules.conditioner import HFEncoder
-
-
-def get_noise(
-    num_samples: int,
-    height: int,
-    width: int,
-    device: torch.device,
-    dtype: torch.dtype,
-    seed: int,
-):
-    # We always generate noise on the same device and dtype then cast to ensure consistency across devices/dtypes.
-    rand_device = "cpu"
-    rand_dtype = torch.float16
-    return torch.randn(
-        num_samples,
-        16,
-        # allow for packing
-        2 * math.ceil(height / 16),
-        2 * math.ceil(width / 16),
-        device=rand_device,
-        dtype=rand_dtype,
-        generator=torch.Generator(device=rand_device).manual_seed(seed),
-    ).to(device=device, dtype=dtype)
-
-
-def prepare(t5: HFEncoder, clip: HFEncoder, img: Tensor, prompt: str | list[str]) -> dict[str, Tensor]:
-    bs, c, h, w = img.shape
-    if bs == 1 and not isinstance(prompt, str):
-        bs = len(prompt)
-
-    img = rearrange(img, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
-    if img.shape[0] == 1 and bs > 1:
-        img = repeat(img, "1 ... -> bs ...", bs=bs)
-
-    img_ids = torch.zeros(h // 2, w // 2, 3)
-    img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2)[:, None]
-    img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2)[None, :]
-    img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
-
-    if isinstance(prompt, str):
-        prompt = [prompt]
-    txt = t5(prompt)
-    if txt.shape[0] == 1 and bs > 1:
-        txt = repeat(txt, "1 ... -> bs ...", bs=bs)
-    txt_ids = torch.zeros(bs, txt.shape[1], 3)
-
-    vec = clip(prompt)
-    if vec.shape[0] == 1 and bs > 1:
-        vec = repeat(vec, "1 ... -> bs ...", bs=bs)
-
-    return {
-        "img": img,
-        "img_ids": img_ids.to(img.device),
-        "txt": txt.to(img.device),
-        "txt_ids": txt_ids.to(img.device),
-        "vec": vec.to(img.device),
-    }
-
-
-def time_shift(mu: float, sigma: float, t: Tensor):
-    return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
-
-
-def get_lin_function(x1: float = 256, y1: float = 0.5, x2: float = 4096, y2: float = 1.15) -> Callable[[float], float]:
-    m = (y2 - y1) / (x2 - x1)
-    b = y1 - m * x1
-    return lambda x: m * x + b
-
-
-def get_schedule(
-    num_steps: int,
-    image_seq_len: int,
-    base_shift: float = 0.5,
-    max_shift: float = 1.15,
-    shift: bool = True,
-) -> list[float]:
-    # extra step for zero
-    timesteps = torch.linspace(1, 0, num_steps + 1)
-
-    # shifting the schedule to favor high timesteps for higher signal images
-    if shift:
-        # eastimate mu based on linear estimation between two points
-        mu = get_lin_function(y1=base_shift, y2=max_shift)(image_seq_len)
-        timesteps = time_shift(mu, 1.0, timesteps)
-
-    return timesteps.tolist()
-
-
-def denoise(
-    model: Flux,
-    # model input
-    img: Tensor,
-    img_ids: Tensor,
-    txt: Tensor,
-    txt_ids: Tensor,
-    vec: Tensor,
-    # sampling parameters
-    timesteps: list[float],
-    step_callback: Callable[[], None],
-    guidance: float = 4.0,
-):
-    # guidance_vec is ignored for schnell.
-    guidance_vec = torch.full((img.shape[0],), guidance, device=img.device, dtype=img.dtype)
-    for t_curr, t_prev in tqdm(list(zip(timesteps[:-1], timesteps[1:], strict=True))):
-        t_vec = torch.full((img.shape[0],), t_curr, dtype=img.dtype, device=img.device)
-        pred = model(
-            img=img,
-            img_ids=img_ids,
-            txt=txt,
-            txt_ids=txt_ids,
-            y=vec,
-            timesteps=t_vec,
-            guidance=guidance_vec,
-        )
-
-        img = img + (t_prev - t_curr) * pred
-        step_callback()
-
-    return img
-
-
-def unpack(x: Tensor, height: int, width: int) -> Tensor:
-    return rearrange(
-        x,
-        "b (h w) (c ph pw) -> b c (h ph) (w pw)",
-        h=math.ceil(height / 16),
-        w=math.ceil(width / 16),
-        ph=2,
-        pw=2,
-    )
-
-
-def prepare_latent_img_patches(latent_img: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
-    """Convert an input image in latent space to patches for diffusion.
-
-    This implementation was extracted from:
-    https://github.com/black-forest-labs/flux/blob/c00d7c60b085fce8058b9df845e036090873f2ce/src/flux/sampling.py#L32
-
-    Returns:
-        tuple[Tensor, Tensor]: (img, img_ids), as defined in the original flux repo.
-    """
-    bs, c, h, w = latent_img.shape
-
-    # Pixel unshuffle with a scale of 2, and flatten the height/width dimensions to get an array of patches.
-    img = rearrange(latent_img, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
-    if img.shape[0] == 1 and bs > 1:
-        img = repeat(img, "1 ... -> bs ...", bs=bs)
-
-    # Generate patch position ids.
-    img_ids = torch.zeros(h // 2, w // 2, 3, device=img.device, dtype=img.dtype)
-    img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2, device=img.device, dtype=img.dtype)[:, None]
-    img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2, device=img.device, dtype=img.dtype)[None, :]
-    img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
-
-    return img, img_ids

