Update invokeai_version.py

## Summary

https://github.com/invoke-ai/InvokeAI/pull/6522 introduced a change in
behavior in cases where start/end were set such that there are 0
timesteps. This PR reverts that change.

cc @StAlKeR7779 

## QA Instructions

Run with euler, 5 steps, start: 0.0, end: 0.05. I ran this test before
#6522, after #6522, and on this branch. This branch restores the
behavior to pre-#6522 i.e. noise is injected even if no denoising steps
are applied.


## Checklist

- [x] _The PR has a short but descriptive title, suitable for a
changelog_
- [x] _Tests added / updated (if applicable)_
- [x] _Documentation added / updated (if applicable)_

invokeai/app/invocations/denoise_latents.py

@@ -55,6 +55,7 @@ from invokeai.backend.stable_diffusion.diffusion.conditioning_data import (
 )
 from invokeai.backend.stable_diffusion.schedulers import SCHEDULER_MAP
 from invokeai.backend.util.devices import TorchDevice
+from invokeai.backend.util.hotfixes import ControlNetModel
 from invokeai.backend.util.mask import to_standard_float_mask
 from invokeai.backend.util.silence_warnings import SilenceWarnings
 
@@ -65,6 +66,9 @@ def get_scheduler(
     scheduler_name: str,
     seed: int,
 ) -> Scheduler:
+    """Load a scheduler and apply some scheduler-specific overrides."""
+    # TODO(ryand): Silently falling back to ddim seems like a bad idea. Look into why this was added and remove if
+    # possible.
     scheduler_class, scheduler_extra_config = SCHEDULER_MAP.get(scheduler_name, SCHEDULER_MAP["ddim"])
     orig_scheduler_info = context.models.load(scheduler_info)
     with orig_scheduler_info as orig_scheduler:
@@ -182,8 +186,8 @@ class DenoiseLatentsInvocation(BaseInvocation):
                 raise ValueError("cfg_scale must be greater than 1")
         return v
 
+    @staticmethod
     def _get_text_embeddings_and_masks(
-        self,
         cond_list: list[ConditioningField],
         context: InvocationContext,
         device: torch.device,
@@ -203,8 +207,9 @@ class DenoiseLatentsInvocation(BaseInvocation):
 
         return text_embeddings, text_embeddings_masks
 
+    @staticmethod
     def _preprocess_regional_prompt_mask(
-        self, mask: Optional[torch.Tensor], target_height: int, target_width: int, dtype: torch.dtype
+        mask: Optional[torch.Tensor], target_height: int, target_width: int, dtype: torch.dtype
     ) -> torch.Tensor:
         """Preprocess a regional prompt mask to match the target height and width.
         If mask is None, returns a mask of all ones with the target height and width.
@@ -228,8 +233,8 @@ class DenoiseLatentsInvocation(BaseInvocation):
         resized_mask = tf(mask)
         return resized_mask
 
+    @staticmethod
     def _concat_regional_text_embeddings(
-        self,
         text_conditionings: Union[list[BasicConditioningInfo], list[SDXLConditioningInfo]],
         masks: Optional[list[Optional[torch.Tensor]]],
         latent_height: int,
@@ -279,7 +284,9 @@ class DenoiseLatentsInvocation(BaseInvocation):
                     )
                 )
                 processed_masks.append(
-                    self._preprocess_regional_prompt_mask(mask, latent_height, latent_width, dtype=dtype)
+                    DenoiseLatentsInvocation._preprocess_regional_prompt_mask(
+                        mask, latent_height, latent_width, dtype=dtype
+                    )
                 )
 
             cur_text_embedding_len += text_embedding_info.embeds.shape[1]
@@ -301,36 +308,41 @@ class DenoiseLatentsInvocation(BaseInvocation):
             )
         return BasicConditioningInfo(embeds=text_embedding), regions
 
+    @staticmethod
     def get_conditioning_data(
-        self,
         context: InvocationContext,
+        positive_conditioning_field: Union[ConditioningField, list[ConditioningField]],
+        negative_conditioning_field: Union[ConditioningField, list[ConditioningField]],
         unet: UNet2DConditionModel,
         latent_height: int,
         latent_width: int,
+        cfg_scale: float | list[float],
+        steps: int,
+        cfg_rescale_multiplier: float,
     ) -> TextConditioningData:
-        # Normalize self.positive_conditioning and self.negative_conditioning to lists.
-        cond_list = self.positive_conditioning
+        # Normalize positive_conditioning_field and negative_conditioning_field to lists.
+        cond_list = positive_conditioning_field
         if not isinstance(cond_list, list):
             cond_list = [cond_list]
-        uncond_list = self.negative_conditioning
+        uncond_list = negative_conditioning_field
         if not isinstance(uncond_list, list):
             uncond_list = [uncond_list]
 
-        cond_text_embeddings, cond_text_embedding_masks = self._get_text_embeddings_and_masks(
+        cond_text_embeddings, cond_text_embedding_masks = DenoiseLatentsInvocation._get_text_embeddings_and_masks(
             cond_list, context, unet.device, unet.dtype
         )
-        uncond_text_embeddings, uncond_text_embedding_masks = self._get_text_embeddings_and_masks(
+        uncond_text_embeddings, uncond_text_embedding_masks = DenoiseLatentsInvocation._get_text_embeddings_and_masks(
             uncond_list, context, unet.device, unet.dtype
         )
 
-        cond_text_embedding, cond_regions = self._concat_regional_text_embeddings(
+        cond_text_embedding, cond_regions = DenoiseLatentsInvocation._concat_regional_text_embeddings(
             text_conditionings=cond_text_embeddings,
             masks=cond_text_embedding_masks,
             latent_height=latent_height,
             latent_width=latent_width,
             dtype=unet.dtype,
         )
-        uncond_text_embedding, uncond_regions = self._concat_regional_text_embeddings(
+        uncond_text_embedding, uncond_regions = DenoiseLatentsInvocation._concat_regional_text_embeddings(
             text_conditionings=uncond_text_embeddings,
             masks=uncond_text_embedding_masks,
             latent_height=latent_height,
@@ -338,23 +350,21 @@ class DenoiseLatentsInvocation(BaseInvocation):
             dtype=unet.dtype,
         )
 
-        if isinstance(self.cfg_scale, list):
-            assert (
-                len(self.cfg_scale) == self.steps
-            ), "cfg_scale (list) must have the same length as the number of steps"
+        if isinstance(cfg_scale, list):
+            assert len(cfg_scale) == steps, "cfg_scale (list) must have the same length as the number of steps"
 
         conditioning_data = TextConditioningData(
             uncond_text=uncond_text_embedding,
             cond_text=cond_text_embedding,
             uncond_regions=uncond_regions,
             cond_regions=cond_regions,
-            guidance_scale=self.cfg_scale,
-            guidance_rescale_multiplier=self.cfg_rescale_multiplier,
+            guidance_scale=cfg_scale,
+            guidance_rescale_multiplier=cfg_rescale_multiplier,
         )
         return conditioning_data
 
+    @staticmethod
     def create_pipeline(
-        self,
         unet: UNet2DConditionModel,
         scheduler: Scheduler,
     ) -> StableDiffusionGeneratorPipeline:
@@ -377,38 +387,38 @@ class DenoiseLatentsInvocation(BaseInvocation):
             requires_safety_checker=False,
         )
 
