Modular backend - inpaint

invokeai/app/invocations/denoise_latents.py

@@ -37,7 +37,7 @@
 from invokeai.app.util.controlnet_utils import prepare_control_image
 from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
 from invokeai.backend.lora import LoRAModelRaw
-from invokeai.backend.model_manager import BaseModelType
+from invokeai.backend.model_manager import BaseModelType, ModelVariantType
 from invokeai.backend.model_patcher import ModelPatcher
 from invokeai.backend.stable_diffusion import PipelineIntermediateState, set_seamless
 from invokeai.backend.stable_diffusion.denoise_context import DenoiseContext, DenoiseInputs
@@ -58,6 +58,8 @@
 from invokeai.backend.stable_diffusion.diffusion.custom_atttention import CustomAttnProcessor2_0
 from invokeai.backend.stable_diffusion.diffusion_backend import StableDiffusionBackend
 from invokeai.backend.stable_diffusion.extension_callback_type import ExtensionCallbackType
+from invokeai.backend.stable_diffusion.extensions.inpaint import InpaintExt
+from invokeai.backend.stable_diffusion.extensions.inpaint_model import InpaintModelExt
 from invokeai.backend.stable_diffusion.extensions.preview import PreviewExt
 from invokeai.backend.stable_diffusion.extensions_manager import ExtensionsManager
 from invokeai.backend.stable_diffusion.schedulers import SCHEDULER_MAP
@@ -790,6 +792,13 @@ def step_callback(state: PipelineIntermediateState) -> None:
 
         ext_manager.add_extension(PreviewExt(step_callback))
 
+        ### inpaint
+        mask, masked_latents, is_gradient_mask = self.prep_inpaint_mask(context, latents)
+        if unet_config.variant == ModelVariantType.Inpaint:
+            ext_manager.add_extension(InpaintModelExt(mask, masked_latents, is_gradient_mask))
+        elif mask is not None:
+            ext_manager.add_extension(InpaintExt(mask, is_gradient_mask))
+
         # ext: t2i/ip adapter
         ext_manager.run_callback(ExtensionCallbackType.SETUP, denoise_ctx)
 

invokeai/backend/stable_diffusion/extensions/inpaint.py

@@ -0,0 +1,92 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Optional
+
+import einops
+import torch
+from diffusers import UNet2DConditionModel
+
+from invokeai.backend.stable_diffusion.extension_callback_type import ExtensionCallbackType
+from invokeai.backend.stable_diffusion.extensions.base import ExtensionBase, callback
+
+if TYPE_CHECKING:
+    from invokeai.backend.stable_diffusion.denoise_context import DenoiseContext
+
+
+class InpaintExt(ExtensionBase):
+    def __init__(
+        self,
+        mask: torch.Tensor,
+        is_gradient_mask: bool,
+    ):
+        super().__init__()
+        self._mask = mask
+        self._is_gradient_mask = is_gradient_mask
+        self._noise: Optional[torch.Tensor] = None
+
+    @staticmethod
+    def _is_normal_model(unet: UNet2DConditionModel):
+        return unet.conv_in.in_channels == 4
+
+    def _apply_mask(self, ctx: DenoiseContext, 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:
+            # some schedulers expect t to be one-dimensional.
+            # TODO: file diffusers bug about inconsistency?
+            t = einops.repeat(t, "-> batch", batch=batch_size)
+        # Noise shouldn't be re-randomized between steps here. The multistep schedulers
+        # get very confused about what is happening from step to step when we do that.
+        mask_latents = ctx.scheduler.add_noise(ctx.inputs.orig_latents, self._noise, t)
+        # TODO: Do we need to also apply scheduler.scale_model_input? Or is add_noise appropriately scaled already?
+        # mask_latents = self.scheduler.scale_model_input(mask_latents, t)
+        mask_latents = einops.repeat(mask_latents, "b c h w -> (repeat b) c h w", repeat=batch_size)
+        if self._is_gradient_mask:
+            threshhold = (t.item()) / ctx.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)
+        else:
+            masked_input = torch.lerp(mask_latents.to(dtype=latents.dtype), latents, mask.to(dtype=latents.dtype))
+        return masked_input
+
+    @callback(ExtensionCallbackType.PRE_DENOISE_LOOP)
+    def init_tensors(self, ctx: DenoiseContext):
+        if not self._is_normal_model(ctx.unet):
+            raise Exception("InpaintExt should be used only on normal models!")
+
+        self._mask = self._mask.to(device=ctx.latents.device, dtype=ctx.latents.dtype)
+
+        self._noise = ctx.inputs.noise
+        if self._noise is None:
+            self._noise = torch.randn(
+                ctx.latents.shape,
+                dtype=torch.float32,
+                device="cpu",
+                generator=torch.Generator(device="cpu").manual_seed(ctx.seed),
+            ).to(device=ctx.latents.device, dtype=ctx.latents.dtype)
+
+    # TODO: order value
+    @callback(ExtensionCallbackType.PRE_STEP, order=-100)
+    def apply_mask_to_initial_latents(self, ctx: DenoiseContext):
+        ctx.latents = self._apply_mask(ctx, ctx.latents, ctx.timestep)
+
+    # TODO: order value
+    # TODO: redo this with preview events rewrite
+    @callback(ExtensionCallbackType.POST_STEP, order=-100)
+    def apply_mask_to_step_output(self, ctx: DenoiseContext):
+        timestep = ctx.scheduler.timesteps[-1]
+        if hasattr(ctx.step_output, "denoised"):
+            ctx.step_output.denoised = self._apply_mask(ctx, ctx.step_output.denoised, timestep)
+        elif hasattr(ctx.step_output, "pred_original_sample"):
+            ctx.step_output.pred_original_sample = self._apply_mask(ctx, ctx.step_output.pred_original_sample, timestep)
+        else:
+            ctx.step_output.pred_original_sample = self._apply_mask(ctx, ctx.step_output.prev_sample, timestep)
+
+    # TODO: should here be used order?
+    # restore unmasked part after the last step is completed
+    @callback(ExtensionCallbackType.POST_DENOISE_LOOP)
+    def restore_unmasked(self, ctx: DenoiseContext):
+        if self._is_gradient_mask:
+            ctx.latents = torch.where(self._mask > 0, ctx.latents, ctx.inputs.orig_latents)
+        else:
+            ctx.latents = torch.lerp(ctx.inputs.orig_latents, ctx.latents, self._mask)

