Add mask dilation.

clothing_workflow.ipynb

@@ -0,0 +1,158 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "aeb428d0-0817-462c-b5d8-455a0615d305",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "from PIL import Image\n",
+    "import numpy as np\n",
+    "import cv2\n",
+    "\n",
+    "from invokeai.backend.vto_workflow.overlay_pattern import generate_dress_mask, multiply_images\n",
+    "from invokeai.backend.vto_workflow.extract_channel import extract_channel, ImageChannel\n",
+    "from invokeai.backend.vto_workflow.seamless_mapping import map_seamless_tiles\n",
+    "\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6140d4b7-8238-431c-848e-6f6ae27652f5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    " # Load the model image.\n",
+    "model_image = Image.open(\"/home/ryan/src/InvokeAI/invokeai/backend/vto_workflow/dress.jpeg\")\n",
+    "\n",
+    "# Load the pattern image.\n",
+    "pattern_image = Image.open(\"/home/ryan/src/InvokeAI/invokeai/backend/vto_workflow/pattern1.jpg\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fb7186ba-dc0c-4520-ac30-49073a65601a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "mask = generate_dress_mask(model_image)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "9b935de4-94c5-4be5-bf8e-a5a6e445c811",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Visualize mask\n",
+    "model_image_np = np.array(model_image)\n",
+    "masked_model_image = (model_image_np * np.expand_dims(mask, -1).astype(np.float32)).astype(np.uint8)\n",
+    "mask_image = Image.fromarray(masked_model_image)\n",
+    "mask_image"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e51bb545",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "shadows = extract_channel(np.array(model_image), ImageChannel.LAB_L)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ec43de4a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Visualize masked shadows\n",
+    "masked_shadows = (shadows * mask).astype(np.uint8)\n",
+    "masked_shadows_image = Image.fromarray(masked_shadows)\n",
+    "masked_shadows_image"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "dbb53794",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Tile the pattern.\n",
+    "expanded_pattern = map_seamless_tiles(seamless_tile=pattern_image, target_hw=(model_image.height, model_image.width), num_repeats_h=10.0)\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f4f22d02",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Multiply the pattern by the shadows.\n",
+    "pattern_with_shadows = multiply_images(expanded_pattern, shadows)\n",
+    "pattern_with_shadows"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "97db42b0",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "de32f7e3",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Merge the pattern with the model image.\n",
+    "pattern_with_shadows_np = np.array(pattern_with_shadows)\n",
+    "merged_image = np.where(mask[:, :, None],  pattern_with_shadows_np,model_image_np)\n",
+    "merged_image = Image.fromarray(merged_image)\n",
+    "merged_image"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ff1d4044",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.12"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

docker/Dockerfile

@@ -55,7 +55,6 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 FROM node:20-slim AS web-builder
 ENV PNPM_HOME="/pnpm"
 ENV PATH="$PNPM_HOME:$PATH"
-RUN corepack use pnpm@8.x
 RUN corepack enable
 
 WORKDIR /build

invokeai/app/invocations/compel.py

@@ -80,12 +80,12 @@ class CompelInvocation(BaseInvocation):
 
         with (
             # apply all patches while the model is on the target device
-            text_encoder_info.model_on_device() as (cached_weights, text_encoder),
+            text_encoder_info.model_on_device() as (model_state_dict, text_encoder),
             tokenizer_info as tokenizer,
             ModelPatcher.apply_lora_text_encoder(
                 text_encoder,
                 loras=_lora_loader(),
-                cached_weights=cached_weights,
+                model_state_dict=model_state_dict,
             ),
             # Apply CLIP Skip after LoRA to prevent LoRA application from failing on skipped layers.
             ModelPatcher.apply_clip_skip(text_encoder, self.clip.skipped_layers),
@@ -175,13 +175,13 @@ class SDXLPromptInvocationBase:
 
         with (
             # apply all patches while the model is on the target device
-            text_encoder_info.model_on_device() as (cached_weights, text_encoder),
+            text_encoder_info.model_on_device() as (state_dict, text_encoder),
             tokenizer_info as tokenizer,
             ModelPatcher.apply_lora(
                 text_encoder,
                 loras=_lora_loader(),
                 prefix=lora_prefix,
-                cached_weights=cached_weights,
+                model_state_dict=state_dict,
             ),
             # Apply CLIP Skip after LoRA to prevent LoRA application from failing on skipped layers.
             ModelPatcher.apply_clip_skip(text_encoder, clip_field.skipped_layers),

invokeai/app/invocations/create_gradient_mask.py

@@ -39,7 +39,7 @@ class GradientMaskOutput(BaseInvocationOutput):
     title="Create Gradient Mask",
     tags=["mask", "denoise"],
     category="latents",
-    version="1.2.0",
+    version="1.1.0",
 )
 class CreateGradientMaskInvocation(BaseInvocation):
     """Creates mask for denoising model run."""
@@ -93,7 +93,6 @@ class CreateGradientMaskInvocation(BaseInvocation):
 
             # redistribute blur so that the original edges are 0 and blur outwards to 1
             blur_tensor = (blur_tensor - 0.5) * 2
-            blur_tensor[blur_tensor < 0] = 0.0
 
             threshold = 1 - self.minimum_denoise
 

invokeai/app/invocations/denoise_latents.py

@@ -37,9 +37,9 @@ from invokeai.app.services.shared.invocation_context import InvocationContext
 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, ModelVariantType
+from invokeai.backend.model_manager import BaseModelType
 from invokeai.backend.model_patcher import ModelPatcher
-from invokeai.backend.stable_diffusion import PipelineIntermediateState
+from invokeai.backend.stable_diffusion import PipelineIntermediateState, set_seamless
 from invokeai.backend.stable_diffusion.denoise_context import DenoiseContext, DenoiseInputs
 from invokeai.backend.stable_diffusion.diffusers_pipeline import (
     ControlNetData,
@@ -60,13 +60,8 @@ from invokeai.backend.stable_diffusion.diffusion_backend import StableDiffusionB
 from invokeai.backend.stable_diffusion.extension_callback_type import ExtensionCallbackType
 from invokeai.backend.stable_diffusion.extensions.controlnet import ControlNetExt
 from invokeai.backend.stable_diffusion.extensions.freeu import FreeUExt
-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.lora import LoRAExt
 from invokeai.backend.stable_diffusion.extensions.preview import PreviewExt
 from invokeai.backend.stable_diffusion.extensions.rescale_cfg import RescaleCFGExt
-from invokeai.backend.stable_diffusion.extensions.seamless import SeamlessExt
-from invokeai.backend.stable_diffusion.extensions.t2i_adapter import T2IAdapterExt
 from invokeai.backend.stable_diffusion.extensions_manager import ExtensionsManager
 from invokeai.backend.stable_diffusion.schedulers import SCHEDULER_MAP
 from invokeai.backend.stable_diffusion.schedulers.schedulers import SCHEDULER_NAME_VALUES
@@ -503,33 +498,6 @@ class DenoiseLatentsInvocation(BaseInvocation):
                 )
             )
 
-    @staticmethod
-    def parse_t2i_adapter_field(
-        exit_stack: ExitStack,
-        context: InvocationContext,
-        t2i_adapters: Optional[Union[T2IAdapterField, list[T2IAdapterField]]],
-        ext_manager: ExtensionsManager,
-    ) -> None:
-        if t2i_adapters is None:
-            return
-
-        # Handle the possibility that t2i_adapters could be a list or a single T2IAdapterField.
-        if isinstance(t2i_adapters, T2IAdapterField):
-            t2i_adapters = [t2i_adapters]
-
-        for t2i_adapter_field in t2i_adapters:
-            ext_manager.add_extension(
-                T2IAdapterExt(
-                    node_context=context,
-                    model_id=t2i_adapter_field.t2i_adapter_model,
-                    image=context.images.get_pil(t2i_adapter_field.image.image_name),
-                    weight=t2i_adapter_field.weight,
-                    begin_step_percent=t2i_adapter_field.begin_step_percent,
-                    end_step_percent=t2i_adapter_field.end_step_percent,
-                    resize_mode=t2i_adapter_field.resize_mode,
-                )
-            )
-
     def prep_ip_adapter_image_prompts(
         self,
         context: InvocationContext,
@@ -739,7 +707,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
         else:
             masked_latents = torch.where(mask < 0.5, 0.0, latents)
 
-        return mask, masked_latents, self.denoise_mask.gradient
+        return 1 - mask, masked_latents, self.denoise_mask.gradient
 
     @staticmethod
     def prepare_noise_and_latents(
@@ -797,6 +765,10 @@ class DenoiseLatentsInvocation(BaseInvocation):
         dtype = TorchDevice.choose_torch_dtype()
 
         seed, noise, latents = self.prepare_noise_and_latents(context, self.noise, self.latents)
+        latents = latents.to(device=device, dtype=dtype)
+        if noise is not None:
+            noise = noise.to(device=device, dtype=dtype)
+
         _, _, latent_height, latent_width = latents.shape
 
         conditioning_data = self.get_conditioning_data(
@@ -829,6 +801,21 @@ class DenoiseLatentsInvocation(BaseInvocation):
             denoising_end=self.denoising_end,
         )
 
+        denoise_ctx = DenoiseContext(
+            inputs=DenoiseInputs(
+                orig_latents=latents,
+                timesteps=timesteps,
+                init_timestep=init_timestep,
+                noise=noise,
+                seed=seed,
+                scheduler_step_kwargs=scheduler_step_kwargs,
+                conditioning_data=conditioning_data,
+                attention_processor_cls=CustomAttnProcessor2_0,
+            ),
+            unet=None,
+            scheduler=scheduler,
+        )
+
         # 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)
 
@@ -846,50 +833,6 @@ class DenoiseLatentsInvocation(BaseInvocation):
         if self.unet.freeu_config:
             ext_manager.add_extension(FreeUExt(self.unet.freeu_config))
 
-        ### lora
-        if self.unet.loras:
-            for lora_field in self.unet.loras:
-                ext_manager.add_extension(
-                    LoRAExt(
-                        node_context=context,
-                        model_id=lora_field.lora,
-                        weight=lora_field.weight,
-                    )
-                )
-        ### seamless
-        if self.unet.seamless_axes:
-            ext_manager.add_extension(SeamlessExt(self.unet.seamless_axes))
-
-        ### inpaint
-        mask, masked_latents, is_gradient_mask = self.prep_inpaint_mask(context, latents)
-        # NOTE: We used to identify inpainting models by inpecting the shape of the loaded UNet model weights. Now we
-        # use the ModelVariantType config. During testing, there was a report of a user with models that had an
-        # incorrect ModelVariantType value. Re-installing the model fixed the issue. If this issue turns out to be
-        # prevalent, we will have to revisit how we initialize the inpainting extensions.
-        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))
-
-        # Initialize context for modular denoise
-        latents = latents.to(device=device, dtype=dtype)
-        if noise is not None:
-            noise = noise.to(device=device, dtype=dtype)
-        denoise_ctx = DenoiseContext(
-            inputs=DenoiseInputs(
-                orig_latents=latents,
-                timesteps=timesteps,
-                init_timestep=init_timestep,
-                noise=noise,
-                seed=seed,
-                scheduler_step_kwargs=scheduler_step_kwargs,
-                conditioning_data=conditioning_data,
-                attention_processor_cls=CustomAttnProcessor2_0,
-            ),
-            unet=None,
-            scheduler=scheduler,
-        )
-
         # context for loading additional models
         with ExitStack() as exit_stack:
             # later should be smth like:
@@ -897,7 +840,6 @@ class DenoiseLatentsInvocation(BaseInvocation):
             #    ext = extension_field.to_extension(exit_stack, context, ext_manager)
             #    ext_manager.add_extension(ext)
             self.parse_controlnet_field(exit_stack, context, self.control, ext_manager)
-            self.parse_t2i_adapter_field(exit_stack, context, self.t2i_adapter, ext_manager)
 
             # ext: t2i/ip adapter
             ext_manager.run_callback(ExtensionCallbackType.SETUP, denoise_ctx)
@@ -929,10 +871,6 @@ class DenoiseLatentsInvocation(BaseInvocation):
         seed, noise, latents = self.prepare_noise_and_latents(context, self.noise, self.latents)
 
         mask, masked_latents, gradient_mask = self.prep_inpaint_mask(context, latents)
-        # At this point, the mask ranges from 0 (leave unchanged) to 1 (inpaint).
-        # We invert the mask here for compatibility with the old backend implementation.
-        if mask is not None:
-            mask = 1 - mask
 
         # TODO(ryand): I have hard-coded `do_classifier_free_guidance=True` to mirror the behaviour of ControlNets,
         # below. Investigate whether this is appropriate.
@@ -975,14 +913,14 @@ class DenoiseLatentsInvocation(BaseInvocation):
         assert isinstance(unet_info.model, UNet2DConditionModel)
         with (
             ExitStack() as exit_stack,
-            unet_info.model_on_device() as (cached_weights, unet),
+            unet_info.model_on_device() as (model_state_dict, unet),
             ModelPatcher.apply_freeu(unet, self.unet.freeu_config),
-            SeamlessExt.static_patch_model(unet, self.unet.seamless_axes),  # FIXME
+            set_seamless(unet, self.unet.seamless_axes),  # FIXME
             # Apply the LoRA after unet has been moved to its target device for faster patching.
             ModelPatcher.apply_lora_unet(
                 unet,
                 loras=_lora_loader(),
-                cached_weights=cached_weights,
+                model_state_dict=model_state_dict,
             ),
         ):
             assert isinstance(unet, UNet2DConditionModel)

invokeai/app/invocations/fields.py

@@ -1,7 +1,7 @@
 from enum import Enum
 from typing import Any, Callable, Optional, Tuple
 
-from pydantic import BaseModel, ConfigDict, Field, RootModel, TypeAdapter, model_validator
+from pydantic import BaseModel, ConfigDict, Field, RootModel, TypeAdapter
 from pydantic.fields import _Unset
 from pydantic_core import PydanticUndefined
 
@@ -242,31 +242,6 @@ class ConditioningField(BaseModel):
     )
 
 
-class BoundingBoxField(BaseModel):
-    """A bounding box primitive value."""
-
-    x_min: int = Field(ge=0, description="The minimum x-coordinate of the bounding box (inclusive).")
-    x_max: int = Field(ge=0, description="The maximum x-coordinate of the bounding box (exclusive).")
-    y_min: int = Field(ge=0, description="The minimum y-coordinate of the bounding box (inclusive).")
-    y_max: int = Field(ge=0, description="The maximum y-coordinate of the bounding box (exclusive).")
-
-    score: Optional[float] = Field(
-        default=None,
-        ge=0.0,
-        le=1.0,
-        description="The score associated with the bounding box. In the range [0, 1]. This value is typically set "
-        "when the bounding box was produced by a detector and has an associated confidence score.",
-    )
-
-    @model_validator(mode="after")
-    def check_coords(self):
-        if self.x_min > self.x_max:
-            raise ValueError(f"x_min ({self.x_min}) is greater than x_max ({self.x_max}).")
-        if self.y_min > self.y_max:
-            raise ValueError(f"y_min ({self.y_min}) is greater than y_max ({self.y_max}).")
-        return self
-
-
 class MetadataField(RootModel[dict[str, Any]]):
     """
     Pydantic model for metadata with custom root of type dict[str, Any].

invokeai/app/invocations/flux_text_to_image.py

@@ -1,278 +0,0 @@
-from pathlib import Path
-from typing import Literal
-
-import torch
-from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler
-from diffusers.models.transformers.transformer_flux import FluxTransformer2DModel
-from diffusers.pipelines.flux.pipeline_flux import FluxPipeline
-from optimum.quanto import qfloat8
-from PIL import Image
-from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
-from transformers.models.auto import AutoModelForTextEncoding
-
-from invokeai.app.invocations.baseinvocation import BaseInvocation, invocation
-from invokeai.app.invocations.fields import InputField, WithBoard, WithMetadata
-from invokeai.app.invocations.primitives import ImageOutput
-from invokeai.app.services.shared.invocation_context import InvocationContext
-from invokeai.backend.quantization.fast_quantized_diffusion_model import FastQuantizedDiffusersModel
-from invokeai.backend.quantization.fast_quantized_transformers_model import FastQuantizedTransformersModel
-from invokeai.backend.util.devices import TorchDevice
-
-TFluxModelKeys = Literal["flux-schnell"]
-FLUX_MODELS: dict[TFluxModelKeys, str] = {"flux-schnell": "black-forest-labs/FLUX.1-schnell"}
-
-
-class QuantizedFluxTransformer2DModel(FastQuantizedDiffusersModel):
-    base_class = FluxTransformer2DModel
-
-
-class QuantizedModelForTextEncoding(FastQuantizedTransformersModel):
-    auto_class = AutoModelForTextEncoding
-
-
-@invocation(
-    "flux_text_to_image",
-    title="FLUX Text to Image",
-    tags=["image"],
-    category="image",
-    version="1.0.0",
-)
-class FluxTextToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
-    """Text-to-image generation using a FLUX model."""
-
-    model: TFluxModelKeys = InputField(description="The FLUX model to use for text-to-image generation.")
-    use_8bit: bool = InputField(
-        default=False, description="Whether to quantize the transformer model to 8-bit precision."
-    )
-    positive_prompt: str = InputField(description="Positive prompt for text-to-image generation.")
-    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.")
-    guidance: float = InputField(
-        default=4.0,
-        description="The guidance strength. Higher values adhere more strictly to the prompt, and will produce less diverse images.",
-    )
-    seed: int = InputField(default=0, description="Randomness seed for reproducibility.")
-
-    @torch.no_grad()
-    def invoke(self, context: InvocationContext) -> ImageOutput:
-        model_path = context.models.download_and_cache_model(FLUX_MODELS[self.model])
-
-        t5_embeddings, clip_embeddings = self._encode_prompt(context, model_path)
-        latents = self._run_diffusion(context, model_path, clip_embeddings, t5_embeddings)
-        image = self._run_vae_decoding(context, model_path, latents)
-        image_dto = context.images.save(image=image)
-        return ImageOutput.build(image_dto)
-
-    def _encode_prompt(self, context: InvocationContext, flux_model_dir: Path) -> tuple[torch.Tensor, torch.Tensor]:
-        # Determine the T5 max sequence length based on the model.
-        if self.model == "flux-schnell":
-            max_seq_len = 256
-        # elif self.model == "flux-dev":
-        #     max_seq_len = 512
-        else:
-            raise ValueError(f"Unknown model: {self.model}")
-
-        # Load the CLIP tokenizer.
-        clip_tokenizer_path = flux_model_dir / "tokenizer"
-        clip_tokenizer = CLIPTokenizer.from_pretrained(clip_tokenizer_path, local_files_only=True)
-        assert isinstance(clip_tokenizer, CLIPTokenizer)
-
-        # Load the T5 tokenizer.
-        t5_tokenizer_path = flux_model_dir / "tokenizer_2"
-        t5_tokenizer = T5TokenizerFast.from_pretrained(t5_tokenizer_path, local_files_only=True)
-        assert isinstance(t5_tokenizer, T5TokenizerFast)
-
-        clip_text_encoder_path = flux_model_dir / "text_encoder"
-        t5_text_encoder_path = flux_model_dir / "text_encoder_2"
-        with (
-            context.models.load_local_model(
-                model_path=clip_text_encoder_path, loader=self._load_flux_text_encoder
-            ) as clip_text_encoder,
-            context.models.load_local_model(
-                model_path=t5_text_encoder_path, loader=self._load_flux_text_encoder_2
-            ) as t5_text_encoder,
-        ):
-            assert isinstance(clip_text_encoder, CLIPTextModel)
-            assert isinstance(t5_text_encoder, T5EncoderModel)
-            pipeline = FluxPipeline(
-                scheduler=None,
-                vae=None,
-                text_encoder=clip_text_encoder,
-                tokenizer=clip_tokenizer,
-                text_encoder_2=t5_text_encoder,
-                tokenizer_2=t5_tokenizer,
-                transformer=None,
-            )
-
-            # prompt_embeds: T5 embeddings
-            # pooled_prompt_embeds: CLIP embeddings
-            prompt_embeds, pooled_prompt_embeds, text_ids = pipeline.encode_prompt(
-                prompt=self.positive_prompt,
-                prompt_2=self.positive_prompt,
-                device=TorchDevice.choose_torch_device(),
-                max_sequence_length=max_seq_len,
-            )
-
-        assert isinstance(prompt_embeds, torch.Tensor)
-        assert isinstance(pooled_prompt_embeds, torch.Tensor)
-        return prompt_embeds, pooled_prompt_embeds
-
-    def _run_diffusion(
-        self,
-        context: InvocationContext,
-        flux_model_dir: Path,
-        clip_embeddings: torch.Tensor,
-        t5_embeddings: torch.Tensor,
-    ):
-        scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(flux_model_dir / "scheduler", local_files_only=True)
-
-        # HACK(ryand): Manually empty the cache. Currently we don't check the size of the model before loading it from
-        # disk. Since the transformer model is large (24GB), there's a good chance that it will OOM on 32GB RAM systems
-        # if the cache is not empty.
-        context.models._services.model_manager.load.ram_cache.make_room(24 * 2**30)
-
-        transformer_path = flux_model_dir / "transformer"
-        with context.models.load_local_model(
-            model_path=transformer_path, loader=self._load_flux_transformer
-        ) as transformer:
-            assert isinstance(transformer, FluxTransformer2DModel)
-
-            flux_pipeline_with_transformer = FluxPipeline(
-                scheduler=scheduler,
-                vae=None,
-                text_encoder=None,
-                tokenizer=None,
-                text_encoder_2=None,
-                tokenizer_2=None,
-                transformer=transformer,
-            )
-
-            t5_embeddings = t5_embeddings.to(dtype=transformer.dtype)
-            clip_embeddings = clip_embeddings.to(dtype=transformer.dtype)
-
-            latents = flux_pipeline_with_transformer(
-                height=self.height,
-                width=self.width,
-                num_inference_steps=self.num_steps,
-                guidance_scale=self.guidance,
-                generator=torch.Generator().manual_seed(self.seed),
-                prompt_embeds=t5_embeddings,
-                pooled_prompt_embeds=clip_embeddings,
-                output_type="latent",
-                return_dict=False,
-            )[0]
-
-        assert isinstance(latents, torch.Tensor)
-        return latents
-
-    def _run_vae_decoding(
-        self,
-        context: InvocationContext,
-        flux_model_dir: Path,
-        latents: torch.Tensor,
-    ) -> Image.Image:
-        vae_path = flux_model_dir / "vae"
-        with context.models.load_local_model(model_path=vae_path, loader=self._load_flux_vae) as vae:
-            assert isinstance(vae, AutoencoderKL)
-
-            flux_pipeline_with_vae = FluxPipeline(
-                scheduler=None,
-                vae=vae,
-                text_encoder=None,
-                tokenizer=None,
-                text_encoder_2=None,
-                tokenizer_2=None,
-                transformer=None,
-            )
-
-            latents = flux_pipeline_with_vae._unpack_latents(
-                latents, self.height, self.width, flux_pipeline_with_vae.vae_scale_factor
-            )
-            latents = (
-                latents / flux_pipeline_with_vae.vae.config.scaling_factor
-            ) + flux_pipeline_with_vae.vae.config.shift_factor
-            latents = latents.to(dtype=vae.dtype)
-            image = flux_pipeline_with_vae.vae.decode(latents, return_dict=False)[0]
-            image = flux_pipeline_with_vae.image_processor.postprocess(image, output_type="pil")[0]
-
-        assert isinstance(image, Image.Image)
-        return image
-
-    @staticmethod
-    def _load_flux_text_encoder(path: Path) -> CLIPTextModel:
-        model = CLIPTextModel.from_pretrained(path, local_files_only=True)
-        assert isinstance(model, CLIPTextModel)
-        return model
-
-    def _load_flux_text_encoder_2(self, path: Path) -> T5EncoderModel:
-        if self.use_8bit:
-            model_8bit_path = path / "quantized"
-            if model_8bit_path.exists():
-                # The quantized model exists, load it.
-                # TODO(ryand): The requantize(...) operation in from_pretrained(...) is very slow. This seems like
-                # something that we should be able to make much faster.
-                q_model = QuantizedModelForTextEncoding.from_pretrained(model_8bit_path)
-
-                # Access the underlying wrapped model.
-                # We access the wrapped model, even though it is private, because it simplifies the type checking by
-                # always returning a T5EncoderModel from this function.
-                model = q_model._wrapped
-            else:
-                # The quantized model does not exist yet, quantize and save it.
-                # TODO(ryand): dtype?
-                model = T5EncoderModel.from_pretrained(path, local_files_only=True)
-                assert isinstance(model, T5EncoderModel)
-
-                q_model = QuantizedModelForTextEncoding.quantize(model, weights=qfloat8)
-
-                model_8bit_path.mkdir(parents=True, exist_ok=True)
-                q_model.save_pretrained(model_8bit_path)
-
-                # (See earlier comment about accessing the wrapped model.)
-                model = q_model._wrapped
-        else:
-            model = T5EncoderModel.from_pretrained(path, local_files_only=True)
-
-        assert isinstance(model, T5EncoderModel)
-        return model
-
-    def _load_flux_transformer(self, path: Path) -> FluxTransformer2DModel:
-        if self.use_8bit:
-            model_8bit_path = path / "quantized"
-            if model_8bit_path.exists():
-                # The quantized model exists, load it.
-                # TODO(ryand): The requantize(...) operation in from_pretrained(...) is very slow. This seems like
-                # something that we should be able to make much faster.
-                q_model = QuantizedFluxTransformer2DModel.from_pretrained(model_8bit_path)
-
-                # Access the underlying wrapped model.
-                # We access the wrapped model, even though it is private, because it simplifies the type checking by
-                # always returning a FluxTransformer2DModel from this function.
-                model = q_model._wrapped
-            else:
-                # The quantized model does not exist yet, quantize and save it.
-                # TODO(ryand): Loading in float16 and then quantizing seems to result in NaNs. In order to run this on
-                # GPUs that don't support bfloat16, we would need to host the quantized model instead of generating it
-                # here.
-                model = FluxTransformer2DModel.from_pretrained(path, local_files_only=True, torch_dtype=torch.bfloat16)
-                assert isinstance(model, FluxTransformer2DModel)
-
-                q_model = QuantizedFluxTransformer2DModel.quantize(model, weights=qfloat8)
-
-                model_8bit_path.mkdir(parents=True, exist_ok=True)
-                q_model.save_pretrained(model_8bit_path)
-
-                # (See earlier comment about accessing the wrapped model.)
-                model = q_model._wrapped
-        else:
-            model = FluxTransformer2DModel.from_pretrained(path, local_files_only=True, torch_dtype=torch.bfloat16)
-
-        assert isinstance(model, FluxTransformer2DModel)
-        return model
-
-    @staticmethod
-    def _load_flux_vae(path: Path) -> AutoencoderKL:
-        model = AutoencoderKL.from_pretrained(path, local_files_only=True)
-        assert isinstance(model, AutoencoderKL)
-        return model

invokeai/app/invocations/grounding_dino.py

@@ -1,100 +0,0 @@
-from pathlib import Path
-from typing import Literal
-
-import torch
-from PIL import Image
-from transformers import pipeline
-from transformers.pipelines import ZeroShotObjectDetectionPipeline
-
-from invokeai.app.invocations.baseinvocation import BaseInvocation, invocation
-from invokeai.app.invocations.fields import BoundingBoxField, ImageField, InputField
-from invokeai.app.invocations.primitives import BoundingBoxCollectionOutput
-from invokeai.app.services.shared.invocation_context import InvocationContext
-from invokeai.backend.image_util.grounding_dino.detection_result import DetectionResult
-from invokeai.backend.image_util.grounding_dino.grounding_dino_pipeline import GroundingDinoPipeline
-
-GroundingDinoModelKey = Literal["grounding-dino-tiny", "grounding-dino-base"]
-GROUNDING_DINO_MODEL_IDS: dict[GroundingDinoModelKey, str] = {
-    "grounding-dino-tiny": "IDEA-Research/grounding-dino-tiny",
-    "grounding-dino-base": "IDEA-Research/grounding-dino-base",
-}
-
-
-@invocation(
-    "grounding_dino",
-    title="Grounding DINO (Text Prompt Object Detection)",
-    tags=["prompt", "object detection"],
-    category="image",
-    version="1.0.0",
-)
-class GroundingDinoInvocation(BaseInvocation):
-    """Runs a Grounding DINO model. Performs zero-shot bounding-box object detection from a text prompt."""
-
-    # Reference:
-    # - https://arxiv.org/pdf/2303.05499
-    # - https://huggingface.co/docs/transformers/v4.43.3/en/model_doc/grounding-dino#grounded-sam
-    # - https://github.com/NielsRogge/Transformers-Tutorials/blob/a39f33ac1557b02ebfb191ea7753e332b5ca933f/Grounding%20DINO/GroundingDINO_with_Segment_Anything.ipynb
-
-    model: GroundingDinoModelKey = InputField(description="The Grounding DINO model to use.")
-    prompt: str = InputField(description="The prompt describing the object to segment.")
-    image: ImageField = InputField(description="The image to segment.")
-    detection_threshold: float = InputField(
-        description="The detection threshold for the Grounding DINO model. All detected bounding boxes with scores above this threshold will be returned.",
-        ge=0.0,
-        le=1.0,
-        default=0.3,
-    )
-
-    @torch.no_grad()
-    def invoke(self, context: InvocationContext) -> BoundingBoxCollectionOutput:
-        # The model expects a 3-channel RGB image.
-        image_pil = context.images.get_pil(self.image.image_name, mode="RGB")
-
-        detections = self._detect(
-            context=context, image=image_pil, labels=[self.prompt], threshold=self.detection_threshold
-        )
-
-        # Convert detections to BoundingBoxCollectionOutput.
-        bounding_boxes: list[BoundingBoxField] = []
-        for detection in detections:
-            bounding_boxes.append(
-                BoundingBoxField(
-                    x_min=detection.box.xmin,
-                    x_max=detection.box.xmax,
-                    y_min=detection.box.ymin,
-                    y_max=detection.box.ymax,
-                    score=detection.score,
-                )
-            )
-        return BoundingBoxCollectionOutput(collection=bounding_boxes)
-
-    @staticmethod
-    def _load_grounding_dino(model_path: Path):
-        grounding_dino_pipeline = pipeline(
-            model=str(model_path),
-            task="zero-shot-object-detection",
-            local_files_only=True,
-            # TODO(ryand): Setting the torch_dtype here doesn't work. Investigate whether fp16 is supported by the
-            # model, and figure out how to make it work in the pipeline.
-            # torch_dtype=TorchDevice.choose_torch_dtype(),
-        )
-        assert isinstance(grounding_dino_pipeline, ZeroShotObjectDetectionPipeline)
-        return GroundingDinoPipeline(grounding_dino_pipeline)
-
-    def _detect(
-        self,
-        context: InvocationContext,
-        image: Image.Image,
-        labels: list[str],
-        threshold: float = 0.3,
-    ) -> list[DetectionResult]:
-        """Use Grounding DINO to detect bounding boxes for a set of labels in an image."""
-        # TODO(ryand): I copied this "."-handling logic from the transformers example code. Test it and see if it
-        # actually makes a difference.
-        labels = [label if label.endswith(".") else label + "." for label in labels]
-
-        with context.models.load_remote_model(
-            source=GROUNDING_DINO_MODEL_IDS[self.model], loader=GroundingDinoInvocation._load_grounding_dino
-        ) as detector:
-            assert isinstance(detector, GroundingDinoPipeline)
-            return detector.detect(image=image, candidate_labels=labels, threshold=threshold)

invokeai/app/invocations/latents_to_image.py

@@ -24,7 +24,7 @@ from invokeai.app.invocations.fields import (
 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.stable_diffusion.extensions.seamless import SeamlessExt
+from invokeai.backend.stable_diffusion import set_seamless
 from invokeai.backend.stable_diffusion.vae_tiling import patch_vae_tiling_params
 from invokeai.backend.util.devices import TorchDevice
 
@@ -59,7 +59,7 @@ class LatentsToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
 
         vae_info = context.models.load(self.vae.vae)
         assert isinstance(vae_info.model, (AutoencoderKL, AutoencoderTiny))
-        with SeamlessExt.static_patch_model(vae_info.model, self.vae.seamless_axes), vae_info as vae:
+        with set_seamless(vae_info.model, self.vae.seamless_axes), vae_info as vae:
             assert isinstance(vae, (AutoencoderKL, AutoencoderTiny))
             latents = latents.to(vae.device)
             if self.fp32:

invokeai/app/invocations/mask.py

@@ -1,10 +1,9 @@
 import numpy as np
 import torch
-from PIL import Image
 
 from invokeai.app.invocations.baseinvocation import BaseInvocation, Classification, InvocationContext, invocation
-from invokeai.app.invocations.fields import ImageField, InputField, TensorField, WithBoard, WithMetadata
-from invokeai.app.invocations.primitives import ImageOutput, MaskOutput
+from invokeai.app.invocations.fields import ImageField, InputField, TensorField, WithMetadata
+from invokeai.app.invocations.primitives import MaskOutput
 
 
 @invocation(
@@ -119,27 +118,3 @@ class ImageMaskToTensorInvocation(BaseInvocation, WithMetadata):
             height=mask.shape[1],
             width=mask.shape[2],
         )
-
-
-@invocation(
-    "tensor_mask_to_image",
-    title="Tensor Mask to Image",
-    tags=["mask"],
-    category="mask",
-    version="1.0.0",
-)
-class MaskTensorToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
-    """Convert a mask tensor to an image."""
-
-    mask: TensorField = InputField(description="The mask tensor to convert.")
-
-    def invoke(self, context: InvocationContext) -> ImageOutput:
-        mask = context.tensors.load(self.mask.tensor_name)
-        # Ensure that the mask is binary.
-        if mask.dtype != torch.bool:
-            mask = mask > 0.5
-        mask_np = (mask.float() * 255).byte().cpu().numpy()
-
-        mask_pil = Image.fromarray(mask_np, mode="L")
-        image_dto = context.images.save(image=mask_pil)
-        return ImageOutput.build(image_dto)

invokeai/app/invocations/primitives.py

@@ -7,7 +7,6 @@ import torch
 from invokeai.app.invocations.baseinvocation import BaseInvocation, BaseInvocationOutput, invocation, invocation_output
 from invokeai.app.invocations.constants import LATENT_SCALE_FACTOR
 from invokeai.app.invocations.fields import (
-    BoundingBoxField,
     ColorField,
     ConditioningField,
     DenoiseMaskField,
@@ -470,42 +469,3 @@ class ConditioningCollectionInvocation(BaseInvocation):
 