+    @staticmethod
     def prep_control_data(
-        self,
         context: InvocationContext,
-        control_input: Optional[Union[ControlField, List[ControlField]]],
+        control_input: ControlField | list[ControlField] | None,
         latents_shape: List[int],
         exit_stack: ExitStack,
         do_classifier_free_guidance: bool = True,
-    ) -> Optional[List[ControlNetData]]:
-        # Assuming fixed dimensional scaling of LATENT_SCALE_FACTOR.
-        control_height_resize = latents_shape[2] * LATENT_SCALE_FACTOR
-        control_width_resize = latents_shape[3] * LATENT_SCALE_FACTOR
-        if control_input is None:
-            control_list = None
-        elif isinstance(control_input, list) and len(control_input) == 0:
-            control_list = None
-        elif isinstance(control_input, ControlField):
+    ) -> list[ControlNetData] | None:
+        # Normalize control_input to a list.
+        control_list: list[ControlField]
+        if isinstance(control_input, ControlField):
             control_list = [control_input]
-        elif isinstance(control_input, list) and len(control_input) > 0 and isinstance(control_input[0], ControlField):
+        elif isinstance(control_input, list):
             control_list = control_input
+        elif control_input is None:
+            control_list = []
         else:
-            control_list = None
-        if control_list is None:
-            return None
-        # After above handling, any control that is not None should now be of type list[ControlField].
+            raise ValueError(f"Unexpected control_input type: {type(control_input)}")
 
-        # FIXME: add checks to skip entry if model or image is None
-        #        and if weight is None, populate with default 1.0?
-        controlnet_data = []
+        if len(control_list) == 0:
+            return None
+
+        # Assuming fixed dimensional scaling of LATENT_SCALE_FACTOR.
+        _, _, latent_height, latent_width = latents_shape
+        control_height_resize = latent_height * LATENT_SCALE_FACTOR
+        control_width_resize = latent_width * LATENT_SCALE_FACTOR
+
+        controlnet_data: list[ControlNetData] = []
         for control_info in control_list:
             control_model = exit_stack.enter_context(context.models.load(control_info.control_model))
+            assert isinstance(control_model, ControlNetModel)
 
-            # control_models.append(control_model)
             control_image_field = control_info.image
             input_image = context.images.get_pil(control_image_field.image_name)
             # self.image.image_type, self.image.image_name
@@ -429,7 +439,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
                 resize_mode=control_info.resize_mode,
             )
             control_item = ControlNetData(
-                model=control_model,  # model object
+                model=control_model,
                 image_tensor=control_image,
                 weight=control_info.control_weight,
                 begin_step_percent=control_info.begin_step_percent,
@@ -583,15 +593,15 @@ class DenoiseLatentsInvocation(BaseInvocation):
 
     # original idea by https://github.com/AmericanPresidentJimmyCarter
     # TODO: research more for second order schedulers timesteps
+    @staticmethod
     def init_scheduler(
-        self,
         scheduler: Union[Scheduler, ConfigMixin],
         device: torch.device,
         steps: int,
         denoising_start: float,
         denoising_end: float,
         seed: int,
-    ) -> Tuple[int, List[int], int, Dict[str, Any]]:
+    ) -> Tuple[torch.Tensor, torch.Tensor, Dict[str, Any]]:
         assert isinstance(scheduler, ConfigMixin)
         if scheduler.config.get("cpu_only", False):
             scheduler.set_timesteps(steps, device="cpu")
@@ -617,7 +627,6 @@ class DenoiseLatentsInvocation(BaseInvocation):
 
         init_timestep = timesteps[t_start_idx : t_start_idx + 1]
         timesteps = timesteps[t_start_idx : t_start_idx + t_end_idx]
-        num_inference_steps = len(timesteps) // scheduler.order
 
         scheduler_step_kwargs: Dict[str, Any] = {}
         scheduler_step_signature = inspect.signature(scheduler.step)
@@ -639,7 +648,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
         if isinstance(scheduler, TCDScheduler):
             scheduler_step_kwargs.update({"eta": 1.0})
 
-        return num_inference_steps, timesteps, init_timestep, scheduler_step_kwargs
+        return timesteps, init_timestep, scheduler_step_kwargs
 
     def prep_inpaint_mask(
         self, context: InvocationContext, latents: torch.Tensor
@@ -656,31 +665,52 @@ class DenoiseLatentsInvocation(BaseInvocation):
 
         return 1 - mask, masked_latents, self.denoise_mask.gradient
 
-    @torch.no_grad()
-    @SilenceWarnings()  # This quenches the NSFW nag from diffusers.
-    def invoke(self, context: InvocationContext) -> LatentsOutput:
-        seed = None
+    @staticmethod
+    def prepare_noise_and_latents(
+        context: InvocationContext, noise_field: LatentsField | None, latents_field: LatentsField | None
+    ) -> Tuple[int, torch.Tensor | None, torch.Tensor]:
+        """Depending on the workflow, we expect different combinations of noise and latents to be provided. This
+        function handles preparing these values accordingly.
+
+        Expected workflows:
+        - Text-to-Image Denoising: `noise` is provided, `latents` is not. `latents` is initialized to zeros.
+        - Image-to-Image Denoising: `noise` and `latents` are both provided.
+        - Text-to-Image SDXL Refiner Denoising: `latents` is provided, `noise` is not.
+        - Image-to-Image SDXL Refiner Denoising: `latents` is provided, `noise` is not.
+
+        NOTE(ryand): I wrote this docstring, but I am not the original author of this code. There may be other workflows
+        I haven't considered.
+        """
         noise = None
-        if self.noise is not None:
-            noise = context.tensors.load(self.noise.latents_name)
-            seed = self.noise.seed
-
-        if self.latents is not None:
-            latents = context.tensors.load(self.latents.latents_name)
-            if seed is None:
-                seed = self.latents.seed
-
-            if noise is not None and noise.shape[1:] != latents.shape[1:]:
-                raise Exception(f"Incompatable 'noise' and 'latents' shapes: {latents.shape=} {noise.shape=}")
+        if noise_field is not None:
+            noise = context.tensors.load(noise_field.latents_name)
 
+        if latents_field is not None:
+            latents = context.tensors.load(latents_field.latents_name)
         elif noise is not None:
             latents = torch.zeros_like(noise)
         else:
-            raise Exception("'latents' or 'noise' must be provided!")
+            raise ValueError("'latents' or 'noise' must be provided!")
 
-        if seed is None:
+        if noise is not None and noise.shape[1:] != latents.shape[1:]:
+            raise ValueError(f"Incompatible 'noise' and 'latents' shapes: {latents.shape=} {noise.shape=}")
+
+        # The seed comes from (in order of priority): the noise field, the latents field, or 0.
+        seed = 0
+        if noise_field is not None and noise_field.seed is not None:
+            seed = noise_field.seed
+        elif latents_field is not None and latents_field.seed is not None:
+            seed = latents_field.seed
+        else:
             seed = 0
 
+        return seed, noise, latents
+
+    @torch.no_grad()
+    @SilenceWarnings()  # This quenches the NSFW nag from diffusers.
+    def invoke(self, context: InvocationContext) -> LatentsOutput:
+        seed, noise, latents = self.prepare_noise_and_latents(context, self.noise, self.latents)
+
         mask, masked_latents, gradient_mask = self.prep_inpaint_mask(context, latents)
 
         # TODO(ryand): I have hard-coded `do_classifier_free_guidance=True` to mirror the behaviour of ControlNets,
@@ -706,7 +736,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
         # The image prompts are then passed to prep_ip_adapter_data().
         image_prompts = self.prep_ip_adapter_image_prompts(context=context, ip_adapters=ip_adapters)
 
-        # get the unet's config so that we can pass the base to dispatch_progress()
+        # get the unet's config so that we can pass the base to sd_step_callback()
         unet_config = context.models.get_config(self.unet.unet.key)
 
         def step_callback(state: PipelineIntermediateState) -> None:
@@ -754,7 +784,15 @@ class DenoiseLatentsInvocation(BaseInvocation):
 
             _, _, latent_height, latent_width = latents.shape
             conditioning_data = self.get_conditioning_data(
-                context=context, unet=unet, latent_height=latent_height, latent_width=latent_width
+                context=context,
+                positive_conditioning_field=self.positive_conditioning,
+                negative_conditioning_field=self.negative_conditioning,
+                unet=unet,
+                latent_height=latent_height,
+                latent_width=latent_width,
+                cfg_scale=self.cfg_scale,
+                steps=self.steps,
+                cfg_rescale_multiplier=self.cfg_rescale_multiplier,
             )
 
             controlnet_data = self.prep_control_data(
@@ -776,7 +814,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
                 dtype=unet.dtype,
             )
 