invokeai/backend/flux/sampling_utils.py

@@ -0,0 +1,135 @@
+# Initially pulled from https://github.com/black-forest-labs/flux
+
+import math
+from typing import Callable
+
+import torch
+from einops import rearrange, repeat
+
+
+def get_noise(
+    num_samples: int,
+    height: int,
+    width: int,
+    device: torch.device,
+    dtype: torch.dtype,
+    seed: int,
+):
+    # We always generate noise on the same device and dtype then cast to ensure consistency across devices/dtypes.
+    rand_device = "cpu"
+    rand_dtype = torch.float16
+    return torch.randn(
+        num_samples,
+        16,
+        # allow for packing
+        2 * math.ceil(height / 16),
+        2 * math.ceil(width / 16),
+        device=rand_device,
+        dtype=rand_dtype,
+        generator=torch.Generator(device=rand_device).manual_seed(seed),
+    ).to(device=device, dtype=dtype)
+
+
+def time_shift(mu: float, sigma: float, t: torch.Tensor) -> torch.Tensor:
+    return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
+
+
+def get_lin_function(x1: float = 256, y1: float = 0.5, x2: float = 4096, y2: float = 1.15) -> Callable[[float], float]:
+    m = (y2 - y1) / (x2 - x1)
+    b = y1 - m * x1
+    return lambda x: m * x + b
+
+
+def get_schedule(
+    num_steps: int,
+    image_seq_len: int,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+    shift: bool = True,
+) -> list[float]:
+    # extra step for zero
+    timesteps = torch.linspace(1, 0, num_steps + 1)
+
+    # shifting the schedule to favor high timesteps for higher signal images
+    if shift:
+        # estimate mu based on linear estimation between two points
+        mu = get_lin_function(y1=base_shift, y2=max_shift)(image_seq_len)
+        timesteps = time_shift(mu, 1.0, timesteps)
+
+    return timesteps.tolist()
+
+
+def _find_last_index_ge_val(timesteps: list[float], val: float, eps: float = 1e-6) -> int:
+    """Find the last index in timesteps that is >= val.
+
+    We use epsilon-close equality to avoid potential floating point errors.
+    """
+    idx = len(list(filter(lambda t: t >= (val - eps), timesteps))) - 1
+    assert idx >= 0
+    return idx
+
+
+def clip_timestep_schedule(timesteps: list[float], denoising_start: float, denoising_end: float) -> list[float]:
+    """Clip the timestep schedule to the denoising range.
+
+    Args:
+        timesteps (list[float]): The original timestep schedule: [1.0, ..., 0.0].
+        denoising_start (float): A value in [0, 1] specifying the start of the denoising process. E.g. a value of 0.2
+            would mean that the denoising process start at the last timestep in the schedule >= 0.8.
+        denoising_end (float): A value in [0, 1] specifying the end of the denoising process. E.g. a value of 0.8 would
+            mean that the denoising process end at the last timestep in the schedule >= 0.2.
+
+    Returns:
+        list[float]: The clipped timestep schedule.
+    """
+    assert 0.0 <= denoising_start <= 1.0
+    assert 0.0 <= denoising_end <= 1.0
+    assert denoising_start <= denoising_end
+
+    t_start_val = 1.0 - denoising_start
+    t_end_val = 1.0 - denoising_end
+
+    t_start_idx = _find_last_index_ge_val(timesteps, t_start_val)
+    t_end_idx = _find_last_index_ge_val(timesteps, t_end_val)
+
+    clipped_timesteps = timesteps[t_start_idx : t_end_idx + 1]
+
+    return clipped_timesteps
+
+
+def unpack(x: torch.Tensor, height: int, width: int) -> torch.Tensor:
+    """Unpack flat array of patch embeddings to latent image."""
+    return rearrange(
+        x,
+        "b (h w) (c ph pw) -> b c (h ph) (w pw)",
+        h=math.ceil(height / 16),
+        w=math.ceil(width / 16),
+        ph=2,
+        pw=2,
+    )
+
+
+def pack(x: torch.Tensor) -> torch.Tensor:
+    """Pack latent image to flattented array of patch embeddings."""
+    # Pixel unshuffle with a scale of 2, and flatten the height/width dimensions to get an array of patches.
+    return rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+
+
+def generate_img_ids(h: int, w: int, batch_size: int, device: torch.device, dtype: torch.dtype) -> torch.Tensor:
+    """Generate tensor of image position ids.
+
+    Args:
+        h (int): Height of image in latent space.
+        w (int): Width of image in latent space.
+        batch_size (int): Batch size.
+        device (torch.device): Device.
+        dtype (torch.dtype): dtype.
+
+    Returns:
+        torch.Tensor: Image position ids.
+    """
+    img_ids = torch.zeros(h // 2, w // 2, 3, device=device, dtype=dtype)
+    img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2, device=device, dtype=dtype)[:, None]
+    img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2, device=device, dtype=dtype)[None, :]
+    img_ids = repeat(img_ids, "h w c -> b (h w) c", b=batch_size)
+    return img_ids