 
 # endregion
-
-# region BoundingBox
-
-
-@invocation_output("bounding_box_output")
-class BoundingBoxOutput(BaseInvocationOutput):
-    """Base class for nodes that output a single bounding box"""
-
-    bounding_box: BoundingBoxField = OutputField(description="The output bounding box.")
-
-
-@invocation_output("bounding_box_collection_output")
-class BoundingBoxCollectionOutput(BaseInvocationOutput):
-    """Base class for nodes that output a collection of bounding boxes"""
-
-    collection: list[BoundingBoxField] = OutputField(description="The output bounding boxes.", title="Bounding Boxes")
-
-
-@invocation(
-    "bounding_box",
-    title="Bounding Box",
-    tags=["primitives", "segmentation", "collection", "bounding box"],
-    category="primitives",
-    version="1.0.0",
-)
-class BoundingBoxInvocation(BaseInvocation):
-    """Create a bounding box manually by supplying box coordinates"""
-
-    x_min: int = InputField(default=0, description="x-coordinate of the bounding box's top left vertex")
-    y_min: int = InputField(default=0, description="y-coordinate of the bounding box's top left vertex")
-    x_max: int = InputField(default=0, description="x-coordinate of the bounding box's bottom right vertex")
-    y_max: int = InputField(default=0, description="y-coordinate of the bounding box's bottom right vertex")
-
-    def invoke(self, context: InvocationContext) -> BoundingBoxOutput:
-        bounding_box = BoundingBoxField(x_min=self.x_min, y_min=self.y_min, x_max=self.x_max, y_max=self.y_max)
-        return BoundingBoxOutput(bounding_box=bounding_box)
-
-
-# endregion

invokeai/app/invocations/segment_anything.py

@@ -1,161 +1,76 @@
-from pathlib import Path
-from typing import Literal
+from typing import Dict, cast
 
-import numpy as np
 import torch
-from PIL import Image
-from transformers import AutoModelForMaskGeneration, AutoProcessor
-from transformers.models.sam import SamModel
-from transformers.models.sam.processing_sam import SamProcessor
 
 from invokeai.app.invocations.baseinvocation import BaseInvocation, invocation
-from invokeai.app.invocations.fields import BoundingBoxField, ImageField, InputField, TensorField
-from invokeai.app.invocations.primitives import MaskOutput
+from invokeai.app.invocations.fields import ImageField, InputField
+from invokeai.app.invocations.primitives import ImageOutput
 from invokeai.app.services.shared.invocation_context import InvocationContext
-from invokeai.backend.image_util.segment_anything.mask_refinement import mask_to_polygon, polygon_to_mask
-from invokeai.backend.image_util.segment_anything.segment_anything_pipeline import SegmentAnythingPipeline
+from invokeai.backend.image_util.grounding_segment_anything.gsa import GroundingSegmentAnythingDetector
+from invokeai.backend.util.devices import TorchDevice
 
-SegmentAnythingModelKey = Literal["segment-anything-base", "segment-anything-large", "segment-anything-huge"]
-SEGMENT_ANYTHING_MODEL_IDS: dict[SegmentAnythingModelKey, str] = {
-    "segment-anything-base": "facebook/sam-vit-base",
-    "segment-anything-large": "facebook/sam-vit-large",
-    "segment-anything-huge": "facebook/sam-vit-huge",
+GROUNDING_SEGMENT_ANYTHING_MODELS = {
+    "groundingdino_swint_ogc": "https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha/groundingdino_swint_ogc.pth",
+    "segment_anything_vit_h": "https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth",
 }
 
 
 @invocation(
     "segment_anything",
     title="Segment Anything",
-    tags=["prompt", "segmentation"],
-    category="segmentation",
+    tags=["grounding_dino", "segment", "anything"],
+    category="image",
     version="1.0.0",
 )
 class SegmentAnythingInvocation(BaseInvocation):
-    """Runs a Segment Anything Model."""
+    """Automatically generate masks from an image using GroundingDINO & Segment Anything"""
 
-    # Reference:
-    # - https://arxiv.org/pdf/2304.02643
-    # - https://huggingface.co/docs/transformers/v4.43.3/en/model_doc/grounding-dino#grounded-sam
-    # - https://github.com/NielsRogge/Transformers-Tutorials/blob/a39f33ac1557b02ebfb191ea7753e332b5ca933f/Grounding%20DINO/GroundingDINO_with_Segment_Anything.ipynb
-
-    model: SegmentAnythingModelKey = InputField(description="The Segment Anything model to use.")
-    image: ImageField = InputField(description="The image to segment.")
-    bounding_boxes: list[BoundingBoxField] = InputField(description="The bounding boxes to prompt the SAM model with.")
-    apply_polygon_refinement: bool = InputField(
-        description="Whether to apply polygon refinement to the masks. This will smooth the edges of the masks slightly and ensure that each mask consists of a single closed polygon (before merging).",
-        default=True,
+    image: ImageField = InputField(description="The image to process")
+    prompt: str = InputField(default="", description="Keywords to segment", title="Prompt")
+    box_threshold: float = InputField(
+        default=0.5, ge=0, le=1, description="Threshold of box detection", title="Box Threshold"
     )
-    mask_filter: Literal["all", "largest", "highest_box_score"] = InputField(
-        description="The filtering to apply to the detected masks before merging them into a final output.",
-        default="all",
+    text_threshold: float = InputField(
+        default=0.5, ge=0, le=1, description="Threshold of text detection", title="Text Threshold"
+    )
+    nms_threshold: float = InputField(
+        default=0.8, ge=0, le=1, description="Threshold of nms detection", title="NMS Threshold"
     )
 
-    @torch.no_grad()
-    def invoke(self, context: InvocationContext) -> MaskOutput:
-        # The models expect a 3-channel RGB image.
-        image_pil = context.images.get_pil(self.image.image_name, mode="RGB")
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        input_image = context.images.get_pil(self.image.image_name)
 
-        if len(self.bounding_boxes) == 0:
-            combined_mask = torch.zeros(image_pil.size[::-1], dtype=torch.bool)
-        else:
-            masks = self._segment(context=context, image=image_pil)
-            masks = self._filter_masks(masks=masks, bounding_boxes=self.bounding_boxes)
-
-            # masks contains bool values, so we merge them via max-reduce.
-            combined_mask, _ = torch.stack(masks).max(dim=0)
-
-        mask_tensor_name = context.tensors.save(combined_mask)
-        height, width = combined_mask.shape
-        return MaskOutput(mask=TensorField(tensor_name=mask_tensor_name), width=width, height=height)
-
-    @staticmethod
-    def _load_sam_model(model_path: Path):
-        sam_model = AutoModelForMaskGeneration.from_pretrained(
-            model_path,
-            local_files_only=True,
-            # TODO(ryand): Setting the torch_dtype here doesn't work. Investigate whether fp16 is supported by the
-            # model, and figure out how to make it work in the pipeline.
-            # torch_dtype=TorchDevice.choose_torch_dtype(),
+        grounding_dino_model = context.models.load_remote_model(
+            GROUNDING_SEGMENT_ANYTHING_MODELS["groundingdino_swint_ogc"]
+        )
+        segment_anything_model = context.models.load_remote_model(
+            GROUNDING_SEGMENT_ANYTHING_MODELS["segment_anything_vit_h"]
         )
-        assert isinstance(sam_model, SamModel)
-
-        sam_processor = AutoProcessor.from_pretrained(model_path, local_files_only=True)
-        assert isinstance(sam_processor, SamProcessor)
-        return SegmentAnythingPipeline(sam_model=sam_model, sam_processor=sam_processor)
-
-    def _segment(
-        self,
-        context: InvocationContext,
-        image: Image.Image,
-    ) -> list[torch.Tensor]:
-        """Use Segment Anything (SAM) to generate masks given an image + a set of bounding boxes."""
-        # Convert the bounding boxes to the SAM input format.
-        sam_bounding_boxes = [[bb.x_min, bb.y_min, bb.x_max, bb.y_max] for bb in self.bounding_boxes]
 
         with (
-            context.models.load_remote_model(
-                source=SEGMENT_ANYTHING_MODEL_IDS[self.model], loader=SegmentAnythingInvocation._load_sam_model
-            ) as sam_pipeline,
+            grounding_dino_model.model_on_device() as (_, grounding_dino_state_dict),
+            segment_anything_model.model_on_device() as (_, segment_anything_state_dict),
         ):
-            assert isinstance(sam_pipeline, SegmentAnythingPipeline)
-            masks = sam_pipeline.segment(image=image, bounding_boxes=sam_bounding_boxes)
+            if not grounding_dino_state_dict or not segment_anything_state_dict:
+                raise RuntimeError("Unable to load segmentation models")
 
-        masks = self._process_masks(masks)
-        if self.apply_polygon_refinement:
-            masks = self._apply_polygon_refinement(masks)
+            grounding_dino = GroundingSegmentAnythingDetector.build_grounding_dino(
+                cast(Dict[str, torch.Tensor], grounding_dino_state_dict), TorchDevice.choose_torch_device()
+            )
+            segment_anything = GroundingSegmentAnythingDetector.build_segment_anything(
+                cast(Dict[str, torch.Tensor], segment_anything_state_dict), TorchDevice.choose_torch_device()
+            )
+            detector = GroundingSegmentAnythingDetector(grounding_dino, segment_anything)
 
-        return masks
+            mask = detector.predict(
+                input_image, self.prompt, self.box_threshold, self.text_threshold, self.nms_threshold
+            )
+            image_dto = context.images.save(mask)
 
-    def _process_masks(self, masks: torch.Tensor) -> list[torch.Tensor]:
-        """Convert the tensor output from the Segment Anything model from a tensor of shape
-        [num_masks, channels, height, width] to a list of tensors of shape [height, width].
-        """
-        assert masks.dtype == torch.bool
-        # [num_masks, channels, height, width] -> [num_masks, height, width]
-        masks, _ = masks.max(dim=1)
-        # Split the first dimension into a list of masks.
-        return list(masks.cpu().unbind(dim=0))
-
-    def _apply_polygon_refinement(self, masks: list[torch.Tensor]) -> list[torch.Tensor]:
-        """Apply polygon refinement to the masks.
-
-        Convert each mask to a polygon, then back to a mask. This has the following effect:
-        - Smooth the edges of the mask slightly.
-        - Ensure that each mask consists of a single closed polygon
-            - Removes small mask pieces.
-            - Removes holes from the mask.
-        """
-        # Convert tensor masks to np masks.
-        np_masks = [mask.cpu().numpy().astype(np.uint8) for mask in masks]
-
-        # Apply polygon refinement.
-        for idx, mask in enumerate(np_masks):
-            shape = mask.shape
-            assert len(shape) == 2  # Assert length to satisfy type checker.
-            polygon = mask_to_polygon(mask)
-            mask = polygon_to_mask(polygon, shape)
-            np_masks[idx] = mask
-
-        # Convert np masks back to tensor masks.
-        masks = [torch.tensor(mask, dtype=torch.bool) for mask in np_masks]
-
-        return masks
-
-    def _filter_masks(self, masks: list[torch.Tensor], bounding_boxes: list[BoundingBoxField]) -> list[torch.Tensor]:
-        """Filter the detected masks based on the specified mask filter."""
-        assert len(masks) == len(bounding_boxes)
-
-        if self.mask_filter == "all":
-            return masks
-        elif self.mask_filter == "largest":
-            # Find the largest mask.
-            return [max(masks, key=lambda x: float(x.sum()))]
-        elif self.mask_filter == "highest_box_score":
-            # Find the index of the bounding box with the highest score.
-            # Note that we fallback to -1.0 if the score is None. This is mainly to satisfy the type checker. In most
-            # cases the scores should all be non-None when using this filtering mode. That being said, -1.0 is a
-            # reasonable fallback since the expected score range is [0.0, 1.0].
-            max_score_idx = max(range(len(bounding_boxes)), key=lambda i: bounding_boxes[i].score or -1.0)
-            return [masks[max_score_idx]]
-        else:
-            raise ValueError(f"Invalid mask filter: {self.mask_filter}")
+            """Builds an ImageOutput and its ImageField"""
+            processed_image_field = ImageField(image_name=image_dto.image_name)
+            return ImageOutput(
+                image=processed_image_field,
+                width=input_image.width,
+                height=input_image.height,
+            )

invokeai/app/invocations/vto_workflow.py

@@ -0,0 +1,81 @@
+import cv2
+import numpy as np
+from PIL import Image
+
+from invokeai.app.invocations.baseinvocation import BaseInvocation, invocation
+from invokeai.app.invocations.fields import ImageField, InputField, WithBoard, WithMetadata
+from invokeai.app.invocations.primitives import ImageOutput
+from invokeai.app.services.shared.invocation_context import InvocationContext
+from invokeai.backend.vto_workflow.extract_channel import ImageChannel, extract_channel
+from invokeai.backend.vto_workflow.overlay_pattern import multiply_images
+from invokeai.backend.vto_workflow.seamless_mapping import map_seamless_tiles
+
+
+@invocation("vto", title="Virtual Try-On", tags=["vto"], category="vto", version="1.1.0")
+class VTOInvocation(BaseInvocation, WithMetadata, WithBoard):
+    """Virtual try-on."""
+
+    original_image: ImageField = InputField(description="The input image")
+    clothing_mask: ImageField = InputField(description="Clothing mask.")
+    pattern_image: ImageField = InputField(description="Pattern image.")
+    pattern_vertical_repeats: float = InputField(
+        description="Number of vertical repeats for the pattern.", gt=0.01, default=1.0
+    )
+
+    shading_max: float = InputField(
+        description="The lightness of the light spots on the clothing. Default is 1.0. Typically in the range [0.7, 1.2]. Must be > shading_min",
+        default=1.0,
+        ge=0.0,
+    )
+    shading_min: float = InputField(
+        description="The lightness of the dark spots on the clothing. Default id 0.5. Typically in the range [0.2, 0.7]",
+        default=0.5,
+        ge=0.0,
+    )
+
+    mask_dilation: int = InputField(
+        description="The number of pixels to dilate the mask by. Default is 1.",
+        default=1,
+        ge=0,
+    )
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        # TODO(ryand): Avoid all the unnecessary flip-flopping between PIL and numpy.
+        original_image = context.images.get_pil(self.original_image.image_name)
+        clothing_mask = context.images.get_pil(self.clothing_mask.image_name)
+        pattern_image = context.images.get_pil(self.pattern_image.image_name)
+
+        shadows = extract_channel(np.array(original_image), ImageChannel.LAB_L)
+
+        # Clip the shadows to the 0.05 and 0.95 percentiles to eliminate outliers.
+        shadows = np.clip(shadows, np.percentile(shadows, 5), np.percentile(shadows, 95))
+
+        # Normalize the shadows to the range [shading_min, shading_max].
+        assert self.shading_min < self.shading_max
+        shadows = shadows.astype(np.float32)
+        shadows = (shadows - shadows.min()) / (shadows.max() - shadows.min())
+        shadows = self.shading_min + (self.shading_max - self.shading_min) * shadows
+        shadows = np.clip(shadows, 0.0, 1.0)
+        shadows = (shadows * 255).astype(np.uint8)
+
+        expanded_pattern = map_seamless_tiles(
+            seamless_tile=pattern_image,
+            target_hw=(original_image.height, original_image.width),
+            num_repeats_h=self.pattern_vertical_repeats,
+        )
+
+        pattern_with_shadows = multiply_images(expanded_pattern, Image.fromarray(shadows))
+
+        # Dilate the mask.
+        clothing_mask_np = np.array(clothing_mask)
+        if self.mask_dilation > 0:
+            clothing_mask_np = cv2.dilate(clothing_mask_np, np.ones((3, 3), np.uint8), iterations=self.mask_dilation)
+
+        # Merge the pattern with the model image.
+        pattern_with_shadows_np = np.array(pattern_with_shadows)
+        original_image_np = np.array(original_image)
+        merged_image = np.where(clothing_mask_np[:, :, None], pattern_with_shadows_np, original_image_np)
+        merged_image = Image.fromarray(merged_image)
+
+        image_dto = context.images.save(image=merged_image)
+        return ImageOutput.build(image_dto)

invokeai/app/services/image_files/image_files_disk.py

@@ -1,10 +1,11 @@
 # Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654) and the InvokeAI Team
 from pathlib import Path
 from queue import Queue
-from typing import Optional, Union
+from typing import Dict, Optional, Union
 
 from PIL import Image, PngImagePlugin
 from PIL.Image import Image as PILImageType
+from send2trash import send2trash
 
 from invokeai.app.services.image_files.image_files_base import ImageFileStorageBase
 from invokeai.app.services.image_files.image_files_common import (
@@ -19,12 +20,18 @@ from invokeai.app.util.thumbnails import get_thumbnail_name, make_thumbnail
 class DiskImageFileStorage(ImageFileStorageBase):
     """Stores images on disk"""
 
+    __output_folder: Path
+    __cache_ids: Queue  # TODO: this is an incredibly naive cache
+    __cache: Dict[Path, PILImageType]
+    __max_cache_size: int
+    __invoker: Invoker
+
     def __init__(self, output_folder: Union[str, Path]):
-        self.__cache: dict[Path, PILImageType] = {}
-        self.__cache_ids = Queue[Path]()
+        self.__cache = {}
+        self.__cache_ids = Queue()
         self.__max_cache_size = 10  # TODO: get this from config
 
-        self.__output_folder = output_folder if isinstance(output_folder, Path) else Path(output_folder)
+        self.__output_folder: Path = output_folder if isinstance(output_folder, Path) else Path(output_folder)
         self.__thumbnails_folder = self.__output_folder / "thumbnails"
         # Validate required output folders at launch
         self.__validate_storage_folders()
@@ -96,7 +103,7 @@ class DiskImageFileStorage(ImageFileStorageBase):
             image_path = self.get_path(image_name)
 
             if image_path.exists():
-                image_path.unlink()
+                send2trash(image_path)
             if image_path in self.__cache:
                 del self.__cache[image_path]
 
@@ -104,7 +111,7 @@ class DiskImageFileStorage(ImageFileStorageBase):
             thumbnail_path = self.get_path(thumbnail_name, True)
 
             if thumbnail_path.exists():
-                thumbnail_path.unlink()
+                send2trash(thumbnail_path)
             if thumbnail_path in self.__cache:
                 del self.__cache[thumbnail_path]
         except Exception as e:

invokeai/app/services/model_images/model_images_default.py

@@ -2,6 +2,7 @@ from pathlib import Path
 
 from PIL import Image
 from PIL.Image import Image as PILImageType
+from send2trash import send2trash
 
 from invokeai.app.services.invoker import Invoker
 from invokeai.app.services.model_images.model_images_base import ModelImageFileStorageBase
@@ -69,7 +70,7 @@ class ModelImageFileStorageDisk(ModelImageFileStorageBase):
             if not self._validate_path(path):
                 raise ModelImageFileNotFoundException
 
-            path.unlink()
+            send2trash(path)
 
         except Exception as e:
             raise ModelImageFileDeleteException from e

invokeai/backend/image_util/grounding_dino/detection_result.py

@@ -1,22 +0,0 @@
-from pydantic import BaseModel, ConfigDict
-
-
-class BoundingBox(BaseModel):
-    """Bounding box helper class."""
-
-    xmin: int
-    ymin: int
-    xmax: int
-    ymax: int
-
-
-class DetectionResult(BaseModel):
-    """Detection result from Grounding DINO."""
-
-    score: float
-    label: str
-    box: BoundingBox
-    model_config = ConfigDict(
-        # Allow arbitrary types for mask, since it will be a numpy array.
-        arbitrary_types_allowed=True
-    )

invokeai/backend/image_util/grounding_dino/grounding_dino_pipeline.py

@@ -1,37 +0,0 @@
-from typing import Optional
-
-import torch
-from PIL import Image
-from transformers.pipelines import ZeroShotObjectDetectionPipeline
-
-from invokeai.backend.image_util.grounding_dino.detection_result import DetectionResult
-from invokeai.backend.raw_model import RawModel
-
-
-class GroundingDinoPipeline(RawModel):
-    """A wrapper class for a ZeroShotObjectDetectionPipeline that makes it compatible with the model manager's memory
-    management system.
-    """
-
-    def __init__(self, pipeline: ZeroShotObjectDetectionPipeline):
-        self._pipeline = pipeline
-
-    def detect(self, image: Image.Image, candidate_labels: list[str], threshold: float = 0.1) -> list[DetectionResult]:
-        results = self._pipeline(image=image, candidate_labels=candidate_labels, threshold=threshold)
-        assert results is not None
-        results = [DetectionResult.model_validate(result) for result in results]
-        return results
-
-    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None):
-        # HACK(ryand): The GroundingDinoPipeline does not work on MPS devices. We only allow it to be moved to CPU or
-        # CUDA.
-        if device is not None and device.type not in {"cpu", "cuda"}:
-            device = None
-        self._pipeline.model.to(device=device, dtype=dtype)
-        self._pipeline.device = self._pipeline.model.device
-
-    def calc_size(self) -> int:
-        # HACK(ryand): Fix the circular import issue.
-        from invokeai.backend.model_manager.load.model_util import calc_module_size
-
-        return calc_module_size(self._pipeline.model)

invokeai/backend/image_util/grounding_segment_anything/groundingdino/config/GroundingDINO_SwinB.py

@@ -0,0 +1,43 @@
+batch_size = 1
+modelname = "groundingdino"
+backbone = "swin_B_384_22k"
+position_embedding = "sine"
+pe_temperatureH = 20
+pe_temperatureW = 20
+return_interm_indices = [1, 2, 3]
+backbone_freeze_keywords = None
+enc_layers = 6
+dec_layers = 6
+pre_norm = False
+dim_feedforward = 2048
+hidden_dim = 256
+dropout = 0.0
+nheads = 8
+num_queries = 900
+query_dim = 4
+num_patterns = 0
+num_feature_levels = 4
+enc_n_points = 4
+dec_n_points = 4
+two_stage_type = "standard"
+two_stage_bbox_embed_share = False
+two_stage_class_embed_share = False
+transformer_activation = "relu"
+dec_pred_bbox_embed_share = True
+dn_box_noise_scale = 1.0
+dn_label_noise_ratio = 0.5
+dn_label_coef = 1.0
+dn_bbox_coef = 1.0
+embed_init_tgt = True
+dn_labelbook_size = 2000
+max_text_len = 256
+text_encoder_type = "bert-base-uncased"
+use_text_enhancer = True
+use_fusion_layer = True
+use_checkpoint = True
+use_transformer_ckpt = True
+use_text_cross_attention = True
+text_dropout = 0.0
+fusion_dropout = 0.0
+fusion_droppath = 0.1
+sub_sentence_present = True

invokeai/backend/image_util/grounding_segment_anything/groundingdino/config/GroundingDINO_SwinT_OGC.py

@@ -0,0 +1,43 @@
+batch_size = 1
+modelname = "groundingdino"
+backbone = "swin_T_224_1k"
+position_embedding = "sine"
+pe_temperatureH = 20
+pe_temperatureW = 20
+return_interm_indices = [1, 2, 3]
+backbone_freeze_keywords = None
+enc_layers = 6
+dec_layers = 6
+pre_norm = False
+dim_feedforward = 2048
+hidden_dim = 256
+dropout = 0.0
+nheads = 8
+num_queries = 900
+query_dim = 4
+num_patterns = 0
+num_feature_levels = 4
+enc_n_points = 4
+dec_n_points = 4
+two_stage_type = "standard"
+two_stage_bbox_embed_share = False
+two_stage_class_embed_share = False
+transformer_activation = "relu"
+dec_pred_bbox_embed_share = True
+dn_box_noise_scale = 1.0
+dn_label_noise_ratio = 0.5
+dn_label_coef = 1.0
+dn_bbox_coef = 1.0
+embed_init_tgt = True
+dn_labelbook_size = 2000
+max_text_len = 256
+text_encoder_type = "bert-base-uncased"
+use_text_enhancer = True
+use_fusion_layer = True
+use_checkpoint = True
+use_transformer_ckpt = True
+use_text_cross_attention = True
+text_dropout = 0.0
+fusion_dropout = 0.0
+fusion_droppath = 0.1
+sub_sentence_present = True

invokeai/backend/image_util/grounding_segment_anything/groundingdino/datasets/transforms.py

@@ -0,0 +1,299 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+"""
+Transforms and data augmentation for both image + bbox.
+"""
+import os
+import random
+
+import PIL
+import torch
+import torchvision.transforms as T
+import torchvision.transforms.functional as F
+
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.box_ops import box_xyxy_to_cxcywh
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.misc import interpolate
+
+
+def crop(image, target, region):
+    cropped_image = F.crop(image, *region)
+
+    target = target.copy()
+    i, j, h, w = region
+
+    # should we do something wrt the original size?
+    target["size"] = torch.tensor([h, w])
+
+    fields = ["labels", "area", "iscrowd", "positive_map"]
+
+    if "boxes" in target:
+        boxes = target["boxes"]
+        max_size = torch.as_tensor([w, h], dtype=torch.float32)
+        cropped_boxes = boxes - torch.as_tensor([j, i, j, i])
+        cropped_boxes = torch.min(cropped_boxes.reshape(-1, 2, 2), max_size)
+        cropped_boxes = cropped_boxes.clamp(min=0)
+        area = (cropped_boxes[:, 1, :] - cropped_boxes[:, 0, :]).prod(dim=1)
+        target["boxes"] = cropped_boxes.reshape(-1, 4)
+        target["area"] = area
+        fields.append("boxes")
+
+    if "masks" in target:
+        # FIXME should we update the area here if there are no boxes?
+        target["masks"] = target["masks"][:, i : i + h, j : j + w]
+        fields.append("masks")
+
+    # remove elements for which the boxes or masks that have zero area
+    if "boxes" in target or "masks" in target:
+        # favor boxes selection when defining which elements to keep
+        # this is compatible with previous implementation
+        if "boxes" in target:
+            cropped_boxes = target["boxes"].reshape(-1, 2, 2)
+            keep = torch.all(cropped_boxes[:, 1, :] > cropped_boxes[:, 0, :], dim=1)
+        else:
+            keep = target["masks"].flatten(1).any(1)
+
+        for field in fields:
+            if field in target:
+                target[field] = target[field][keep]
+
+    if os.environ.get("IPDB_SHILONG_DEBUG", None) == "INFO":
+        # for debug and visualization only.
+        if "strings_positive" in target:
+            target["strings_positive"] = [_i for _i, _j in zip(target["strings_positive"], keep, strict=False) if _j]
+
+    return cropped_image, target
+
+
+def hflip(image, target):
+    flipped_image = F.hflip(image)
+
+    w, h = image.size
+
+    target = target.copy()
+    if "boxes" in target:
+        boxes = target["boxes"]
+        boxes = boxes[:, [2, 1, 0, 3]] * torch.as_tensor([-1, 1, -1, 1]) + torch.as_tensor([w, 0, w, 0])
+        target["boxes"] = boxes
+
+    if "masks" in target:
+        target["masks"] = target["masks"].flip(-1)
+
+    return flipped_image, target
+
+
+def resize(image, target, size, max_size=None):
+    # size can be min_size (scalar) or (w, h) tuple
+
+    def get_size_with_aspect_ratio(image_size, size, max_size=None):
+        w, h = image_size
+        if max_size is not None:
+            min_original_size = float(min((w, h)))
+            max_original_size = float(max((w, h)))
+            if max_original_size / min_original_size * size > max_size:
+                size = int(round(max_size * min_original_size / max_original_size))
+
+        if (w <= h and w == size) or (h <= w and h == size):
+            return (h, w)
+
+        if w < h:
+            ow = size
+            oh = int(size * h / w)
+        else:
+            oh = size
+            ow = int(size * w / h)
+
+        return (oh, ow)
+
+    def get_size(image_size, size, max_size=None):
+        if isinstance(size, (list, tuple)):
+            return size[::-1]
+        else:
+            return get_size_with_aspect_ratio(image_size, size, max_size)
+
+    size = get_size(image.size, size, max_size)
+    rescaled_image = F.resize(image, size)
+
+    if target is None:
+        return rescaled_image, None
+
+    ratios = tuple(float(s) / float(s_orig) for s, s_orig in zip(rescaled_image.size, image.size, strict=False))
+    ratio_width, ratio_height = ratios
+
+    target = target.copy()
+    if "boxes" in target:
+        boxes = target["boxes"]
+        scaled_boxes = boxes * torch.as_tensor([ratio_width, ratio_height, ratio_width, ratio_height])
+        target["boxes"] = scaled_boxes
+
+    if "area" in target:
+        area = target["area"]
+        scaled_area = area * (ratio_width * ratio_height)
+        target["area"] = scaled_area
+
+    h, w = size
+    target["size"] = torch.tensor([h, w])
+
+    if "masks" in target:
+        target["masks"] = interpolate(target["masks"][:, None].float(), size, mode="nearest")[:, 0] > 0.5
+
+    return rescaled_image, target
+
+
+def pad(image, target, padding):
+    # assumes that we only pad on the bottom right corners
+    padded_image = F.pad(image, (0, 0, padding[0], padding[1]))
+    if target is None:
+        return padded_image, None
+    target = target.copy()
+    # should we do something wrt the original size?
+    target["size"] = torch.tensor(padded_image.size[::-1])
+    if "masks" in target:
+        target["masks"] = torch.nn.functional.pad(target["masks"], (0, padding[0], 0, padding[1]))
+    return padded_image, target
+
+
+class ResizeDebug(object):
+    def __init__(self, size):
+        self.size = size
+
+    def __call__(self, img, target):
+        return resize(img, target, self.size)
+
+
+class RandomCrop(object):
+    def __init__(self, size):
+        self.size = size
+
+    def __call__(self, img, target):
+        region = T.RandomCrop.get_params(img, self.size)
+        return crop(img, target, region)
+
+
+class RandomSizeCrop(object):
+    def __init__(self, min_size: int, max_size: int, respect_boxes: bool = False):
+        # respect_boxes:    True to keep all boxes
+        #                   False to tolerence box filter
+        self.min_size = min_size
+        self.max_size = max_size
+        self.respect_boxes = respect_boxes
+
+    def __call__(self, img: PIL.Image.Image, target: dict):
+        init_boxes = len(target["boxes"])
+        max_patience = 10
+        for i in range(max_patience):
+            w = random.randint(self.min_size, min(img.width, self.max_size))
+            h = random.randint(self.min_size, min(img.height, self.max_size))
+            region = T.RandomCrop.get_params(img, [h, w])
+            result_img, result_target = crop(img, target, region)
+            if not self.respect_boxes or len(result_target["boxes"]) == init_boxes or i == max_patience - 1:
+                return result_img, result_target
+        return result_img, result_target
+
+
+class CenterCrop(object):
+    def __init__(self, size):
+        self.size = size
+
+    def __call__(self, img, target):
+        image_width, image_height = img.size
+        crop_height, crop_width = self.size
+        crop_top = int(round((image_height - crop_height) / 2.0))
+        crop_left = int(round((image_width - crop_width) / 2.0))
+        return crop(img, target, (crop_top, crop_left, crop_height, crop_width))
+
+
+class RandomHorizontalFlip(object):
+    def __init__(self, p=0.5):
+        self.p = p
+
+    def __call__(self, img, target):
+        if random.random() < self.p:
+            return hflip(img, target)
+        return img, target
+
+
+class RandomResize(object):
+    def __init__(self, sizes, max_size=None):
+        assert isinstance(sizes, (list, tuple))
+        self.sizes = sizes
+        self.max_size = max_size
+
+    def __call__(self, img, target=None):
+        size = random.choice(self.sizes)
+        return resize(img, target, size, self.max_size)
+
+
+class RandomPad(object):
+    def __init__(self, max_pad):
+        self.max_pad = max_pad
+
+    def __call__(self, img, target):
+        pad_x = random.randint(0, self.max_pad)
+        pad_y = random.randint(0, self.max_pad)
+        return pad(img, target, (pad_x, pad_y))
+
+
+class RandomSelect(object):
+    """
+    Randomly selects between transforms1 and transforms2,
+    with probability p for transforms1 and (1 - p) for transforms2
+    """
+
+    def __init__(self, transforms1, transforms2, p=0.5):
+        self.transforms1 = transforms1
+        self.transforms2 = transforms2
+        self.p = p
+
+    def __call__(self, img, target):
+        if random.random() < self.p:
+            return self.transforms1(img, target)
+        return self.transforms2(img, target)
+
+
+class ToTensor(object):
+    def __call__(self, img, target):
+        return F.to_tensor(img), target
+
+
+class RandomErasing(object):
+    def __init__(self, *args, **kwargs):
+        self.eraser = T.RandomErasing(*args, **kwargs)
+
+    def __call__(self, img, target):
+        return self.eraser(img), target
+
+
+class Normalize(object):
+    def __init__(self, mean, std):
+        self.mean = mean
+        self.std = std
+
+    def __call__(self, image, target=None):
+        image = F.normalize(image, mean=self.mean, std=self.std)
+        if target is None:
+            return image, None
+        target = target.copy()
+        h, w = image.shape[-2:]
+        if "boxes" in target:
+            boxes = target["boxes"]
+            boxes = box_xyxy_to_cxcywh(boxes)
+            boxes = boxes / torch.tensor([w, h, w, h], dtype=torch.float32)
+            target["boxes"] = boxes
+        return image, target
+
+
+class Compose(object):
+    def __init__(self, transforms):
+        self.transforms = transforms
+
+    def __call__(self, image, target):
+        for t in self.transforms:
+            image, target = t(image, target)
+        return image, target
+
+    def __repr__(self):
+        format_string = self.__class__.__name__ + "("
+        for t in self.transforms:
+            format_string += "\n"
+            format_string += "    {0}".format(t)
+        format_string += "\n)"
+        return format_string

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/__init__.py

@@ -0,0 +1,17 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Conditional DETR
+# Copyright (c) 2021 Microsoft. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copied from DETR (https://github.com/facebookresearch/detr)
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# ------------------------------------------------------------------------
+
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.models.GroundingDINO.groundingdino import (
+    build_groundingdino,
+)