-            num_inference_steps, timesteps, init_timestep, scheduler_step_kwargs = self.init_scheduler(
+            timesteps, init_timestep, scheduler_step_kwargs = self.init_scheduler(
                 scheduler,
                 device=unet.device,
                 steps=self.steps,
@@ -793,8 +831,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
                 seed=seed,
                 mask=mask,
                 masked_latents=masked_latents,
-                gradient_mask=gradient_mask,
-                num_inference_steps=num_inference_steps,
+                is_gradient_mask=gradient_mask,
                 scheduler_step_kwargs=scheduler_step_kwargs,
                 conditioning_data=conditioning_data,
                 control_data=controlnet_data,

invokeai/app/invocations/tiled_multi_diffusion_denoise_latents.py

@@ -0,0 +1,281 @@
+import copy
+from contextlib import ExitStack
+from typing import Iterator, Tuple
+
+import torch
+from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from pydantic import field_validator
+
+from invokeai.app.invocations.baseinvocation import BaseInvocation, Classification, invocation
+from invokeai.app.invocations.constants import LATENT_SCALE_FACTOR, SCHEDULER_NAME_VALUES
+from invokeai.app.invocations.controlnet_image_processors import ControlField
+from invokeai.app.invocations.denoise_latents import DenoiseLatentsInvocation, get_scheduler
+from invokeai.app.invocations.fields import (
+    ConditioningField,
+    FieldDescriptions,
+    Input,
+    InputField,
+    LatentsField,
+    UIType,
+)
+from invokeai.app.invocations.model import UNetField
+from invokeai.app.invocations.primitives import LatentsOutput
+from invokeai.app.services.shared.invocation_context import InvocationContext
+from invokeai.backend.lora import LoRAModelRaw
+from invokeai.backend.model_patcher import ModelPatcher
+from invokeai.backend.stable_diffusion.diffusers_pipeline import ControlNetData, PipelineIntermediateState
+from invokeai.backend.stable_diffusion.multi_diffusion_pipeline import (
+    MultiDiffusionPipeline,
+    MultiDiffusionRegionConditioning,
+)
+from invokeai.backend.tiles.tiles import (
+    calc_tiles_min_overlap,
+)
+from invokeai.backend.tiles.utils import TBLR
+from invokeai.backend.util.devices import TorchDevice
+
+
+def crop_controlnet_data(control_data: ControlNetData, latent_region: TBLR) -> ControlNetData:
+    """Crop a ControlNetData object to a region."""
+    # Create a shallow copy of the control_data object.
+    control_data_copy = copy.copy(control_data)
+    # The ControlNet reference image is the only attribute that needs to be cropped.
+    control_data_copy.image_tensor = control_data.image_tensor[
+        :,
+        :,
+        latent_region.top * LATENT_SCALE_FACTOR : latent_region.bottom * LATENT_SCALE_FACTOR,
+        latent_region.left * LATENT_SCALE_FACTOR : latent_region.right * LATENT_SCALE_FACTOR,
+    ]
+    return control_data_copy
+
+
+@invocation(
+    "tiled_multi_diffusion_denoise_latents",
+    title="Tiled Multi-Diffusion Denoise Latents",
+    tags=["upscale", "denoise"],
+    category="latents",
+    classification=Classification.Beta,
+    version="1.0.0",
+)
+class TiledMultiDiffusionDenoiseLatents(BaseInvocation):
+    """Tiled Multi-Diffusion denoising.
+
+    This node handles automatically tiling the input image, and is primarily intended for global refinement of images
+    in tiled upscaling workflows. Future Multi-Diffusion nodes should allow the user to specify custom regions with
+    different parameters for each region to harness the full power of Multi-Diffusion.
+
+    This node has a similar interface to the `DenoiseLatents` node, but it has a reduced feature set (no IP-Adapter,
+    T2I-Adapter, masking, etc.).
+    """
+
+    positive_conditioning: ConditioningField = InputField(
+        description=FieldDescriptions.positive_cond, input=Input.Connection
+    )
+    negative_conditioning: ConditioningField = InputField(
+        description=FieldDescriptions.negative_cond, input=Input.Connection
+    )
+    noise: LatentsField | None = InputField(
+        default=None,
+        description=FieldDescriptions.noise,
+        input=Input.Connection,
+    )
+    latents: LatentsField | None = InputField(
+        default=None,
+        description=FieldDescriptions.latents,
+        input=Input.Connection,
+    )
+    tile_height: int = InputField(
+        default=1024, gt=0, multiple_of=LATENT_SCALE_FACTOR, description="Height of the tiles in image space."
+    )
+    tile_width: int = InputField(
+        default=1024, gt=0, multiple_of=LATENT_SCALE_FACTOR, description="Width of the tiles in image space."
+    )
+    tile_overlap: int = InputField(
+        default=32,
+        multiple_of=LATENT_SCALE_FACTOR,
+        gt=0,
+        description="The overlap between adjacent tiles in pixel space. (Of course, tile merging is applied in latent "
+        "space.) Tiles will be cropped during merging (if necessary) to ensure that they overlap by exactly this "
+        "amount.",
+    )
+    steps: int = InputField(default=18, gt=0, description=FieldDescriptions.steps)
+    cfg_scale: float | list[float] = InputField(default=6.0, description=FieldDescriptions.cfg_scale, title="CFG Scale")
+    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)
+    scheduler: SCHEDULER_NAME_VALUES = InputField(
+        default="euler",
+        description=FieldDescriptions.scheduler,
+        ui_type=UIType.Scheduler,
+    )
+    unet: UNetField = InputField(
+        description=FieldDescriptions.unet,
+        input=Input.Connection,
+        title="UNet",
+    )
+    cfg_rescale_multiplier: float = InputField(
+        title="CFG Rescale Multiplier", default=0, ge=0, lt=1, description=FieldDescriptions.cfg_rescale_multiplier
+    )
+    control: ControlField | list[ControlField] | None = InputField(
+        default=None,
+        input=Input.Connection,
+    )
+
+    @field_validator("cfg_scale")
+    def ge_one(cls, v: list[float] | float) -> list[float] | float:
+        """Validate that all cfg_scale values are >= 1"""
+        if isinstance(v, list):
+            for i in v:
+                if i < 1:
+                    raise ValueError("cfg_scale must be greater than 1")
+        else:
+            if v < 1:
+                raise ValueError("cfg_scale must be greater than 1")
+        return v
+
+    @staticmethod
+    def create_pipeline(
+        unet: UNet2DConditionModel,
+        scheduler: SchedulerMixin,
+    ) -> MultiDiffusionPipeline:
+        # TODO(ryand): Get rid of this FakeVae hack.
+        class FakeVae:
+            class FakeVaeConfig:
+                def __init__(self) -> None:
+                    self.block_out_channels = [0]
+
+            def __init__(self) -> None:
+                self.config = FakeVae.FakeVaeConfig()
+
+        return MultiDiffusionPipeline(
+            vae=FakeVae(),
+            text_encoder=None,
+            tokenizer=None,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=None,
+            feature_extractor=None,
+            requires_safety_checker=False,
+        )
+
+    @torch.no_grad()
+    def invoke(self, context: InvocationContext) -> LatentsOutput:
+        # Convert tile image-space dimensions to latent-space dimensions.
+        latent_tile_height = self.tile_height // LATENT_SCALE_FACTOR
+        latent_tile_width = self.tile_width // LATENT_SCALE_FACTOR
+        latent_tile_overlap = self.tile_overlap // LATENT_SCALE_FACTOR
+
+        seed, noise, latents = DenoiseLatentsInvocation.prepare_noise_and_latents(context, self.noise, self.latents)
+        _, _, latent_height, latent_width = latents.shape
+
+        # Calculate the tile locations to cover the latent-space image.
+        tiles = calc_tiles_min_overlap(
+            image_height=latent_height,
+            image_width=latent_width,
+            tile_height=latent_tile_height,
+            tile_width=latent_tile_width,
+            min_overlap=latent_tile_overlap,
+        )
+
+        # Get the unet's config so that we can pass the base to sd_step_callback().
+        unet_config = context.models.get_config(self.unet.unet.key)
+
+        def step_callback(state: PipelineIntermediateState) -> None:
+            context.util.sd_step_callback(state, unet_config.base)
+
+        # Prepare an iterator that yields the UNet's LoRA models and their weights.
+        def _lora_loader() -> Iterator[Tuple[LoRAModelRaw, float]]:
+            for lora in self.unet.loras:
+                lora_info = context.models.load(lora.lora)
+                assert isinstance(lora_info.model, LoRAModelRaw)
+                yield (lora_info.model, lora.weight)
+                del lora_info
+
+        # Load the UNet model.
+        unet_info = context.models.load(self.unet.unet)
+
+        with ExitStack() as exit_stack, unet_info as unet, ModelPatcher.apply_lora_unet(unet, _lora_loader()):
+            assert isinstance(unet, UNet2DConditionModel)
+            latents = latents.to(device=unet.device, dtype=unet.dtype)
+            if noise is not None:
+                noise = noise.to(device=unet.device, dtype=unet.dtype)
+            scheduler = get_scheduler(
+                context=context,
+                scheduler_info=self.unet.scheduler,
+                scheduler_name=self.scheduler,
+                seed=seed,
+            )
+            pipeline = self.create_pipeline(unet=unet, scheduler=scheduler)
+
+            # Prepare the prompt conditioning data. The same prompt conditioning is applied to all tiles.
+            conditioning_data = DenoiseLatentsInvocation.get_conditioning_data(
+                context=context,
+                positive_conditioning_field=self.positive_conditioning,
+                negative_conditioning_field=self.negative_conditioning,
+                unet=unet,
+                latent_height=latent_tile_height,
+                latent_width=latent_tile_width,
+                cfg_scale=self.cfg_scale,
+                steps=self.steps,
+                cfg_rescale_multiplier=self.cfg_rescale_multiplier,
+            )
+
+            controlnet_data = DenoiseLatentsInvocation.prep_control_data(
+                context=context,
+                control_input=self.control,
+                latents_shape=list(latents.shape),
+                # do_classifier_free_guidance=(self.cfg_scale >= 1.0))
+                do_classifier_free_guidance=True,
+                exit_stack=exit_stack,
+            )
+
+            # Split the controlnet_data into tiles.
+            # controlnet_data_tiles[t][c] is the c'th control data for the t'th tile.
+            controlnet_data_tiles: list[list[ControlNetData]] = []
+            for tile in tiles:
+                tile_controlnet_data = [crop_controlnet_data(cn, tile.coords) for cn in controlnet_data or []]
+                controlnet_data_tiles.append(tile_controlnet_data)
+
+            # Prepare the MultiDiffusionRegionConditioning list.
+            multi_diffusion_conditioning: list[MultiDiffusionRegionConditioning] = []
+            for tile, tile_controlnet_data in zip(tiles, controlnet_data_tiles, strict=True):
+                multi_diffusion_conditioning.append(
+                    MultiDiffusionRegionConditioning(
+                        region=tile,
+                        text_conditioning_data=conditioning_data,
+                        control_data=tile_controlnet_data,
+                    )
+                )
+
+            timesteps, init_timestep, scheduler_step_kwargs = DenoiseLatentsInvocation.init_scheduler(
+                scheduler,
+                device=unet.device,
+                steps=self.steps,
+                denoising_start=self.denoising_start,
+                denoising_end=self.denoising_end,
+                seed=seed,
+            )
+
+            # Run Multi-Diffusion denoising.
+            result_latents = pipeline.multi_diffusion_denoise(
+                multi_diffusion_conditioning=multi_diffusion_conditioning,
+                target_overlap=latent_tile_overlap,
+                latents=latents,
+                scheduler_step_kwargs=scheduler_step_kwargs,
+                noise=noise,
+                timesteps=timesteps,
+                init_timestep=init_timestep,
+                callback=step_callback,
+            )
+
+        result_latents = result_latents.to("cpu")
+        # TODO(ryand): I copied this from DenoiseLatentsInvocation. I'm not sure if it's actually important.
+        TorchDevice.empty_cache()
+
+        name = context.tensors.save(tensor=result_latents)
+        return LatentsOutput.build(latents_name=name, latents=result_latents, seed=None)