invokeai/backend/model_manager/load/load_default.py

@@ -66,8 +66,9 @@ class ModelLoader(ModelLoaderBase):
         return (model_base / config.path).resolve()
 
     def _load_and_cache(self, config: AnyModelConfig, submodel_type: Optional[SubModelType] = None) -> ModelLockerBase:
+        stats_name = ":".join([config.base, config.type, config.name, (submodel_type or "")])
         try:
-            return self._ram_cache.get(config.key, submodel_type)
+            return self._ram_cache.get(config.key, submodel_type, stats_name=stats_name)
         except IndexError:
             pass
 
@@ -84,7 +85,7 @@ class ModelLoader(ModelLoaderBase):
         return self._ram_cache.get(
             key=config.key,
             submodel_type=submodel_type,
-            stats_name=":".join([config.base, config.type, config.name, (submodel_type or "")]),
+            stats_name=stats_name,
         )
 
     def get_size_fs(

invokeai/backend/model_manager/load/model_cache/model_cache_base.py

@@ -128,7 +128,24 @@ class ModelCacheBase(ABC, Generic[T]):
     @property
     @abstractmethod
     def max_cache_size(self) -> float:
-        """Return true if the cache is configured to lazily offload models in VRAM."""
+        """Return the maximum size the RAM cache can grow to."""
+        pass
+
+    @max_cache_size.setter
+    @abstractmethod
+    def max_cache_size(self, value: float) -> None:
+        """Set the cap on vram cache size."""
+
+    @property
+    @abstractmethod
+    def max_vram_cache_size(self) -> float:
+        """Return the maximum size the VRAM cache can grow to."""
+        pass
+
+    @max_vram_cache_size.setter
+    @abstractmethod
+    def max_vram_cache_size(self, value: float) -> float:
+        """Set the maximum size the VRAM cache can grow to."""
         pass
 