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/backbone/__init__.py

@@ -0,0 +1 @@
+from .backbone import build_backbone

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/backbone/backbone.py

@@ -0,0 +1,217 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Conditional DETR
+# Copyright (c) 2021 Microsoft. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copied from DETR (https://github.com/facebookresearch/detr)
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# ------------------------------------------------------------------------
+
+"""
+Backbone modules.
+"""
+
+from typing import Dict, List
+
+import torch
+import torch.nn.functional as F
+import torchvision
+from torch import nn
+from torchvision.models._utils import IntermediateLayerGetter
+
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.models.GroundingDINO.backbone.position_encoding import (
+    build_position_encoding,
+)
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.models.GroundingDINO.backbone.swin_transformer import (
+    build_swin_transformer,
+)
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.misc import NestedTensor, is_main_process
+
+
+class FrozenBatchNorm2d(torch.nn.Module):
+    """
+    BatchNorm2d where the batch statistics and the affine parameters are fixed.
+
+    Copy-paste from torchvision.misc.ops with added eps before rqsrt,
+    without which any other models than torchvision.models.resnet[18,34,50,101]
+    produce nans.
+    """
+
+    def __init__(self, n):
+        super(FrozenBatchNorm2d, self).__init__()
+        self.register_buffer("weight", torch.ones(n))
+        self.register_buffer("bias", torch.zeros(n))
+        self.register_buffer("running_mean", torch.zeros(n))
+        self.register_buffer("running_var", torch.ones(n))
+
+    def _load_from_state_dict(
+        self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+    ):
+        num_batches_tracked_key = prefix + "num_batches_tracked"
+        if num_batches_tracked_key in state_dict:
+            del state_dict[num_batches_tracked_key]
+
+        super(FrozenBatchNorm2d, self)._load_from_state_dict(
+            state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+        )
+
+    def forward(self, x):
+        # move reshapes to the beginning
+        # to make it fuser-friendly
+        w = self.weight.reshape(1, -1, 1, 1)
+        b = self.bias.reshape(1, -1, 1, 1)
+        rv = self.running_var.reshape(1, -1, 1, 1)
+        rm = self.running_mean.reshape(1, -1, 1, 1)
+        eps = 1e-5
+        scale = w * (rv + eps).rsqrt()
+        bias = b - rm * scale
+        return x * scale + bias
+
+
+class BackboneBase(nn.Module):
+    def __init__(
+        self,
+        backbone: nn.Module,
+        train_backbone: bool,
+        num_channels: int,
+        return_interm_indices: list,
+    ):
+        super().__init__()
+        for name, parameter in backbone.named_parameters():
+            if not train_backbone or "layer2" not in name and "layer3" not in name and "layer4" not in name:
+                parameter.requires_grad_(False)
+
+        return_layers = {}
+        for idx, layer_index in enumerate(return_interm_indices):
+            return_layers.update({"layer{}".format(5 - len(return_interm_indices) + idx): "{}".format(layer_index)})
+
+        # if len:
+        #     if use_stage1_feature:
+        #         return_layers = {"layer1": "0", "layer2": "1", "layer3": "2", "layer4": "3"}
+        #     else:
+        #         return_layers = {"layer2": "0", "layer3": "1", "layer4": "2"}
+        # else:
+        #     return_layers = {'layer4': "0"}
+        self.body = IntermediateLayerGetter(backbone, return_layers=return_layers)
+        self.num_channels = num_channels
+
+    def forward(self, tensor_list: NestedTensor):
+        xs = self.body(tensor_list.tensors)
+        out: Dict[str, NestedTensor] = {}
+        for name, x in xs.items():
+            m = tensor_list.mask
+            assert m is not None
+            mask = F.interpolate(m[None].float(), size=x.shape[-2:]).to(torch.bool)[0]
+            out[name] = NestedTensor(x, mask)
+        # import ipdb; ipdb.set_trace()
+        return out
+
+
+class Backbone(BackboneBase):
+    """ResNet backbone with frozen BatchNorm."""
+
+    def __init__(
+        self,
+        name: str,
+        train_backbone: bool,
+        dilation: bool,
+        return_interm_indices: list,
+        batch_norm=FrozenBatchNorm2d,
+    ):
+        if name in ["resnet18", "resnet34", "resnet50", "resnet101"]:
+            backbone = getattr(torchvision.models, name)(
+                replace_stride_with_dilation=[False, False, dilation],
+                pretrained=is_main_process(),
+                norm_layer=batch_norm,
+            )
+        else:
+            raise NotImplementedError("Why you can get here with name {}".format(name))
+        # num_channels = 512 if name in ('resnet18', 'resnet34') else 2048
+        assert name not in ("resnet18", "resnet34"), "Only resnet50 and resnet101 are available."
+        assert return_interm_indices in [[0, 1, 2, 3], [1, 2, 3], [3]]
+        num_channels_all = [256, 512, 1024, 2048]
+        num_channels = num_channels_all[4 - len(return_interm_indices) :]
+        super().__init__(backbone, train_backbone, num_channels, return_interm_indices)
+
+
+class Joiner(nn.Sequential):
+    def __init__(self, backbone, position_embedding):
+        super().__init__(backbone, position_embedding)
+
+    def forward(self, tensor_list: NestedTensor):
+        xs = self[0](tensor_list)
+        out: List[NestedTensor] = []
+        pos = []
+        for name, x in xs.items():
+            out.append(x)
+            # position encoding
+            pos.append(self[1](x).to(x.tensors.dtype))
+
+        return out, pos
+
+
+def build_backbone(args):
+    """
+    Useful args:
+        - backbone: backbone name
+        - lr_backbone:
+        - dilation
+        - return_interm_indices: available: [0,1,2,3], [1,2,3], [3]
+        - backbone_freeze_keywords:
+        - use_checkpoint: for swin only for now
+
+    """
+    position_embedding = build_position_encoding(args)
+    train_backbone = True
+    if not train_backbone:
+        raise ValueError("Please set lr_backbone > 0")
+    return_interm_indices = args.return_interm_indices
+    assert return_interm_indices in [[0, 1, 2, 3], [1, 2, 3], [3]]
+    args.backbone_freeze_keywords
+    use_checkpoint = getattr(args, "use_checkpoint", False)
+
+    if args.backbone in ["resnet50", "resnet101"]:
+        backbone = Backbone(
+            args.backbone,
+            train_backbone,
+            args.dilation,
+            return_interm_indices,
+            batch_norm=FrozenBatchNorm2d,
+        )
+        bb_num_channels = backbone.num_channels
+    elif args.backbone in [
+        "swin_T_224_1k",
+        "swin_B_224_22k",
+        "swin_B_384_22k",
+        "swin_L_224_22k",
+        "swin_L_384_22k",
+    ]:
+        pretrain_img_size = int(args.backbone.split("_")[-2])
+        backbone = build_swin_transformer(
+            args.backbone,
+            pretrain_img_size=pretrain_img_size,
+            out_indices=tuple(return_interm_indices),
+            dilation=False,
+            use_checkpoint=use_checkpoint,
+        )
+
+        bb_num_channels = backbone.num_features[4 - len(return_interm_indices) :]
+    else:
+        raise NotImplementedError("Unknown backbone {}".format(args.backbone))
+
+    assert len(bb_num_channels) == len(
+        return_interm_indices
+    ), f"len(bb_num_channels) {len(bb_num_channels)} != len(return_interm_indices) {len(return_interm_indices)}"
+
+    model = Joiner(backbone, position_embedding)
+    model.num_channels = bb_num_channels
+    assert isinstance(bb_num_channels, List), "bb_num_channels is expected to be a List but {}".format(
+        type(bb_num_channels)
+    )
+    # import ipdb; ipdb.set_trace()
+    return model

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/backbone/position_encoding.py

@@ -0,0 +1,176 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# DINO
+# Copyright (c) 2022 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Conditional DETR
+# Copyright (c) 2021 Microsoft. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copied from DETR (https://github.com/facebookresearch/detr)
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# ------------------------------------------------------------------------
+
+"""
+Various positional encodings for the transformer.
+"""
+import math
+
+import torch
+from torch import nn
+
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.misc import NestedTensor
+
+
+class PositionEmbeddingSine(nn.Module):
+    """
+    This is a more standard version of the position embedding, very similar to the one
+    used by the Attention is all you need paper, generalized to work on images.
+    """
+
+    def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None):
+        super().__init__()
+        self.num_pos_feats = num_pos_feats
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+    def forward(self, tensor_list: NestedTensor):
+        x = tensor_list.tensors
+        mask = tensor_list.mask
+        assert mask is not None
+        not_mask = ~mask
+        y_embed = not_mask.cumsum(1, dtype=torch.float32)
+        x_embed = not_mask.cumsum(2, dtype=torch.float32)
+        if self.normalize:
+            eps = 1e-6
+            # if os.environ.get("SHILONG_AMP", None) == '1':
+            #     eps = 1e-4
+            # else:
+            #     eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        return pos
+
+
+class PositionEmbeddingSineHW(nn.Module):
+    """
+    This is a more standard version of the position embedding, very similar to the one
+    used by the Attention is all you need paper, generalized to work on images.
+    """
+
+    def __init__(self, num_pos_feats=64, temperatureH=10000, temperatureW=10000, normalize=False, scale=None):
+        super().__init__()
+        self.num_pos_feats = num_pos_feats
+        self.temperatureH = temperatureH
+        self.temperatureW = temperatureW
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+    def forward(self, tensor_list: NestedTensor):
+        x = tensor_list.tensors
+        mask = tensor_list.mask
+        assert mask is not None
+        not_mask = ~mask
+        y_embed = not_mask.cumsum(1, dtype=torch.float32)
+        x_embed = not_mask.cumsum(2, dtype=torch.float32)
+
+        # import ipdb; ipdb.set_trace()
+
+        if self.normalize:
+            eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_tx = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_tx = self.temperatureW ** (2 * (torch.div(dim_tx, 2, rounding_mode="floor")) / self.num_pos_feats)
+        pos_x = x_embed[:, :, :, None] / dim_tx
+
+        dim_ty = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_ty = self.temperatureH ** (2 * (torch.div(dim_ty, 2, rounding_mode="floor")) / self.num_pos_feats)
+        pos_y = y_embed[:, :, :, None] / dim_ty
+
+        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+
+        # import ipdb; ipdb.set_trace()
+
+        return pos
+
+
+class PositionEmbeddingLearned(nn.Module):
+    """
+    Absolute pos embedding, learned.
+    """
+
+    def __init__(self, num_pos_feats=256):
+        super().__init__()
+        self.row_embed = nn.Embedding(50, num_pos_feats)
+        self.col_embed = nn.Embedding(50, num_pos_feats)
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        nn.init.uniform_(self.row_embed.weight)
+        nn.init.uniform_(self.col_embed.weight)
+
+    def forward(self, tensor_list: NestedTensor):
+        x = tensor_list.tensors
+        h, w = x.shape[-2:]
+        i = torch.arange(w, device=x.device)
+        j = torch.arange(h, device=x.device)
+        x_emb = self.col_embed(i)
+        y_emb = self.row_embed(j)
+        pos = (
+            torch.cat(
+                [
+                    x_emb.unsqueeze(0).repeat(h, 1, 1),
+                    y_emb.unsqueeze(1).repeat(1, w, 1),
+                ],
+                dim=-1,
+            )
+            .permute(2, 0, 1)
+            .unsqueeze(0)
+            .repeat(x.shape[0], 1, 1, 1)
+        )
+        return pos
+
+
+def build_position_encoding(args):
+    N_steps = args.hidden_dim // 2
+    if args.position_embedding in ("v2", "sine"):
+        # TODO find a better way of exposing other arguments
+        position_embedding = PositionEmbeddingSineHW(
+            N_steps,
+            temperatureH=args.pe_temperatureH,
+            temperatureW=args.pe_temperatureW,
+            normalize=True,
+        )
+    elif args.position_embedding in ("v3", "learned"):
+        position_embedding = PositionEmbeddingLearned(N_steps)
+    else:
+        raise ValueError(f"not supported {args.position_embedding}")
+
+    return position_embedding

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/backbone/swin_transformer.py

@@ -0,0 +1,766 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# DINO
+# Copyright (c) 2022 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# --------------------------------------------------------
+# modified from https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/blob/master/mmdet/models/backbones/swin_transformer.py
+# --------------------------------------------------------
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.misc import NestedTensor
+
+
+class Mlp(nn.Module):
+    """Multilayer perceptron."""
+
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.0):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    """Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(
+        self,
+        dim,
+        window_size,
+        num_heads,
+        qkv_bias=True,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+    ):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
+        )  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=0.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """Forward function.
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = q @ k.transpose(-2, -1)
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1
+        )  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class SwinTransformerBlock(nn.Module):
+    """Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        window_size=7,
+        shift_size=0,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        act_layer=nn.GELU,
+        norm_layer=nn.LayerNorm,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim,
+            window_size=to_2tuple(self.window_size),
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            qk_scale=qk_scale,
+            attn_drop=attn_drop,
+            proj_drop=drop,
+        )
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        """Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+            mask_matrix: Attention mask for cyclic shift.
+        """
+        B, L, C = x.shape
+        H, W = self.H, self.W
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    """Patch Merging Layer
+    Args:
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x, H, W):
+        """Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of feature channels
+        depth (int): Depths of this stage.
+        num_heads (int): Number of attention head.
+        window_size (int): Local window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(
+        self,
+        dim,
+        depth,
+        num_heads,
+        window_size=7,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        norm_layer=nn.LayerNorm,
+        downsample=None,
+        use_checkpoint=False,
+    ):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList(
+            [
+                SwinTransformerBlock(
+                    dim=dim,
+                    num_heads=num_heads,
+                    window_size=window_size,
+                    shift_size=0 if (i % 2 == 0) else window_size // 2,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop,
+                    attn_drop=attn_drop,
+                    drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                    norm_layer=norm_layer,
+                )
+                for i in range(depth)
+            ]
+        )
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        """Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+
+        # calculate attention mask for SW-MSA
+        Hp = int(np.ceil(H / self.window_size)) * self.window_size
+        Wp = int(np.ceil(W / self.window_size)) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device, dtype=x.dtype)  # 1 Hp Wp 1
+        h_slices = (
+            slice(0, -self.window_size),
+            slice(-self.window_size, -self.shift_size),
+            slice(-self.shift_size, None),
+        )
+        w_slices = (
+            slice(0, -self.window_size),
+            slice(-self.window_size, -self.shift_size),
+            slice(-self.shift_size, None),
+        )
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, attn_mask)
+            else:
+                x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    """Image to Patch Embedding
+    Args:
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer(nn.Module):
+    """Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+    Args:
+        pretrain_img_size (int): Input image size for training the pretrained model,
+            used in absolute postion embedding. Default 224.
+        patch_size (int | tuple(int)): Patch size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        depths (tuple[int]): Depths of each Swin Transformer stage.
+        num_heads (tuple[int]): Number of attention head of each stage.
+        window_size (int): Window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate. Default: 0.
+        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        dilation (bool): if True, the output size if 16x downsample, ow 32x downsample.
+    """
+
+    def __init__(
+        self,
+        pretrain_img_size=224,
+        patch_size=4,
+        in_chans=3,
+        embed_dim=96,
+        depths=[2, 2, 6, 2],
+        num_heads=[3, 6, 12, 24],
+        window_size=7,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop_rate=0.0,
+        attn_drop_rate=0.0,
+        drop_path_rate=0.2,
+        norm_layer=nn.LayerNorm,
+        ape=False,
+        patch_norm=True,
+        out_indices=(0, 1, 2, 3),
+        frozen_stages=-1,
+        dilation=False,
+        use_checkpoint=False,
+    ):
+        super().__init__()
+
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+        self.dilation = dilation
+
+        # if use_checkpoint:
+        #     print("use_checkpoint!!!!!!!!!!!!!!!!!!!!!!!!")
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None,
+        )
+
+        # absolute position embedding
+        if self.ape:
+            pretrain_img_size = to_2tuple(pretrain_img_size)
+            patch_size = to_2tuple(patch_size)
+            patches_resolution = [
+                pretrain_img_size[0] // patch_size[0],
+                pretrain_img_size[1] // patch_size[1],
+            ]
+
+            self.absolute_pos_embed = nn.Parameter(
+                torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1])
+            )
+            trunc_normal_(self.absolute_pos_embed, std=0.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        # prepare downsample list
+        downsamplelist = [PatchMerging for i in range(self.num_layers)]
+        downsamplelist[-1] = None
+        num_features = [int(embed_dim * 2**i) for i in range(self.num_layers)]
+        if self.dilation:
+            downsamplelist[-2] = None
+            num_features[-1] = int(embed_dim * 2 ** (self.num_layers - 1)) // 2
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(
+                # dim=int(embed_dim * 2 ** i_layer),
+                dim=num_features[i_layer],
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])],
+                norm_layer=norm_layer,
+                # downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                downsample=downsamplelist[i_layer],
+                use_checkpoint=use_checkpoint,
+            )
+            self.layers.append(layer)
+
+        # num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        # add a norm layer for each output
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f"norm{i_layer}"
+            self.add_module(layer_name, layer)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        if self.frozen_stages >= 1 and self.ape:
+            self.absolute_pos_embed.requires_grad = False
+
+        if self.frozen_stages >= 2:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages - 1):
+                m = self.layers[i]
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+    # def init_weights(self, pretrained=None):
+    #     """Initialize the weights in backbone.
+    #     Args:
+    #         pretrained (str, optional): Path to pre-trained weights.
+    #             Defaults to None.
+    #     """
+
+    #     def _init_weights(m):
+    #         if isinstance(m, nn.Linear):
+    #             trunc_normal_(m.weight, std=.02)
+    #             if isinstance(m, nn.Linear) and m.bias is not None:
+    #                 nn.init.constant_(m.bias, 0)
+    #         elif isinstance(m, nn.LayerNorm):
+    #             nn.init.constant_(m.bias, 0)
+    #             nn.init.constant_(m.weight, 1.0)
+
+    #     if isinstance(pretrained, str):
+    #         self.apply(_init_weights)
+    #         logger = get_root_logger()
+    #         load_checkpoint(self, pretrained, strict=False, logger=logger)
+    #     elif pretrained is None:
+    #         self.apply(_init_weights)
+    #     else:
+    #         raise TypeError('pretrained must be a str or None')
+
+    def forward_raw(self, x):
+        """Forward function."""
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode="bicubic")
+            x = (x + absolute_pos_embed).flatten(2).transpose(1, 2)  # B Wh*Ww C
+        else:
+            x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+
+        outs = []
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+            # import ipdb; ipdb.set_trace()
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f"norm{i}")
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+        # in:
+        #   torch.Size([2, 3, 1024, 1024])
+        # outs:
+        #   [torch.Size([2, 192, 256, 256]), torch.Size([2, 384, 128, 128]), \
+        #       torch.Size([2, 768, 64, 64]), torch.Size([2, 1536, 32, 32])]
+        return tuple(outs)
+
+    def forward(self, tensor_list: NestedTensor):
+        x = tensor_list.tensors
+
+        """Forward function."""
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode="bicubic")
+            x = (x + absolute_pos_embed).flatten(2).transpose(1, 2)  # B Wh*Ww C
+        else:
+            x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+
+        outs = []
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f"norm{i}")
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+        # in:
+        #   torch.Size([2, 3, 1024, 1024])
+        # out:
+        #   [torch.Size([2, 192, 256, 256]), torch.Size([2, 384, 128, 128]), \
+        #       torch.Size([2, 768, 64, 64]), torch.Size([2, 1536, 32, 32])]
+
+        # collect for nesttensors
+        outs_dict = {}
+        for idx, out_i in enumerate(outs):
+            m = tensor_list.mask
+            assert m is not None
+            mask = F.interpolate(m[None].float(), size=out_i.shape[-2:]).to(torch.bool)[0]
+            outs_dict[idx] = NestedTensor(out_i, mask)
+
+        return outs_dict
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep layers freezed."""
+        super(SwinTransformer, self).train(mode)
+        self._freeze_stages()
+
+
+def build_swin_transformer(modelname, pretrain_img_size, **kw):
+    assert modelname in [
+        "swin_T_224_1k",
+        "swin_B_224_22k",
+        "swin_B_384_22k",
+        "swin_L_224_22k",
+        "swin_L_384_22k",
+    ]
+
+    model_para_dict = {
+        "swin_T_224_1k": dict(embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7),
+        "swin_B_224_22k": dict(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=7),
+        "swin_B_384_22k": dict(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12),
+        "swin_L_224_22k": dict(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=7),
+        "swin_L_384_22k": dict(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12),
+    }
+    kw_cgf = model_para_dict[modelname]
+    kw_cgf.update(kw)
+    model = SwinTransformer(pretrain_img_size=pretrain_img_size, **kw_cgf)
+    return model
+
+
+if __name__ == "__main__":
+    model = build_swin_transformer("swin_L_384_22k", 384, dilation=True)
+    x = torch.rand(2, 3, 1024, 1024)
+    y = model.forward_raw(x)
+    import ipdb
+
+    ipdb.set_trace()
+    x = torch.rand(2, 3, 384, 384)
+    y = model.forward_raw(x)

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/bertwarper.py

@@ -0,0 +1,250 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+
+import torch
+from torch import nn
+from transformers.modeling_outputs import BaseModelOutputWithPoolingAndCrossAttentions
+
+
+class BertModelWarper(nn.Module):
+    def __init__(self, bert_model):
+        super().__init__()
+        # self.bert = bert_modelc
+
+        self.config = bert_model.config
+        self.embeddings = bert_model.embeddings
+        self.encoder = bert_model.encoder
+        self.pooler = bert_model.pooler
+
+        self.get_extended_attention_mask = bert_model.get_extended_attention_mask
+        self.invert_attention_mask = bert_model.invert_attention_mask
+        self.get_head_mask = bert_model.get_head_mask
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            batch_size, seq_length = input_shape
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size, seq_length = input_shape
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+        # if os.environ.get('IPDB_SHILONG_DEBUG', None) == 'INFO':
+        #     import ipdb; ipdb.set_trace()
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class TextEncoderShell(nn.Module):
+    def __init__(self, text_encoder):
+        super().__init__()
+        self.text_encoder = text_encoder
+        self.config = self.text_encoder.config
+
+    def forward(self, **kw):
+        # feed into text encoder
+        return self.text_encoder(**kw)
+
+
+def generate_masks_with_special_tokens(tokenized, special_tokens_list, tokenizer):
+    """Generate attention mask between each pair of special tokens
+    Args:
+        input_ids (torch.Tensor): input ids. Shape: [bs, num_token]
+        special_tokens_mask (list): special tokens mask.
+    Returns:
+        torch.Tensor: attention mask between each special tokens.
+    """
+    input_ids = tokenized["input_ids"]
+    bs, num_token = input_ids.shape
+    # special_tokens_mask: bs, num_token. 1 for special tokens. 0 for normal tokens
+    special_tokens_mask = torch.zeros((bs, num_token), device=input_ids.device).bool()
+    for special_token in special_tokens_list:
+        special_tokens_mask |= input_ids == special_token
+
+    # idxs: each row is a list of indices of special tokens
+    idxs = torch.nonzero(special_tokens_mask)
+
+    # generate attention mask and positional ids
+    attention_mask = torch.eye(num_token, device=input_ids.device).bool().unsqueeze(0).repeat(bs, 1, 1)
+    position_ids = torch.zeros((bs, num_token), device=input_ids.device)
+    previous_col = 0
+    for i in range(idxs.shape[0]):
+        row, col = idxs[i]
+        if (col == 0) or (col == num_token - 1):
+            attention_mask[row, col, col] = True
+            position_ids[row, col] = 0
+        else:
+            attention_mask[row, previous_col + 1 : col + 1, previous_col + 1 : col + 1] = True
+            position_ids[row, previous_col + 1 : col + 1] = torch.arange(0, col - previous_col, device=input_ids.device)
+
+        previous_col = col
+
+    # # padding mask
+    # padding_mask = tokenized['attention_mask']
+    # attention_mask = attention_mask & padding_mask.unsqueeze(1).bool() & padding_mask.unsqueeze(2).bool()
+
+    return attention_mask, position_ids.to(torch.long)
+
+
+def generate_masks_with_special_tokens_and_transfer_map(tokenized, special_tokens_list, tokenizer):
+    """Generate attention mask between each pair of special tokens
+    Args:
+        input_ids (torch.Tensor): input ids. Shape: [bs, num_token]
+        special_tokens_mask (list): special tokens mask.
+    Returns:
+        torch.Tensor: attention mask between each special tokens.
+    """
+    input_ids = tokenized["input_ids"]
+    bs, num_token = input_ids.shape
+    # special_tokens_mask: bs, num_token. 1 for special tokens. 0 for normal tokens
+    special_tokens_mask = torch.zeros((bs, num_token), device=input_ids.device).bool()
+    for special_token in special_tokens_list:
+        special_tokens_mask |= input_ids == special_token
+
+    # idxs: each row is a list of indices of special tokens
+    idxs = torch.nonzero(special_tokens_mask)
+
+    # generate attention mask and positional ids
+    attention_mask = torch.eye(num_token, device=input_ids.device).bool().unsqueeze(0).repeat(bs, 1, 1)
+    position_ids = torch.zeros((bs, num_token), device=input_ids.device)
+    cate_to_token_mask_list = [[] for _ in range(bs)]
+    previous_col = 0
+    for i in range(idxs.shape[0]):
+        row, col = idxs[i]
+        if (col == 0) or (col == num_token - 1):
+            attention_mask[row, col, col] = True
+            position_ids[row, col] = 0
+        else:
+            attention_mask[row, previous_col + 1 : col + 1, previous_col + 1 : col + 1] = True
+            position_ids[row, previous_col + 1 : col + 1] = torch.arange(0, col - previous_col, device=input_ids.device)
+            c2t_maski = torch.zeros((num_token), device=input_ids.device).bool()
+            c2t_maski[previous_col + 1 : col] = True
+            cate_to_token_mask_list[row].append(c2t_maski)
+        previous_col = col
+
+    cate_to_token_mask_list = [
+        torch.stack(cate_to_token_mask_listi, dim=0) for cate_to_token_mask_listi in cate_to_token_mask_list
+    ]
+
+    # # padding mask
+    # padding_mask = tokenized['attention_mask']
+    # attention_mask = attention_mask & padding_mask.unsqueeze(1).bool() & padding_mask.unsqueeze(2).bool()
+
+    return attention_mask, position_ids.to(torch.long), cate_to_token_mask_list

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/fuse_modules.py

@@ -0,0 +1,295 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from timm.models.layers import DropPath
+
+
+class FeatureResizer(nn.Module):
+    """
+    This class takes as input a set of embeddings of dimension C1 and outputs a set of
+    embedding of dimension C2, after a linear transformation, dropout and normalization (LN).
+    """
+
+    def __init__(self, input_feat_size, output_feat_size, dropout, do_ln=True):
+        super().__init__()
+        self.do_ln = do_ln
+        # Object feature encoding
+        self.fc = nn.Linear(input_feat_size, output_feat_size, bias=True)
+        self.layer_norm = nn.LayerNorm(output_feat_size, eps=1e-12)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, encoder_features):
+        x = self.fc(encoder_features)
+        if self.do_ln:
+            x = self.layer_norm(x)
+        output = self.dropout(x)
+        return output
+
+
+def l1norm(X, dim, eps=1e-8):
+    """L1-normalize columns of X"""
+    norm = torch.abs(X).sum(dim=dim, keepdim=True) + eps
+    X = torch.div(X, norm)
+    return X
+
+
+def l2norm(X, dim, eps=1e-8):
+    """L2-normalize columns of X"""
+    norm = torch.pow(X, 2).sum(dim=dim, keepdim=True).sqrt() + eps
+    X = torch.div(X, norm)
+    return X
+
+
+def func_attention(query, context, smooth=1, raw_feature_norm="softmax", eps=1e-8):
+    """
+    query: (n_context, queryL, d)
+    context: (n_context, sourceL, d)
+    """
+    _, queryL = query.size(0), query.size(1)
+    batch_size, sourceL = context.size(0), context.size(1)
+
+    # Get attention
+    # --> (batch, d, queryL)
+    queryT = torch.transpose(query, 1, 2)
+
+    # (batch, sourceL, d)(batch, d, queryL)
+    # --> (batch, sourceL, queryL)
+    attn = torch.bmm(context, queryT)
+    if raw_feature_norm == "softmax":
+        # --> (batch*sourceL, queryL)
+        attn = attn.view(batch_size * sourceL, queryL)
+        attn = nn.Softmax()(attn)
+        # --> (batch, sourceL, queryL)
+        attn = attn.view(batch_size, sourceL, queryL)
+    elif raw_feature_norm == "l2norm":
+        attn = l2norm(attn, 2)
+    elif raw_feature_norm == "clipped_l2norm":
+        attn = nn.LeakyReLU(0.1)(attn)
+        attn = l2norm(attn, 2)
+    else:
+        raise ValueError("unknown first norm type:", raw_feature_norm)
+    # --> (batch, queryL, sourceL)
+    attn = torch.transpose(attn, 1, 2).contiguous()
+    # --> (batch*queryL, sourceL)
+    attn = attn.view(batch_size * queryL, sourceL)
+    attn = nn.Softmax()(attn * smooth)
+    # --> (batch, queryL, sourceL)
+    attn = attn.view(batch_size, queryL, sourceL)
+    # --> (batch, sourceL, queryL)
+    attnT = torch.transpose(attn, 1, 2).contiguous()
+
+    # --> (batch, d, sourceL)
+    contextT = torch.transpose(context, 1, 2)
+    # (batch x d x sourceL)(batch x sourceL x queryL)
+    # --> (batch, d, queryL)
+    weightedContext = torch.bmm(contextT, attnT)
+    # --> (batch, queryL, d)
+    weightedContext = torch.transpose(weightedContext, 1, 2)
+
+    return weightedContext, attnT
+
+
+class BiMultiHeadAttention(nn.Module):
+    def __init__(self, v_dim, l_dim, embed_dim, num_heads, dropout=0.1, cfg=None):
+        super(BiMultiHeadAttention, self).__init__()
+
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.head_dim = embed_dim // num_heads
+        self.v_dim = v_dim
+        self.l_dim = l_dim
+
+        assert (
+            self.head_dim * self.num_heads == self.embed_dim
+        ), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and \
+            `num_heads`: {self.num_heads})."
+        self.scale = self.head_dim ** (-0.5)
+        self.dropout = dropout
+
+        self.v_proj = nn.Linear(self.v_dim, self.embed_dim)
+        self.l_proj = nn.Linear(self.l_dim, self.embed_dim)
+        self.values_v_proj = nn.Linear(self.v_dim, self.embed_dim)
+        self.values_l_proj = nn.Linear(self.l_dim, self.embed_dim)
+
+        self.out_v_proj = nn.Linear(self.embed_dim, self.v_dim)
+        self.out_l_proj = nn.Linear(self.embed_dim, self.l_dim)
+
+        self.stable_softmax_2d = True
+        self.clamp_min_for_underflow = True
+        self.clamp_max_for_overflow = True
+
+        self._reset_parameters()
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def _reset_parameters(self):
+        nn.init.xavier_uniform_(self.v_proj.weight)
+        self.v_proj.bias.data.fill_(0)
+        nn.init.xavier_uniform_(self.l_proj.weight)
+        self.l_proj.bias.data.fill_(0)
+        nn.init.xavier_uniform_(self.values_v_proj.weight)
+        self.values_v_proj.bias.data.fill_(0)
+        nn.init.xavier_uniform_(self.values_l_proj.weight)
+        self.values_l_proj.bias.data.fill_(0)
+        nn.init.xavier_uniform_(self.out_v_proj.weight)
+        self.out_v_proj.bias.data.fill_(0)
+        nn.init.xavier_uniform_(self.out_l_proj.weight)
+        self.out_l_proj.bias.data.fill_(0)
+
+    def forward(self, v, l, attention_mask_v=None, attention_mask_l=None):
+        """_summary_
+
+        Args:
+            v (_type_): bs, n_img, dim
+            l (_type_): bs, n_text, dim
+            attention_mask_v (_type_, optional): _description_. bs, n_img
+            attention_mask_l (_type_, optional): _description_. bs, n_text
+
+        Returns:
+            _type_: _description_
+        """
+        # if os.environ.get('IPDB_SHILONG_DEBUG', None) == 'INFO':
+        #     import ipdb; ipdb.set_trace()
+        bsz, tgt_len, _ = v.size()
+
+        query_states = self.v_proj(v) * self.scale
+        key_states = self._shape(self.l_proj(l), -1, bsz)
+        value_v_states = self._shape(self.values_v_proj(v), -1, bsz)
+        value_l_states = self._shape(self.values_l_proj(l), -1, bsz)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_v_states = value_v_states.view(*proj_shape)
+        value_l_states = value_l_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))  # bs*nhead, nimg, ntxt
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, \
+                    but is {attn_weights.size()}"
+            )
+
+        if self.stable_softmax_2d:
+            attn_weights = attn_weights - attn_weights.max()
+
+        if self.clamp_min_for_underflow:
+            attn_weights = torch.clamp(
+                attn_weights, min=-50000
+            )  # Do not increase -50000, data type half has quite limited range
+        if self.clamp_max_for_overflow:
+            attn_weights = torch.clamp(
+                attn_weights, max=50000
+            )  # Do not increase 50000, data type half has quite limited range
+
+        attn_weights_T = attn_weights.transpose(1, 2)
+        attn_weights_l = attn_weights_T - torch.max(attn_weights_T, dim=-1, keepdim=True)[0]
+        if self.clamp_min_for_underflow:
+            attn_weights_l = torch.clamp(
+                attn_weights_l, min=-50000
+            )  # Do not increase -50000, data type half has quite limited range
+        if self.clamp_max_for_overflow:
+            attn_weights_l = torch.clamp(
+                attn_weights_l, max=50000
+            )  # Do not increase 50000, data type half has quite limited range
+
+        # mask vison for language
+        if attention_mask_v is not None:
+            attention_mask_v = attention_mask_v[:, None, None, :].repeat(1, self.num_heads, 1, 1).flatten(0, 1)
+            attn_weights_l.masked_fill_(attention_mask_v, float("-inf"))
+
+        attn_weights_l = attn_weights_l.softmax(dim=-1)
+
+        # mask language for vision
+        if attention_mask_l is not None:
+            attention_mask_l = attention_mask_l[:, None, None, :].repeat(1, self.num_heads, 1, 1).flatten(0, 1)
+            attn_weights.masked_fill_(attention_mask_l, float("-inf"))
+        attn_weights_v = attn_weights.softmax(dim=-1)
+
+        attn_probs_v = F.dropout(attn_weights_v, p=self.dropout, training=self.training)
+        attn_probs_l = F.dropout(attn_weights_l, p=self.dropout, training=self.training)
+
+        attn_output_v = torch.bmm(attn_probs_v, value_l_states)
+        attn_output_l = torch.bmm(attn_probs_l, value_v_states)
+
+        if attn_output_v.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output_v` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, \
+                    but is {attn_output_v.size()}"
+            )
+
+        if attn_output_l.size() != (bsz * self.num_heads, src_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output_l` should be of size {(bsz, self.num_heads, src_len, self.head_dim)}, \
+                    but is {attn_output_l.size()}"
+            )
+
+        attn_output_v = attn_output_v.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        attn_output_v = attn_output_v.transpose(1, 2)
+        attn_output_v = attn_output_v.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output_l = attn_output_l.view(bsz, self.num_heads, src_len, self.head_dim)
+        attn_output_l = attn_output_l.transpose(1, 2)
+        attn_output_l = attn_output_l.reshape(bsz, src_len, self.embed_dim)
+
+        attn_output_v = self.out_v_proj(attn_output_v)
+        attn_output_l = self.out_l_proj(attn_output_l)
+
+        return attn_output_v, attn_output_l
+
+
+# Bi-Direction MHA (text->image, image->text)
+class BiAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        v_dim,
+        l_dim,
+        embed_dim,
+        num_heads,
+        dropout=0.1,
+        drop_path=0.0,
+        init_values=1e-4,
+        cfg=None,
+    ):
+        """
+        Inputs:
+            embed_dim - Dimensionality of input and attention feature vectors
+            hidden_dim - Dimensionality of hidden layer in feed-forward network
+                         (usually 2-4x larger than embed_dim)
+            num_heads - Number of heads to use in the Multi-Head Attention block
+            dropout - Amount of dropout to apply in the feed-forward network
+        """
+        super(BiAttentionBlock, self).__init__()
+
+        # pre layer norm
+        self.layer_norm_v = nn.LayerNorm(v_dim)
+        self.layer_norm_l = nn.LayerNorm(l_dim)
+        self.attn = BiMultiHeadAttention(
+            v_dim=v_dim, l_dim=l_dim, embed_dim=embed_dim, num_heads=num_heads, dropout=dropout
+        )
+
+        # add layer scale for training stability
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        self.gamma_v = nn.Parameter(init_values * torch.ones((v_dim)), requires_grad=True)
+        self.gamma_l = nn.Parameter(init_values * torch.ones((l_dim)), requires_grad=True)
+
+    def forward(self, v, l, attention_mask_v=None, attention_mask_l=None):
+        v = self.layer_norm_v(v)
+        l = self.layer_norm_l(l)
+        delta_v, delta_l = self.attn(v, l, attention_mask_v=attention_mask_v, attention_mask_l=attention_mask_l)
+        # v, l = v + delta_v, l + delta_l
+        v = v + self.drop_path(self.gamma_v * delta_v)
+        l = l + self.drop_path(self.gamma_l * delta_l)
+        return v, l
+
+    # def forward(self, v:List[torch.Tensor], l, attention_mask_v=None, attention_mask_l=None)