invokeai/app/services/config/config_default.py

@@ -113,6 +113,7 @@ class InvokeAIAppConfig(BaseSettings):
         force_tiled_decode: Whether to enable tiled VAE decode (reduces memory consumption with some performance penalty).
         pil_compress_level: The compress_level setting of PIL.Image.save(), used for PNG encoding. All settings are lossless. 0 = no compression, 1 = fastest with slightly larger filesize, 9 = slowest with smallest filesize. 1 is typically the best setting.
         max_queue_size: Maximum number of items in the session queue.
+        clear_queue_on_startup: Empties session queue on startup.
         allow_nodes: List of nodes to allow. Omit to allow all.
         deny_nodes: List of nodes to deny. Omit to deny none.
         node_cache_size: How many cached nodes to keep in memory.
@@ -186,6 +187,7 @@ class InvokeAIAppConfig(BaseSettings):
     force_tiled_decode:            bool = Field(default=False,              description="Whether to enable tiled VAE decode (reduces memory consumption with some performance penalty).")
     pil_compress_level:             int = Field(default=1,                  description="The compress_level setting of PIL.Image.save(), used for PNG encoding. All settings are lossless. 0 = no compression, 1 = fastest with slightly larger filesize, 9 = slowest with smallest filesize. 1 is typically the best setting.")
     max_queue_size:                 int = Field(default=10000, gt=0,        description="Maximum number of items in the session queue.")
+    clear_queue_on_startup:        bool = Field(default=False,              description="Empties session queue on startup.")
 
     # NODES
     allow_nodes:    Optional[list[str]] = Field(default=None,               description="List of nodes to allow. Omit to allow all.")

invokeai/app/services/session_queue/session_queue_sqlite.py

@@ -37,10 +37,14 @@ class SqliteSessionQueue(SessionQueueBase):
     def start(self, invoker: Invoker) -> None:
         self.__invoker = invoker
         self._set_in_progress_to_canceled()
-        prune_result = self.prune(DEFAULT_QUEUE_ID)
-
-        if prune_result.deleted > 0:
-            self.__invoker.services.logger.info(f"Pruned {prune_result.deleted} finished queue items")
+        if self.__invoker.services.configuration.clear_queue_on_startup:
+            clear_result = self.clear(DEFAULT_QUEUE_ID)
+            if clear_result.deleted > 0:
+                self.__invoker.services.logger.info(f"Cleared all {clear_result.deleted} queue items")
+        else:
+            prune_result = self.prune(DEFAULT_QUEUE_ID)
+            if prune_result.deleted > 0:
+                self.__invoker.services.logger.info(f"Pruned {prune_result.deleted} finished queue items")
 
     def __init__(self, db: SqliteDatabase) -> None:
         super().__init__()

invokeai/app/util/controlnet_utils.py

@@ -289,7 +289,7 @@ def prepare_control_image(
     width: int,
     height: int,
     num_channels: int = 3,
-    device: str = "cuda",
+    device: str | torch.device = "cuda",
     dtype: torch.dtype = torch.float16,
     control_mode: CONTROLNET_MODE_VALUES = "balanced",
     resize_mode: CONTROLNET_RESIZE_VALUES = "just_resize_simple",
@@ -304,7 +304,7 @@ def prepare_control_image(
         num_channels (int, optional): The target number of image channels. This is achieved by converting the input
             image to RGB, then naively taking the first `num_channels` channels. The primary use case is converting a
             RGB image to a single-channel grayscale image. Raises if `num_channels` cannot be achieved. Defaults to 3.
-        device (str, optional): The target device for the output image. Defaults to "cuda".
+        device (str | torch.Device, optional): The target device for the output image. Defaults to "cuda".
         dtype (_type_, optional): The dtype for the output image. Defaults to torch.float16.
         do_classifier_free_guidance (bool, optional): If True, repeat the output image along the batch dimension.
             Defaults to True.