invokeai/backend/stable_diffusion/extensions/inpaint_model.py

@@ -0,0 +1,66 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Optional
+
+import torch
+from diffusers import UNet2DConditionModel
+
+from invokeai.backend.stable_diffusion.extension_callback_type import ExtensionCallbackType
+from invokeai.backend.stable_diffusion.extensions.base import ExtensionBase, callback
+
+if TYPE_CHECKING:
+    from invokeai.backend.stable_diffusion.denoise_context import DenoiseContext
+
+
+class InpaintModelExt(ExtensionBase):
+    def __init__(
+        self,
+        mask: Optional[torch.Tensor],
+        masked_latents: Optional[torch.Tensor],
+        is_gradient_mask: bool,
+    ):
+        super().__init__()
+        if mask is not None and masked_latents is None:
+            raise ValueError("Source image required for inpaint mask when inpaint model used!")
+
+        self._mask = mask
+        self._masked_latents = masked_latents
+        self._is_gradient_mask = is_gradient_mask
+
+    @staticmethod
+    def _is_inpaint_model(unet: UNet2DConditionModel):
+        return unet.conv_in.in_channels == 9
+
+    @callback(ExtensionCallbackType.PRE_DENOISE_LOOP)
+    def init_tensors(self, ctx: DenoiseContext):
+        if not self._is_inpaint_model(ctx.unet):
+            raise Exception("InpaintModelExt should be used only on inpaint models!")
+
+        if self._mask is None:
+            self._mask = torch.ones_like(ctx.latents[:1, :1])
+        self._mask = self._mask.to(device=ctx.latents.device, dtype=ctx.latents.dtype)
+
+        if self._masked_latents is None:
+            self._masked_latents = torch.zeros_like(ctx.latents[:1])
+        self._masked_latents = self._masked_latents.to(device=ctx.latents.device, dtype=ctx.latents.dtype)
+
+    # TODO: any ideas about order value?
+    # do last so that other extensions works with normal latents
+    @callback(ExtensionCallbackType.PRE_UNET, order=1000)
+    def append_inpaint_layers(self, ctx: DenoiseContext):
+        batch_size = ctx.unet_kwargs.sample.shape[0]
+        b_mask = torch.cat([self._mask] * batch_size)
+        b_masked_latents = torch.cat([self._masked_latents] * batch_size)
+        ctx.unet_kwargs.sample = torch.cat(
+            [ctx.unet_kwargs.sample, b_mask, b_masked_latents],
+            dim=1,
+        )
+
+    # TODO: should here be used order?
+    # restore unmasked part as inpaint model can change unmasked part slightly
+    @callback(ExtensionCallbackType.POST_DENOISE_LOOP)
+    def restore_unmasked(self, ctx: DenoiseContext):
+        if self._is_gradient_mask:
+            ctx.latents = torch.where(self._mask > 0, ctx.latents, ctx.inputs.orig_latents)
+        else:
+            ctx.latents = torch.lerp(ctx.inputs.orig_latents, ctx.latents, self._mask)