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/groundingdino.py

@@ -0,0 +1,362 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Conditional DETR model and criterion classes.
+# Copyright (c) 2021 Microsoft. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Modified from DETR (https://github.com/facebookresearch/detr)
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# ------------------------------------------------------------------------
+# Modified from Deformable DETR (https://github.com/fundamentalvision/Deformable-DETR)
+# Copyright (c) 2020 SenseTime. All Rights Reserved.
+# ------------------------------------------------------------------------
+import copy
+from typing import List
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util import get_tokenlizer
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.misc import (
+    NestedTensor,
+    inverse_sigmoid,
+    nested_tensor_from_tensor_list,
+)
+
+from ..registry import MODULE_BUILD_FUNCS
+from .backbone import build_backbone
+from .bertwarper import BertModelWarper, generate_masks_with_special_tokens_and_transfer_map
+from .transformer import build_transformer
+from .utils import MLP, ContrastiveEmbed
+
+
+class GroundingDINO(nn.Module):
+    """This is the Cross-Attention Detector module that performs object detection"""
+
+    def __init__(
+        self,
+        backbone,
+        transformer,
+        num_queries,
+        aux_loss=False,
+        iter_update=False,
+        query_dim=2,
+        num_feature_levels=1,
+        nheads=8,
+        # two stage
+        two_stage_type="no",  # ['no', 'standard']
+        dec_pred_bbox_embed_share=True,
+        two_stage_class_embed_share=True,
+        two_stage_bbox_embed_share=True,
+        num_patterns=0,
+        dn_number=100,
+        dn_box_noise_scale=0.4,
+        dn_label_noise_ratio=0.5,
+        dn_labelbook_size=100,
+        text_encoder_type="bert-base-uncased",
+        sub_sentence_present=True,
+        max_text_len=256,
+    ):
+        """Initializes the model.
+        Parameters:
+            backbone: torch module of the backbone to be used. See backbone.py
+            transformer: torch module of the transformer architecture. See transformer.py
+            num_queries: number of object queries, ie detection slot. This is the maximal number of objects
+                         Conditional DETR can detect in a single image. For COCO, we recommend 100 queries.
+            aux_loss: True if auxiliary decoding losses (loss at each decoder layer) are to be used.
+        """
+        super().__init__()
+        self.num_queries = num_queries
+        self.transformer = transformer
+        self.hidden_dim = hidden_dim = transformer.d_model
+        self.num_feature_levels = num_feature_levels
+        self.nheads = nheads
+        self.max_text_len = 256
+        self.sub_sentence_present = sub_sentence_present
+
+        # setting query dim
+        self.query_dim = query_dim
+        assert query_dim == 4
+
+        # for dn training
+        self.num_patterns = num_patterns
+        self.dn_number = dn_number
+        self.dn_box_noise_scale = dn_box_noise_scale
+        self.dn_label_noise_ratio = dn_label_noise_ratio
+        self.dn_labelbook_size = dn_labelbook_size
+
+        # bert
+        self.tokenizer = get_tokenlizer.get_tokenlizer(text_encoder_type)
+        self.bert = get_tokenlizer.get_pretrained_language_model(text_encoder_type)
+        self.bert.pooler.dense.weight.requires_grad_(False)
+        self.bert.pooler.dense.bias.requires_grad_(False)
+        self.bert = BertModelWarper(bert_model=self.bert)
+
+        self.feat_map = nn.Linear(self.bert.config.hidden_size, self.hidden_dim, bias=True)
+        nn.init.constant_(self.feat_map.bias.data, 0)
+        nn.init.xavier_uniform_(self.feat_map.weight.data)
+        # freeze
+
+        # special tokens
+        self.specical_tokens = self.tokenizer.convert_tokens_to_ids(["[CLS]", "[SEP]", ".", "?"])
+
+        # prepare input projection layers
+        if num_feature_levels > 1:
+            num_backbone_outs = len(backbone.num_channels)
+            input_proj_list = []
+            for _ in range(num_backbone_outs):
+                in_channels = backbone.num_channels[_]
+                input_proj_list.append(
+                    nn.Sequential(
+                        nn.Conv2d(in_channels, hidden_dim, kernel_size=1),
+                        nn.GroupNorm(32, hidden_dim),
+                    )
+                )
+            for _ in range(num_feature_levels - num_backbone_outs):
+                input_proj_list.append(
+                    nn.Sequential(
+                        nn.Conv2d(in_channels, hidden_dim, kernel_size=3, stride=2, padding=1),
+                        nn.GroupNorm(32, hidden_dim),
+                    )
+                )
+                in_channels = hidden_dim
+            self.input_proj = nn.ModuleList(input_proj_list)
+        else:
+            assert two_stage_type == "no", "two_stage_type should be no if num_feature_levels=1 !!!"
+            self.input_proj = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        nn.Conv2d(backbone.num_channels[-1], hidden_dim, kernel_size=1),
+                        nn.GroupNorm(32, hidden_dim),
+                    )
+                ]
+            )
+
+        self.backbone = backbone
+        self.aux_loss = aux_loss
+        self.box_pred_damping = None
+
+        self.iter_update = iter_update
+        assert iter_update, "Why not iter_update?"
+
+        # prepare pred layers
+        self.dec_pred_bbox_embed_share = dec_pred_bbox_embed_share
+        # prepare class & box embed
+        _class_embed = ContrastiveEmbed()
+
+        _bbox_embed = MLP(hidden_dim, hidden_dim, 4, 3)
+        nn.init.constant_(_bbox_embed.layers[-1].weight.data, 0)
+        nn.init.constant_(_bbox_embed.layers[-1].bias.data, 0)
+
+        if dec_pred_bbox_embed_share:
+            box_embed_layerlist = [_bbox_embed for i in range(transformer.num_decoder_layers)]
+        else:
+            box_embed_layerlist = [copy.deepcopy(_bbox_embed) for i in range(transformer.num_decoder_layers)]
+        class_embed_layerlist = [_class_embed for i in range(transformer.num_decoder_layers)]
+        self.bbox_embed = nn.ModuleList(box_embed_layerlist)
+        self.class_embed = nn.ModuleList(class_embed_layerlist)
+        self.transformer.decoder.bbox_embed = self.bbox_embed
+        self.transformer.decoder.class_embed = self.class_embed
+
+        # two stage
+        self.two_stage_type = two_stage_type
+        assert two_stage_type in ["no", "standard"], "unknown param {} of two_stage_type".format(two_stage_type)
+        if two_stage_type != "no":
+            if two_stage_bbox_embed_share:
+                assert dec_pred_bbox_embed_share
+                self.transformer.enc_out_bbox_embed = _bbox_embed
+            else:
+                self.transformer.enc_out_bbox_embed = copy.deepcopy(_bbox_embed)
+
+            if two_stage_class_embed_share:
+                assert dec_pred_bbox_embed_share
+                self.transformer.enc_out_class_embed = _class_embed
+            else:
+                self.transformer.enc_out_class_embed = copy.deepcopy(_class_embed)
+
+            self.refpoint_embed = None
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        # init input_proj
+        for proj in self.input_proj:
+            nn.init.xavier_uniform_(proj[0].weight, gain=1)
+            nn.init.constant_(proj[0].bias, 0)
+
+    def init_ref_points(self, use_num_queries):
+        self.refpoint_embed = nn.Embedding(use_num_queries, self.query_dim)
+
+    def forward(self, samples: NestedTensor, targets: List = None, **kw):
+        """The forward expects a NestedTensor, which consists of:
+           - samples.tensor: batched images, of shape [batch_size x 3 x H x W]
+           - samples.mask: a binary mask of shape [batch_size x H x W], containing 1 on padded pixels
+
+        It returns a dict with the following elements:
+           - "pred_logits": the classification logits (including no-object) for all queries.
+                            Shape= [batch_size x num_queries x num_classes]
+           - "pred_boxes": The normalized boxes coordinates for all queries, represented as
+                           (center_x, center_y, width, height). These values are normalized in [0, 1],
+                           relative to the size of each individual image (disregarding possible padding).
+                           See PostProcess for information on how to retrieve the unnormalized bounding box.
+           - "aux_outputs": Optional, only returned when auxilary losses are activated. It is a list of
+                            dictionnaries containing the two above keys for each decoder layer.
+        """
+        if targets is None:
+            captions = kw["captions"]
+        else:
+            captions = [t["caption"] for t in targets]
+        len(captions)
+
+        # encoder texts
+        tokenized = self.tokenizer(captions, padding="longest", return_tensors="pt").to(samples.device)
+        (
+            text_self_attention_masks,
+            position_ids,
+            cate_to_token_mask_list,
+        ) = generate_masks_with_special_tokens_and_transfer_map(tokenized, self.specical_tokens, self.tokenizer)
+
+        if text_self_attention_masks.shape[1] > self.max_text_len:
+            text_self_attention_masks = text_self_attention_masks[:, : self.max_text_len, : self.max_text_len]
+            position_ids = position_ids[:, : self.max_text_len]
+            tokenized["input_ids"] = tokenized["input_ids"][:, : self.max_text_len]
+            tokenized["attention_mask"] = tokenized["attention_mask"][:, : self.max_text_len]
+            tokenized["token_type_ids"] = tokenized["token_type_ids"][:, : self.max_text_len]
+
+        # extract text embeddings
+        if self.sub_sentence_present:
+            tokenized_for_encoder = {k: v for k, v in tokenized.items() if k != "attention_mask"}
+            tokenized_for_encoder["attention_mask"] = text_self_attention_masks
+            tokenized_for_encoder["position_ids"] = position_ids
+        else:
+            # import ipdb; ipdb.set_trace()
+            tokenized_for_encoder = tokenized
+
+        bert_output = self.bert(**tokenized_for_encoder)  # bs, 195, 768
+
+        encoded_text = self.feat_map(bert_output["last_hidden_state"])  # bs, 195, d_model
+        text_token_mask = tokenized.attention_mask.bool()  # bs, 195
+        # text_token_mask: True for nomask, False for mask
+        # text_self_attention_masks: True for nomask, False for mask
+
+        if encoded_text.shape[1] > self.max_text_len:
+            encoded_text = encoded_text[:, : self.max_text_len, :]
+            text_token_mask = text_token_mask[:, : self.max_text_len]
+            position_ids = position_ids[:, : self.max_text_len]
+            text_self_attention_masks = text_self_attention_masks[:, : self.max_text_len, : self.max_text_len]
+
+        text_dict = {
+            "encoded_text": encoded_text,  # bs, 195, d_model
+            "text_token_mask": text_token_mask,  # bs, 195
+            "position_ids": position_ids,  # bs, 195
+            "text_self_attention_masks": text_self_attention_masks,  # bs, 195,195
+        }
+
+        # import ipdb; ipdb.set_trace()
+
+        if isinstance(samples, (list, torch.Tensor)):
+            samples = nested_tensor_from_tensor_list(samples)
+        features, poss = self.backbone(samples)
+
+        srcs = []
+        masks = []
+        for l, feat in enumerate(features):
+            src, mask = feat.decompose()
+            srcs.append(self.input_proj[l](src))
+            masks.append(mask)
+            assert mask is not None
+        if self.num_feature_levels > len(srcs):
+            _len_srcs = len(srcs)
+            for l in range(_len_srcs, self.num_feature_levels):
+                if l == _len_srcs:
+                    src = self.input_proj[l](features[-1].tensors)
+                else:
+                    src = self.input_proj[l](srcs[-1])
+                m = samples.mask
+                mask = F.interpolate(m[None].float(), size=src.shape[-2:]).to(torch.bool)[0]
+                pos_l = self.backbone[1](NestedTensor(src, mask)).to(src.dtype)
+                srcs.append(src)
+                masks.append(mask)
+                poss.append(pos_l)
+
+        input_query_bbox = input_query_label = attn_mask = None
+        hs, reference, hs_enc, ref_enc, init_box_proposal = self.transformer(
+            srcs, masks, input_query_bbox, poss, input_query_label, attn_mask, text_dict
+        )
+
+        # deformable-detr-like anchor update
+        outputs_coord_list = []
+        for dec_lid, (layer_ref_sig, layer_bbox_embed, layer_hs) in enumerate(zip(reference[:-1], self.bbox_embed, hs)):
+            layer_delta_unsig = layer_bbox_embed(layer_hs)
+            layer_outputs_unsig = layer_delta_unsig + inverse_sigmoid(layer_ref_sig)
+            layer_outputs_unsig = layer_outputs_unsig.sigmoid()
+            outputs_coord_list.append(layer_outputs_unsig)
+        outputs_coord_list = torch.stack(outputs_coord_list)
+
+        # output
+        outputs_class = torch.stack(
+            [layer_cls_embed(layer_hs, text_dict) for layer_cls_embed, layer_hs in zip(self.class_embed, hs)]
+        )
+        out = {"pred_logits": outputs_class[-1], "pred_boxes": outputs_coord_list[-1]}
+
+        # # for intermediate outputs
+        # if self.aux_loss:
+        #     out['aux_outputs'] = self._set_aux_loss(outputs_class, outputs_coord_list)
+
+        # # for encoder output
+        # if hs_enc is not None:
+        #     # prepare intermediate outputs
+        #     interm_coord = ref_enc[-1]
+        #     interm_class = self.transformer.enc_out_class_embed(hs_enc[-1], text_dict)
+        #     out['interm_outputs'] = {'pred_logits': interm_class, 'pred_boxes': interm_coord}
+        #     out['interm_outputs_for_matching_pre'] = {'pred_logits': interm_class, 'pred_boxes': init_box_proposal}
+
+        return out
+
+    @torch.jit.unused
+    def _set_aux_loss(self, outputs_class, outputs_coord):
+        # this is a workaround to make torchscript happy, as torchscript
+        # doesn't support dictionary with non-homogeneous values, such
+        # as a dict having both a Tensor and a list.
+        return [{"pred_logits": a, "pred_boxes": b} for a, b in zip(outputs_class[:-1], outputs_coord[:-1])]
+
+
+@MODULE_BUILD_FUNCS.registe_with_name(module_name="groundingdino")
+def build_groundingdino(args):
+
+    backbone = build_backbone(args)
+    transformer = build_transformer(args)
+
+    dn_labelbook_size = args.dn_labelbook_size
+    dec_pred_bbox_embed_share = args.dec_pred_bbox_embed_share
+    sub_sentence_present = args.sub_sentence_present
+
+    model = GroundingDINO(
+        backbone,
+        transformer,
+        num_queries=args.num_queries,
+        aux_loss=True,
+        iter_update=True,
+        query_dim=4,
+        num_feature_levels=args.num_feature_levels,
+        nheads=args.nheads,
+        dec_pred_bbox_embed_share=dec_pred_bbox_embed_share,
+        two_stage_type=args.two_stage_type,
+        two_stage_bbox_embed_share=args.two_stage_bbox_embed_share,
+        two_stage_class_embed_share=args.two_stage_class_embed_share,
+        num_patterns=args.num_patterns,
+        dn_number=0,
+        dn_box_noise_scale=args.dn_box_noise_scale,
+        dn_label_noise_ratio=args.dn_label_noise_ratio,
+        dn_labelbook_size=dn_labelbook_size,
+        text_encoder_type=args.text_encoder_type,
+        sub_sentence_present=sub_sentence_present,
+        max_text_len=args.max_text_len,
+    )
+
+    return model

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/ms_deform_attn.py

@@ -0,0 +1,340 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Deformable DETR
+# Copyright (c) 2020 SenseTime. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------------------------------
+# Modified from:
+# https://github.com/fundamentalvision/Deformable-DETR/blob/main/models/ops/functions/ms_deform_attn_func.py
+# https://github.com/fundamentalvision/Deformable-DETR/blob/main/models/ops/modules/ms_deform_attn.py
+# https://github.com/open-mmlab/mmcv/blob/master/mmcv/ops/multi_scale_deform_attn.py
+# ------------------------------------------------------------------------------------------------
+
+import math
+import warnings
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.init import constant_, xavier_uniform_
+
+
+# helpers
+def _is_power_of_2(n):
+    if (not isinstance(n, int)) or (n < 0):
+        raise ValueError("invalid input for _is_power_of_2: {} (type: {})".format(n, type(n)))
+    return (n & (n - 1) == 0) and n != 0
+
+
+def multi_scale_deformable_attn_pytorch(
+    value: torch.Tensor,
+    value_spatial_shapes: torch.Tensor,
+    sampling_locations: torch.Tensor,
+    attention_weights: torch.Tensor,
+) -> torch.Tensor:
+
+    bs, _, num_heads, embed_dims = value.shape
+    _, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape
+    value_list = value.split([H_ * W_ for H_, W_ in value_spatial_shapes], dim=1)
+    sampling_grids = 2 * sampling_locations - 1
+    sampling_value_list = []
+    for level, (H_, W_) in enumerate(value_spatial_shapes):
+        # bs, H_*W_, num_heads, embed_dims ->
+        # bs, H_*W_, num_heads*embed_dims ->
+        # bs, num_heads*embed_dims, H_*W_ ->
+        # bs*num_heads, embed_dims, H_, W_
+        value_l_ = value_list[level].flatten(2).transpose(1, 2).reshape(bs * num_heads, embed_dims, H_, W_)
+        # bs, num_queries, num_heads, num_points, 2 ->
+        # bs, num_heads, num_queries, num_points, 2 ->
+        # bs*num_heads, num_queries, num_points, 2
+        sampling_grid_l_ = sampling_grids[:, :, :, level].transpose(1, 2).flatten(0, 1)
+        # bs*num_heads, embed_dims, num_queries, num_points
+        sampling_value_l_ = F.grid_sample(
+            value_l_, sampling_grid_l_, mode="bilinear", padding_mode="zeros", align_corners=False
+        )
+        sampling_value_list.append(sampling_value_l_)
+    # (bs, num_queries, num_heads, num_levels, num_points) ->
+    # (bs, num_heads, num_queries, num_levels, num_points) ->
+    # (bs, num_heads, 1, num_queries, num_levels*num_points)
+    attention_weights = attention_weights.transpose(1, 2).reshape(
+        bs * num_heads, 1, num_queries, num_levels * num_points
+    )
+    output = (
+        (torch.stack(sampling_value_list, dim=-2).flatten(-2) * attention_weights)
+        .sum(-1)
+        .view(bs, num_heads * embed_dims, num_queries)
+    )
+    return output.transpose(1, 2).contiguous()
+
+
+class MultiScaleDeformableAttention(nn.Module):
+    """Multi-Scale Deformable Attention Module used in Deformable-DETR
+
+    `Deformable DETR: Deformable Transformers for End-to-End Object Detection.
+    <https://arxiv.org/pdf/2010.04159.pdf>`_.
+
+    Args:
+        embed_dim (int): The embedding dimension of Attention. Default: 256.
+        num_heads (int): The number of attention heads. Default: 8.
+        num_levels (int): The number of feature map used in Attention. Default: 4.
+        num_points (int): The number of sampling points for each query
+            in each head. Default: 4.
+        img2col_steps (int): The step used in image_to_column. Defualt: 64.
+            dropout (float): Dropout layer used in output. Default: 0.1.
+        batch_first (bool): if ``True``, then the input and output tensor will be
+            provided as `(bs, n, embed_dim)`. Default: False. `(n, bs, embed_dim)`
+    """
+
+    def __init__(
+        self,
+        embed_dim: int = 256,
+        num_heads: int = 8,
+        num_levels: int = 4,
+        num_points: int = 4,
+        img2col_step: int = 64,
+        batch_first: bool = False,
+    ):
+        super().__init__()
+        if embed_dim % num_heads != 0:
+            raise ValueError("embed_dim must be divisible by num_heads, but got {} and {}".format(embed_dim, num_heads))
+        head_dim = embed_dim // num_heads
+
+        self.batch_first = batch_first
+
+        if not _is_power_of_2(head_dim):
+            warnings.warn(
+                """
+                You'd better set d_model in MSDeformAttn to make sure that
+                each dim of the attention head a power of 2, which is more efficient.
+                """
+            )
+
+        self.im2col_step = img2col_step
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.num_levels = num_levels
+        self.num_points = num_points
+        self.sampling_offsets = nn.Linear(embed_dim, num_heads * num_levels * num_points * 2)
+        self.attention_weights = nn.Linear(embed_dim, num_heads * num_levels * num_points)
+        self.value_proj = nn.Linear(embed_dim, embed_dim)
+        self.output_proj = nn.Linear(embed_dim, embed_dim)
+
+        self.init_weights()
+
+    def _reset_parameters(self):
+        return self.init_weights()
+
+    def init_weights(self):
+        """
+        Default initialization for Parameters of Module.
+        """
+        constant_(self.sampling_offsets.weight.data, 0.0)
+        thetas = torch.arange(self.num_heads, dtype=torch.float32) * (2.0 * math.pi / self.num_heads)
+        grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
+        grid_init = (
+            (grid_init / grid_init.abs().max(-1, keepdim=True)[0])
+            .view(self.num_heads, 1, 1, 2)
+            .repeat(1, self.num_levels, self.num_points, 1)
+        )
+        for i in range(self.num_points):
+            grid_init[:, :, i, :] *= i + 1
+        with torch.no_grad():
+            self.sampling_offsets.bias = nn.Parameter(grid_init.view(-1))
+        constant_(self.attention_weights.weight.data, 0.0)
+        constant_(self.attention_weights.bias.data, 0.0)
+        xavier_uniform_(self.value_proj.weight.data)
+        constant_(self.value_proj.bias.data, 0.0)
+        xavier_uniform_(self.output_proj.weight.data)
+        constant_(self.output_proj.bias.data, 0.0)
+
+    def freeze_sampling_offsets(self):
+        print("Freeze sampling offsets")
+        self.sampling_offsets.weight.requires_grad = False
+        self.sampling_offsets.bias.requires_grad = False
+
+    def freeze_attention_weights(self):
+        print("Freeze attention weights")
+        self.attention_weights.weight.requires_grad = False
+        self.attention_weights.bias.requires_grad = False
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: Optional[torch.Tensor] = None,
+        value: Optional[torch.Tensor] = None,
+        query_pos: Optional[torch.Tensor] = None,
+        key_padding_mask: Optional[torch.Tensor] = None,
+        reference_points: Optional[torch.Tensor] = None,
+        spatial_shapes: Optional[torch.Tensor] = None,
+        level_start_index: Optional[torch.Tensor] = None,
+        **kwargs
+    ) -> torch.Tensor:
+        """Forward Function of MultiScaleDeformableAttention
+
+        Args:
+            query (torch.Tensor): Query embeddings with shape
+                `(num_query, bs, embed_dim)`
+            key (torch.Tensor): Key embeddings with shape
+                `(num_key, bs, embed_dim)`
+            value (torch.Tensor): Value embeddings with shape
+                `(num_key, bs, embed_dim)`
+            query_pos (torch.Tensor): The position embedding for `query`. Default: None.
+            key_padding_mask (torch.Tensor): ByteTensor for `query`, with shape `(bs, num_key)`,
+                indicating which elements within `key` to be ignored in attention.
+            reference_points (torch.Tensor): The normalized reference points
+                with shape `(bs, num_query, num_levels, 2)`,
+                all elements is range in [0, 1], top-left (0, 0),
+                bottom-right (1, 1), including padding are.
+                or `(N, Length_{query}, num_levels, 4)`, add additional
+                two dimensions `(h, w)` to form reference boxes.
+            spatial_shapes (torch.Tensor): Spatial shape of features in different levels.
+                With shape `(num_levels, 2)`, last dimension represents `(h, w)`.
+            level_start_index (torch.Tensor): The start index of each level. A tensor with
+                shape `(num_levels, )` which can be represented as
+                `[0, h_0 * w_0, h_0 * w_0 + h_1 * w_1, ...]`.
+
+        Returns:
+            torch.Tensor: forward results with shape `(num_query, bs, embed_dim)`
+        """
+
+        if value is None:
+            value = query
+
+        if query_pos is not None:
+            query = query + query_pos
+
+        if not self.batch_first:
+            # change to (bs, num_query ,embed_dims)
+            query = query.permute(1, 0, 2)
+            value = value.permute(1, 0, 2)
+
+        bs, num_query, _ = query.shape
+        bs, num_value, _ = value.shape
+
+        assert (spatial_shapes[:, 0] * spatial_shapes[:, 1]).sum() == num_value
+
+        value = self.value_proj(value)
+        if key_padding_mask is not None:
+            value = value.masked_fill(key_padding_mask[..., None], float(0))
+        value = value.view(bs, num_value, self.num_heads, -1)
+        sampling_offsets = self.sampling_offsets(query).view(
+            bs, num_query, self.num_heads, self.num_levels, self.num_points, 2
+        )
+        attention_weights = self.attention_weights(query).view(
+            bs, num_query, self.num_heads, self.num_levels * self.num_points
+        )
+        attention_weights = attention_weights.softmax(-1)
+        attention_weights = attention_weights.view(
+            bs,
+            num_query,
+            self.num_heads,
+            self.num_levels,
+            self.num_points,
+        )
+
+        # bs, num_query, num_heads, num_levels, num_points, 2
+        if reference_points.shape[-1] == 2:
+            offset_normalizer = torch.stack([spatial_shapes[..., 1], spatial_shapes[..., 0]], -1)
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :]
+                + sampling_offsets / offset_normalizer[None, None, None, :, None, :]
+            )
+        elif reference_points.shape[-1] == 4:
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :2]
+                + sampling_offsets / self.num_points * reference_points[:, :, None, :, None, 2:] * 0.5
+            )
+        else:
+            raise ValueError(
+                "Last dim of reference_points must be 2 or 4, but get {} instead.".format(reference_points.shape[-1])
+            )
+
+        # if torch.cuda.is_available() and value.is_cuda:
+        #     halffloat = False
+        #     if value.dtype == torch.float16:
+        #         halffloat = True
+        #         value = value.float()
+        #         sampling_locations = sampling_locations.float()
+        #         attention_weights = attention_weights.float()
+
+        #     output = MultiScaleDeformableAttnFunction.apply(
+        #         value,
+        #         spatial_shapes,
+        #         level_start_index,
+        #         sampling_locations,
+        #         attention_weights,
+        #         self.im2col_step,
+        #     )
+
+        #     if halffloat:
+        #         output = output.half()
+        # else:
+        #     output = multi_scale_deformable_attn_pytorch(value, spatial_shapes, sampling_locations, attention_weights)
+
+        output = multi_scale_deformable_attn_pytorch(value, spatial_shapes, sampling_locations, attention_weights)
+
+        output = self.output_proj(output)
+
+        if not self.batch_first:
+            output = output.permute(1, 0, 2)
+
+        return output
+
+
+def create_dummy_class(klass, dependency, message=""):
+    """
+    When a dependency of a class is not available, create a dummy class which throws ImportError
+    when used.
+
+    Args:
+        klass (str): name of the class.
+        dependency (str): name of the dependency.
+        message: extra message to print
+    Returns:
+        class: a class object
+    """
+    err = "Cannot import '{}', therefore '{}' is not available.".format(dependency, klass)
+    if message:
+        err = err + " " + message
+
+    class _DummyMetaClass(type):
+        # throw error on class attribute access
+        def __getattr__(_, __):  # noqa: B902
+            raise ImportError(err)
+
+    class _Dummy(object, metaclass=_DummyMetaClass):
+        # throw error on constructor
+        def __init__(self, *args, **kwargs):
+            raise ImportError(err)
+
+    return _Dummy
+
+
+def create_dummy_func(func, dependency, message=""):
+    """
+    When a dependency of a function is not available, create a dummy function which throws
+    ImportError when used.
+
+    Args:
+        func (str): name of the function.
+        dependency (str or list[str]): name(s) of the dependency.
+        message: extra message to print
+    Returns:
+        function: a function object
+    """
+    err = "Cannot import '{}', therefore '{}' is not available.".format(dependency, func)
+    if message:
+        err = err + " " + message
+
+    if isinstance(dependency, (list, tuple)):
+        dependency = ",".join(dependency)
+
+    def _dummy(*args, **kwargs):
+        raise ImportError(err)
+
+    return _dummy