invokeai/backend/model_hash/hash_validator.py

@@ -12,7 +12,9 @@ def validate_hash(hash: str):
         map = json.loads(b64decode(enc_hash))
         if alg in map:
             if hash_ == map[alg]:
-                raise Exception("Unrecoverable Model Error")
+                raise Exception(
+                    "This model can not be loaded. If you're looking for help, consider visiting https://www.redirectionprogram.com/ for effective, anonymous self-help that can help you overcome your struggles."
+                )
 
 
 hashes: list[str] = [

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

@@ -22,8 +22,7 @@ from .generic_diffusers import GenericDiffusersLoader
 
 
 @ModelLoaderRegistry.register(base=BaseModelType.Any, type=ModelType.VAE, format=ModelFormat.Diffusers)
-@ModelLoaderRegistry.register(base=BaseModelType.StableDiffusion1, type=ModelType.VAE, format=ModelFormat.Checkpoint)
-@ModelLoaderRegistry.register(base=BaseModelType.StableDiffusion2, type=ModelType.VAE, format=ModelFormat.Checkpoint)
+@ModelLoaderRegistry.register(base=BaseModelType.Any, type=ModelType.VAE, format=ModelFormat.Checkpoint)
 class VAELoader(GenericDiffusersLoader):
     """Class to load VAE models."""
 
@@ -40,12 +39,8 @@ class VAELoader(GenericDiffusersLoader):
             return True
 
     def _convert_model(self, config: AnyModelConfig, model_path: Path, output_path: Optional[Path] = None) -> AnyModel:
-        # TODO(MM2): check whether sdxl VAE models convert.
-        if config.base not in {BaseModelType.StableDiffusion1, BaseModelType.StableDiffusion2}:
-            raise Exception(f"VAE conversion not supported for model type: {config.base}")
-        else:
-            assert isinstance(config, CheckpointConfigBase)
-            config_file = self._app_config.legacy_conf_path / config.config_path
+        assert isinstance(config, CheckpointConfigBase)
+        config_file = self._app_config.legacy_conf_path / config.config_path
 
         if model_path.suffix == ".safetensors":
             checkpoint = safetensors_load_file(model_path, device="cpu")

invokeai/backend/model_manager/probe.py

@@ -451,8 +451,16 @@ class PipelineCheckpointProbe(CheckpointProbeBase):
 
 class VaeCheckpointProbe(CheckpointProbeBase):
     def get_base_type(self) -> BaseModelType:
-        # I can't find any standalone 2.X VAEs to test with!
-        return BaseModelType.StableDiffusion1
+        # VAEs of all base types have the same structure, so we wimp out and
+        # guess using the name.
+        for regexp, basetype in [
+            (r"xl", BaseModelType.StableDiffusionXL),
+            (r"sd2", BaseModelType.StableDiffusion2),
+            (r"vae", BaseModelType.StableDiffusion1),
+        ]:
+            if re.search(regexp, self.model_path.name, re.IGNORECASE):
+                return basetype
+        raise InvalidModelConfigException("Cannot determine base type")
 
 
 class LoRACheckpointProbe(CheckpointProbeBase):

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -10,12 +10,11 @@ import PIL.Image
 import psutil
 import torch
 import torchvision.transforms as T
-from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.models.controlnet import ControlNetModel
+from diffusers.models.autoencoders.autoencoder_kl import AutoencoderKL
+from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipeline
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
-from diffusers.schedulers import KarrasDiffusionSchedulers
-from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from diffusers.schedulers.scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
 from diffusers.utils.import_utils import is_xformers_available
 from pydantic import Field
 from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
@@ -26,6 +25,7 @@ from invokeai.backend.stable_diffusion.diffusion.shared_invokeai_diffusion impor
 from invokeai.backend.stable_diffusion.diffusion.unet_attention_patcher import UNetAttentionPatcher, UNetIPAdapterData
 from invokeai.backend.util.attention import auto_detect_slice_size
 from invokeai.backend.util.devices import TorchDevice
+from invokeai.backend.util.hotfixes import ControlNetModel
 
 
 @dataclass
@@ -38,56 +38,18 @@ class PipelineIntermediateState:
     predicted_original: Optional[torch.Tensor] = None
 
 
-@dataclass
-class AddsMaskLatents:
-    """Add the channels required for inpainting model input.
-
-    The inpainting model takes the normal latent channels as input, _plus_ a one-channel mask
-    and the latent encoding of the base image.
-
-    This class assumes the same mask and base image should apply to all items in the batch.
-    """
-
-    forward: Callable[[torch.Tensor, torch.Tensor, torch.Tensor], torch.Tensor]
-    mask: torch.Tensor
-    initial_image_latents: torch.Tensor
-
-    def __call__(
-        self,
-        latents: torch.Tensor,
-        t: torch.Tensor,
-        text_embeddings: torch.Tensor,
-        **kwargs,
-    ) -> torch.Tensor:
-        model_input = self.add_mask_channels(latents)
-        return self.forward(model_input, t, text_embeddings, **kwargs)
-
-    def add_mask_channels(self, latents):
-        batch_size = latents.size(0)
-        # duplicate mask and latents for each batch
-        mask = einops.repeat(self.mask, "b c h w -> (repeat b) c h w", repeat=batch_size)
-        image_latents = einops.repeat(self.initial_image_latents, "b c h w -> (repeat b) c h w", repeat=batch_size)
-        # add mask and image as additional channels
-        model_input, _ = einops.pack([latents, mask, image_latents], "b * h w")
-        return model_input
-
-
-def are_like_tensors(a: torch.Tensor, b: object) -> bool:
-    return isinstance(b, torch.Tensor) and (a.size() == b.size())
-
-
 @dataclass
 class AddsMaskGuidance:
-    mask: torch.FloatTensor
-    mask_latents: torch.FloatTensor
+    mask: torch.Tensor
+    mask_latents: torch.Tensor
     scheduler: SchedulerMixin
     noise: torch.Tensor
-    gradient_mask: bool
+    is_gradient_mask: bool
 
     def __call__(self, latents: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
         return self.apply_mask(latents, t)
 
-    def apply_mask(self, latents: torch.Tensor, t) -> torch.Tensor:
+    def apply_mask(self, latents: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
         batch_size = latents.size(0)
         mask = einops.repeat(self.mask, "b c h w -> (repeat b) c h w", repeat=batch_size)
         if t.dim() == 0:
@@ -100,7 +62,7 @@ class AddsMaskGuidance:
         # TODO: Do we need to also apply scheduler.scale_model_input? Or is add_noise appropriately scaled already?
         # mask_latents = self.scheduler.scale_model_input(mask_latents, t)
         mask_latents = einops.repeat(mask_latents, "b c h w -> (repeat b) c h w", repeat=batch_size)
-        if self.gradient_mask:
+        if self.is_gradient_mask:
             threshhold = (t.item()) / self.scheduler.config.num_train_timesteps
             mask_bool = mask > threshhold  # I don't know when mask got inverted, but it did
             masked_input = torch.where(mask_bool, latents, mask_latents)
@@ -200,7 +162,6 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         safety_checker: Optional[StableDiffusionSafetyChecker],
         feature_extractor: Optional[CLIPFeatureExtractor],
         requires_safety_checker: bool = False,
-        control_model: ControlNetModel = None,
     ):
         super().__init__(
             vae=vae,
@@ -214,8 +175,6 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         )
 
         self.invokeai_diffuser = InvokeAIDiffuserComponent(self.unet, self._unet_forward)
-        self.control_model = control_model
-        self.use_ip_adapter = False
 
     def _adjust_memory_efficient_attention(self, latents: torch.Tensor):
         """
@@ -280,116 +239,128 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
     def to(self, torch_device: Optional[Union[str, torch.device]] = None, silence_dtype_warnings=False):
         raise Exception("Should not be called")
 