     @abstractmethod

invokeai/backend/model_manager/load/model_cache/model_cache_default.py

@@ -70,6 +70,7 @@ class ModelCache(ModelCacheBase[AnyModel]):
         max_vram_cache_size: float,
         execution_device: torch.device = torch.device("cuda"),
         storage_device: torch.device = torch.device("cpu"),
+        precision: torch.dtype = torch.float16,
         lazy_offloading: bool = True,
         log_memory_usage: bool = False,
         logger: Optional[Logger] = None,
@@ -81,11 +82,13 @@ class ModelCache(ModelCacheBase[AnyModel]):
         :param max_vram_cache_size: Maximum size of the execution_device cache in GBs.
         :param execution_device: Torch device to load active model into [torch.device('cuda')]
         :param storage_device: Torch device to save inactive model in [torch.device('cpu')]
-        :param lazy_offloading: Keep model in VRAM until another model needs to be loaded.
+        :param precision: Precision for loaded models [torch.float16]
+        :param lazy_offloading: Keep model in VRAM until another model needs to be loaded
         :param log_memory_usage: If True, a memory snapshot will be captured before and after every model cache
             operation, and the result will be logged (at debug level). There is a time cost to capturing the memory
             snapshots, so it is recommended to disable this feature unless you are actively inspecting the model cache's
             behaviour.
+        :param logger: InvokeAILogger to use (otherwise creates one)
         """
         # allow lazy offloading only when vram cache enabled
         self._lazy_offloading = lazy_offloading and max_vram_cache_size > 0
@@ -130,6 +133,16 @@ class ModelCache(ModelCacheBase[AnyModel]):
         """Set the cap on cache size."""
         self._max_cache_size = value
 
+    @property
+    def max_vram_cache_size(self) -> float:
+        """Return the cap on vram cache size."""
+        return self._max_vram_cache_size
+
+    @max_vram_cache_size.setter
+    def max_vram_cache_size(self, value: float) -> None:
+        """Set the cap on vram cache size."""
+        self._max_vram_cache_size = value
+
     @property
     def stats(self) -> Optional[CacheStats]:
         """Return collected CacheStats object."""

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

@@ -32,6 +32,7 @@ from invokeai.backend.model_manager.config import (
 )
 from invokeai.backend.model_manager.load.load_default import ModelLoader
 from invokeai.backend.model_manager.load.model_loader_registry import ModelLoaderRegistry
+from invokeai.backend.model_manager.util.model_util import convert_bundle_to_flux_transformer_checkpoint
 from invokeai.backend.util.silence_warnings import SilenceWarnings
 
 try:
@@ -190,6 +191,8 @@ class FluxCheckpointModel(ModelLoader):
         with SilenceWarnings():
             model = Flux(params[config.config_path])
             sd = load_file(model_path)
+            if "model.diffusion_model.double_blocks.0.img_attn.norm.key_norm.scale" in sd:
+                sd = convert_bundle_to_flux_transformer_checkpoint(sd)
             model.load_state_dict(sd, assign=True)
         return model
 
@@ -230,5 +233,7 @@ class FluxBnbQuantizednf4bCheckpointModel(ModelLoader):
                 model = Flux(params[config.config_path])
                 model = quantize_model_nf4(model, modules_to_not_convert=set(), compute_dtype=torch.bfloat16)
             sd = load_file(model_path)
+            if "model.diffusion_model.double_blocks.0.img_attn.norm.key_norm.scale" in sd:
+                sd = convert_bundle_to_flux_transformer_checkpoint(sd)
             model.load_state_dict(sd, assign=True)
         return model

invokeai/backend/model_manager/probe.py

@@ -108,6 +108,8 @@ class ModelProbe(object):
         "CLIPVisionModelWithProjection": ModelType.CLIPVision,
         "T2IAdapter": ModelType.T2IAdapter,
         "CLIPModel": ModelType.CLIPEmbed,
+        "CLIPTextModel": ModelType.CLIPEmbed,
+        "T5EncoderModel": ModelType.T5Encoder,
     }
 