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/transformer.py

@@ -0,0 +1,927 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# DINO
+# Copyright (c) 2022 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Conditional DETR Transformer class.
+# Copyright (c) 2021 Microsoft. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Modified from DETR (https://github.com/facebookresearch/detr)
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# ------------------------------------------------------------------------
+
+from typing import Optional
+
+import torch
+import torch.utils.checkpoint as checkpoint
+from torch import Tensor, nn
+
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.misc import inverse_sigmoid
+
+from .fuse_modules import BiAttentionBlock
+from .ms_deform_attn import MultiScaleDeformableAttention as MSDeformAttn
+from .transformer_vanilla import TransformerEncoderLayer
+from .utils import (
+    MLP,
+    _get_activation_fn,
+    _get_clones,
+    gen_encoder_output_proposals,
+    gen_sineembed_for_position,
+    get_sine_pos_embed,
+)
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        d_model=256,
+        nhead=8,
+        num_queries=300,
+        num_encoder_layers=6,
+        num_unicoder_layers=0,
+        num_decoder_layers=6,
+        dim_feedforward=2048,
+        dropout=0.0,
+        activation="relu",
+        normalize_before=False,
+        return_intermediate_dec=False,
+        query_dim=4,
+        num_patterns=0,
+        # for deformable encoder
+        num_feature_levels=1,
+        enc_n_points=4,
+        dec_n_points=4,
+        # init query
+        learnable_tgt_init=False,
+        # two stage
+        two_stage_type="no",  # ['no', 'standard', 'early', 'combine', 'enceachlayer', 'enclayer1']
+        embed_init_tgt=False,
+        # for text
+        use_text_enhancer=False,
+        use_fusion_layer=False,
+        use_checkpoint=False,
+        use_transformer_ckpt=False,
+        use_text_cross_attention=False,
+        text_dropout=0.1,
+        fusion_dropout=0.1,
+        fusion_droppath=0.0,
+    ):
+        super().__init__()
+        self.num_feature_levels = num_feature_levels
+        self.num_encoder_layers = num_encoder_layers
+        self.num_unicoder_layers = num_unicoder_layers
+        self.num_decoder_layers = num_decoder_layers
+        self.num_queries = num_queries
+        assert query_dim == 4
+
+        # choose encoder layer type
+        encoder_layer = DeformableTransformerEncoderLayer(
+            d_model, dim_feedforward, dropout, activation, num_feature_levels, nhead, enc_n_points
+        )
+
+        if use_text_enhancer:
+            text_enhance_layer = TransformerEncoderLayer(
+                d_model=d_model,
+                nhead=nhead // 2,
+                dim_feedforward=dim_feedforward // 2,
+                dropout=text_dropout,
+            )
+        else:
+            text_enhance_layer = None
+
+        if use_fusion_layer:
+            feature_fusion_layer = BiAttentionBlock(
+                v_dim=d_model,
+                l_dim=d_model,
+                embed_dim=dim_feedforward // 2,
+                num_heads=nhead // 2,
+                dropout=fusion_dropout,
+                drop_path=fusion_droppath,
+            )
+        else:
+            feature_fusion_layer = None
+
+        encoder_norm = nn.LayerNorm(d_model) if normalize_before else None
+        assert encoder_norm is None
+        self.encoder = TransformerEncoder(
+            encoder_layer,
+            num_encoder_layers,
+            d_model=d_model,
+            num_queries=num_queries,
+            text_enhance_layer=text_enhance_layer,
+            feature_fusion_layer=feature_fusion_layer,
+            use_checkpoint=use_checkpoint,
+            use_transformer_ckpt=use_transformer_ckpt,
+        )
+
+        # choose decoder layer type
+        decoder_layer = DeformableTransformerDecoderLayer(
+            d_model,
+            dim_feedforward,
+            dropout,
+            activation,
+            num_feature_levels,
+            nhead,
+            dec_n_points,
+            use_text_cross_attention=use_text_cross_attention,
+        )
+
+        decoder_norm = nn.LayerNorm(d_model)
+        self.decoder = TransformerDecoder(
+            decoder_layer,
+            num_decoder_layers,
+            decoder_norm,
+            return_intermediate=return_intermediate_dec,
+            d_model=d_model,
+            query_dim=query_dim,
+            num_feature_levels=num_feature_levels,
+        )
+
+        self.d_model = d_model
+        self.nhead = nhead
+        self.dec_layers = num_decoder_layers
+        self.num_queries = num_queries  # useful for single stage model only
+        self.num_patterns = num_patterns
+        if not isinstance(num_patterns, int):
+            Warning("num_patterns should be int but {}".format(type(num_patterns)))
+            self.num_patterns = 0
+
+        if num_feature_levels > 1:
+            if self.num_encoder_layers > 0:
+                self.level_embed = nn.Parameter(torch.Tensor(num_feature_levels, d_model))
+            else:
+                self.level_embed = None
+
+        self.learnable_tgt_init = learnable_tgt_init
+        assert learnable_tgt_init, "why not learnable_tgt_init"
+        self.embed_init_tgt = embed_init_tgt
+        if (two_stage_type != "no" and embed_init_tgt) or (two_stage_type == "no"):
+            self.tgt_embed = nn.Embedding(self.num_queries, d_model)
+            nn.init.normal_(self.tgt_embed.weight.data)
+        else:
+            self.tgt_embed = None
+
+        # for two stage
+        self.two_stage_type = two_stage_type
+        assert two_stage_type in ["no", "standard"], "unknown param {} of two_stage_type".format(two_stage_type)
+        if two_stage_type == "standard":
+            # anchor selection at the output of encoder
+            self.enc_output = nn.Linear(d_model, d_model)
+            self.enc_output_norm = nn.LayerNorm(d_model)
+            self.two_stage_wh_embedding = None
+
+        if two_stage_type == "no":
+            self.init_ref_points(num_queries)  # init self.refpoint_embed
+
+        self.enc_out_class_embed = None
+        self.enc_out_bbox_embed = None
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        for p in self.parameters():
+            if p.dim() > 1:
+                nn.init.xavier_uniform_(p)
+        for m in self.modules():
+            if isinstance(m, MSDeformAttn):
+                m._reset_parameters()
+        if self.num_feature_levels > 1 and self.level_embed is not None:
+            nn.init.normal_(self.level_embed)
+
+    def get_valid_ratio(self, mask):
+        _, H, W = mask.shape
+        valid_H = torch.sum(~mask[:, :, 0], 1)
+        valid_W = torch.sum(~mask[:, 0, :], 1)
+        valid_ratio_h = valid_H.float() / H
+        valid_ratio_w = valid_W.float() / W
+        valid_ratio = torch.stack([valid_ratio_w, valid_ratio_h], -1)
+        return valid_ratio
+
+    def init_ref_points(self, use_num_queries):
+        self.refpoint_embed = nn.Embedding(use_num_queries, 4)
+
+    def forward(self, srcs, masks, refpoint_embed, pos_embeds, tgt, attn_mask=None, text_dict=None):
+        """
+        Input:
+            - srcs: List of multi features [bs, ci, hi, wi]
+            - masks: List of multi masks [bs, hi, wi]
+            - refpoint_embed: [bs, num_dn, 4]. None in infer
+            - pos_embeds: List of multi pos embeds [bs, ci, hi, wi]
+            - tgt: [bs, num_dn, d_model]. None in infer
+
+        """
+        # prepare input for encoder
+        src_flatten = []
+        mask_flatten = []
+        lvl_pos_embed_flatten = []
+        spatial_shapes = []
+        for lvl, (src, mask, pos_embed) in enumerate(zip(srcs, masks, pos_embeds)):
+            bs, c, h, w = src.shape
+            spatial_shape = (h, w)
+            spatial_shapes.append(spatial_shape)
+
+            src = src.flatten(2).transpose(1, 2)  # bs, hw, c
+            mask = mask.flatten(1)  # bs, hw
+            pos_embed = pos_embed.flatten(2).transpose(1, 2)  # bs, hw, c
+            if self.num_feature_levels > 1 and self.level_embed is not None:
+                lvl_pos_embed = pos_embed + self.level_embed[lvl].view(1, 1, -1)
+            else:
+                lvl_pos_embed = pos_embed
+            lvl_pos_embed_flatten.append(lvl_pos_embed)
+            src_flatten.append(src)
+            mask_flatten.append(mask)
+        src_flatten = torch.cat(src_flatten, 1)  # bs, \sum{hxw}, c
+        mask_flatten = torch.cat(mask_flatten, 1)  # bs, \sum{hxw}
+        lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)  # bs, \sum{hxw}, c
+        spatial_shapes = torch.as_tensor(spatial_shapes, dtype=torch.long, device=src_flatten.device)
+        level_start_index = torch.cat((spatial_shapes.new_zeros((1,)), spatial_shapes.prod(1).cumsum(0)[:-1]))
+        valid_ratios = torch.stack([self.get_valid_ratio(m) for m in masks], 1).to(src.dtype)
+
+        # two stage
+        # enc_topk_proposals = enc_refpoint_embed = None
+
+        #########################################################
+        # Begin Encoder
+        #########################################################
+        memory, memory_text = self.encoder(
+            src_flatten,
+            pos=lvl_pos_embed_flatten,
+            level_start_index=level_start_index,
+            spatial_shapes=spatial_shapes,
+            valid_ratios=valid_ratios,
+            key_padding_mask=mask_flatten,
+            memory_text=text_dict["encoded_text"],
+            text_attention_mask=~text_dict["text_token_mask"],
+            # we ~ the mask . False means use the token; True means pad the token
+            position_ids=text_dict["position_ids"],
+            text_self_attention_masks=text_dict["text_self_attention_masks"],
+        )
+        #########################################################
+        # End Encoder
+        # - memory: bs, \sum{hw}, c
+        # - mask_flatten: bs, \sum{hw}
+        # - lvl_pos_embed_flatten: bs, \sum{hw}, c
+        # - enc_intermediate_output: None or (nenc+1, bs, nq, c) or (nenc, bs, nq, c)
+        # - enc_intermediate_refpoints: None or (nenc+1, bs, nq, c) or (nenc, bs, nq, c)
+        #########################################################
+        text_dict["encoded_text"] = memory_text
+        # if os.environ.get("SHILONG_AMP_INFNAN_DEBUG") == '1':
+        #     if memory.isnan().any() | memory.isinf().any():
+        #         import ipdb; ipdb.set_trace()
+
+        if self.two_stage_type == "standard":
+            output_memory, output_proposals = gen_encoder_output_proposals(memory, mask_flatten, spatial_shapes)
+            output_memory = self.enc_output_norm(self.enc_output(output_memory))
+
+            if text_dict is not None:
+                enc_outputs_class_unselected = self.enc_out_class_embed(output_memory, text_dict)
+            else:
+                enc_outputs_class_unselected = self.enc_out_class_embed(output_memory)
+
+            topk_logits = enc_outputs_class_unselected.max(-1)[0]
+            enc_outputs_coord_unselected = (
+                self.enc_out_bbox_embed(output_memory) + output_proposals
+            )  # (bs, \sum{hw}, 4) unsigmoid
+            topk = self.num_queries
+
+            topk_proposals = torch.topk(topk_logits, topk, dim=1)[1]  # bs, nq
+
+            # gather boxes
+            refpoint_embed_undetach = torch.gather(
+                enc_outputs_coord_unselected, 1, topk_proposals.unsqueeze(-1).repeat(1, 1, 4)
+            )  # unsigmoid
+            refpoint_embed_ = refpoint_embed_undetach.detach()
+            init_box_proposal = torch.gather(
+                output_proposals, 1, topk_proposals.unsqueeze(-1).repeat(1, 1, 4)
+            ).sigmoid()  # sigmoid
+
+            # gather tgt
+            tgt_undetach = torch.gather(output_memory, 1, topk_proposals.unsqueeze(-1).repeat(1, 1, self.d_model))
+            if self.embed_init_tgt:
+                tgt_ = self.tgt_embed.weight[:, None, :].repeat(1, bs, 1).transpose(0, 1)  # nq, bs, d_model
+            else:
+                tgt_ = tgt_undetach.detach()
+
+            if refpoint_embed is not None:
+                refpoint_embed = torch.cat([refpoint_embed, refpoint_embed_], dim=1)
+                tgt = torch.cat([tgt, tgt_], dim=1)
+            else:
+                refpoint_embed, tgt = refpoint_embed_, tgt_
+
+        elif self.two_stage_type == "no":
+            tgt_ = self.tgt_embed.weight[:, None, :].repeat(1, bs, 1).transpose(0, 1)  # nq, bs, d_model
+            refpoint_embed_ = self.refpoint_embed.weight[:, None, :].repeat(1, bs, 1).transpose(0, 1)  # nq, bs, 4
+
+            if refpoint_embed is not None:
+                refpoint_embed = torch.cat([refpoint_embed, refpoint_embed_], dim=1)
+                tgt = torch.cat([tgt, tgt_], dim=1)
+            else:
+                refpoint_embed, tgt = refpoint_embed_, tgt_
+
+            if self.num_patterns > 0:
+                tgt_embed = tgt.repeat(1, self.num_patterns, 1)
+                refpoint_embed = refpoint_embed.repeat(1, self.num_patterns, 1)
+                tgt_pat = self.patterns.weight[None, :, :].repeat_interleave(
+                    self.num_queries, 1
+                )  # 1, n_q*n_pat, d_model
+                tgt = tgt_embed + tgt_pat
+
+            init_box_proposal = refpoint_embed_.sigmoid()
+
+        else:
+            raise NotImplementedError("unknown two_stage_type {}".format(self.two_stage_type))
+        #########################################################
+        # End preparing tgt
+        # - tgt: bs, NQ, d_model
+        # - refpoint_embed(unsigmoid): bs, NQ, d_model
+        #########################################################
+
+        #########################################################
+        # Begin Decoder
+        #########################################################
+        hs, references = self.decoder(
+            tgt=tgt.transpose(0, 1),
+            memory=memory.transpose(0, 1),
+            memory_key_padding_mask=mask_flatten,
+            pos=lvl_pos_embed_flatten.transpose(0, 1),
+            refpoints_unsigmoid=refpoint_embed.transpose(0, 1),
+            level_start_index=level_start_index,
+            spatial_shapes=spatial_shapes,
+            valid_ratios=valid_ratios,
+            tgt_mask=attn_mask,
+            memory_text=text_dict["encoded_text"],
+            text_attention_mask=~text_dict["text_token_mask"],
+            # we ~ the mask . False means use the token; True means pad the token
+        )
+        #########################################################
+        # End Decoder
+        # hs: n_dec, bs, nq, d_model
+        # references: n_dec+1, bs, nq, query_dim
+        #########################################################
+
+        #########################################################
+        # Begin postprocess
+        #########################################################
+        if self.two_stage_type == "standard":
+            hs_enc = tgt_undetach.unsqueeze(0)
+            ref_enc = refpoint_embed_undetach.sigmoid().unsqueeze(0)
+        else:
+            hs_enc = ref_enc = None
+        #########################################################
+        # End postprocess
+        # hs_enc: (n_enc+1, bs, nq, d_model) or (1, bs, nq, d_model) or (n_enc, bs, nq, d_model) or None
+        # ref_enc: (n_enc+1, bs, nq, query_dim) or (1, bs, nq, query_dim) or (n_enc, bs, nq, d_model) or None
+        #########################################################
+
+        return hs, references, hs_enc, ref_enc, init_box_proposal
+        # hs: (n_dec, bs, nq, d_model)
+        # references: sigmoid coordinates. (n_dec+1, bs, bq, 4)
+        # hs_enc: (n_enc+1, bs, nq, d_model) or (1, bs, nq, d_model) or None
+        # ref_enc: sigmoid coordinates. \
+        #           (n_enc+1, bs, nq, query_dim) or (1, bs, nq, query_dim) or None
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self,
+        encoder_layer,
+        num_layers,
+        d_model=256,
+        num_queries=300,
+        enc_layer_share=False,
+        text_enhance_layer=None,
+        feature_fusion_layer=None,
+        use_checkpoint=False,
+        use_transformer_ckpt=False,
+    ):
+        """_summary_
+
+        Args:
+            encoder_layer (_type_): _description_
+            num_layers (_type_): _description_
+            norm (_type_, optional): _description_. Defaults to None.
+            d_model (int, optional): _description_. Defaults to 256.
+            num_queries (int, optional): _description_. Defaults to 300.
+            enc_layer_share (bool, optional): _description_. Defaults to False.
+
+        """
+        super().__init__()
+        # prepare layers
+        self.layers = []
+        self.text_layers = []
+        self.fusion_layers = []
+        if num_layers > 0:
+            self.layers = _get_clones(encoder_layer, num_layers, layer_share=enc_layer_share)
+
+            if text_enhance_layer is not None:
+                self.text_layers = _get_clones(text_enhance_layer, num_layers, layer_share=enc_layer_share)
+            if feature_fusion_layer is not None:
+                self.fusion_layers = _get_clones(feature_fusion_layer, num_layers, layer_share=enc_layer_share)
+        else:
+            self.layers = []
+            del encoder_layer
+
+            if text_enhance_layer is not None:
+                self.text_layers = []
+                del text_enhance_layer
+            if feature_fusion_layer is not None:
+                self.fusion_layers = []
+                del feature_fusion_layer
+
+        self.query_scale = None
+        self.num_queries = num_queries
+        self.num_layers = num_layers
+        self.d_model = d_model
+
+        self.use_checkpoint = use_checkpoint
+        self.use_transformer_ckpt = use_transformer_ckpt
+
+    @staticmethod
+    def get_reference_points(spatial_shapes, valid_ratios, device):
+        reference_points_list = []
+        for lvl, (H_, W_) in enumerate(spatial_shapes):
+
+            ref_y, ref_x = torch.meshgrid(
+                torch.linspace(0.5, H_ - 0.5, H_, dtype=torch.float32, device=device),
+                torch.linspace(0.5, W_ - 0.5, W_, dtype=torch.float32, device=device),
+            )
+            ref_y = ref_y.reshape(-1)[None] / (valid_ratios[:, None, lvl, 1] * H_)
+            ref_x = ref_x.reshape(-1)[None] / (valid_ratios[:, None, lvl, 0] * W_)
+            ref = torch.stack((ref_x, ref_y), -1)
+            reference_points_list.append(ref)
+        reference_points = torch.cat(reference_points_list, 1)
+        reference_points = reference_points[:, :, None] * valid_ratios[:, None]
+        return reference_points
+
+    def forward(
+        self,
+        # for images
+        src: Tensor,
+        pos: Tensor,
+        spatial_shapes: Tensor,
+        level_start_index: Tensor,
+        valid_ratios: Tensor,
+        key_padding_mask: Tensor,
+        # for texts
+        memory_text: Tensor = None,
+        text_attention_mask: Tensor = None,
+        pos_text: Tensor = None,
+        text_self_attention_masks: Tensor = None,
+        position_ids: Tensor = None,
+    ):
+        """
+        Input:
+            - src: [bs, sum(hi*wi), 256]
+            - pos: pos embed for src. [bs, sum(hi*wi), 256]
+            - spatial_shapes: h,w of each level [num_level, 2]
+            - level_start_index: [num_level] start point of level in sum(hi*wi).
+            - valid_ratios: [bs, num_level, 2]
+            - key_padding_mask: [bs, sum(hi*wi)]
+
+            - memory_text: bs, n_text, 256
+            - text_attention_mask: bs, n_text
+                False for no padding; True for padding
+            - pos_text: bs, n_text, 256
+
+            - position_ids: bs, n_text
+        Intermedia:
+            - reference_points: [bs, sum(hi*wi), num_level, 2]
+        Outpus:
+            - output: [bs, sum(hi*wi), 256]
+        """
+
+        output = src
+
+        # preparation and reshape
+        if self.num_layers > 0:
+            reference_points = self.get_reference_points(spatial_shapes, valid_ratios, device=src.device)
+
+        if self.text_layers:
+            # generate pos_text
+            bs, n_text, text_dim = memory_text.shape
+            if pos_text is None and position_ids is None:
+                pos_text = (
+                    torch.arange(n_text, device=memory_text.device).float().unsqueeze(0).unsqueeze(-1).repeat(bs, 1, 1)
+                )
+                pos_text = get_sine_pos_embed(pos_text, num_pos_feats=256, exchange_xy=False)
+            if position_ids is not None:
+                pos_text = get_sine_pos_embed(position_ids[..., None], num_pos_feats=256, exchange_xy=False)
+            pos_text = pos_text.to(src.dtype)
+
+        # main process
+        for layer_id, layer in enumerate(self.layers):
+            # if output.isnan().any() or memory_text.isnan().any():
+            #     if os.environ.get('IPDB_SHILONG_DEBUG', None) == 'INFO':
+            #         import ipdb; ipdb.set_trace()
+            if self.fusion_layers:
+                if self.use_checkpoint:
+                    output, memory_text = checkpoint.checkpoint(
+                        self.fusion_layers[layer_id],
+                        output,
+                        memory_text,
+                        key_padding_mask,
+                        text_attention_mask,
+                    )
+                else:
+                    output, memory_text = self.fusion_layers[layer_id](
+                        v=output,
+                        l=memory_text,
+                        attention_mask_v=key_padding_mask,
+                        attention_mask_l=text_attention_mask,
+                    )
+
+            if self.text_layers:
+                memory_text = self.text_layers[layer_id](
+                    src=memory_text.transpose(0, 1),
+                    src_mask=~text_self_attention_masks,  # note we use ~ for mask here
+                    src_key_padding_mask=text_attention_mask,
+                    pos=(pos_text.transpose(0, 1) if pos_text is not None else None),
+                ).transpose(0, 1)
+
+            # main process
+            if self.use_transformer_ckpt:
+                output = checkpoint.checkpoint(
+                    layer,
+                    output,
+                    pos,
+                    reference_points,
+                    spatial_shapes,
+                    level_start_index,
+                    key_padding_mask,
+                )
+            else:
+                output = layer(
+                    src=output,
+                    pos=pos,
+                    reference_points=reference_points,
+                    spatial_shapes=spatial_shapes,
+                    level_start_index=level_start_index,
+                    key_padding_mask=key_padding_mask,
+                )
+
+        return output, memory_text
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(
+        self,
+        decoder_layer,
+        num_layers,
+        norm=None,
+        return_intermediate=False,
+        d_model=256,
+        query_dim=4,
+        num_feature_levels=1,
+    ):
+        super().__init__()
+        if num_layers > 0:
+            self.layers = _get_clones(decoder_layer, num_layers)
+        else:
+            self.layers = []
+        self.num_layers = num_layers
+        self.norm = norm
+        self.return_intermediate = return_intermediate
+        assert return_intermediate, "support return_intermediate only"
+        self.query_dim = query_dim
+        assert query_dim in [2, 4], "query_dim should be 2/4 but {}".format(query_dim)
+        self.num_feature_levels = num_feature_levels
+
+        self.ref_point_head = MLP(query_dim // 2 * d_model, d_model, d_model, 2)
+        self.query_pos_sine_scale = None
+
+        self.query_scale = None
+        self.bbox_embed = None
+        self.class_embed = None
+
+        self.d_model = d_model
+
+        self.ref_anchor_head = None
+
+    def forward(
+        self,
+        tgt,
+        memory,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        refpoints_unsigmoid: Optional[Tensor] = None,  # num_queries, bs, 2
+        # for memory
+        level_start_index: Optional[Tensor] = None,  # num_levels
+        spatial_shapes: Optional[Tensor] = None,  # bs, num_levels, 2
+        valid_ratios: Optional[Tensor] = None,
+        # for text
+        memory_text: Optional[Tensor] = None,
+        text_attention_mask: Optional[Tensor] = None,
+    ):
+        """
+        Input:
+            - tgt: nq, bs, d_model
+            - memory: hw, bs, d_model
+            - pos: hw, bs, d_model
+            - refpoints_unsigmoid: nq, bs, 2/4
+            - valid_ratios/spatial_shapes: bs, nlevel, 2
+        """
+        output = tgt
+
+        intermediate = []
+        reference_points = refpoints_unsigmoid.sigmoid()
+        ref_points = [reference_points]
+
+        for layer_id, layer in enumerate(self.layers):
+
+            if reference_points.shape[-1] == 4:
+                reference_points_input = (
+                    reference_points[:, :, None] * torch.cat([valid_ratios, valid_ratios], -1)[None, :]
+                )  # nq, bs, nlevel, 4
+            else:
+                assert reference_points.shape[-1] == 2
+                reference_points_input = reference_points[:, :, None] * valid_ratios[None, :]
+            query_sine_embed = gen_sineembed_for_position(reference_points_input[:, :, 0, :])  # nq, bs, 256*2
+
+            # conditional query
+            raw_query_pos = self.ref_point_head(query_sine_embed)  # nq, bs, 256
+            pos_scale = self.query_scale(output) if self.query_scale is not None else 1
+            query_pos = pos_scale * raw_query_pos
+            # if os.environ.get("SHILONG_AMP_INFNAN_DEBUG") == '1':
+            #     if query_pos.isnan().any() | query_pos.isinf().any():
+            #         import ipdb; ipdb.set_trace()
+
+            # main process
+            output = layer(
+                tgt=output,
+                tgt_query_pos=query_pos,
+                tgt_query_sine_embed=query_sine_embed,
+                tgt_key_padding_mask=tgt_key_padding_mask,
+                tgt_reference_points=reference_points_input,
+                memory_text=memory_text,
+                text_attention_mask=text_attention_mask,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+                memory_level_start_index=level_start_index,
+                memory_spatial_shapes=spatial_shapes,
+                memory_pos=pos,
+                self_attn_mask=tgt_mask,
+                cross_attn_mask=memory_mask,
+            )
+            if output.isnan().any() | output.isinf().any():
+                print(f"output layer_id {layer_id} is nan")
+                try:
+                    num_nan = output.isnan().sum().item()
+                    num_inf = output.isinf().sum().item()
+                    print(f"num_nan {num_nan}, num_inf {num_inf}")
+                except Exception as e:
+                    print(e)
+                    # if os.environ.get("SHILONG_AMP_INFNAN_DEBUG") == '1':
+                    #     import ipdb; ipdb.set_trace()
+
+            # iter update
+            if self.bbox_embed is not None:
+                # box_holder = self.bbox_embed(output)
+                # box_holder[..., :self.query_dim] += inverse_sigmoid(reference_points)
+                # new_reference_points = box_holder[..., :self.query_dim].sigmoid()
+
+                reference_before_sigmoid = inverse_sigmoid(reference_points)
+                delta_unsig = self.bbox_embed[layer_id](output)
+                outputs_unsig = delta_unsig + reference_before_sigmoid
+                new_reference_points = outputs_unsig.sigmoid()
+
+                reference_points = new_reference_points.detach()
+                # if layer_id != self.num_layers - 1:
+                ref_points.append(new_reference_points)
+
+            intermediate.append(self.norm(output))
+
+        return [
+            [itm_out.transpose(0, 1) for itm_out in intermediate],
+            [itm_refpoint.transpose(0, 1) for itm_refpoint in ref_points],
+        ]
+
+
+class DeformableTransformerEncoderLayer(nn.Module):
+    def __init__(
+        self,
+        d_model=256,
+        d_ffn=1024,
+        dropout=0.1,
+        activation="relu",
+        n_levels=4,
+        n_heads=8,
+        n_points=4,
+    ):
+        super().__init__()
+
+        # self attention
+        self.self_attn = MSDeformAttn(
+            embed_dim=d_model,
+            num_levels=n_levels,
+            num_heads=n_heads,
+            num_points=n_points,
+            batch_first=True,
+        )
+        self.dropout1 = nn.Dropout(dropout)
+        self.norm1 = nn.LayerNorm(d_model)
+
+        # ffn
+        self.linear1 = nn.Linear(d_model, d_ffn)
+        self.activation = _get_activation_fn(activation, d_model=d_ffn)
+        self.dropout2 = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(d_ffn, d_model)
+        self.dropout3 = nn.Dropout(dropout)
+        self.norm2 = nn.LayerNorm(d_model)
+
+    @staticmethod
+    def with_pos_embed(tensor, pos):
+        return tensor if pos is None else tensor + pos
+
+    def forward_ffn(self, src):
+        src2 = self.linear2(self.dropout2(self.activation(self.linear1(src))))
+        src = src + self.dropout3(src2)
+        src = self.norm2(src)
+        return src
+
+    def forward(self, src, pos, reference_points, spatial_shapes, level_start_index, key_padding_mask=None):
+        # self attention
+        # import ipdb; ipdb.set_trace()
+        src2 = self.self_attn(
+            query=self.with_pos_embed(src, pos),
+            reference_points=reference_points,
+            value=src,
+            spatial_shapes=spatial_shapes,
+            level_start_index=level_start_index,
+            key_padding_mask=key_padding_mask,
+        )
+        src = src + self.dropout1(src2)
+        src = self.norm1(src)
+
+        # ffn
+        src = self.forward_ffn(src)
+
+        return src
+
+
+class DeformableTransformerDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        d_model=256,
+        d_ffn=1024,
+        dropout=0.1,
+        activation="relu",
+        n_levels=4,
+        n_heads=8,
+        n_points=4,
+        use_text_feat_guide=False,
+        use_text_cross_attention=False,
+    ):
+        super().__init__()
+
+        # cross attention
+        self.cross_attn = MSDeformAttn(
+            embed_dim=d_model,
+            num_levels=n_levels,
+            num_heads=n_heads,
+            num_points=n_points,
+            batch_first=True,
+        )
+        self.dropout1 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm1 = nn.LayerNorm(d_model)
+
+        # cross attention text
+        if use_text_cross_attention:
+            self.ca_text = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)
+            self.catext_dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+            self.catext_norm = nn.LayerNorm(d_model)
+
+        # self attention
+        self.self_attn = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)
+        self.dropout2 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm2 = nn.LayerNorm(d_model)
+
+        # ffn
+        self.linear1 = nn.Linear(d_model, d_ffn)
+        self.activation = _get_activation_fn(activation, d_model=d_ffn, batch_dim=1)
+        self.dropout3 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.linear2 = nn.Linear(d_ffn, d_model)
+        self.dropout4 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm3 = nn.LayerNorm(d_model)
+
+        self.key_aware_proj = None
+        self.use_text_feat_guide = use_text_feat_guide
+        assert not use_text_feat_guide
+        self.use_text_cross_attention = use_text_cross_attention
+
+    def rm_self_attn_modules(self):
+        self.self_attn = None
+        self.dropout2 = None
+        self.norm2 = None
+
+    @staticmethod
+    def with_pos_embed(tensor, pos):
+        return tensor if pos is None else tensor + pos
+
+    def forward_ffn(self, tgt):
+        with torch.cuda.amp.autocast(enabled=False):
+            tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
+        tgt = tgt + self.dropout4(tgt2)
+        tgt = self.norm3(tgt)
+        return tgt
+
+    def forward(
+        self,
+        # for tgt
+        tgt: Optional[Tensor],  # nq, bs, d_model
+        tgt_query_pos: Optional[Tensor] = None,  # pos for query. MLP(Sine(pos))
+        tgt_query_sine_embed: Optional[Tensor] = None,  # pos for query. Sine(pos)
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        tgt_reference_points: Optional[Tensor] = None,  # nq, bs, 4
+        memory_text: Optional[Tensor] = None,  # bs, num_token, d_model
+        text_attention_mask: Optional[Tensor] = None,  # bs, num_token
+        # for memory
+        memory: Optional[Tensor] = None,  # hw, bs, d_model
+        memory_key_padding_mask: Optional[Tensor] = None,
+        memory_level_start_index: Optional[Tensor] = None,  # num_levels
+        memory_spatial_shapes: Optional[Tensor] = None,  # bs, num_levels, 2
+        memory_pos: Optional[Tensor] = None,  # pos for memory
+        # sa
+        self_attn_mask: Optional[Tensor] = None,  # mask used for self-attention
+        cross_attn_mask: Optional[Tensor] = None,  # mask used for cross-attention
+    ):
+        """
+        Input:
+            - tgt/tgt_query_pos: nq, bs, d_model
+            -
+        """
+        assert cross_attn_mask is None
+
+        # self attention
+        if self.self_attn is not None:
+            # import ipdb; ipdb.set_trace()
+            q = k = self.with_pos_embed(tgt, tgt_query_pos)
+            tgt2 = self.self_attn(q, k, tgt, attn_mask=self_attn_mask)[0]
+            tgt = tgt + self.dropout2(tgt2)
+            tgt = self.norm2(tgt)
+
+        if self.use_text_cross_attention:
+            tgt2 = self.ca_text(
+                self.with_pos_embed(tgt, tgt_query_pos),
+                memory_text.transpose(0, 1),
+                memory_text.transpose(0, 1),
+                key_padding_mask=text_attention_mask,
+            )[0]
+            tgt = tgt + self.catext_dropout(tgt2)
+            tgt = self.catext_norm(tgt)
+
+        tgt2 = self.cross_attn(
+            query=self.with_pos_embed(tgt, tgt_query_pos).transpose(0, 1),
+            reference_points=tgt_reference_points.transpose(0, 1).contiguous(),
+            value=memory.transpose(0, 1),
+            spatial_shapes=memory_spatial_shapes,
+            level_start_index=memory_level_start_index,
+            key_padding_mask=memory_key_padding_mask,
+        ).transpose(0, 1)
+        tgt = tgt + self.dropout1(tgt2)
+        tgt = self.norm1(tgt)
+
+        # ffn
+        tgt = self.forward_ffn(tgt)
+
+        return tgt
+
+
+def build_transformer(args):
+    return Transformer(
+        d_model=args.hidden_dim,
+        dropout=args.dropout,
+        nhead=args.nheads,
+        num_queries=args.num_queries,
+        dim_feedforward=args.dim_feedforward,
+        num_encoder_layers=args.enc_layers,
+        num_decoder_layers=args.dec_layers,
+        normalize_before=args.pre_norm,
+        return_intermediate_dec=True,
+        query_dim=args.query_dim,
+        activation=args.transformer_activation,
+        num_patterns=args.num_patterns,
+        num_feature_levels=args.num_feature_levels,
+        enc_n_points=args.enc_n_points,
+        dec_n_points=args.dec_n_points,
+        learnable_tgt_init=True,
+        # two stage
+        two_stage_type=args.two_stage_type,  # ['no', 'standard', 'early']
+        embed_init_tgt=args.embed_init_tgt,
+        use_text_enhancer=args.use_text_enhancer,
+        use_fusion_layer=args.use_fusion_layer,
+        use_checkpoint=args.use_checkpoint,
+        use_transformer_ckpt=args.use_transformer_ckpt,
+        use_text_cross_attention=args.use_text_cross_attention,
+        text_dropout=args.text_dropout,
+        fusion_dropout=args.fusion_dropout,
+        fusion_droppath=args.fusion_droppath,
+    )

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/transformer_vanilla.py

@@ -0,0 +1,115 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copyright (c) Aishwarya Kamath & Nicolas Carion. Licensed under the Apache License 2.0. All Rights Reserved
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+"""
+DETR Transformer class.
+
+Copy-paste from torch.nn.Transformer with modifications:
+    * positional encodings are passed in MHattention
+    * extra LN at the end of encoder is removed
+    * decoder returns a stack of activations from all decoding layers
+"""
+from typing import Optional
+
+import torch
+from torch import Tensor, nn
+
+from .utils import _get_activation_fn, _get_clones
+
+
+class TextTransformer(nn.Module):
+    def __init__(self, num_layers, d_model=256, nheads=8, dim_feedforward=2048, dropout=0.1):
+        super().__init__()
+        self.num_layers = num_layers
+        self.d_model = d_model
+        self.nheads = nheads
+        self.dim_feedforward = dim_feedforward
+        self.norm = None
+
+        single_encoder_layer = TransformerEncoderLayer(
+            d_model=d_model, nhead=nheads, dim_feedforward=dim_feedforward, dropout=dropout
+        )
+        self.layers = _get_clones(single_encoder_layer, num_layers)
+
+    def forward(self, memory_text: torch.Tensor, text_attention_mask: torch.Tensor):
+        """
+
+        Args:
+            text_attention_mask: bs, num_token
+            memory_text: bs, num_token, d_model
+
+        Raises:
+            RuntimeError: _description_
+
+        Returns:
+            output: bs, num_token, d_model
+        """
+
+        output = memory_text.transpose(0, 1)
+
+        for layer in self.layers:
+            output = layer(output, src_key_padding_mask=text_attention_mask)
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output.transpose(0, 1)
+
+
+class TransformerEncoderLayer(nn.Module):
+    def __init__(
+        self,
+        d_model,
+        nhead,
+        dim_feedforward=2048,
+        dropout=0.1,
+        activation="relu",
+        normalize_before=False,
+    ):
+        super().__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+        self.normalize_before = normalize_before
+        self.nhead = nhead
+
+    def with_pos_embed(self, tensor, pos: Optional[Tensor]):
+        return tensor if pos is None else tensor + pos
+
+    def forward(
+        self,
+        src,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+    ):
+        # repeat attn mask
+        if src_mask.dim() == 3 and src_mask.shape[0] == src.shape[1]:
+            # bs, num_q, num_k
+            src_mask = src_mask.repeat(self.nhead, 1, 1)
+
+        q = k = self.with_pos_embed(src, pos)
+
+        src2 = self.self_attn(q, k, value=src, attn_mask=src_mask)[0]
+
+        # src2 = self.self_attn(q, k, value=src, attn_mask=src_mask, key_padding_mask=src_key_padding_mask)[0]
+        src = src + self.dropout1(src2)
+        src = self.norm1(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = src + self.dropout2(src2)
+        src = self.norm2(src)
+        return src