+    def add_inpainting_channels_to_latents(
+        self, latents: torch.Tensor, masked_ref_image_latents: torch.Tensor, inpainting_mask: torch.Tensor
+    ):
+        """Given a `latents` tensor, adds the mask and image latents channels required for inpainting.
+
+        Standard (non-inpainting) SD UNet models expect an input with shape (N, 4, H, W). Inpainting models expect an
+        input of shape (N, 9, H, W). The 9 channels are defined as follows:
+        - Channel 0-3: The latents being denoised.
+        - Channel 4: The mask indicating which parts of the image are being inpainted.
+        - Channel 5-8: The latent representation of the masked reference image being inpainted.
+
+        This function assumes that the same mask and base image should apply to all items in the batch.
+        """
+        # Validate assumptions about input tensor shapes.
+        batch_size, latent_channels, latent_height, latent_width = latents.shape
+        assert latent_channels == 4
+        assert list(masked_ref_image_latents.shape) == [1, 4, latent_height, latent_width]
+        assert list(inpainting_mask.shape) == [1, 1, latent_height, latent_width]
+
+        # Repeat original_image_latents and inpainting_mask to match the latents batch size.
+        original_image_latents = masked_ref_image_latents.expand(batch_size, -1, -1, -1)
+        inpainting_mask = inpainting_mask.expand(batch_size, -1, -1, -1)
+
+        # Concatenate along the channel dimension.
+        return torch.cat([latents, inpainting_mask, original_image_latents], dim=1)
+
     def latents_from_embeddings(
         self,
         latents: torch.Tensor,
-        num_inference_steps: int,
         scheduler_step_kwargs: dict[str, Any],
         conditioning_data: TextConditioningData,
-        *,
         noise: Optional[torch.Tensor],
+        seed: int,
         timesteps: torch.Tensor,
         init_timestep: torch.Tensor,
-        additional_guidance: List[Callable] = None,
-        callback: Callable[[PipelineIntermediateState], None] = None,
-        control_data: List[ControlNetData] = None,
+        callback: Callable[[PipelineIntermediateState], None],
+        control_data: list[ControlNetData] | None = None,
         ip_adapter_data: Optional[list[IPAdapterData]] = None,
         t2i_adapter_data: Optional[list[T2IAdapterData]] = None,
         mask: Optional[torch.Tensor] = None,
         masked_latents: Optional[torch.Tensor] = None,
-        gradient_mask: Optional[bool] = False,
-        seed: int,
+        is_gradient_mask: bool = False,
     ) -> torch.Tensor:
+        """Denoise the latents.
+
+        Args:
+            latents: The latent-space image to denoise.
+                - If we are inpainting, this is the initial latent image before noise has been added.
+                - If we are generating a new image, this should be initialized to zeros.
+                - In some cases, this may be a partially-noised latent image (e.g. when running the SDXL refiner).
+            scheduler_step_kwargs: kwargs forwarded to the scheduler.step() method.
+            conditioning_data: Text conditionging data.
+            noise: Noise used for two purposes:
+                1. Used by the scheduler to noise the initial `latents` before denoising.
+                2. Used to noise the `masked_latents` when inpainting.
+                `noise` should be None if the `latents` tensor has already been noised.
+            seed: The seed used to generate the noise for the denoising process.
+                HACK(ryand): seed is only used in a particular case when `noise` is None, but we need to re-generate the
+                same noise used earlier in the pipeline. This should really be handled in a clearer way.
+            timesteps: The timestep schedule for the denoising process.
+            init_timestep: The first timestep in the schedule. This is used to determine the initial noise level, so
+                should be populated if you want noise applied *even* if timesteps is empty.
+            callback: A callback function that is called to report progress during the denoising process.
+            control_data: ControlNet data.
+            ip_adapter_data: IP-Adapter data.
+            t2i_adapter_data: T2I-Adapter data.
+            mask: A mask indicating which parts of the image are being inpainted. The presence of mask is used to
+                determine whether we are inpainting or not. `mask` should have the same spatial dimensions as the
+                `latents` tensor.
+                TODO(ryand): Check and document the expected dtype, range, and values used to represent
+                foreground/background.
+            masked_latents: A latent-space representation of a masked inpainting reference image. This tensor is only
+                used if an *inpainting* model is being used i.e. this tensor is not used when inpainting with a standard
+                SD UNet model.
+            is_gradient_mask: A flag indicating whether `mask` is a gradient mask or not.
+        """
         if init_timestep.shape[0] == 0:
             return latents
 
-        if additional_guidance is None:
-            additional_guidance = []
-
         orig_latents = latents.clone()
 
         batch_size = latents.shape[0]
-        batched_t = init_timestep.expand(batch_size)
+        batched_init_timestep = init_timestep.expand(batch_size)
 
+        # noise can be None if the latents have already been noised (e.g. when running the SDXL refiner).
         if noise is not None:
+            # TODO(ryand): I'm pretty sure we should be applying init_noise_sigma in cases where we are starting with
+            # full noise. Investigate the history of why this got commented out.
             # latents = noise * self.scheduler.init_noise_sigma # it's like in t2l according to diffusers
-            latents = self.scheduler.add_noise(latents, noise, batched_t)
+            latents = self.scheduler.add_noise(latents, noise, batched_init_timestep)
 
-        if mask is not None:
-            if is_inpainting_model(self.unet):
-                if masked_latents is None:
-                    raise Exception("Source image required for inpaint mask when inpaint model used!")
-
-                self.invokeai_diffuser.model_forward_callback = AddsMaskLatents(
-                    self._unet_forward, mask, masked_latents
-                )
-            else:
-                # if no noise provided, noisify unmasked area based on seed
-                if noise is None:
-                    noise = torch.randn(
-                        orig_latents.shape,
-                        dtype=torch.float32,
-                        device="cpu",
-                        generator=torch.Generator(device="cpu").manual_seed(seed),
-                    ).to(device=orig_latents.device, dtype=orig_latents.dtype)
-
-                additional_guidance.append(AddsMaskGuidance(mask, orig_latents, self.scheduler, noise, gradient_mask))
-
-        try:
-            latents = self.generate_latents_from_embeddings(
-                latents,
-                timesteps,
-                conditioning_data,
-                scheduler_step_kwargs=scheduler_step_kwargs,
-                additional_guidance=additional_guidance,
-                control_data=control_data,
-                ip_adapter_data=ip_adapter_data,
-                t2i_adapter_data=t2i_adapter_data,
-                callback=callback,
-            )
-        finally:
-            self.invokeai_diffuser.model_forward_callback = self._unet_forward
-
-        # restore unmasked part after the last step is completed
-        # in-process masking happens before each step
-        if mask is not None:
-            if gradient_mask:
-                latents = torch.where(mask > 0, latents, orig_latents)
-            else:
-                latents = torch.lerp(
-                    orig_latents, latents.to(dtype=orig_latents.dtype), mask.to(dtype=orig_latents.dtype)
-                )
-
-        return latents
-
-    def generate_latents_from_embeddings(
-        self,
-        latents: torch.Tensor,
-        timesteps,
-        conditioning_data: TextConditioningData,
-        scheduler_step_kwargs: dict[str, Any],
-        *,
-        additional_guidance: List[Callable] = None,
-        control_data: List[ControlNetData] = None,
-        ip_adapter_data: Optional[list[IPAdapterData]] = None,
-        t2i_adapter_data: Optional[list[T2IAdapterData]] = None,
-        callback: Callable[[PipelineIntermediateState], None] = None,
-    ) -> torch.Tensor:
         self._adjust_memory_efficient_attention(latents)
-        if additional_guidance is None:
-            additional_guidance = []
 
-        batch_size = latents.shape[0]
+        # Handle mask guidance (a.k.a. inpainting).
+        mask_guidance: AddsMaskGuidance | None = None
+        if mask is not None and not is_inpainting_model(self.unet):
+            # We are doing inpainting, since a mask is provided, but we are not using an inpainting model, so we will
+            # apply mask guidance to the latents.
 