     @classmethod
@@ -224,7 +226,18 @@ class ModelProbe(object):
         ckpt = ckpt.get("state_dict", ckpt)
 
         for key in [str(k) for k in ckpt.keys()]:
-            if key.startswith(("cond_stage_model.", "first_stage_model.", "model.diffusion_model.", "double_blocks.")):
+            if key.startswith(
+                (
+                    "cond_stage_model.",
+                    "first_stage_model.",
+                    "model.diffusion_model.",
+                    # FLUX models in the official BFL format contain keys with the "double_blocks." prefix.
+                    "double_blocks.",
+                    # Some FLUX checkpoint files contain transformer keys prefixed with "model.diffusion_model".
+                    # This prefix is typically used to distinguish between multiple models bundled in a single file.
+                    "model.diffusion_model.double_blocks.",
+                )
+            ):
                 # Keys starting with double_blocks are associated with Flux models
                 return ModelType.Main
             elif key.startswith(("encoder.conv_in", "decoder.conv_in")):
@@ -283,9 +296,16 @@ class ModelProbe(object):
         if (folder_path / "image_encoder.txt").exists():
             return ModelType.IPAdapter
 
-        i = folder_path / "model_index.json"
-        c = folder_path / "config.json"
-        config_path = i if i.exists() else c if c.exists() else None
+        config_path = None
+        for p in [
+            folder_path / "model_index.json",  # pipeline
+            folder_path / "config.json",  # most diffusers
+            folder_path / "text_encoder_2" / "config.json",  # T5 text encoder
+            folder_path / "text_encoder" / "config.json",  # T5 CLIP
+        ]:
+            if p.exists():
+                config_path = p
+                break
 
         if config_path:
             with open(config_path, "r") as file:
@@ -328,7 +348,10 @@ class ModelProbe(object):
                 # TODO: Decide between dev/schnell
                 checkpoint = ModelProbe._scan_and_load_checkpoint(model_path)
                 state_dict = checkpoint.get("state_dict") or checkpoint
-                if "guidance_in.out_layer.weight" in state_dict:
+                if (
+                    "guidance_in.out_layer.weight" in state_dict
+                    or "model.diffusion_model.guidance_in.out_layer.weight" in state_dict
+                ):
                     # For flux, this is a key in invokeai.backend.flux.util.params
                     #   Due to model type and format being the descriminator for model configs this
                     #   is used rather than attempting to support flux with separate model types and format
@@ -336,7 +359,7 @@ class ModelProbe(object):
                     config_file = "flux-dev"
                 else:
                     # For flux, this is a key in invokeai.backend.flux.util.params
-                    #   Due to model type and format being the descriminator for model configs this
+                    #   Due to model type and format being the discriminator for model configs this
                     #   is used rather than attempting to support flux with separate model types and format
                     #   If changed in the future, please fix me
                     config_file = "flux-schnell"
@@ -443,7 +466,10 @@ class CheckpointProbeBase(ProbeBase):
 
     def get_format(self) -> ModelFormat:
         state_dict = self.checkpoint.get("state_dict") or self.checkpoint
-        if "double_blocks.0.img_attn.proj.weight.quant_state.bitsandbytes__nf4" in state_dict:
+        if (
+            "double_blocks.0.img_attn.proj.weight.quant_state.bitsandbytes__nf4" in state_dict
+            or "model.diffusion_model.double_blocks.0.img_attn.proj.weight.quant_state.bitsandbytes__nf4" in state_dict
+        ):
             return ModelFormat.BnbQuantizednf4b
         return ModelFormat("checkpoint")
 