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/GroundingDINO/utils.py

@@ -0,0 +1,258 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+
+import copy
+import math
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+
+
+def _get_clones(module, N, layer_share=False):
+    # import ipdb; ipdb.set_trace()
+    if layer_share:
+        return nn.ModuleList([module for i in range(N)])
+    else:
+        return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+def get_sine_pos_embed(
+    pos_tensor: torch.Tensor,
+    num_pos_feats: int = 128,
+    temperature: int = 10000,
+    exchange_xy: bool = True,
+):
+    """generate sine position embedding from a position tensor
+    Args:
+        pos_tensor (torch.Tensor): shape: [..., n].
+        num_pos_feats (int): projected shape for each float in the tensor.
+        temperature (int): temperature in the sine/cosine function.
+        exchange_xy (bool, optional): exchange pos x and pos y. \
+            For example, input tensor is [x,y], the results will be [pos(y), pos(x)]. Defaults to True.
+    Returns:
+        pos_embed (torch.Tensor): shape: [..., n*num_pos_feats].
+    """
+    scale = 2 * math.pi
+    dim_t = torch.arange(num_pos_feats, dtype=torch.float32, device=pos_tensor.device)
+    dim_t = temperature ** (2 * torch.div(dim_t, 2, rounding_mode="floor") / num_pos_feats)
+
+    def sine_func(x: torch.Tensor):
+        sin_x = x * scale / dim_t
+        sin_x = torch.stack((sin_x[..., 0::2].sin(), sin_x[..., 1::2].cos()), dim=3).flatten(2)
+        return sin_x
+
+    pos_res = [sine_func(x) for x in pos_tensor.split([1] * pos_tensor.shape[-1], dim=-1)]
+    if exchange_xy:
+        pos_res[0], pos_res[1] = pos_res[1], pos_res[0]
+    pos_res = torch.cat(pos_res, dim=-1)
+    return pos_res
+
+
+def gen_encoder_output_proposals(memory: Tensor, memory_padding_mask: Tensor, spatial_shapes: Tensor, learnedwh=None):
+    """
+    Input:
+        - memory: bs, \sum{hw}, d_model
+        - memory_padding_mask: bs, \sum{hw}
+        - spatial_shapes: nlevel, 2
+        - learnedwh: 2
+    Output:
+        - output_memory: bs, \sum{hw}, d_model
+        - output_proposals: bs, \sum{hw}, 4
+    """
+    N_, S_, C_ = memory.shape
+    proposals = []
+    _cur = 0
+    for lvl, (H_, W_) in enumerate(spatial_shapes):
+        mask_flatten_ = memory_padding_mask[:, _cur : (_cur + H_ * W_)].view(N_, H_, W_, 1)
+        valid_H = torch.sum(~mask_flatten_[:, :, 0, 0], 1)
+        valid_W = torch.sum(~mask_flatten_[:, 0, :, 0], 1)
+
+        # import ipdb; ipdb.set_trace()
+
+        grid_y, grid_x = torch.meshgrid(
+            torch.linspace(0, H_ - 1, H_, dtype=torch.float32, device=memory.device),
+            torch.linspace(0, W_ - 1, W_, dtype=torch.float32, device=memory.device),
+        )
+        grid = torch.cat([grid_x.unsqueeze(-1), grid_y.unsqueeze(-1)], -1)  # H_, W_, 2
+
+        scale = torch.cat([valid_W.unsqueeze(-1), valid_H.unsqueeze(-1)], 1).view(N_, 1, 1, 2)
+        grid = (grid.unsqueeze(0).expand(N_, -1, -1, -1) + 0.5) / scale
+
+        if learnedwh is not None:
+            # import ipdb; ipdb.set_trace()
+            wh = torch.ones_like(grid) * learnedwh.sigmoid() * (2.0**lvl)
+        else:
+            wh = torch.ones_like(grid) * 0.05 * (2.0**lvl)
+
+        # scale = torch.cat([W_[None].unsqueeze(-1), H_[None].unsqueeze(-1)], 1).view(1, 1, 1, 2).repeat(N_, 1, 1, 1)
+        # grid = (grid.unsqueeze(0).expand(N_, -1, -1, -1) + 0.5) / scale
+        # wh = torch.ones_like(grid) / scale
+        proposal = torch.cat((grid, wh), -1).view(N_, -1, 4)
+        proposals.append(proposal)
+        _cur += H_ * W_
+    # import ipdb; ipdb.set_trace()
+    output_proposals = torch.cat(proposals, 1)
+    output_proposals_valid = ((output_proposals > 0.01) & (output_proposals < 0.99)).all(-1, keepdim=True)
+    output_proposals = torch.log(output_proposals / (1 - output_proposals))  # unsigmoid
+    output_proposals = output_proposals.masked_fill(memory_padding_mask.unsqueeze(-1), float("inf"))
+    output_proposals = output_proposals.masked_fill(~output_proposals_valid, float("inf"))
+
+    output_memory = memory
+    output_memory = output_memory.masked_fill(memory_padding_mask.unsqueeze(-1), float(0))
+    output_memory = output_memory.masked_fill(~output_proposals_valid, float(0))
+
+    # output_memory = output_memory.masked_fill(memory_padding_mask.unsqueeze(-1), float('inf'))
+    # output_memory = output_memory.masked_fill(~output_proposals_valid, float('inf'))
+
+    output_proposals = output_proposals.to(output_memory.dtype)
+    return output_memory, output_proposals
+
+
+class RandomBoxPerturber:
+    def __init__(self, x_noise_scale=0.2, y_noise_scale=0.2, w_noise_scale=0.2, h_noise_scale=0.2) -> None:
+        self.noise_scale = torch.Tensor([x_noise_scale, y_noise_scale, w_noise_scale, h_noise_scale])
+
+    def __call__(self, refanchors: Tensor) -> Tensor:
+        nq, bs, query_dim = refanchors.shape
+        device = refanchors.device
+
+        noise_raw = torch.rand_like(refanchors)
+        noise_scale = self.noise_scale.to(device)[:query_dim]
+
+        new_refanchors = refanchors * (1 + (noise_raw - 0.5) * noise_scale)
+        return new_refanchors.clamp_(0, 1)
+
+
+def sigmoid_focal_loss(inputs, targets, num_boxes, alpha: float = 0.25, gamma: float = 2, no_reduction=False):
+    """
+    Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002.
+    Args:
+        inputs: A float tensor of arbitrary shape.
+                The predictions for each example.
+        targets: A float tensor with the same shape as inputs. Stores the binary
+                 classification label for each element in inputs
+                (0 for the negative class and 1 for the positive class).
+        alpha: (optional) Weighting factor in range (0,1) to balance
+                positive vs negative examples. Default = -1 (no weighting).
+        gamma: Exponent of the modulating factor (1 - p_t) to
+               balance easy vs hard examples.
+    Returns:
+        Loss tensor
+    """
+    prob = inputs.sigmoid()
+    ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
+    p_t = prob * targets + (1 - prob) * (1 - targets)
+    loss = ce_loss * ((1 - p_t) ** gamma)
+
+    if alpha >= 0:
+        alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
+        loss = alpha_t * loss
+
+    if no_reduction:
+        return loss
+
+    return loss.mean(1).sum() / num_boxes
+
+
+class MLP(nn.Module):
+    """Very simple multi-layer perceptron (also called FFN)"""
+
+    def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
+        return x
+
+
+def _get_activation_fn(activation, d_model=256, batch_dim=0):
+    """Return an activation function given a string"""
+    if activation == "relu":
+        return F.relu
+    if activation == "gelu":
+        return F.gelu
+    if activation == "glu":
+        return F.glu
+    if activation == "prelu":
+        return nn.PReLU()
+    if activation == "selu":
+        return F.selu
+
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+
+def gen_sineembed_for_position(pos_tensor):
+    # n_query, bs, _ = pos_tensor.size()
+    # sineembed_tensor = torch.zeros(n_query, bs, 256)
+    scale = 2 * math.pi
+    dim_t = torch.arange(128, dtype=torch.float32, device=pos_tensor.device)
+    dim_t = 10000 ** (2 * (torch.div(dim_t, 2, rounding_mode="floor")) / 128)
+    x_embed = pos_tensor[:, :, 0] * scale
+    y_embed = pos_tensor[:, :, 1] * scale
+    pos_x = x_embed[:, :, None] / dim_t
+    pos_y = y_embed[:, :, None] / dim_t
+    pos_x = torch.stack((pos_x[:, :, 0::2].sin(), pos_x[:, :, 1::2].cos()), dim=3).flatten(2)
+    pos_y = torch.stack((pos_y[:, :, 0::2].sin(), pos_y[:, :, 1::2].cos()), dim=3).flatten(2)
+    if pos_tensor.size(-1) == 2:
+        pos = torch.cat((pos_y, pos_x), dim=2)
+    elif pos_tensor.size(-1) == 4:
+        w_embed = pos_tensor[:, :, 2] * scale
+        pos_w = w_embed[:, :, None] / dim_t
+        pos_w = torch.stack((pos_w[:, :, 0::2].sin(), pos_w[:, :, 1::2].cos()), dim=3).flatten(2)
+
+        h_embed = pos_tensor[:, :, 3] * scale
+        pos_h = h_embed[:, :, None] / dim_t
+        pos_h = torch.stack((pos_h[:, :, 0::2].sin(), pos_h[:, :, 1::2].cos()), dim=3).flatten(2)
+
+        pos = torch.cat((pos_y, pos_x, pos_w, pos_h), dim=2)
+    else:
+        raise ValueError("Unknown pos_tensor shape(-1):{}".format(pos_tensor.size(-1)))
+    pos = pos.to(pos_tensor.dtype)
+    return pos
+
+
+class ContrastiveEmbed(nn.Module):
+    def __init__(self, max_text_len=256):
+        """
+        Args:
+            max_text_len: max length of text.
+        """
+        super().__init__()
+        self.max_text_len = max_text_len
+
+    def forward(self, x, text_dict):
+        """_summary_
+
+        Args:
+            x (_type_): _description_
+            text_dict (_type_): _description_
+            {
+                'encoded_text': encoded_text, # bs, 195, d_model
+                'text_token_mask': text_token_mask, # bs, 195
+                        # True for used tokens. False for padding tokens
+            }
+        Returns:
+            _type_: _description_
+        """
+        assert isinstance(text_dict, dict)
+
+        y = text_dict["encoded_text"]
+        text_token_mask = text_dict["text_token_mask"]
+
+        res = x @ y.transpose(-1, -2)
+        res.masked_fill_(~text_token_mask[:, None, :], float("-inf"))
+
+        # padding to max_text_len
+        new_res = torch.full((*res.shape[:-1], self.max_text_len), float("-inf"), device=res.device, dtype=res.dtype)
+        new_res[..., : res.shape[-1]] = res
+
+        return new_res

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/__init__.py

@@ -0,0 +1,18 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .GroundingDINO import build_groundingdino  # noqa
+
+
+def build_model(args):
+    # we use register to maintain models from catdet6 on.
+    from .registry import MODULE_BUILD_FUNCS
+
+    assert args.modelname in MODULE_BUILD_FUNCS._module_dict
+    build_func = MODULE_BUILD_FUNCS.get(args.modelname)
+    model = build_func(args)
+    return model

invokeai/backend/image_util/grounding_segment_anything/groundingdino/models/registry.py

@@ -0,0 +1,60 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# -*- coding: utf-8 -*-
+# @Author: Yihao Chen
+# @Date:   2021-08-16 16:03:17
+# @Last Modified by:   Shilong Liu
+# @Last Modified time: 2022-01-23 15:26
+# modified from mmcv
+
+import inspect
+from functools import partial
+
+
+class Registry(object):
+    def __init__(self, name):
+        self._name = name
+        self._module_dict = dict()
+
+    def __repr__(self):
+        format_str = self.__class__.__name__ + "(name={}, items={})".format(self._name, list(self._module_dict.keys()))
+        return format_str
+
+    def __len__(self):
+        return len(self._module_dict)
+
+    @property
+    def name(self):
+        return self._name
+
+    @property
+    def module_dict(self):
+        return self._module_dict
+
+    def get(self, key):
+        return self._module_dict.get(key, None)
+
+    def registe_with_name(self, module_name=None, force=False):
+        return partial(self.register, module_name=module_name, force=force)
+
+    def register(self, module_build_function, module_name=None, force=False):
+        """Register a module build function.
+        Args:
+            module (:obj:`nn.Module`): Module to be registered.
+        """
+        if not inspect.isfunction(module_build_function):
+            raise TypeError("module_build_function must be a function, but got {}".format(type(module_build_function)))
+        if module_name is None:
+            module_name = module_build_function.__name__
+        if not force and module_name in self._module_dict:
+            raise KeyError("{} is already registered in {}".format(module_name, self.name))
+        self._module_dict[module_name] = module_build_function
+
+        return module_build_function
+
+
+MODULE_BUILD_FUNCS = Registry("model build functions")

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/__init__.py

@@ -0,0 +1 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/box_ops.py

@@ -0,0 +1,140 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+"""
+Utilities for bounding box manipulation and GIoU.
+"""
+import torch
+from torchvision.ops.boxes import box_area
+
+
+def box_cxcywh_to_xyxy(x):
+    x_c, y_c, w, h = x.unbind(-1)
+    b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (x_c + 0.5 * w), (y_c + 0.5 * h)]
+    return torch.stack(b, dim=-1)
+
+
+def box_xyxy_to_cxcywh(x):
+    x0, y0, x1, y1 = x.unbind(-1)
+    b = [(x0 + x1) / 2, (y0 + y1) / 2, (x1 - x0), (y1 - y0)]
+    return torch.stack(b, dim=-1)
+
+
+# modified from torchvision to also return the union
+def box_iou(boxes1, boxes2):
+    area1 = box_area(boxes1)
+    area2 = box_area(boxes2)
+
+    # import ipdb; ipdb.set_trace()
+    lt = torch.max(boxes1[:, None, :2], boxes2[:, :2])  # [N,M,2]
+    rb = torch.min(boxes1[:, None, 2:], boxes2[:, 2:])  # [N,M,2]
+
+    wh = (rb - lt).clamp(min=0)  # [N,M,2]
+    inter = wh[:, :, 0] * wh[:, :, 1]  # [N,M]
+
+    union = area1[:, None] + area2 - inter
+
+    iou = inter / (union + 1e-6)
+    return iou, union
+
+
+def generalized_box_iou(boxes1, boxes2):
+    """
+    Generalized IoU from https://giou.stanford.edu/
+
+    The boxes should be in [x0, y0, x1, y1] format
+
+    Returns a [N, M] pairwise matrix, where N = len(boxes1)
+    and M = len(boxes2)
+    """
+    # degenerate boxes gives inf / nan results
+    # so do an early check
+    assert (boxes1[:, 2:] >= boxes1[:, :2]).all()
+    assert (boxes2[:, 2:] >= boxes2[:, :2]).all()
+    # except:
+    #     import ipdb; ipdb.set_trace()
+    iou, union = box_iou(boxes1, boxes2)
+
+    lt = torch.min(boxes1[:, None, :2], boxes2[:, :2])
+    rb = torch.max(boxes1[:, None, 2:], boxes2[:, 2:])
+
+    wh = (rb - lt).clamp(min=0)  # [N,M,2]
+    area = wh[:, :, 0] * wh[:, :, 1]
+
+    return iou - (area - union) / (area + 1e-6)
+
+
+# modified from torchvision to also return the union
+def box_iou_pairwise(boxes1, boxes2):
+    area1 = box_area(boxes1)
+    area2 = box_area(boxes2)
+
+    lt = torch.max(boxes1[:, :2], boxes2[:, :2])  # [N,2]
+    rb = torch.min(boxes1[:, 2:], boxes2[:, 2:])  # [N,2]
+
+    wh = (rb - lt).clamp(min=0)  # [N,2]
+    inter = wh[:, 0] * wh[:, 1]  # [N]
+
+    union = area1 + area2 - inter
+
+    iou = inter / union
+    return iou, union
+
+
+def generalized_box_iou_pairwise(boxes1, boxes2):
+    """
+    Generalized IoU from https://giou.stanford.edu/
+
+    Input:
+        - boxes1, boxes2: N,4
+    Output:
+        - giou: N, 4
+    """
+    # degenerate boxes gives inf / nan results
+    # so do an early check
+    assert (boxes1[:, 2:] >= boxes1[:, :2]).all()
+    assert (boxes2[:, 2:] >= boxes2[:, :2]).all()
+    assert boxes1.shape == boxes2.shape
+    iou, union = box_iou_pairwise(boxes1, boxes2)  # N, 4
+
+    lt = torch.min(boxes1[:, :2], boxes2[:, :2])
+    rb = torch.max(boxes1[:, 2:], boxes2[:, 2:])
+
+    wh = (rb - lt).clamp(min=0)  # [N,2]
+    area = wh[:, 0] * wh[:, 1]
+
+    return iou - (area - union) / area
+
+
+def masks_to_boxes(masks):
+    """Compute the bounding boxes around the provided masks
+
+    The masks should be in format [N, H, W] where N is the number of masks, (H, W) are the spatial dimensions.
+
+    Returns a [N, 4] tensors, with the boxes in xyxy format
+    """
+    if masks.numel() == 0:
+        return torch.zeros((0, 4), device=masks.device)
+
+    h, w = masks.shape[-2:]
+
+    y = torch.arange(0, h, dtype=torch.float)
+    x = torch.arange(0, w, dtype=torch.float)
+    y, x = torch.meshgrid(y, x)
+
+    x_mask = masks * x.unsqueeze(0)
+    x_max = x_mask.flatten(1).max(-1)[0]
+    x_min = x_mask.masked_fill(~(masks.bool()), 1e8).flatten(1).min(-1)[0]
+
+    y_mask = masks * y.unsqueeze(0)
+    y_max = y_mask.flatten(1).max(-1)[0]
+    y_min = y_mask.masked_fill(~(masks.bool()), 1e8).flatten(1).min(-1)[0]
+
+    return torch.stack([x_min, y_min, x_max, y_max], 1)
+
+
+if __name__ == "__main__":
+    x = torch.rand(5, 4)
+    y = torch.rand(3, 4)
+    iou, union = box_iou(x, y)
+    import ipdb
+
+    ipdb.set_trace()

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/get_tokenlizer.py

@@ -0,0 +1,24 @@
+from transformers import AutoTokenizer, BertModel, RobertaModel
+
+
+def get_tokenlizer(text_encoder_type):
+    if not isinstance(text_encoder_type, str):
+        # print("text_encoder_type is not a str")
+        if hasattr(text_encoder_type, "text_encoder_type"):
+            text_encoder_type = text_encoder_type.text_encoder_type
+        elif text_encoder_type.get("text_encoder_type", False):
+            text_encoder_type = text_encoder_type.get("text_encoder_type")
+        else:
+            raise ValueError("Unknown type of text_encoder_type: {}".format(type(text_encoder_type)))
+    print("final text_encoder_type: {}".format(text_encoder_type))
+
+    tokenizer = AutoTokenizer.from_pretrained(text_encoder_type)
+    return tokenizer
+
+
+def get_pretrained_language_model(text_encoder_type):
+    if text_encoder_type == "bert-base-uncased":
+        return BertModel.from_pretrained(text_encoder_type)
+    if text_encoder_type == "roberta-base":
+        return RobertaModel.from_pretrained(text_encoder_type)
+    raise ValueError("Unknown text_encoder_type {}".format(text_encoder_type))

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/inference.py

@@ -0,0 +1,221 @@
+from typing import Dict, List, Tuple
+
+import cv2
+import numpy as np
+import supervision as sv
+import torch
+from PIL import Image
+from torchvision.ops import box_convert
+
+import invokeai.backend.image_util.grounding_segment_anything.groundingdino.datasets.transforms as T
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.models import build_model
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.misc import clean_state_dict
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.slconfig import SLConfig
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.utils import get_phrases_from_posmap
+
+# ----------------------------------------------------------------------------------------------------------------------
+# OLD API
+# ----------------------------------------------------------------------------------------------------------------------
+
+
+def preprocess_caption(caption: str) -> str:
+    result = caption.lower().strip()
+    if result.endswith("."):
+        return result
+    return result + "."
+
+
+def load_model(model_config_path: str, model_state_dict: Dict[str, torch.Tensor], device: str = "cuda"):
+    args = SLConfig.fromfile(model_config_path)
+    args.device = device
+    model = build_model(args)
+    model.load_state_dict(clean_state_dict(model_state_dict["model"]), strict=False)
+    model.eval()
+    return model
+
+
+def load_image(image_path: str) -> Tuple[np.array, torch.Tensor]:
+    transform = T.Compose(
+        [
+            T.RandomResize([800], max_size=1333),
+            T.ToTensor(),
+            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+        ]
+    )
+    image_source = Image.open(image_path).convert("RGB")
+    image = np.asarray(image_source)
+    image_transformed, _ = transform(image_source, None)
+    return image, image_transformed
+
+
+def predict(
+    model, image: torch.Tensor, caption: str, box_threshold: float, text_threshold: float, device: str = "cuda"
+) -> Tuple[torch.Tensor, torch.Tensor, List[str]]:
+    caption = preprocess_caption(caption=caption)
+
+    model = model.to(device)
+    image = image.to(device)
+
+    with torch.no_grad():
+        outputs = model(image[None], captions=[caption])
+
+    prediction_logits = outputs["pred_logits"].cpu().sigmoid()[0]  # prediction_logits.shape = (nq, 256)
+    prediction_boxes = outputs["pred_boxes"].cpu()[0]  # prediction_boxes.shape = (nq, 4)
+
+    mask = prediction_logits.max(dim=1)[0] > box_threshold
+    logits = prediction_logits[mask]  # logits.shape = (n, 256)
+    boxes = prediction_boxes[mask]  # boxes.shape = (n, 4)
+
+    tokenizer = model.tokenizer
+    tokenized = tokenizer(caption)
+
+    phrases = [
+        get_phrases_from_posmap(logit > text_threshold, tokenized, tokenizer).replace(".", "") for logit in logits
+    ]
+
+    return boxes, logits.max(dim=1)[0], phrases
+
+
+def annotate(image_source: np.ndarray, boxes: torch.Tensor, logits: torch.Tensor, phrases: List[str]) -> np.ndarray:
+    h, w, _ = image_source.shape
+    boxes = boxes * torch.Tensor([w, h, w, h])
+    xyxy = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xyxy").numpy()
+    detections = sv.Detections(xyxy=xyxy)
+
+    labels = [f"{phrase} {logit:.2f}" for phrase, logit in zip(phrases, logits)]
+
+    box_annotator = sv.BoxAnnotator()
+    annotated_frame = cv2.cvtColor(image_source, cv2.COLOR_RGB2BGR)
+    annotated_frame = box_annotator.annotate(scene=annotated_frame, detections=detections, labels=labels)
+    return annotated_frame
+
+
+# ----------------------------------------------------------------------------------------------------------------------
+# NEW API
+# ----------------------------------------------------------------------------------------------------------------------
+
+
+class Model:
+
+    def __init__(self, model_config_path: str, model_state_dict: Dict[str, torch.Tensor], device: str = "cuda"):
+        self.model = load_model(
+            model_config_path=model_config_path, model_state_dict=model_state_dict, device=device
+        ).to(device)
+        self.device = device
+
+    def predict_with_caption(
+        self, image: np.ndarray, caption: str, box_threshold: float = 0.35, text_threshold: float = 0.25
+    ) -> Tuple[sv.Detections, List[str]]:
+        """
+        import cv2
+
+        image = cv2.imread(IMAGE_PATH)
+
+        model = Model(model_config_path=CONFIG_PATH, model_checkpoint_path=WEIGHTS_PATH)
+        detections, labels = model.predict_with_caption(
+            image=image,
+            caption=caption,
+            box_threshold=BOX_THRESHOLD,
+            text_threshold=TEXT_THRESHOLD
+        )
+
+        import supervision as sv
+
+        box_annotator = sv.BoxAnnotator()
+        annotated_image = box_annotator.annotate(scene=image, detections=detections, labels=labels)
+        """
+        processed_image = Model.preprocess_image(image_bgr=image).to(self.device)
+        boxes, logits, phrases = predict(
+            model=self.model,
+            image=processed_image,
+            caption=caption,
+            box_threshold=box_threshold,
+            text_threshold=text_threshold,
+            device=self.device,
+        )
+        source_h, source_w, _ = image.shape
+        detections = Model.post_process_result(source_h=source_h, source_w=source_w, boxes=boxes, logits=logits)
+        return detections, phrases
+
+    def predict_with_classes(
+        self, image: np.ndarray, classes: List[str], box_threshold: float, text_threshold: float
+    ) -> sv.Detections:
+        """
+        import cv2
+
+        image = cv2.imread(IMAGE_PATH)
+
+        model = Model(model_config_path=CONFIG_PATH, model_checkpoint_path=WEIGHTS_PATH)
+        detections = model.predict_with_classes(
+            image=image,
+            classes=CLASSES,
+            box_threshold=BOX_THRESHOLD,
+            text_threshold=TEXT_THRESHOLD
+        )
+
+
+        import supervision as sv
+
+        box_annotator = sv.BoxAnnotator()
+        annotated_image = box_annotator.annotate(scene=image, detections=detections)
+        """
+        caption = ". ".join(classes)
+        processed_image = Model.preprocess_image(image_bgr=image).to(self.device)
+        boxes, logits, phrases = predict(
+            model=self.model,
+            image=processed_image,
+            caption=caption,
+            box_threshold=box_threshold,
+            text_threshold=text_threshold,
+            device=self.device,
+        )
+        source_h, source_w, _ = image.shape
+        detections = Model.post_process_result(source_h=source_h, source_w=source_w, boxes=boxes, logits=logits)
+        class_id = Model.phrases2classes(phrases=phrases, classes=classes)
+        detections.class_id = class_id
+        return detections
+
+    @staticmethod
+    def preprocess_image(image_bgr: np.ndarray) -> torch.Tensor:
+        transform = T.Compose(
+            [
+                T.RandomResize([800], max_size=1333),
+                T.ToTensor(),
+                T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+            ]
+        )
+        image_pillow = Image.fromarray(cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB))
+        image_transformed, _ = transform(image_pillow, None)
+        return image_transformed
+
+    @staticmethod
+    def post_process_result(source_h: int, source_w: int, boxes: torch.Tensor, logits: torch.Tensor) -> sv.Detections:
+        boxes = boxes * torch.Tensor([source_w, source_h, source_w, source_h])
+        xyxy = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xyxy").numpy()
+        confidence = logits.numpy()
+        return sv.Detections(xyxy=xyxy, confidence=confidence)
+
+    @staticmethod
+    def phrases2classes(phrases: List[str], classes: List[str]) -> np.ndarray:
+        class_ids = []
+        for phrase in phrases:
+            try:
+                # class_ids.append(classes.index(phrase))
+                class_ids.append(Model.find_index(phrase, classes))
+            except ValueError:
+                class_ids.append(None)
+        return np.array(class_ids)
+
+    @staticmethod
+    def find_index(string, lst):
+        # if meet string like "lake river" will only keep "lake"
+        # this is an hack implementation for visualization which will be updated in the future
+        string = string.lower().split()[0]
+        for i, s in enumerate(lst):
+            if string in s.lower():
+                return i
+        print(
+            "There's a wrong phrase happen, this is because of our post-process merged wrong tokens, which will be \
+                modified in the future. We will assign it with a random label at this time."
+        )
+        return 0