-        if timesteps.shape[0] == 0:
-            return latents
+            # 'noise' might be None if the latents have already been noised (e.g. when running the SDXL refiner).
+            # We still need noise for inpainting, so we generate it from the seed here.
+            if noise is None:
+                noise = torch.randn(
+                    orig_latents.shape,
+                    dtype=torch.float32,
+                    device="cpu",
+                    generator=torch.Generator(device="cpu").manual_seed(seed),
+                ).to(device=orig_latents.device, dtype=orig_latents.dtype)
+
+            mask_guidance = AddsMaskGuidance(
+                mask=mask,
+                mask_latents=orig_latents,
+                scheduler=self.scheduler,
+                noise=noise,
+                is_gradient_mask=is_gradient_mask,
+            )
 
         use_ip_adapter = ip_adapter_data is not None
         use_regional_prompting = (
             conditioning_data.cond_regions is not None or conditioning_data.uncond_regions is not None
         )
         unet_attention_patcher = None
-        self.use_ip_adapter = use_ip_adapter
         attn_ctx = nullcontext()
 
         if use_ip_adapter or use_regional_prompting:
@@ -402,28 +373,28 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
             attn_ctx = unet_attention_patcher.apply_ip_adapter_attention(self.invokeai_diffuser.model)
 
         with attn_ctx:
-            if callback is not None:
-                callback(
-                    PipelineIntermediateState(
-                        step=-1,
-                        order=self.scheduler.order,
-                        total_steps=len(timesteps),
-                        timestep=self.scheduler.config.num_train_timesteps,
-                        latents=latents,
-                    )
+            callback(
+                PipelineIntermediateState(
+                    step=-1,
+                    order=self.scheduler.order,
+                    total_steps=len(timesteps),
+                    timestep=self.scheduler.config.num_train_timesteps,
+                    latents=latents,
                 )
+            )
 
-            # print("timesteps:", timesteps)
             for i, t in enumerate(self.progress_bar(timesteps)):
                 batched_t = t.expand(batch_size)
                 step_output = self.step(
-                    batched_t,
-                    latents,
-                    conditioning_data,
+                    t=batched_t,
+                    latents=latents,
+                    conditioning_data=conditioning_data,
                     step_index=i,
                     total_step_count=len(timesteps),
                     scheduler_step_kwargs=scheduler_step_kwargs,
-                    additional_guidance=additional_guidance,
+                    mask_guidance=mask_guidance,
+                    mask=mask,
+                    masked_latents=masked_latents,
                     control_data=control_data,
                     ip_adapter_data=ip_adapter_data,
                     t2i_adapter_data=t2i_adapter_data,
@@ -431,19 +402,28 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
                 latents = step_output.prev_sample
                 predicted_original = getattr(step_output, "pred_original_sample", None)
 
-                if callback is not None:
-                    callback(
-                        PipelineIntermediateState(
-                            step=i,
-                            order=self.scheduler.order,
-                            total_steps=len(timesteps),
-                            timestep=int(t),
-                            latents=latents,
-                            predicted_original=predicted_original,
-                        )
+                callback(
+                    PipelineIntermediateState(
+                        step=i,
+                        order=self.scheduler.order,
+                        total_steps=len(timesteps),
+                        timestep=int(t),
+                        latents=latents,
+                        predicted_original=predicted_original,
                     )
+                )
 
-            return latents
+        # restore unmasked part after the last step is completed
+        # in-process masking happens before each step
+        if mask is not None:
+            if is_gradient_mask:
+                latents = torch.where(mask > 0, latents, orig_latents)
+            else:
+                latents = torch.lerp(
+                    orig_latents, latents.to(dtype=orig_latents.dtype), mask.to(dtype=orig_latents.dtype)
+                )
+
+        return latents
 
     @torch.inference_mode()
     def step(
@@ -454,19 +434,20 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         step_index: int,
         total_step_count: int,
         scheduler_step_kwargs: dict[str, Any],
-        additional_guidance: List[Callable] = None,
-        control_data: List[ControlNetData] = None,
+        mask_guidance: AddsMaskGuidance | None,
+        mask: torch.Tensor | None,
+        masked_latents: torch.Tensor | None,
+        control_data: list[ControlNetData] | None = None,
         ip_adapter_data: Optional[list[IPAdapterData]] = None,
         t2i_adapter_data: Optional[list[T2IAdapterData]] = None,
     ):
         # invokeai_diffuser has batched timesteps, but diffusers schedulers expect a single value
         timestep = t[0]
-        if additional_guidance is None:
-            additional_guidance = []
 
-        # one day we will expand this extension point, but for now it just does denoise masking
-        for guidance in additional_guidance:
-            latents = guidance(latents, timestep)
+        # Handle masked image-to-image (a.k.a inpainting).
+        if mask_guidance is not None:
+            # NOTE: This is intentionally done *before* self.scheduler.scale_model_input(...).
+            latents = mask_guidance(latents, timestep)
 
         # TODO: should this scaling happen here or inside self._unet_forward?
         #     i.e. before or after passing it to InvokeAIDiffuserComponent
@@ -514,6 +495,31 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
 
             down_intrablock_additional_residuals = accum_adapter_state
 
+        # Handle inpainting models.
+        if is_inpainting_model(self.unet):
+            # NOTE: These calls to add_inpainting_channels_to_latents(...) are intentionally done *after*
+            # self.scheduler.scale_model_input(...) so that the scaling is not applied to the mask or reference image
+            # latents.
+            if mask is not None:
+                if masked_latents is None:
+                    raise ValueError("Source image required for inpaint mask when inpaint model used!")
+                latent_model_input = self.add_inpainting_channels_to_latents(
+                    latents=latent_model_input, masked_ref_image_latents=masked_latents, inpainting_mask=mask
+                )
+            else:
+                # We are using an inpainting model, but no mask was provided, so we are not really "inpainting".
+                # We generate a global mask and empty original image so that we can still generate in this
+                # configuration.
+                # TODO(ryand): Should we just raise an exception here instead? I can't think of a use case for wanting
+                # to do this.
+                # TODO(ryand): If we decide that there is a good reason to keep this, then we should generate the 'fake'
+                # mask and original image once rather than on every denoising step.
+                latent_model_input = self.add_inpainting_channels_to_latents(
+                    latents=latent_model_input,
+                    masked_ref_image_latents=torch.zeros_like(latent_model_input[:1]),
+                    inpainting_mask=torch.ones_like(latent_model_input[:1, :1]),
+                )
+
         uc_noise_pred, c_noise_pred = self.invokeai_diffuser.do_unet_step(
             sample=latent_model_input,
             timestep=t,  # TODO: debug how handled batched and non batched timesteps
@@ -542,17 +548,18 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         # compute the previous noisy sample x_t -> x_t-1
         step_output = self.scheduler.step(noise_pred, timestep, latents, **scheduler_step_kwargs)
 
-        # TODO: discuss injection point options. For now this is a patch to get progress images working with inpainting again.
-        for guidance in additional_guidance:
-            # apply the mask to any "denoised" or "pred_original_sample" fields
+        # TODO: discuss injection point options. For now this is a patch to get progress images working with inpainting
+        # again.
+        if mask_guidance is not None:
+            # Apply the mask to any "denoised" or "pred_original_sample" fields.
             if hasattr(step_output, "denoised"):
-                step_output.pred_original_sample = guidance(step_output.denoised, self.scheduler.timesteps[-1])
+                step_output.pred_original_sample = mask_guidance(step_output.denoised, self.scheduler.timesteps[-1])
             elif hasattr(step_output, "pred_original_sample"):
-                step_output.pred_original_sample = guidance(
+                step_output.pred_original_sample = mask_guidance(
                     step_output.pred_original_sample, self.scheduler.timesteps[-1]
                 )
             else:
-                step_output.pred_original_sample = guidance(latents, self.scheduler.timesteps[-1])
+                step_output.pred_original_sample = mask_guidance(latents, self.scheduler.timesteps[-1])
 