@@ -470,7 +496,10 @@ class PipelineCheckpointProbe(CheckpointProbeBase):
     def get_base_type(self) -> BaseModelType:
         checkpoint = self.checkpoint
         state_dict = self.checkpoint.get("state_dict") or checkpoint
-        if "double_blocks.0.img_attn.norm.key_norm.scale" in state_dict:
+        if (
+            "double_blocks.0.img_attn.norm.key_norm.scale" in state_dict
+            or "model.diffusion_model.double_blocks.0.img_attn.norm.key_norm.scale" in state_dict
+        ):
             return BaseModelType.Flux
         key_name = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight"
         if key_name in state_dict and state_dict[key_name].shape[-1] == 768:
@@ -747,8 +776,27 @@ class TextualInversionFolderProbe(FolderProbeBase):
 
 
 class T5EncoderFolderProbe(FolderProbeBase):
+    def get_base_type(self) -> BaseModelType:
+        return BaseModelType.Any
+
     def get_format(self) -> ModelFormat:
-        return ModelFormat.T5Encoder
+        path = self.model_path / "text_encoder_2"
+        if (path / "model.safetensors.index.json").exists():
+            return ModelFormat.T5Encoder
+        files = list(path.glob("*.safetensors"))
+        if len(files) == 0:
+            raise InvalidModelConfigException(f"{self.model_path.as_posix()}: no .safetensors files found")
+
+        # shortcut: look for the quantization in the name
+        if any(x for x in files if "llm_int8" in x.as_posix()):
+            return ModelFormat.BnbQuantizedLlmInt8b
+
+        # more reliable path: probe contents for a 'SCB' key
+        ckpt = read_checkpoint_meta(files[0], scan=True)
+        if any("SCB" in x for x in ckpt.keys()):
+            return ModelFormat.BnbQuantizedLlmInt8b
+
+        raise InvalidModelConfigException(f"{self.model_path.as_posix()}: unknown model format")
 
 
 class ONNXFolderProbe(PipelineFolderProbe):

invokeai/backend/model_manager/util/model_util.py

@@ -133,3 +133,29 @@ def lora_token_vector_length(checkpoint: Dict[str, torch.Tensor]) -> Optional[in
             break
 
     return lora_token_vector_length
+
+
+def convert_bundle_to_flux_transformer_checkpoint(
+    transformer_state_dict: dict[str, torch.Tensor],
+) -> dict[str, torch.Tensor]:
+    original_state_dict: dict[str, torch.Tensor] = {}
+    keys_to_remove: list[str] = []
+
+    for k, v in transformer_state_dict.items():
+        if not k.startswith("model.diffusion_model"):
+            keys_to_remove.append(k)  # This can be removed in the future if we only want to delete transformer keys
+            continue
+        if k.endswith("scale"):
+            # Scale math must be done at bfloat16 due to our current flux model
+            # support limitations at inference time
+            v = v.to(dtype=torch.bfloat16)
+        new_key = k.replace("model.diffusion_model.", "")
+        original_state_dict[new_key] = v
+        keys_to_remove.append(k)
+
+    # Remove processed keys from the original dictionary, leaving others in case
+    # other model state dicts need to be pulled
+    for k in keys_to_remove:
+        del transformer_state_dict[k]
+
+    return original_state_dict

invokeai/version/invokeai_version.py

@@ -1 +1 @@
-__version__ = "4.2.9rc1"
+__version__ = "4.2.9rc2"