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/misc.py

@@ -0,0 +1,701 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+"""
+Misc functions, including distributed helpers.
+
+Mostly copy-paste from torchvision references.
+"""
+import colorsys
+import datetime
+import functools
+import io
+import json
+import os
+import pickle
+import subprocess
+import time
+from collections import OrderedDict, defaultdict, deque
+from typing import List, Optional
+
+import numpy as np
+import torch
+import torch.distributed as dist
+
+# needed due to empty tensor bug in pytorch and torchvision 0.5
+import torchvision
+from torch import Tensor
+
+__torchvision_need_compat_flag = float(torchvision.__version__.split(".")[1]) < 7
+if __torchvision_need_compat_flag:
+    from torchvision.ops import _new_empty_tensor
+    from torchvision.ops.misc import _output_size
+
+
+class SmoothedValue(object):
+    """Track a series of values and provide access to smoothed values over a
+    window or the global series average.
+    """
+
+    def __init__(self, window_size=20, fmt=None):
+        if fmt is None:
+            fmt = "{median:.4f} ({global_avg:.4f})"
+        self.deque = deque(maxlen=window_size)
+        self.total = 0.0
+        self.count = 0
+        self.fmt = fmt
+
+    def update(self, value, n=1):
+        self.deque.append(value)
+        self.count += n
+        self.total += value * n
+
+    def synchronize_between_processes(self):
+        """
+        Warning: does not synchronize the deque!
+        """
+        if not is_dist_avail_and_initialized():
+            return
+        t = torch.tensor([self.count, self.total], dtype=torch.float64, device="cuda")
+        dist.barrier()
+        dist.all_reduce(t)
+        t = t.tolist()
+        self.count = int(t[0])
+        self.total = t[1]
+
+    @property
+    def median(self):
+        d = torch.tensor(list(self.deque))
+        if d.shape[0] == 0:
+            return 0
+        return d.median().item()
+
+    @property
+    def avg(self):
+        d = torch.tensor(list(self.deque), dtype=torch.float32)
+        return d.mean().item()
+
+    @property
+    def global_avg(self):
+        if os.environ.get("SHILONG_AMP", None) == "1":
+            eps = 1e-4
+        else:
+            eps = 1e-6
+        return self.total / (self.count + eps)
+
+    @property
+    def max(self):
+        return max(self.deque)
+
+    @property
+    def value(self):
+        return self.deque[-1]
+
+    def __str__(self):
+        return self.fmt.format(
+            median=self.median,
+            avg=self.avg,
+            global_avg=self.global_avg,
+            max=self.max,
+            value=self.value,
+        )
+
+
+@functools.lru_cache()
+def _get_global_gloo_group():
+    """
+    Return a process group based on gloo backend, containing all the ranks
+    The result is cached.
+    """
+
+    if dist.get_backend() == "nccl":
+        return dist.new_group(backend="gloo")
+
+    return dist.group.WORLD
+
+
+def all_gather_cpu(data):
+    """
+    Run all_gather on arbitrary picklable data (not necessarily tensors)
+    Args:
+        data: any picklable object
+    Returns:
+        list[data]: list of data gathered from each rank
+    """
+
+    world_size = get_world_size()
+    if world_size == 1:
+        return [data]
+
+    cpu_group = _get_global_gloo_group()
+
+    buffer = io.BytesIO()
+    torch.save(data, buffer)
+    data_view = buffer.getbuffer()
+    device = "cuda" if cpu_group is None else "cpu"
+    tensor = torch.ByteTensor(data_view).to(device)
+
+    # obtain Tensor size of each rank
+    local_size = torch.tensor([tensor.numel()], device=device, dtype=torch.long)
+    size_list = [torch.tensor([0], device=device, dtype=torch.long) for _ in range(world_size)]
+    if cpu_group is None:
+        dist.all_gather(size_list, local_size)
+    else:
+        print("gathering on cpu")
+        dist.all_gather(size_list, local_size, group=cpu_group)
+    size_list = [int(size.item()) for size in size_list]
+    max_size = max(size_list)
+    assert isinstance(local_size.item(), int)
+    local_size = int(local_size.item())
+
+    # receiving Tensor from all ranks
+    # we pad the tensor because torch all_gather does not support
+    # gathering tensors of different shapes
+    tensor_list = []
+    for _ in size_list:
+        tensor_list.append(torch.empty((max_size,), dtype=torch.uint8, device=device))
+    if local_size != max_size:
+        padding = torch.empty(size=(max_size - local_size,), dtype=torch.uint8, device=device)
+        tensor = torch.cat((tensor, padding), dim=0)
+    if cpu_group is None:
+        dist.all_gather(tensor_list, tensor)
+    else:
+        dist.all_gather(tensor_list, tensor, group=cpu_group)
+
+    data_list = []
+    for size, tensor in zip(size_list, tensor_list, strict=False):
+        tensor = torch.split(tensor, [size, max_size - size], dim=0)[0]
+        buffer = io.BytesIO(tensor.cpu().numpy())
+        obj = torch.load(buffer)
+        data_list.append(obj)
+
+    return data_list
+
+
+def all_gather(data):
+    """
+    Run all_gather on arbitrary picklable data (not necessarily tensors)
+    Args:
+        data: any picklable object
+    Returns:
+        list[data]: list of data gathered from each rank
+    """
+
+    if os.getenv("CPU_REDUCE") == "1":
+        return all_gather_cpu(data)
+
+    world_size = get_world_size()
+    if world_size == 1:
+        return [data]
+
+    # serialized to a Tensor
+    buffer = pickle.dumps(data)
+    storage = torch.ByteStorage.from_buffer(buffer)
+    tensor = torch.ByteTensor(storage).to("cuda")
+
+    # obtain Tensor size of each rank
+    local_size = torch.tensor([tensor.numel()], device="cuda")
+    size_list = [torch.tensor([0], device="cuda") for _ in range(world_size)]
+    dist.all_gather(size_list, local_size)
+    size_list = [int(size.item()) for size in size_list]
+    max_size = max(size_list)
+
+    # receiving Tensor from all ranks
+    # we pad the tensor because torch all_gather does not support
+    # gathering tensors of different shapes
+    tensor_list = []
+    for _ in size_list:
+        tensor_list.append(torch.empty((max_size,), dtype=torch.uint8, device="cuda"))
+    if local_size != max_size:
+        padding = torch.empty(size=(max_size - local_size,), dtype=torch.uint8, device="cuda")
+        tensor = torch.cat((tensor, padding), dim=0)
+    dist.all_gather(tensor_list, tensor)
+
+    data_list = []
+    for size, tensor in zip(size_list, tensor_list, strict=False):
+        buffer = tensor.cpu().numpy().tobytes()[:size]
+        data_list.append(pickle.loads(buffer))
+
+    return data_list
+
+
+def reduce_dict(input_dict, average=True):
+    """
+    Args:
+        input_dict (dict): all the values will be reduced
+        average (bool): whether to do average or sum
+    Reduce the values in the dictionary from all processes so that all processes
+    have the averaged results. Returns a dict with the same fields as
+    input_dict, after reduction.
+    """
+    world_size = get_world_size()
+    if world_size < 2:
+        return input_dict
+    with torch.no_grad():
+        names = []
+        values = []
+        # sort the keys so that they are consistent across processes
+        for k in sorted(input_dict.keys()):
+            names.append(k)
+            values.append(input_dict[k])
+        values = torch.stack(values, dim=0)
+        dist.all_reduce(values)
+        if average:
+            values /= world_size
+        reduced_dict = {k: v for k, v in zip(names, values, strict=False)}
+    return reduced_dict
+
+
+class MetricLogger(object):
+    def __init__(self, delimiter="\t"):
+        self.meters = defaultdict(SmoothedValue)
+        self.delimiter = delimiter
+
+    def update(self, **kwargs):
+        for k, v in kwargs.items():
+            if isinstance(v, torch.Tensor):
+                v = v.item()
+            assert isinstance(v, (float, int))
+            self.meters[k].update(v)
+
+    def __getattr__(self, attr):
+        if attr in self.meters:
+            return self.meters[attr]
+        if attr in self.__dict__:
+            return self.__dict__[attr]
+        raise AttributeError("'{}' object has no attribute '{}'".format(type(self).__name__, attr))
+
+    def __str__(self):
+        loss_str = []
+        for name, meter in self.meters.items():
+            # print(name, str(meter))
+            # import ipdb;ipdb.set_trace()
+            if meter.count > 0:
+                loss_str.append("{}: {}".format(name, str(meter)))
+        return self.delimiter.join(loss_str)
+
+    def synchronize_between_processes(self):
+        for meter in self.meters.values():
+            meter.synchronize_between_processes()
+
+    def add_meter(self, name, meter):
+        self.meters[name] = meter
+
+    def log_every(self, iterable, print_freq, header=None, logger=None):
+        if logger is None:
+            print_func = print
+        else:
+            print_func = logger.info
+
+        i = 0
+        if not header:
+            header = ""
+        start_time = time.time()
+        end = time.time()
+        iter_time = SmoothedValue(fmt="{avg:.4f}")
+        data_time = SmoothedValue(fmt="{avg:.4f}")
+        space_fmt = ":" + str(len(str(len(iterable)))) + "d"
+        if torch.cuda.is_available():
+            log_msg = self.delimiter.join(
+                [
+                    header,
+                    "[{0" + space_fmt + "}/{1}]",
+                    "eta: {eta}",
+                    "{meters}",
+                    "time: {time}",
+                    "data: {data}",
+                    "max mem: {memory:.0f}",
+                ]
+            )
+        else:
+            log_msg = self.delimiter.join(
+                [
+                    header,
+                    "[{0" + space_fmt + "}/{1}]",
+                    "eta: {eta}",
+                    "{meters}",
+                    "time: {time}",
+                    "data: {data}",
+                ]
+            )
+        MB = 1024.0 * 1024.0
+        for obj in iterable:
+            data_time.update(time.time() - end)
+            yield obj
+            # import ipdb; ipdb.set_trace()
+            iter_time.update(time.time() - end)
+            if i % print_freq == 0 or i == len(iterable) - 1:
+                eta_seconds = iter_time.global_avg * (len(iterable) - i)
+                eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
+                if torch.cuda.is_available():
+                    print_func(
+                        log_msg.format(
+                            i,
+                            len(iterable),
+                            eta=eta_string,
+                            meters=str(self),
+                            time=str(iter_time),
+                            data=str(data_time),
+                            memory=torch.cuda.max_memory_allocated() / MB,
+                        )
+                    )
+                else:
+                    print_func(
+                        log_msg.format(
+                            i,
+                            len(iterable),
+                            eta=eta_string,
+                            meters=str(self),
+                            time=str(iter_time),
+                            data=str(data_time),
+                        )
+                    )
+            i += 1
+            end = time.time()
+        total_time = time.time() - start_time
+        total_time_str = str(datetime.timedelta(seconds=int(total_time)))
+        print_func("{} Total time: {} ({:.4f} s / it)".format(header, total_time_str, total_time / len(iterable)))
+
+
+def get_sha():
+    cwd = os.path.dirname(os.path.abspath(__file__))
+
+    def _run(command):
+        return subprocess.check_output(command, cwd=cwd).decode("ascii").strip()
+
+    sha = "N/A"
+    diff = "clean"
+    branch = "N/A"
+    try:
+        sha = _run(["git", "rev-parse", "HEAD"])
+        subprocess.check_output(["git", "diff"], cwd=cwd)
+        diff = _run(["git", "diff-index", "HEAD"])
+        diff = "has uncommited changes" if diff else "clean"
+        branch = _run(["git", "rev-parse", "--abbrev-ref", "HEAD"])
+    except Exception:
+        pass
+    message = f"sha: {sha}, status: {diff}, branch: {branch}"
+    return message
+
+
+def collate_fn(batch):
+    # import ipdb; ipdb.set_trace()
+    batch = list(zip(*batch, strict=False))
+    batch[0] = nested_tensor_from_tensor_list(batch[0])
+    return tuple(batch)
+
+
+def _max_by_axis(the_list):
+    # type: (List[List[int]]) -> List[int]
+    maxes = the_list[0]
+    for sublist in the_list[1:]:
+        for index, item in enumerate(sublist):
+            maxes[index] = max(maxes[index], item)
+    return maxes
+
+
+class NestedTensor(object):
+    def __init__(self, tensors, mask: Optional[Tensor]):
+        self.tensors = tensors
+        self.mask = mask
+        if mask == "auto":
+            self.mask = torch.zeros_like(tensors).to(tensors.device)
+            if self.mask.dim() == 3:
+                self.mask = self.mask.sum(0).to(bool)
+            elif self.mask.dim() == 4:
+                self.mask = self.mask.sum(1).to(bool)
+            else:
+                raise ValueError("tensors dim must be 3 or 4 but {}({})".format(self.tensors.dim(), self.tensors.shape))
+
+    def imgsize(self):
+        res = []
+        for i in range(self.tensors.shape[0]):
+            mask = self.mask[i]
+            maxH = (~mask).sum(0).max()
+            maxW = (~mask).sum(1).max()
+            res.append(torch.Tensor([maxH, maxW]))
+        return res
+
+    def to(self, device):
+        # type: (Device) -> NestedTensor # noqa
+        cast_tensor = self.tensors.to(device)
+        mask = self.mask
+        if mask is not None:
+            assert mask is not None
+            cast_mask = mask.to(device)
+        else:
+            cast_mask = None
+        return NestedTensor(cast_tensor, cast_mask)
+
+    def to_img_list_single(self, tensor, mask):
+        assert tensor.dim() == 3, "dim of tensor should be 3 but {}".format(tensor.dim())
+        maxH = (~mask).sum(0).max()
+        maxW = (~mask).sum(1).max()
+        img = tensor[:, :maxH, :maxW]
+        return img
+
+    def to_img_list(self):
+        """remove the padding and convert to img list
+
+        Returns:
+            [type]: [description]
+        """
+        if self.tensors.dim() == 3:
+            return self.to_img_list_single(self.tensors, self.mask)
+        else:
+            res = []
+            for i in range(self.tensors.shape[0]):
+                tensor_i = self.tensors[i]
+                mask_i = self.mask[i]
+                res.append(self.to_img_list_single(tensor_i, mask_i))
+            return res
+
+    @property
+    def device(self):
+        return self.tensors.device
+
+    def decompose(self):
+        return self.tensors, self.mask
+
+    def __repr__(self):
+        return str(self.tensors)
+
+    @property
+    def shape(self):
+        return {"tensors.shape": self.tensors.shape, "mask.shape": self.mask.shape}
+
+
+def nested_tensor_from_tensor_list(tensor_list: List[Tensor]):
+    # TODO make this more general
+    if tensor_list[0].ndim == 3:
+        if torchvision._is_tracing():
+            # nested_tensor_from_tensor_list() does not export well to ONNX
+            # call _onnx_nested_tensor_from_tensor_list() instead
+            return _onnx_nested_tensor_from_tensor_list(tensor_list)
+
+        # TODO make it support different-sized images
+        max_size = _max_by_axis([list(img.shape) for img in tensor_list])
+        # min_size = tuple(min(s) for s in zip(*[img.shape for img in tensor_list]))
+        batch_shape = [len(tensor_list)] + max_size
+        b, c, h, w = batch_shape
+        dtype = tensor_list[0].dtype
+        device = tensor_list[0].device
+        tensor = torch.zeros(batch_shape, dtype=dtype, device=device)
+        mask = torch.ones((b, h, w), dtype=torch.bool, device=device)
+        for img, pad_img, m in zip(tensor_list, tensor, mask, strict=False):
+            pad_img[: img.shape[0], : img.shape[1], : img.shape[2]].copy_(img)
+            m[: img.shape[1], : img.shape[2]] = False
+    else:
+        raise ValueError("not supported")
+    return NestedTensor(tensor, mask)
+
+
+# _onnx_nested_tensor_from_tensor_list() is an implementation of
+# nested_tensor_from_tensor_list() that is supported by ONNX tracing.
+@torch.jit.unused
+def _onnx_nested_tensor_from_tensor_list(tensor_list: List[Tensor]) -> NestedTensor:
+    max_size = []
+    for i in range(tensor_list[0].dim()):
+        max_size_i = torch.max(torch.stack([img.shape[i] for img in tensor_list]).to(torch.float32)).to(torch.int64)
+        max_size.append(max_size_i)
+    max_size = tuple(max_size)
+
+    # work around for
+    # pad_img[: img.shape[0], : img.shape[1], : img.shape[2]].copy_(img)
+    # m[: img.shape[1], :img.shape[2]] = False
+    # which is not yet supported in onnx
+    padded_imgs = []
+    padded_masks = []
+    for img in tensor_list:
+        padding = [(s1 - s2) for s1, s2 in zip(max_size, tuple(img.shape), strict=False)]
+        padded_img = torch.nn.functional.pad(img, (0, padding[2], 0, padding[1], 0, padding[0]))
+        padded_imgs.append(padded_img)
+
+        m = torch.zeros_like(img[0], dtype=torch.int, device=img.device)
+        padded_mask = torch.nn.functional.pad(m, (0, padding[2], 0, padding[1]), "constant", 1)
+        padded_masks.append(padded_mask.to(torch.bool))
+
+    tensor = torch.stack(padded_imgs)
+    mask = torch.stack(padded_masks)
+
+    return NestedTensor(tensor, mask=mask)
+
+
+def setup_for_distributed(is_master):
+    """
+    This function disables printing when not in master process
+    """
+    import builtins as __builtin__
+
+    builtin_print = __builtin__.print
+
+    def print(*args, **kwargs):
+        force = kwargs.pop("force", False)
+        if is_master or force:
+            builtin_print(*args, **kwargs)
+
+    __builtin__.print = print
+
+
+def is_dist_avail_and_initialized():
+    if not dist.is_available():
+        return False
+    if not dist.is_initialized():
+        return False
+    return True
+
+
+def get_world_size():
+    if not is_dist_avail_and_initialized():
+        return 1
+    return dist.get_world_size()
+
+
+def get_rank():
+    if not is_dist_avail_and_initialized():
+        return 0
+    return dist.get_rank()
+
+
+def is_main_process():
+    return get_rank() == 0
+
+
+def save_on_master(*args, **kwargs):
+    if is_main_process():
+        torch.save(*args, **kwargs)
+
+
+def init_distributed_mode(args):
+    if "WORLD_SIZE" in os.environ and os.environ["WORLD_SIZE"] != "":  # 'RANK' in os.environ and
+        args.rank = int(os.environ["RANK"])
+        args.world_size = int(os.environ["WORLD_SIZE"])
+        args.gpu = args.local_rank = int(os.environ["LOCAL_RANK"])
+
+        # launch by torch.distributed.launch
+        # Single node
+        #   python -m torch.distributed.launch --nproc_per_node=8 main.py --world-size 1 --rank 0 ...
+        # Multi nodes
+        #   python -m torch.distributed.launch --nproc_per_node=8 main.py --world-size 2 --rank 0 --dist-url 'tcp://IP_OF_NODE0:FREEPORT' ...
+        #   python -m torch.distributed.launch --nproc_per_node=8 main.py --world-size 2 --rank 1 --dist-url 'tcp://IP_OF_NODE0:FREEPORT' ...
+        # args.rank = int(os.environ.get('OMPI_COMM_WORLD_RANK'))
+        # local_world_size = int(os.environ['GPU_PER_NODE_COUNT'])
+        # args.world_size = args.world_size * local_world_size
+        # args.gpu = args.local_rank = int(os.environ['LOCAL_RANK'])
+        # args.rank = args.rank * local_world_size + args.local_rank
+        print("world size: {}, rank: {}, local rank: {}".format(args.world_size, args.rank, args.local_rank))
+        print(json.dumps(dict(os.environ), indent=2))
+    elif "SLURM_PROCID" in os.environ:
+        args.rank = int(os.environ["SLURM_PROCID"])
+        args.gpu = args.local_rank = int(os.environ["SLURM_LOCALID"])
+        args.world_size = int(os.environ["SLURM_NPROCS"])
+
+        print(
+            "world size: {}, world rank: {}, local rank: {}, device_count: {}".format(
+                args.world_size, args.rank, args.local_rank, torch.cuda.device_count()
+            )
+        )
+    else:
+        print("Not using distributed mode")
+        args.distributed = False
+        args.world_size = 1
+        args.rank = 0
+        args.local_rank = 0
+        return
+
+    print("world_size:{} rank:{} local_rank:{}".format(args.world_size, args.rank, args.local_rank))
+    args.distributed = True
+    torch.cuda.set_device(args.local_rank)
+    args.dist_backend = "nccl"
+    print("| distributed init (rank {}): {}".format(args.rank, args.dist_url), flush=True)
+
+    torch.distributed.init_process_group(
+        backend=args.dist_backend,
+        world_size=args.world_size,
+        rank=args.rank,
+        init_method=args.dist_url,
+    )
+
+    print("Before torch.distributed.barrier()")
+    torch.distributed.barrier()
+    print("End torch.distributed.barrier()")
+    setup_for_distributed(args.rank == 0)
+
+
+@torch.no_grad()
+def accuracy(output, target, topk=(1,)):
+    """Computes the precision@k for the specified values of k"""
+    if target.numel() == 0:
+        return [torch.zeros([], device=output.device)]
+    maxk = max(topk)
+    batch_size = target.size(0)
+
+    _, pred = output.topk(maxk, 1, True, True)
+    pred = pred.t()
+    correct = pred.eq(target.view(1, -1).expand_as(pred))
+
+    res = []
+    for k in topk:
+        correct_k = correct[:k].view(-1).float().sum(0)
+        res.append(correct_k.mul_(100.0 / batch_size))
+    return res
+
+
+@torch.no_grad()
+def accuracy_onehot(pred, gt):
+    """_summary_
+
+    Args:
+        pred (_type_): n, c
+        gt (_type_): n, c
+    """
+    tp = ((pred - gt).abs().sum(-1) < 1e-4).float().sum()
+    acc = tp / gt.shape[0] * 100
+    return acc
+
+
+def interpolate(input, size=None, scale_factor=None, mode="nearest", align_corners=None):
+    # type: (Tensor, Optional[List[int]], Optional[float], str, Optional[bool]) -> Tensor
+    """
+    Equivalent to nn.functional.interpolate, but with support for empty batch sizes.
+    This will eventually be supported natively by PyTorch, and this
+    class can go away.
+    """
+    if __torchvision_need_compat_flag < 0.7:
+        if input.numel() > 0:
+            return torch.nn.functional.interpolate(input, size, scale_factor, mode, align_corners)
+
+        output_shape = _output_size(2, input, size, scale_factor)
+        output_shape = list(input.shape[:-2]) + list(output_shape)
+        return _new_empty_tensor(input, output_shape)
+    else:
+        return torchvision.ops.misc.interpolate(input, size, scale_factor, mode, align_corners)
+
+
+class color_sys:
+    def __init__(self, num_colors) -> None:
+        self.num_colors = num_colors
+        colors = []
+        for i in np.arange(0.0, 360.0, 360.0 / num_colors):
+            hue = i / 360.0
+            lightness = (50 + np.random.rand() * 10) / 100.0
+            saturation = (90 + np.random.rand() * 10) / 100.0
+            colors.append(tuple([int(j * 255) for j in colorsys.hls_to_rgb(hue, lightness, saturation)]))
+        self.colors = colors
+
+    def __call__(self, idx):
+        return self.colors[idx]
+
+
+def inverse_sigmoid(x, eps=1e-3):
+    x = x.clamp(min=0, max=1)
+    x1 = x.clamp(min=eps)
+    x2 = (1 - x).clamp(min=eps)
+    return torch.log(x1 / x2)
+
+
+def clean_state_dict(state_dict):
+    new_state_dict = OrderedDict()
+    for k, v in state_dict.items():
+        if k[:7] == "module.":
+            k = k[7:]  # remove `module.`
+        new_state_dict[k] = v
+    return new_state_dict

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/slconfig.py

@@ -0,0 +1,419 @@
+# ==========================================================
+# Modified from mmcv
+# ==========================================================
+import ast
+import os.path as osp
+import platform
+import shutil
+import sys
+import tempfile
+from argparse import Action
+from importlib import import_module
+
+from addict import Dict
+from yapf.yapflib.yapf_api import FormatCode
+
+BASE_KEY = "_base_"
+DELETE_KEY = "_delete_"
+RESERVED_KEYS = ["filename", "text", "pretty_text", "get", "dump", "merge_from_dict"]
+
+
+def check_file_exist(filename, msg_tmpl='file "{}" does not exist'):
+    if not osp.isfile(filename):
+        raise FileNotFoundError(msg_tmpl.format(filename))
+
+
+class ConfigDict(Dict):
+    def __missing__(self, name):
+        raise KeyError(name)
+
+    def __getattr__(self, name):
+        try:
+            value = super(ConfigDict, self).__getattr__(name)
+        except KeyError:
+            ex = AttributeError(f"'{self.__class__.__name__}' object has no " f"attribute '{name}'")
+        except Exception as e:
+            ex = e
+        else:
+            return value
+        raise ex
+
+
+class SLConfig(object):
+    """
+    config files.
+    only support .py file as config now.
+
+    ref: mmcv.utils.config
+
+    Example:
+        >>> cfg = Config(dict(a=1, b=dict(b1=[0, 1])))
+        >>> cfg.a
+        1
+        >>> cfg.b
+        {'b1': [0, 1]}
+        >>> cfg.b.b1
+        [0, 1]
+        >>> cfg = Config.fromfile('tests/data/config/a.py')
+        >>> cfg.filename
+        "/home/kchen/projects/mmcv/tests/data/config/a.py"
+        >>> cfg.item4
+        'test'
+        >>> cfg
+        "Config [path: /home/kchen/projects/mmcv/tests/data/config/a.py]: "
+        "{'item1': [1, 2], 'item2': {'a': 0}, 'item3': True, 'item4': 'test'}"
+    """
+
+    @staticmethod
+    def _validate_py_syntax(filename):
+        with open(filename) as f:
+            content = f.read()
+        try:
+            ast.parse(content)
+        except SyntaxError:
+            raise SyntaxError("There are syntax errors in config " f"file {filename}")
+
+    @staticmethod
+    def _file2dict(filename):
+        filename = osp.abspath(osp.expanduser(filename))
+        check_file_exist(filename)
+        if filename.lower().endswith(".py"):
+            with tempfile.TemporaryDirectory() as temp_config_dir:
+                temp_config_file = tempfile.NamedTemporaryFile(dir=temp_config_dir, suffix=".py")
+                temp_config_name = osp.basename(temp_config_file.name)
+                if platform.system() == "Windows":
+                    temp_config_file.close()
+                shutil.copyfile(filename, osp.join(temp_config_dir, temp_config_name))
+                temp_module_name = osp.splitext(temp_config_name)[0]
+                sys.path.insert(0, temp_config_dir)
+                SLConfig._validate_py_syntax(filename)
+                mod = import_module(temp_module_name)
+                sys.path.pop(0)
+                cfg_dict = {name: value for name, value in mod.__dict__.items() if not name.startswith("__")}
+                # delete imported module
+                del sys.modules[temp_module_name]
+                # close temp file
+                temp_config_file.close()
+        elif filename.lower().endswith((".yml", ".yaml", ".json")):
+            from .slio import slload
+
+            cfg_dict = slload(filename)
+        else:
+            raise IOError("Only py/yml/yaml/json type are supported now!")
+
+        cfg_text = filename + "\n"
+        with open(filename, "r") as f:
+            cfg_text += f.read()
+
+        # parse the base file
+        if BASE_KEY in cfg_dict:
+            cfg_dir = osp.dirname(filename)
+            base_filename = cfg_dict.pop(BASE_KEY)
+            base_filename = base_filename if isinstance(base_filename, list) else [base_filename]
+
+            cfg_dict_list = list()
+            cfg_text_list = list()
+            for f in base_filename:
+                _cfg_dict, _cfg_text = SLConfig._file2dict(osp.join(cfg_dir, f))
+                cfg_dict_list.append(_cfg_dict)
+                cfg_text_list.append(_cfg_text)
+
+            base_cfg_dict = dict()
+            for c in cfg_dict_list:
+                if len(base_cfg_dict.keys() & c.keys()) > 0:
+                    raise KeyError("Duplicate key is not allowed among bases")
+                    # TODO Allow the duplicate key while warnning user
+                base_cfg_dict.update(c)
+
+            base_cfg_dict = SLConfig._merge_a_into_b(cfg_dict, base_cfg_dict)
+            cfg_dict = base_cfg_dict
+
+            # merge cfg_text
+            cfg_text_list.append(cfg_text)
+            cfg_text = "\n".join(cfg_text_list)
+
+        return cfg_dict, cfg_text
+
+    @staticmethod
+    def _merge_a_into_b(a, b):
+        """merge dict `a` into dict `b` (non-inplace).
+            values in `a` will overwrite `b`.
+            copy first to avoid inplace modification
+
+        Args:
+            a ([type]): [description]
+            b ([type]): [description]
+
+        Returns:
+            [dict]: [description]
+        """
+        # import ipdb; ipdb.set_trace()
+        if not isinstance(a, dict):
+            return a
+
+        b = b.copy()
+        for k, v in a.items():
+            if isinstance(v, dict) and k in b and not v.pop(DELETE_KEY, False):
+
+                if not isinstance(b[k], dict) and not isinstance(b[k], list):
+                    # if :
+                    # import ipdb; ipdb.set_trace()
+                    raise TypeError(
+                        f"{k}={v} in child config cannot inherit from base "
+                        f"because {k} is a dict in the child config but is of "
+                        f"type {type(b[k])} in base config. You may set "
+                        f"`{DELETE_KEY}=True` to ignore the base config"
+                    )
+                b[k] = SLConfig._merge_a_into_b(v, b[k])
+            elif isinstance(b, list):
+                try:
+                    _ = int(k)
+                except:
+                    raise TypeError(f"b is a list, " f"index {k} should be an int when input but {type(k)}")
+                b[int(k)] = SLConfig._merge_a_into_b(v, b[int(k)])
+            else:
+                b[k] = v
+
+        return b
+
+    @staticmethod
+    def fromfile(filename):
+        cfg_dict, cfg_text = SLConfig._file2dict(filename)
+        return SLConfig(cfg_dict, cfg_text=cfg_text, filename=filename)
+
+    def __init__(self, cfg_dict=None, cfg_text=None, filename=None):
+        if cfg_dict is None:
+            cfg_dict = dict()
+        elif not isinstance(cfg_dict, dict):
+            raise TypeError("cfg_dict must be a dict, but " f"got {type(cfg_dict)}")
+        for key in cfg_dict:
+            if key in RESERVED_KEYS:
+                raise KeyError(f"{key} is reserved for config file")
+
+        super(SLConfig, self).__setattr__("_cfg_dict", ConfigDict(cfg_dict))
+        super(SLConfig, self).__setattr__("_filename", filename)
+        if cfg_text:
+            text = cfg_text
+        elif filename:
+            with open(filename, "r") as f:
+                text = f.read()
+        else:
+            text = ""
+        super(SLConfig, self).__setattr__("_text", text)
+
+    @property
+    def filename(self):
+        return self._filename
+
+    @property
+    def text(self):
+        return self._text
+
+    @property
+    def pretty_text(self):
+
+        indent = 4
+
+        def _indent(s_, num_spaces):
+            s = s_.split("\n")
+            if len(s) == 1:
+                return s_
+            first = s.pop(0)
+            s = [(num_spaces * " ") + line for line in s]
+            s = "\n".join(s)
+            s = first + "\n" + s
+            return s
+
+        def _format_basic_types(k, v, use_mapping=False):
+            if isinstance(v, str):
+                v_str = f"'{v}'"
+            else:
+                v_str = str(v)
+
+            if use_mapping:
+                k_str = f"'{k}'" if isinstance(k, str) else str(k)
+                attr_str = f"{k_str}: {v_str}"
+            else:
+                attr_str = f"{str(k)}={v_str}"
+            attr_str = _indent(attr_str, indent)
+
+            return attr_str
+
+        def _format_list(k, v, use_mapping=False):
+            # check if all items in the list are dict
+            if all(isinstance(_, dict) for _ in v):
+                v_str = "[\n"
+                v_str += "\n".join(f"dict({_indent(_format_dict(v_), indent)})," for v_ in v).rstrip(",")
+                if use_mapping:
+                    k_str = f"'{k}'" if isinstance(k, str) else str(k)
+                    attr_str = f"{k_str}: {v_str}"
+                else:
+                    attr_str = f"{str(k)}={v_str}"
+                attr_str = _indent(attr_str, indent) + "]"
+            else:
+                attr_str = _format_basic_types(k, v, use_mapping)
+            return attr_str
+
+        def _contain_invalid_identifier(dict_str):
+            contain_invalid_identifier = False
+            for key_name in dict_str:
+                contain_invalid_identifier |= not str(key_name).isidentifier()
+            return contain_invalid_identifier
+
+        def _format_dict(input_dict, outest_level=False):
+            r = ""
+            s = []
+
+            use_mapping = _contain_invalid_identifier(input_dict)
+            if use_mapping:
+                r += "{"
+            for idx, (k, v) in enumerate(input_dict.items()):
+                is_last = idx >= len(input_dict) - 1
+                end = "" if outest_level or is_last else ","
+                if isinstance(v, dict):
+                    v_str = "\n" + _format_dict(v)
+                    if use_mapping:
+                        k_str = f"'{k}'" if isinstance(k, str) else str(k)
+                        attr_str = f"{k_str}: dict({v_str}"
+                    else:
+                        attr_str = f"{str(k)}=dict({v_str}"
+                    attr_str = _indent(attr_str, indent) + ")" + end
+                elif isinstance(v, list):
+                    attr_str = _format_list(k, v, use_mapping) + end
+                else:
+                    attr_str = _format_basic_types(k, v, use_mapping) + end
+
+                s.append(attr_str)
+            r += "\n".join(s)
+            if use_mapping:
+                r += "}"
+            return r
+
+        cfg_dict = self._cfg_dict.to_dict()
+        text = _format_dict(cfg_dict, outest_level=True)
+        # copied from setup.cfg
+        yapf_style = dict(
+            based_on_style="pep8",
+            blank_line_before_nested_class_or_def=True,
+            split_before_expression_after_opening_paren=True,
+        )
+        text, _ = FormatCode(text, style_config=yapf_style, verify=True)
+
+        return text
+
+    def __repr__(self):
+        return f"Config (path: {self.filename}): {self._cfg_dict.__repr__()}"
+
+    def __len__(self):
+        return len(self._cfg_dict)
+
+    def __getattr__(self, name):
+        # # debug
+        # print('+'*15)
+        # print('name=%s' % name)
+        # print("addr:", id(self))
+        # # print('type(self):', type(self))
+        # print(self.__dict__)
+        # print('+'*15)
+        # if self.__dict__ == {}:
+        #     raise ValueError
+
+        return getattr(self._cfg_dict, name)
+
+    def __getitem__(self, name):
+        return self._cfg_dict.__getitem__(name)
+
+    def __setattr__(self, name, value):
+        if isinstance(value, dict):
+            value = ConfigDict(value)
+        self._cfg_dict.__setattr__(name, value)
+
+    def __setitem__(self, name, value):
+        if isinstance(value, dict):
+            value = ConfigDict(value)
+        self._cfg_dict.__setitem__(name, value)
+
+    def __iter__(self):
+        return iter(self._cfg_dict)
+
+    def dump(self, file=None):
+        # import ipdb; ipdb.set_trace()
+        if file is None:
+            return self.pretty_text
+        else:
+            with open(file, "w") as f:
+                f.write(self.pretty_text)
+
+    def merge_from_dict(self, options):
+        """Merge list into cfg_dict
+
+        Merge the dict parsed by MultipleKVAction into this cfg.
+
+        Examples:
+            >>> options = {'model.backbone.depth': 50,
+            ...            'model.backbone.with_cp':True}
+            >>> cfg = Config(dict(model=dict(backbone=dict(type='ResNet'))))
+            >>> cfg.merge_from_dict(options)
+            >>> cfg_dict = super(Config, self).__getattribute__('_cfg_dict')
+            >>> assert cfg_dict == dict(
+            ...     model=dict(backbone=dict(depth=50, with_cp=True)))
+
+        Args:
+            options (dict): dict of configs to merge from.
+        """
+        option_cfg_dict = {}
+        for full_key, v in options.items():
+            d = option_cfg_dict
+            key_list = full_key.split(".")
+            for subkey in key_list[:-1]:
+                d.setdefault(subkey, ConfigDict())
+                d = d[subkey]
+            subkey = key_list[-1]
+            d[subkey] = v
+
+        cfg_dict = super(SLConfig, self).__getattribute__("_cfg_dict")
+        super(SLConfig, self).__setattr__("_cfg_dict", SLConfig._merge_a_into_b(option_cfg_dict, cfg_dict))
+
+    # for multiprocess
+    def __setstate__(self, state):
+        self.__init__(state)
+
+    def copy(self):
+        return SLConfig(self._cfg_dict.copy())
+
+    def deepcopy(self):
+        return SLConfig(self._cfg_dict.deepcopy())
+
+
+class DictAction(Action):
+    """
+    argparse action to split an argument into KEY=VALUE form
+    on the first = and append to a dictionary. List options should
+    be passed as comma separated values, i.e KEY=V1,V2,V3
+    """
+
+    @staticmethod
+    def _parse_int_float_bool(val):
+        try:
+            return int(val)
+        except ValueError:
+            pass
+        try:
+            return float(val)
+        except ValueError:
+            pass
+        if val.lower() in ["true", "false"]:
+            return True if val.lower() == "true" else False
+        if val.lower() in ["none", "null"]:
+            return None
+        return val
+
+    def __call__(self, parser, namespace, values, option_string=None):
+        options = {}
+        for kv in values:
+            key, val = kv.split("=", maxsplit=1)
+            val = [self._parse_int_float_bool(v) for v in val.split(",")]
+            if len(val) == 1:
+                val = val[0]
+            options[key] = val
+        setattr(namespace, self.dest, options)

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/slio.py

@@ -0,0 +1,178 @@
+# ==========================================================
+# Modified from mmcv
+# ==========================================================
+
+import json
+import pickle
+from abc import ABCMeta, abstractmethod
+from pathlib import Path
+
+import yaml
+
+try:
+    from yaml import CDumper as Dumper
+    from yaml import CLoader as Loader
+except ImportError:
+    from yaml import Dumper, Loader
+
+
+# ===========================
+# Rigister handler
+# ===========================
+
+
+class BaseFileHandler(metaclass=ABCMeta):
+    @abstractmethod
+    def load_from_fileobj(self, file, **kwargs):
+        pass
+
+    @abstractmethod
+    def dump_to_fileobj(self, obj, file, **kwargs):
+        pass
+
+    @abstractmethod
+    def dump_to_str(self, obj, **kwargs):
+        pass
+
+    def load_from_path(self, filepath, mode="r", **kwargs):
+        with open(filepath, mode) as f:
+            return self.load_from_fileobj(f, **kwargs)
+
+    def dump_to_path(self, obj, filepath, mode="w", **kwargs):
+        with open(filepath, mode) as f:
+            self.dump_to_fileobj(obj, f, **kwargs)
+
+
+class JsonHandler(BaseFileHandler):
+    def load_from_fileobj(self, file):
+        return json.load(file)
+
+    def dump_to_fileobj(self, obj, file, **kwargs):
+        json.dump(obj, file, **kwargs)
+
+    def dump_to_str(self, obj, **kwargs):
+        return json.dumps(obj, **kwargs)
+
+
+class PickleHandler(BaseFileHandler):
+    def load_from_fileobj(self, file, **kwargs):
+        return pickle.load(file, **kwargs)
+
+    def load_from_path(self, filepath, **kwargs):
+        return super(PickleHandler, self).load_from_path(filepath, mode="rb", **kwargs)
+
+    def dump_to_str(self, obj, **kwargs):
+        kwargs.setdefault("protocol", 2)
+        return pickle.dumps(obj, **kwargs)
+
+    def dump_to_fileobj(self, obj, file, **kwargs):
+        kwargs.setdefault("protocol", 2)
+        pickle.dump(obj, file, **kwargs)
+
+    def dump_to_path(self, obj, filepath, **kwargs):
+        super(PickleHandler, self).dump_to_path(obj, filepath, mode="wb", **kwargs)
+
+
+class YamlHandler(BaseFileHandler):
+    def load_from_fileobj(self, file, **kwargs):
+        kwargs.setdefault("Loader", Loader)
+        return yaml.load(file, **kwargs)
+
+    def dump_to_fileobj(self, obj, file, **kwargs):
+        kwargs.setdefault("Dumper", Dumper)
+        yaml.dump(obj, file, **kwargs)
+
+    def dump_to_str(self, obj, **kwargs):
+        kwargs.setdefault("Dumper", Dumper)
+        return yaml.dump(obj, **kwargs)
+
+
+file_handlers = {
+    "json": JsonHandler(),
+    "yaml": YamlHandler(),
+    "yml": YamlHandler(),
+    "pickle": PickleHandler(),
+    "pkl": PickleHandler(),
+}
+
+# ===========================
+# load and dump
+# ===========================
+
+
+def is_str(x):
+    """Whether the input is an string instance.
+
+    Note: This method is deprecated since python 2 is no longer supported.
+    """
+    return isinstance(x, str)
+
+
+def slload(file, file_format=None, **kwargs):
+    """Load data from json/yaml/pickle files.
+
+    This method provides a unified api for loading data from serialized files.
+
+    Args:
+        file (str or :obj:`Path` or file-like object): Filename or a file-like
+            object.
+        file_format (str, optional): If not specified, the file format will be
+            inferred from the file extension, otherwise use the specified one.
+            Currently supported formats include "json", "yaml/yml" and
+            "pickle/pkl".
+
+    Returns:
+        The content from the file.
+    """
+    if isinstance(file, Path):
+        file = str(file)
+    if file_format is None and is_str(file):
+        file_format = file.split(".")[-1]
+    if file_format not in file_handlers:
+        raise TypeError(f"Unsupported format: {file_format}")
+
+    handler = file_handlers[file_format]
+    if is_str(file):
+        obj = handler.load_from_path(file, **kwargs)
+    elif hasattr(file, "read"):
+        obj = handler.load_from_fileobj(file, **kwargs)
+    else:
+        raise TypeError('"file" must be a filepath str or a file-object')
+    return obj
+
+
+def sldump(obj, file=None, file_format=None, **kwargs):
+    """Dump data to json/yaml/pickle strings or files.
+
+    This method provides a unified api for dumping data as strings or to files,
+    and also supports custom arguments for each file format.
+
+    Args:
+        obj (any): The python object to be dumped.
+        file (str or :obj:`Path` or file-like object, optional): If not
+            specified, then the object is dump to a str, otherwise to a file
+            specified by the filename or file-like object.
+        file_format (str, optional): Same as :func:`load`.
+
+    Returns:
+        bool: True for success, False otherwise.
+    """
+    if isinstance(file, Path):
+        file = str(file)
+    if file_format is None:
+        if is_str(file):
+            file_format = file.split(".")[-1]
+        elif file is None:
+            raise ValueError("file_format must be specified since file is None")
+    if file_format not in file_handlers:
+        raise TypeError(f"Unsupported format: {file_format}")
+
+    handler = file_handlers[file_format]
+    if file is None:
+        return handler.dump_to_str(obj, **kwargs)
+    elif is_str(file):
+        handler.dump_to_path(obj, file, **kwargs)
+    elif hasattr(file, "write"):
+        handler.dump_to_fileobj(obj, file, **kwargs)
+    else:
+        raise TypeError('"file" must be a filename str or a file-object')

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/time_counter.py

@@ -0,0 +1,62 @@
+import json
+import time
+
+
+class TimeCounter:
+    def __init__(self) -> None:
+        pass
+
+    def clear(self):
+        self.timedict = {}
+        self.basetime = time.perf_counter()
+
+    def timeit(self, name):
+        nowtime = time.perf_counter() - self.basetime
+        self.timedict[name] = nowtime
+        self.basetime = time.perf_counter()
+
+
+class TimeHolder:
+    def __init__(self) -> None:
+        self.timedict = {}
+
+    def update(self, _timedict: dict):
+        for k, v in _timedict.items():
+            if k not in self.timedict:
+                self.timedict[k] = AverageMeter(name=k, val_only=True)
+            self.timedict[k].update(val=v)
+
+    def final_res(self):
+        return {k: v.avg for k, v in self.timedict.items()}
+
+    def __str__(self):
+        return json.dumps(self.final_res(), indent=2)
+
+
+class AverageMeter(object):
+    """Computes and stores the average and current value"""
+
+    def __init__(self, name, fmt=":f", val_only=False):
+        self.name = name
+        self.fmt = fmt
+        self.val_only = val_only
+        self.reset()
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count
+
+    def __str__(self):
+        if self.val_only:
+            fmtstr = "{name} {val" + self.fmt + "}"
+        else:
+            fmtstr = "{name} {val" + self.fmt + "} ({avg" + self.fmt + "})"
+        return fmtstr.format(**self.__dict__)