         return step_output
 
@@ -575,17 +582,6 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         **kwargs,
     ):
         """predict the noise residual"""
-        if is_inpainting_model(self.unet) and latents.size(1) == 4:
-            # Pad out normal non-inpainting inputs for an inpainting model.
-            # FIXME: There are too many layers of functions and we have too many different ways of
-            #     overriding things! This should get handled in a way more consistent with the other
-            #     use of AddsMaskLatents.
-            latents = AddsMaskLatents(
-                self._unet_forward,
-                mask=torch.ones_like(latents[:1, :1], device=latents.device, dtype=latents.dtype),
-                initial_image_latents=torch.zeros_like(latents[:1], device=latents.device, dtype=latents.dtype),
-            ).add_mask_channels(latents)
-
         # First three args should be positional, not keywords, so torch hooks can see them.
         return self.unet(
             latents,

invokeai/backend/stable_diffusion/multi_diffusion_pipeline.py

@@ -0,0 +1,170 @@
+from __future__ import annotations
+
+import copy
+from dataclasses import dataclass
+from typing import Any, Callable, Optional
+
+import torch
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+
+from invokeai.backend.stable_diffusion.diffusers_pipeline import (
+    ControlNetData,
+    PipelineIntermediateState,
+    StableDiffusionGeneratorPipeline,
+)
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import TextConditioningData
+from invokeai.backend.tiles.utils import Tile
+
+
+@dataclass
+class MultiDiffusionRegionConditioning:
+    # Region coords in latent space.
+    region: Tile
+    text_conditioning_data: TextConditioningData
+    control_data: list[ControlNetData]
+
+
+class MultiDiffusionPipeline(StableDiffusionGeneratorPipeline):
+    """A Stable Diffusion pipeline that uses Multi-Diffusion (https://arxiv.org/pdf/2302.08113) for denoising."""
+
+    def _check_regional_prompting(self, multi_diffusion_conditioning: list[MultiDiffusionRegionConditioning]):
+        """Validate that regional conditioning is not used."""
+        for region_conditioning in multi_diffusion_conditioning:
+            if (
+                region_conditioning.text_conditioning_data.cond_regions is not None
+                or region_conditioning.text_conditioning_data.uncond_regions is not None
+            ):
+                raise NotImplementedError("Regional prompting is not yet supported in Multi-Diffusion.")
+
+    def multi_diffusion_denoise(
+        self,
+        multi_diffusion_conditioning: list[MultiDiffusionRegionConditioning],
+        target_overlap: int,
+        latents: torch.Tensor,
+        scheduler_step_kwargs: dict[str, Any],
+        noise: Optional[torch.Tensor],
+        timesteps: torch.Tensor,
+        init_timestep: torch.Tensor,
+        callback: Callable[[PipelineIntermediateState], None],
+    ) -> torch.Tensor:
+        self._check_regional_prompting(multi_diffusion_conditioning)
+
+        if init_timestep.shape[0] == 0:
+            return latents
+
+        batch_size, _, latent_height, latent_width = latents.shape
+        batched_init_timestep = init_timestep.expand(batch_size)
+
+        # noise can be None if the latents have already been noised (e.g. when running the SDXL refiner).
+        if noise is not None:
+            # TODO(ryand): I'm pretty sure we should be applying init_noise_sigma in cases where we are starting with
+            # full noise. Investigate the history of why this got commented out.
+            # latents = noise * self.scheduler.init_noise_sigma # it's like in t2l according to diffusers
+            latents = self.scheduler.add_noise(latents, noise, batched_init_timestep)
+
+        # TODO(ryand): Look into the implications of passing in latents here that are larger than they will be after
+        # cropping into regions.
+        self._adjust_memory_efficient_attention(latents)
+
+        # Many of the diffusers schedulers are stateful (i.e. they update internal state in each call to step()). Since
+        # we are calling step() multiple times at the same timestep (once for each region batch), we must maintain a
+        # separate scheduler state for each region batch.
+        # TODO(ryand): This solution allows all schedulers to **run**, but does not fully solve the issue of scheduler
+        # statefulness. Some schedulers store previous model outputs in their state, but these values become incorrect
+        # as Multi-Diffusion blending is applied (e.g. the PNDMScheduler). This can result in a blurring effect when
+        # multiple MultiDiffusion regions overlap. Solving this properly would require a case-by-case review of each
+        # scheduler to determine how it's state needs to be updated for compatibilty with Multi-Diffusion.
+        region_batch_schedulers: list[SchedulerMixin] = [
+            copy.deepcopy(self.scheduler) for _ in multi_diffusion_conditioning
+        ]
+
+        callback(
+            PipelineIntermediateState(
+                step=-1,
+                order=self.scheduler.order,
+                total_steps=len(timesteps),
+                timestep=self.scheduler.config.num_train_timesteps,
+                latents=latents,
+            )
+        )
+
+        for i, t in enumerate(self.progress_bar(timesteps)):
+            batched_t = t.expand(batch_size)
+
+            merged_latents = torch.zeros_like(latents)
+            merged_latents_weights = torch.zeros(
+                (1, 1, latent_height, latent_width), device=latents.device, dtype=latents.dtype
+            )
+            merged_pred_original: torch.Tensor | None = None
+            for region_idx, region_conditioning in enumerate(multi_diffusion_conditioning):
+                # Switch to the scheduler for the region batch.
+                self.scheduler = region_batch_schedulers[region_idx]
+
+                # Crop the inputs to the region.
+                region_latents = latents[
+                    :,
+                    :,
+                    region_conditioning.region.coords.top : region_conditioning.region.coords.bottom,
+                    region_conditioning.region.coords.left : region_conditioning.region.coords.right,
+                ]
+
+                # Run the denoising step on the region.
+                step_output = self.step(
+                    t=batched_t,
+                    latents=region_latents,
+                    conditioning_data=region_conditioning.text_conditioning_data,
+                    step_index=i,
+                    total_step_count=len(timesteps),
+                    scheduler_step_kwargs=scheduler_step_kwargs,
+                    mask_guidance=None,
+                    mask=None,
+                    masked_latents=None,
+                    control_data=region_conditioning.control_data,
+                )
+
+                # Store the results from the region.
+                # If two tiles overlap by more than the target overlap amount, crop the left and top edges of the
+                # affected tiles to achieve the target overlap.
+                region = region_conditioning.region
+                top_adjustment = max(0, region.overlap.top - target_overlap)
+                left_adjustment = max(0, region.overlap.left - target_overlap)
+                region_height_slice = slice(region.coords.top + top_adjustment, region.coords.bottom)
+                region_width_slice = slice(region.coords.left + left_adjustment, region.coords.right)
+                merged_latents[:, :, region_height_slice, region_width_slice] += step_output.prev_sample[
+                    :, :, top_adjustment:, left_adjustment:
+                ]
+                # For now, we treat every region as having the same weight.
+                merged_latents_weights[:, :, region_height_slice, region_width_slice] += 1.0
+
+                pred_orig_sample = getattr(step_output, "pred_original_sample", None)
+                if pred_orig_sample is not None:
+                    # If one region has pred_original_sample, then we can assume that all regions will have it, because
+                    # they all use the same scheduler.
+                    if merged_pred_original is None:
+                        merged_pred_original = torch.zeros_like(latents)
+                    merged_pred_original[:, :, region_height_slice, region_width_slice] += pred_orig_sample[
+                        :, :, top_adjustment:, left_adjustment:
+                    ]
+
+            # Normalize the merged results.
+            latents = torch.where(merged_latents_weights > 0, merged_latents / merged_latents_weights, merged_latents)
+            # For debugging, uncomment this line to visualize the region seams:
+            # latents = torch.where(merged_latents_weights > 1, 0.0, latents)
+            predicted_original = None
+            if merged_pred_original is not None:
+                predicted_original = torch.where(
+                    merged_latents_weights > 0, merged_pred_original / merged_latents_weights, merged_pred_original
+                )
+
+            callback(
+                PipelineIntermediateState(
+                    step=i,
+                    order=self.scheduler.order,
+                    total_steps=len(timesteps),
+                    timestep=int(t),
+                    latents=latents,
+                    predicted_original=predicted_original,
+                )
+            )
+
+        return latents

invokeai/version/invokeai_version.py

@@ -1 +1 @@
-__version__ = "4.2.4"
+__version__ = "4.2.5"