scripts/allocate_vram.py

@@ -0,0 +1,63 @@
+import argparse
+
+import torch
+
+
+def display_vram_usage():
+    """Displays the total, allocated, and free VRAM on the current CUDA device."""
+
+    assert torch.cuda.is_available(), "CUDA is not available"
+    device = torch.device("cuda")
+
+    total_vram = torch.cuda.get_device_properties(device).total_memory
+    allocated_vram = torch.cuda.memory_allocated(device)
+    free_vram = total_vram - allocated_vram
+
+    print(f"Total VRAM: {total_vram / (1024 * 1024 * 1024):.2f} GB")
+    print(f"Allocated VRAM: {allocated_vram / (1024 * 1024 * 1024):.2f} GB")
+    print(f"Free VRAM: {free_vram / (1024 * 1024 * 1024):.2f} GB")
+
+
+def allocate_vram(target_gb: float, target_free: bool = False):
+    """Allocates VRAM on the current CUDA device. After allocation, the script will pause until the user presses Enter
+    or ends the script, at which point the VRAM will be released.
+
+    Args:
+        target_gb (float): Amount of VRAM to allocate in GB.
+        target_free (bool, optional): Instead of allocating <target_gb> VRAM, enough VRAM will be allocated so the system has <target_gb> of VRAM free. For example, if <target_gb> is 2 GB, the script will allocate VRAM until the free VRAM is 2 GB.
+    """
+    assert torch.cuda.is_available(), "CUDA is not available"
+    device = torch.device("cuda")
+
+    if target_free:
+        total_vram = torch.cuda.get_device_properties(device).total_memory
+        free_vram = total_vram - torch.cuda.memory_allocated(device)
+        target_free_bytes = target_gb * 1024 * 1024 * 1024
+        bytes_to_allocate = free_vram - target_free_bytes
+
+        if bytes_to_allocate <= 0:
+            print(f"Already at or below the target free VRAM of {target_gb} GB")
+            return
+    else:
+        bytes_to_allocate = target_gb * 1024 * 1024 * 1024
+
+    # FloatTensor (4 bytes per element)
+    _tensor = torch.empty(int(bytes_to_allocate / 4), dtype=torch.float, device="cuda")
+
+    display_vram_usage()
+
+    input("Press Enter to release VRAM allocation and exit...")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Allocate VRAM for testing purposes. Only works on CUDA devices.")
+    parser.add_argument("target_gb", type=float, help="Amount of VRAM to allocate in GB.")
+    parser.add_argument(
+        "--target-free",
+        action="store_true",
+        help="Instead of allocating <target_gb> VRAM, enough VRAM will be allocated so the system has <target_gb> of VRAM free. For example, if <target_gb> is 2 GB, the script will allocate VRAM until the free VRAM is 2 GB.",
+    )
+
+    args = parser.parse_args()
+
+    allocate_vram(target_gb=args.target_gb, target_free=args.target_free)

tests/backend/flux/test_sampling_utils.py

@@ -0,0 +1,42 @@
+import pytest
+import torch
+
+from invokeai.backend.flux.sampling_utils import clip_timestep_schedule
+
+
+def float_lists_almost_equal(list1: list[float], list2: list[float], tol: float = 1e-6) -> bool:
+    return all(abs(a - b) < tol for a, b in zip(list1, list2, strict=True))
+
+
+@pytest.mark.parametrize(
+    ["denoising_start", "denoising_end", "expected_timesteps", "raises"],
+    [
+        (0.0, 1.0, [1.0, 0.75, 0.5, 0.25, 0.0], False),  # Default case.
+        (-0.1, 1.0, [], True),  # Negative denoising_start should raise.
+        (0.0, 1.1, [], True),  # denoising_end > 1 should raise.
+        (0.5, 0.0, [], True),  # denoising_start > denoising_end should raise.
+        (0.0, 0.0, [1.0], False),  # denoising_end == 0.
+        (1.0, 1.0, [0.0], False),  # denoising_start == 1.
+        (0.2, 0.8, [1.0, 0.75, 0.5, 0.25], False),  # Middle of the schedule.
+        # If we denoise from 0.0 to x, then from x to 1.0, it is important that denoise_end = x and denoise_start = x
+        # map to the same timestep. We test this first when x is equal to a timestep, then when it falls between two
+        # timesteps.
+        # x = 0.5
+        (0.0, 0.5, [1.0, 0.75, 0.5], False),
+        (0.5, 1.0, [0.5, 0.25, 0.0], False),
+        # x = 0.3
+        (0.0, 0.3, [1.0, 0.75], False),
+        (0.3, 1.0, [0.75, 0.5, 0.25, 0.0], False),
+    ],
+)
+def test_clip_timestep_schedule(
+    denoising_start: float, denoising_end: float, expected_timesteps: list[float], raises: bool
+):
+    timesteps = torch.linspace(1, 0, 5).tolist()
+    if raises:
+        with pytest.raises(AssertionError):
+            clip_timestep_schedule(timesteps, denoising_start, denoising_end)
+    else:
+        assert float_lists_almost_equal(
+            clip_timestep_schedule(timesteps, denoising_start, denoising_end), expected_timesteps
+        )