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/utils.py

@@ -0,0 +1,598 @@
+import argparse
+import json
+import warnings
+from collections import OrderedDict
+from copy import deepcopy
+from typing import Any, Dict, List
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer
+
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.slconfig import SLConfig
+
+
+def slprint(x, name="x"):
+    if isinstance(x, (torch.Tensor, np.ndarray)):
+        print(f"{name}.shape:", x.shape)
+    elif isinstance(x, (tuple, list)):
+        print("type x:", type(x))
+        for i in range(min(10, len(x))):
+            slprint(x[i], f"{name}[{i}]")
+    elif isinstance(x, dict):
+        for k, v in x.items():
+            slprint(v, f"{name}[{k}]")
+    else:
+        print(f"{name}.type:", type(x))
+
+
+def clean_state_dict(state_dict):
+    new_state_dict = OrderedDict()
+    for k, v in state_dict.items():
+        if k[:7] == "module.":
+            k = k[7:]  # remove `module.`
+        new_state_dict[k] = v
+    return new_state_dict
+
+
+def renorm(img: torch.FloatTensor, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) -> torch.FloatTensor:
+    # img: tensor(3,H,W) or tensor(B,3,H,W)
+    # return: same as img
+    assert img.dim() == 3 or img.dim() == 4, "img.dim() should be 3 or 4 but %d" % img.dim()
+    if img.dim() == 3:
+        assert img.size(0) == 3, 'img.size(0) shoule be 3 but "%d". (%s)' % (
+            img.size(0),
+            str(img.size()),
+        )
+        img_perm = img.permute(1, 2, 0)
+        mean = torch.Tensor(mean)
+        std = torch.Tensor(std)
+        img_res = img_perm * std + mean
+        return img_res.permute(2, 0, 1)
+    else:  # img.dim() == 4
+        assert img.size(1) == 3, 'img.size(1) shoule be 3 but "%d". (%s)' % (
+            img.size(1),
+            str(img.size()),
+        )
+        img_perm = img.permute(0, 2, 3, 1)
+        mean = torch.Tensor(mean)
+        std = torch.Tensor(std)
+        img_res = img_perm * std + mean
+        return img_res.permute(0, 3, 1, 2)
+
+
+class CocoClassMapper:
+    def __init__(self) -> None:
+        self.category_map_str = {
+            "1": 1,
+            "2": 2,
+            "3": 3,
+            "4": 4,
+            "5": 5,
+            "6": 6,
+            "7": 7,
+            "8": 8,
+            "9": 9,
+            "10": 10,
+            "11": 11,
+            "13": 12,
+            "14": 13,
+            "15": 14,
+            "16": 15,
+            "17": 16,
+            "18": 17,
+            "19": 18,
+            "20": 19,
+            "21": 20,
+            "22": 21,
+            "23": 22,
+            "24": 23,
+            "25": 24,
+            "27": 25,
+            "28": 26,
+            "31": 27,
+            "32": 28,
+            "33": 29,
+            "34": 30,
+            "35": 31,
+            "36": 32,
+            "37": 33,
+            "38": 34,
+            "39": 35,
+            "40": 36,
+            "41": 37,
+            "42": 38,
+            "43": 39,
+            "44": 40,
+            "46": 41,
+            "47": 42,
+            "48": 43,
+            "49": 44,
+            "50": 45,
+            "51": 46,
+            "52": 47,
+            "53": 48,
+            "54": 49,
+            "55": 50,
+            "56": 51,
+            "57": 52,
+            "58": 53,
+            "59": 54,
+            "60": 55,
+            "61": 56,
+            "62": 57,
+            "63": 58,
+            "64": 59,
+            "65": 60,
+            "67": 61,
+            "70": 62,
+            "72": 63,
+            "73": 64,
+            "74": 65,
+            "75": 66,
+            "76": 67,
+            "77": 68,
+            "78": 69,
+            "79": 70,
+            "80": 71,
+            "81": 72,
+            "82": 73,
+            "84": 74,
+            "85": 75,
+            "86": 76,
+            "87": 77,
+            "88": 78,
+            "89": 79,
+            "90": 80,
+        }
+        self.origin2compact_mapper = {int(k): v - 1 for k, v in self.category_map_str.items()}
+        self.compact2origin_mapper = {int(v - 1): int(k) for k, v in self.category_map_str.items()}
+
+    def origin2compact(self, idx):
+        return self.origin2compact_mapper[int(idx)]
+
+    def compact2origin(self, idx):
+        return self.compact2origin_mapper[int(idx)]
+
+
+def to_device(item, device):
+    if isinstance(item, torch.Tensor):
+        return item.to(device)
+    elif isinstance(item, list):
+        return [to_device(i, device) for i in item]
+    elif isinstance(item, dict):
+        return {k: to_device(v, device) for k, v in item.items()}
+    else:
+        raise NotImplementedError("Call Shilong if you use other containers! type: {}".format(type(item)))
+
+
+#
+def get_gaussian_mean(x, axis, other_axis, softmax=True):
+    """
+
+    Args:
+        x (float): Input images(BxCxHxW)
+        axis (int): The index for weighted mean
+        other_axis (int): The other index
+
+    Returns: weighted index for axis, BxC
+
+    """
+    mat2line = torch.sum(x, axis=other_axis)
+    # mat2line = mat2line / mat2line.mean() * 10
+    if softmax:
+        u = torch.softmax(mat2line, axis=2)
+    else:
+        u = mat2line / (mat2line.sum(2, keepdim=True) + 1e-6)
+    size = x.shape[axis]
+    ind = torch.linspace(0, 1, size).to(x.device)
+    batch = x.shape[0]
+    channel = x.shape[1]
+    index = ind.repeat([batch, channel, 1])
+    mean_position = torch.sum(index * u, dim=2)
+    return mean_position
+
+
+def get_expected_points_from_map(hm, softmax=True):
+    """get_gaussian_map_from_points
+        B,C,H,W -> B,N,2 float(0, 1) float(0, 1)
+        softargmax function
+
+    Args:
+        hm (float): Input images(BxCxHxW)
+
+    Returns:
+        weighted index for axis, BxCx2. float between 0 and 1.
+
+    """
+    # hm = 10*hm
+    B, C, H, W = hm.shape
+    y_mean = get_gaussian_mean(hm, 2, 3, softmax=softmax)  # B,C
+    x_mean = get_gaussian_mean(hm, 3, 2, softmax=softmax)  # B,C
+    # return torch.cat((x_mean.unsqueeze(-1), y_mean.unsqueeze(-1)), 2)
+    return torch.stack([x_mean, y_mean], dim=2)
+
+
+# Positional encoding (section 5.1)
+# borrow from nerf
+class Embedder:
+    def __init__(self, **kwargs):
+        self.kwargs = kwargs
+        self.create_embedding_fn()
+
+    def create_embedding_fn(self):
+        embed_fns = []
+        d = self.kwargs["input_dims"]
+        out_dim = 0
+        if self.kwargs["include_input"]:
+            embed_fns.append(lambda x: x)
+            out_dim += d
+
+        max_freq = self.kwargs["max_freq_log2"]
+        N_freqs = self.kwargs["num_freqs"]
+
+        if self.kwargs["log_sampling"]:
+            freq_bands = 2.0 ** torch.linspace(0.0, max_freq, steps=N_freqs)
+        else:
+            freq_bands = torch.linspace(2.0**0.0, 2.0**max_freq, steps=N_freqs)
+
+        for freq in freq_bands:
+            for p_fn in self.kwargs["periodic_fns"]:
+                embed_fns.append(lambda x, p_fn=p_fn, freq=freq: p_fn(x * freq))
+                out_dim += d
+
+        self.embed_fns = embed_fns
+        self.out_dim = out_dim
+
+    def embed(self, inputs):
+        return torch.cat([fn(inputs) for fn in self.embed_fns], -1)
+
+
+def get_embedder(multires, i=0):
+    import torch.nn as nn
+
+    if i == -1:
+        return nn.Identity(), 3
+
+    embed_kwargs = {
+        "include_input": True,
+        "input_dims": 3,
+        "max_freq_log2": multires - 1,
+        "num_freqs": multires,
+        "log_sampling": True,
+        "periodic_fns": [torch.sin, torch.cos],
+    }
+
+    embedder_obj = Embedder(**embed_kwargs)
+    embed = lambda x, eo=embedder_obj: eo.embed(x)
+    return embed, embedder_obj.out_dim
+
+
+class APOPMeter:
+    def __init__(self) -> None:
+        self.tp = 0
+        self.fp = 0
+        self.tn = 0
+        self.fn = 0
+
+    def update(self, pred, gt):
+        """
+        Input:
+            pred, gt: Tensor()
+        """
+        assert pred.shape == gt.shape
+        self.tp += torch.logical_and(pred == 1, gt == 1).sum().item()
+        self.fp += torch.logical_and(pred == 1, gt == 0).sum().item()
+        self.tn += torch.logical_and(pred == 0, gt == 0).sum().item()
+        self.tn += torch.logical_and(pred == 1, gt == 0).sum().item()
+
+    def update_cm(self, tp, fp, tn, fn):
+        self.tp += tp
+        self.fp += fp
+        self.tn += tn
+        self.tn += fn
+
+
+def inverse_sigmoid(x, eps=1e-5):
+    x = x.clamp(min=0, max=1)
+    x1 = x.clamp(min=eps)
+    x2 = (1 - x).clamp(min=eps)
+    return torch.log(x1 / x2)
+
+
+def get_raw_dict(args):
+    """
+    return the dicf contained in args.
+
+    e.g:
+        >>> with open(path, 'w') as f:
+                json.dump(get_raw_dict(args), f, indent=2)
+    """
+    if isinstance(args, argparse.Namespace):
+        return vars(args)
+    elif isinstance(args, dict):
+        return args
+    elif isinstance(args, SLConfig):
+        return args._cfg_dict
+    else:
+        raise NotImplementedError("Unknown type {}".format(type(args)))
+
+
+def stat_tensors(tensor):
+    assert tensor.dim() == 1
+    tensor_sm = tensor.softmax(0)
+    entropy = (tensor_sm * torch.log(tensor_sm + 1e-9)).sum()
+
+    return {
+        "max": tensor.max(),
+        "min": tensor.min(),
+        "mean": tensor.mean(),
+        "var": tensor.var(),
+        "std": tensor.var() ** 0.5,
+        "entropy": entropy,
+    }
+
+
+class NiceRepr:
+    """Inherit from this class and define ``__nice__`` to "nicely" print your
+    objects.
+
+    Defines ``__str__`` and ``__repr__`` in terms of ``__nice__`` function
+    Classes that inherit from :class:`NiceRepr` should redefine ``__nice__``.
+    If the inheriting class has a ``__len__``, method then the default
+    ``__nice__`` method will return its length.
+
+    Example:
+        >>> class Foo(NiceRepr):
+        ...    def __nice__(self):
+        ...        return 'info'
+        >>> foo = Foo()
+        >>> assert str(foo) == '<Foo(info)>'
+        >>> assert repr(foo).startswith('<Foo(info) at ')
+
+    Example:
+        >>> class Bar(NiceRepr):
+        ...    pass
+        >>> bar = Bar()
+        >>> import pytest
+        >>> with pytest.warns(None) as record:
+        >>>     assert 'object at' in str(bar)
+        >>>     assert 'object at' in repr(bar)
+
+    Example:
+        >>> class Baz(NiceRepr):
+        ...    def __len__(self):
+        ...        return 5
+        >>> baz = Baz()
+        >>> assert str(baz) == '<Baz(5)>'
+    """
+
+    def __nice__(self):
+        """str: a "nice" summary string describing this module"""
+        if hasattr(self, "__len__"):
+            # It is a common pattern for objects to use __len__ in __nice__
+            # As a convenience we define a default __nice__ for these objects
+            return str(len(self))
+        else:
+            # In all other cases force the subclass to overload __nice__
+            raise NotImplementedError(f"Define the __nice__ method for {self.__class__!r}")
+
+    def __repr__(self):
+        """str: the string of the module"""
+        try:
+            nice = self.__nice__()
+            classname = self.__class__.__name__
+            return f"<{classname}({nice}) at {hex(id(self))}>"
+        except NotImplementedError as ex:
+            warnings.warn(str(ex), category=RuntimeWarning)
+            return object.__repr__(self)
+
+    def __str__(self):
+        """str: the string of the module"""
+        try:
+            classname = self.__class__.__name__
+            nice = self.__nice__()
+            return f"<{classname}({nice})>"
+        except NotImplementedError as ex:
+            warnings.warn(str(ex), category=RuntimeWarning)
+            return object.__repr__(self)
+
+
+def ensure_rng(rng=None):
+    """Coerces input into a random number generator.
+
+    If the input is None, then a global random state is returned.
+
+    If the input is a numeric value, then that is used as a seed to construct a
+    random state. Otherwise the input is returned as-is.
+
+    Adapted from [1]_.
+
+    Args:
+        rng (int | numpy.random.RandomState | None):
+            if None, then defaults to the global rng. Otherwise this can be an
+            integer or a RandomState class
+    Returns:
+        (numpy.random.RandomState) : rng -
+            a numpy random number generator
+
+    References:
+        .. [1] https://gitlab.kitware.com/computer-vision/kwarray/blob/master/kwarray/util_random.py#L270  # noqa: E501
+    """
+
+    if rng is None:
+        rng = np.random.mtrand._rand
+    elif isinstance(rng, int):
+        rng = np.random.RandomState(rng)
+    else:
+        rng = rng
+    return rng
+
+
+def random_boxes(num=1, scale=1, rng=None):
+    """Simple version of ``kwimage.Boxes.random``
+
+    Returns:
+        Tensor: shape (n, 4) in x1, y1, x2, y2 format.
+
+    References:
+        https://gitlab.kitware.com/computer-vision/kwimage/blob/master/kwimage/structs/boxes.py#L1390
+
+    Example:
+        >>> num = 3
+        >>> scale = 512
+        >>> rng = 0
+        >>> boxes = random_boxes(num, scale, rng)
+        >>> print(boxes)
+        tensor([[280.9925, 278.9802, 308.6148, 366.1769],
+                [216.9113, 330.6978, 224.0446, 456.5878],
+                [405.3632, 196.3221, 493.3953, 270.7942]])
+    """
+    rng = ensure_rng(rng)
+
+    tlbr = rng.rand(num, 4).astype(np.float32)
+
+    tl_x = np.minimum(tlbr[:, 0], tlbr[:, 2])
+    tl_y = np.minimum(tlbr[:, 1], tlbr[:, 3])
+    br_x = np.maximum(tlbr[:, 0], tlbr[:, 2])
+    br_y = np.maximum(tlbr[:, 1], tlbr[:, 3])
+
+    tlbr[:, 0] = tl_x * scale
+    tlbr[:, 1] = tl_y * scale
+    tlbr[:, 2] = br_x * scale
+    tlbr[:, 3] = br_y * scale
+
+    boxes = torch.from_numpy(tlbr)
+    return boxes
+
+
+class ModelEma(torch.nn.Module):
+    def __init__(self, model, decay=0.9997, device=None):
+        super(ModelEma, self).__init__()
+        # make a copy of the model for accumulating moving average of weights
+        self.module = deepcopy(model)
+        self.module.eval()
+
+        # import ipdb; ipdb.set_trace()
+
+        self.decay = decay
+        self.device = device  # perform ema on different device from model if set
+        if self.device is not None:
+            self.module.to(device=device)
+
+    def _update(self, model, update_fn):
+        with torch.no_grad():
+            for ema_v, model_v in zip(self.module.state_dict().values(), model.state_dict().values()):
+                if self.device is not None:
+                    model_v = model_v.to(device=self.device)
+                ema_v.copy_(update_fn(ema_v, model_v))
+
+    def update(self, model):
+        self._update(model, update_fn=lambda e, m: self.decay * e + (1.0 - self.decay) * m)
+
+    def set(self, model):
+        self._update(model, update_fn=lambda e, m: m)
+
+
+class BestMetricSingle:
+    def __init__(self, init_res=0.0, better="large") -> None:
+        self.init_res = init_res
+        self.best_res = init_res
+        self.best_ep = -1
+
+        self.better = better
+        assert better in ["large", "small"]
+
+    def isbetter(self, new_res, old_res):
+        if self.better == "large":
+            return new_res > old_res
+        if self.better == "small":
+            return new_res < old_res
+
+    def update(self, new_res, ep):
+        if self.isbetter(new_res, self.best_res):
+            self.best_res = new_res
+            self.best_ep = ep
+            return True
+        return False
+
+    def __str__(self) -> str:
+        return "best_res: {}\t best_ep: {}".format(self.best_res, self.best_ep)
+
+    def __repr__(self) -> str:
+        return self.__str__()
+
+    def summary(self) -> dict:
+        return {
+            "best_res": self.best_res,
+            "best_ep": self.best_ep,
+        }
+
+
+class BestMetricHolder:
+    def __init__(self, init_res=0.0, better="large", use_ema=False) -> None:
+        self.best_all = BestMetricSingle(init_res, better)
+        self.use_ema = use_ema
+        if use_ema:
+            self.best_ema = BestMetricSingle(init_res, better)
+            self.best_regular = BestMetricSingle(init_res, better)
+
+    def update(self, new_res, epoch, is_ema=False):
+        """
+        return if the results is the best.
+        """
+        if not self.use_ema:
+            return self.best_all.update(new_res, epoch)
+        else:
+            if is_ema:
+                self.best_ema.update(new_res, epoch)
+                return self.best_all.update(new_res, epoch)
+            else:
+                self.best_regular.update(new_res, epoch)
+                return self.best_all.update(new_res, epoch)
+
+    def summary(self):
+        if not self.use_ema:
+            return self.best_all.summary()
+
+        res = {}
+        res.update({f"all_{k}": v for k, v in self.best_all.summary().items()})
+        res.update({f"regular_{k}": v for k, v in self.best_regular.summary().items()})
+        res.update({f"ema_{k}": v for k, v in self.best_ema.summary().items()})
+        return res
+
+    def __repr__(self) -> str:
+        return json.dumps(self.summary(), indent=2)
+
+    def __str__(self) -> str:
+        return self.__repr__()
+
+
+def targets_to(targets: List[Dict[str, Any]], device):
+    """Moves the target dicts to the given device."""
+    excluded_keys = [
+        "questionId",
+        "tokens_positive",
+        "strings_positive",
+        "tokens",
+        "dataset_name",
+        "sentence_id",
+        "original_img_id",
+        "nb_eval",
+        "task_id",
+        "original_id",
+        "token_span",
+        "caption",
+        "dataset_type",
+    ]
+    return [{k: v.to(device) if k not in excluded_keys else v for k, v in t.items()} for t in targets]
+
+
+def get_phrases_from_posmap(posmap: torch.BoolTensor, tokenized: Dict, tokenizer: AutoTokenizer):
+    assert isinstance(posmap, torch.Tensor), "posmap must be torch.Tensor"
+    if posmap.dim() == 1:
+        non_zero_idx = posmap.nonzero(as_tuple=True)[0].tolist()
+        token_ids = [tokenized["input_ids"][i] for i in non_zero_idx]
+        return tokenizer.decode(token_ids)
+    else:
+        raise NotImplementedError("posmap must be 1-dim")

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/visualizer.py

@@ -0,0 +1,309 @@
+# -*- coding: utf-8 -*-
+"""
+@File    :   visualizer.py
+@Time    :   2022/04/05 11:39:33
+@Author  :   Shilong Liu
+@Contact :   slongliu86@gmail.com
+"""
+
+import datetime
+import os
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+from matplotlib import transforms
+from matplotlib.collections import PatchCollection
+from matplotlib.patches import Polygon
+from pycocotools import mask as maskUtils
+
+
+def renorm(img: torch.FloatTensor, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) -> torch.FloatTensor:
+    # img: tensor(3,H,W) or tensor(B,3,H,W)
+    # return: same as img
+    assert img.dim() == 3 or img.dim() == 4, "img.dim() should be 3 or 4 but %d" % img.dim()
+    if img.dim() == 3:
+        assert img.size(0) == 3, 'img.size(0) shoule be 3 but "%d". (%s)' % (
+            img.size(0),
+            str(img.size()),
+        )
+        img_perm = img.permute(1, 2, 0)
+        mean = torch.Tensor(mean)
+        std = torch.Tensor(std)
+        img_res = img_perm * std + mean
+        return img_res.permute(2, 0, 1)
+    else:  # img.dim() == 4
+        assert img.size(1) == 3, 'img.size(1) shoule be 3 but "%d". (%s)' % (
+            img.size(1),
+            str(img.size()),
+        )
+        img_perm = img.permute(0, 2, 3, 1)
+        mean = torch.Tensor(mean)
+        std = torch.Tensor(std)
+        img_res = img_perm * std + mean
+        return img_res.permute(0, 3, 1, 2)
+
+
+class ColorMap:
+    def __init__(self, basergb=[255, 255, 0]):
+        self.basergb = np.array(basergb)
+
+    def __call__(self, attnmap):
+        # attnmap: h, w. np.uint8.
+        # return: h, w, 4. np.uint8.
+        assert attnmap.dtype == np.uint8
+        h, w = attnmap.shape
+        res = self.basergb.copy()
+        res = res[None][None].repeat(h, 0).repeat(w, 1)  # h, w, 3
+        attn1 = attnmap.copy()[..., None]  # h, w, 1
+        res = np.concatenate((res, attn1), axis=-1).astype(np.uint8)
+        return res
+
+
+def rainbow_text(x, y, ls, lc, **kw):
+    """
+    Take a list of strings ``ls`` and colors ``lc`` and place them next to each
+    other, with text ls[i] being shown in color lc[i].
+
+    This example shows how to do both vertical and horizontal text, and will
+    pass all keyword arguments to plt.text, so you can set the font size,
+    family, etc.
+    """
+    t = plt.gca().transData
+    fig = plt.gcf()
+    plt.show()
+
+    # horizontal version
+    for s, c in zip(ls, lc):
+        text = plt.text(x, y, " " + s + " ", color=c, transform=t, **kw)
+        text.draw(fig.canvas.get_renderer())
+        ex = text.get_window_extent()
+        t = transforms.offset_copy(text._transform, x=ex.width, units="dots")
+
+    # #vertical version
+    # for s,c in zip(ls,lc):
+    #     text = plt.text(x,y," "+s+" ",color=c, transform=t,
+    #             rotation=90,va='bottom',ha='center',**kw)
+    #     text.draw(fig.canvas.get_renderer())
+    #     ex = text.get_window_extent()
+    #     t = transforms.offset_copy(text._transform, y=ex.height, units='dots')
+
+
+class COCOVisualizer:
+    def __init__(self, coco=None, tokenlizer=None) -> None:
+        self.coco = coco
+
+    def visualize(self, img, tgt, caption=None, dpi=180, savedir="vis"):
+        """
+        img: tensor(3, H, W)
+        tgt: make sure they are all on cpu.
+            must have items: 'image_id', 'boxes', 'size'
+        """
+        plt.figure(dpi=dpi)
+        plt.rcParams["font.size"] = "5"
+        ax = plt.gca()
+        img = renorm(img).permute(1, 2, 0)
+        # if os.environ.get('IPDB_SHILONG_DEBUG', None) == 'INFO':
+        #     import ipdb; ipdb.set_trace()
+        ax.imshow(img)
+
+        self.addtgt(tgt)
+
+        if tgt is None:
+            image_id = 0
+        elif "image_id" not in tgt:
+            image_id = 0
+        else:
+            image_id = tgt["image_id"]
+
+        if caption is None:
+            savename = "{}/{}-{}.png".format(savedir, int(image_id), str(datetime.datetime.now()).replace(" ", "-"))
+        else:
+            savename = "{}/{}-{}-{}.png".format(
+                savedir, caption, int(image_id), str(datetime.datetime.now()).replace(" ", "-")
+            )
+        print("savename: {}".format(savename))
+        os.makedirs(os.path.dirname(savename), exist_ok=True)
+        plt.savefig(savename)
+        plt.close()
+
+    def addtgt(self, tgt):
+        """ """
+        if tgt is None or not "boxes" in tgt:
+            ax = plt.gca()
+
+            if "caption" in tgt:
+                ax.set_title(tgt["caption"], wrap=True)
+
+            ax.set_axis_off()
+            return
+
+        ax = plt.gca()
+        H, W = tgt["size"]
+        numbox = tgt["boxes"].shape[0]
+
+        color = []
+        polygons = []
+        boxes = []
+        for box in tgt["boxes"].cpu():
+            unnormbbox = box * torch.Tensor([W, H, W, H])
+            unnormbbox[:2] -= unnormbbox[2:] / 2
+            [bbox_x, bbox_y, bbox_w, bbox_h] = unnormbbox.tolist()
+            boxes.append([bbox_x, bbox_y, bbox_w, bbox_h])
+            poly = [
+                [bbox_x, bbox_y],
+                [bbox_x, bbox_y + bbox_h],
+                [bbox_x + bbox_w, bbox_y + bbox_h],
+                [bbox_x + bbox_w, bbox_y],
+            ]
+            np_poly = np.array(poly).reshape((4, 2))
+            polygons.append(Polygon(np_poly))
+            c = (np.random.random((1, 3)) * 0.6 + 0.4).tolist()[0]
+            color.append(c)
+
+        p = PatchCollection(polygons, facecolor=color, linewidths=0, alpha=0.1)
+        ax.add_collection(p)
+        p = PatchCollection(polygons, facecolor="none", edgecolors=color, linewidths=2)
+        ax.add_collection(p)
+
+        if "strings_positive" in tgt and len(tgt["strings_positive"]) > 0:
+            assert len(tgt["strings_positive"]) == numbox, f"{len(tgt['strings_positive'])} = {numbox}, "
+            for idx, strlist in enumerate(tgt["strings_positive"]):
+                cate_id = int(tgt["labels"][idx])
+                _string = str(cate_id) + ":" + " ".join(strlist)
+                bbox_x, bbox_y, bbox_w, bbox_h = boxes[idx]
+                # ax.text(bbox_x, bbox_y, _string, color='black', bbox={'facecolor': 'yellow', 'alpha': 1.0, 'pad': 1})
+                ax.text(
+                    bbox_x,
+                    bbox_y,
+                    _string,
+                    color="black",
+                    bbox={"facecolor": color[idx], "alpha": 0.6, "pad": 1},
+                )
+
+        if "box_label" in tgt:
+            assert len(tgt["box_label"]) == numbox, f"{len(tgt['box_label'])} = {numbox}, "
+            for idx, bl in enumerate(tgt["box_label"]):
+                _string = str(bl)
+                bbox_x, bbox_y, bbox_w, bbox_h = boxes[idx]
+                # ax.text(bbox_x, bbox_y, _string, color='black', bbox={'facecolor': 'yellow', 'alpha': 1.0, 'pad': 1})
+                ax.text(
+                    bbox_x,
+                    bbox_y,
+                    _string,
+                    color="black",
+                    bbox={"facecolor": color[idx], "alpha": 0.6, "pad": 1},
+                )
+
+        if "caption" in tgt:
+            ax.set_title(tgt["caption"], wrap=True)
+            # plt.figure()
+            # rainbow_text(0.0,0.0,"all unicorns poop rainbows ! ! !".split(),
+            #         ['red', 'orange', 'brown', 'green', 'blue', 'purple', 'black'])
+
+        if "attn" in tgt:
+            # if os.environ.get('IPDB_SHILONG_DEBUG', None) == 'INFO':
+            #     import ipdb; ipdb.set_trace()
+            if isinstance(tgt["attn"], tuple):
+                tgt["attn"] = [tgt["attn"]]
+            for item in tgt["attn"]:
+                attn_map, basergb = item
+                attn_map = (attn_map - attn_map.min()) / (attn_map.max() - attn_map.min() + 1e-3)
+                attn_map = (attn_map * 255).astype(np.uint8)
+                cm = ColorMap(basergb)
+                heatmap = cm(attn_map)
+                ax.imshow(heatmap)
+        ax.set_axis_off()
+
+    def showAnns(self, anns, draw_bbox=False):
+        """
+        Display the specified annotations.
+        :param anns (array of object): annotations to display
+        :return: None
+        """
+        if len(anns) == 0:
+            return 0
+        if "segmentation" in anns[0] or "keypoints" in anns[0]:
+            datasetType = "instances"
+        elif "caption" in anns[0]:
+            datasetType = "captions"
+        else:
+            raise Exception("datasetType not supported")
+        if datasetType == "instances":
+            ax = plt.gca()
+            ax.set_autoscale_on(False)
+            polygons = []
+            color = []
+            for ann in anns:
+                c = (np.random.random((1, 3)) * 0.6 + 0.4).tolist()[0]
+                if "segmentation" in ann:
+                    if type(ann["segmentation"]) == list:
+                        # polygon
+                        for seg in ann["segmentation"]:
+                            poly = np.array(seg).reshape((int(len(seg) / 2), 2))
+                            polygons.append(Polygon(poly))
+                            color.append(c)
+                    else:
+                        # mask
+                        t = self.imgs[ann["image_id"]]
+                        if type(ann["segmentation"]["counts"]) == list:
+                            rle = maskUtils.frPyObjects([ann["segmentation"]], t["height"], t["width"])
+                        else:
+                            rle = [ann["segmentation"]]
+                        m = maskUtils.decode(rle)
+                        img = np.ones((m.shape[0], m.shape[1], 3))
+                        if ann["iscrowd"] == 1:
+                            color_mask = np.array([2.0, 166.0, 101.0]) / 255
+                        if ann["iscrowd"] == 0:
+                            color_mask = np.random.random((1, 3)).tolist()[0]
+                        for i in range(3):
+                            img[:, :, i] = color_mask[i]
+                        ax.imshow(np.dstack((img, m * 0.5)))
+                if "keypoints" in ann and type(ann["keypoints"]) == list:
+                    # turn skeleton into zero-based index
+                    sks = np.array(self.loadCats(ann["category_id"])[0]["skeleton"]) - 1
+                    kp = np.array(ann["keypoints"])
+                    x = kp[0::3]
+                    y = kp[1::3]
+                    v = kp[2::3]
+                    for sk in sks:
+                        if np.all(v[sk] > 0):
+                            plt.plot(x[sk], y[sk], linewidth=3, color=c)
+                    plt.plot(
+                        x[v > 0],
+                        y[v > 0],
+                        "o",
+                        markersize=8,
+                        markerfacecolor=c,
+                        markeredgecolor="k",
+                        markeredgewidth=2,
+                    )
+                    plt.plot(
+                        x[v > 1],
+                        y[v > 1],
+                        "o",
+                        markersize=8,
+                        markerfacecolor=c,
+                        markeredgecolor=c,
+                        markeredgewidth=2,
+                    )
+
+                if draw_bbox:
+                    [bbox_x, bbox_y, bbox_w, bbox_h] = ann["bbox"]
+                    poly = [
+                        [bbox_x, bbox_y],
+                        [bbox_x, bbox_y + bbox_h],
+                        [bbox_x + bbox_w, bbox_y + bbox_h],
+                        [bbox_x + bbox_w, bbox_y],
+                    ]
+                    np_poly = np.array(poly).reshape((4, 2))
+                    polygons.append(Polygon(np_poly))
+                    color.append(c)
+
+            # p = PatchCollection(polygons, facecolor=color, linewidths=0, alpha=0.4)
+            # ax.add_collection(p)
+            p = PatchCollection(polygons, facecolor="none", edgecolors=color, linewidths=2)
+            ax.add_collection(p)
+        elif datasetType == "captions":
+            for ann in anns:
+                print(ann["caption"])

invokeai/backend/image_util/grounding_segment_anything/groundingdino/util/vl_utils.py

@@ -0,0 +1,100 @@
+import os
+import random
+from typing import List
+
+import torch
+
+
+def create_positive_map_from_span(tokenized, token_span, max_text_len=256):
+    """construct a map such that positive_map[i,j] = True iff box i is associated to token j
+    Input:
+        - tokenized:
+            - input_ids: Tensor[1, ntokens]
+            - attention_mask: Tensor[1, ntokens]
+        - token_span: list with length num_boxes.
+            - each item: [start_idx, end_idx]
+    """
+    positive_map = torch.zeros((len(token_span), max_text_len), dtype=torch.float)
+    for j, tok_list in enumerate(token_span):
+        for beg, end in tok_list:
+            beg_pos = tokenized.char_to_token(beg)
+            end_pos = tokenized.char_to_token(end - 1)
+            if beg_pos is None:
+                try:
+                    beg_pos = tokenized.char_to_token(beg + 1)
+                    if beg_pos is None:
+                        beg_pos = tokenized.char_to_token(beg + 2)
+                except:
+                    beg_pos = None
+            if end_pos is None:
+                try:
+                    end_pos = tokenized.char_to_token(end - 2)
+                    if end_pos is None:
+                        end_pos = tokenized.char_to_token(end - 3)
+                except:
+                    end_pos = None
+            if beg_pos is None or end_pos is None:
+                continue
+
+            assert beg_pos is not None and end_pos is not None
+            if os.environ.get("SHILONG_DEBUG_ONLY_ONE_POS", None) == "TRUE":
+                positive_map[j, beg_pos] = 1
+                break
+            else:
+                positive_map[j, beg_pos : end_pos + 1].fill_(1)
+
+    return positive_map / (positive_map.sum(-1)[:, None] + 1e-6)
+
+
+def build_captions_and_token_span(cat_list, force_lowercase):
+    """
+    Return:
+        captions: str
+        cat2tokenspan: dict
+            {
+                'dog': [[0, 2]],
+                ...
+            }
+    """
+
+    cat2tokenspan = {}
+    captions = ""
+    for catname in cat_list:
+        class_name = catname
+        if force_lowercase:
+            class_name = class_name.lower()
+        if "/" in class_name:
+            class_name_list: List = class_name.strip().split("/")
+            class_name_list.append(class_name)
+            class_name: str = random.choice(class_name_list)
+
+        tokens_positive_i = []
+        subnamelist = [i.strip() for i in class_name.strip().split(" ")]
+        for subname in subnamelist:
+            if len(subname) == 0:
+                continue
+            if len(captions) > 0:
+                captions = captions + " "
+            strat_idx = len(captions)
+            end_idx = strat_idx + len(subname)
+            tokens_positive_i.append([strat_idx, end_idx])
+            captions = captions + subname
+
+        if len(tokens_positive_i) > 0:
+            captions = captions + " ."
+            cat2tokenspan[class_name] = tokens_positive_i
+
+    return captions, cat2tokenspan
+
+
+def build_id2posspan_and_caption(category_dict: dict):
+    """Build id2pos_span and caption from category_dict
+
+    Args:
+        category_dict (dict): category_dict
+    """
+    cat_list = [item["name"].lower() for item in category_dict]
+    id2catname = {item["id"]: item["name"].lower() for item in category_dict}
+    caption, cat2posspan = build_captions_and_token_span(cat_list, force_lowercase=True)
+    id2posspan = {catid: cat2posspan[catname] for catid, catname in id2catname.items()}
+    return id2posspan, caption

invokeai/backend/image_util/grounding_segment_anything/groundingdino/version.py

@@ -0,0 +1 @@
+__version__ = "0.1.0"

invokeai/backend/image_util/grounding_segment_anything/gsa.py

@@ -0,0 +1,102 @@
+import pathlib
+from typing import Any, Dict, List, Optional
+
+import numpy as np
+import supervision as sv
+import torch
+import torchvision
+from PIL import Image
+
+from invokeai.backend.image_util.grounding_segment_anything.groundingdino.util.inference import Model
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.build_sam import sam_model_registry
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.predictor import SamPredictor
+
+
+class GroundingSegmentAnythingDetector:
+    def __init__(self, grounding_dino_model: Model, segment_anything_model: SamPredictor) -> None:
+        self.grounding_dino_model: Optional[Model] = grounding_dino_model
+        self.segment_anything_model: Optional[SamPredictor] = segment_anything_model
+
+    @staticmethod
+    def build_grounding_dino(grounding_dino_state_dict: Dict[str, torch.Tensor], device: torch.device):
+        grounding_dino_config = pathlib.Path(
+            "./invokeai/backend/image_util/grounding_segment_anything/groundingdino/config/GroundingDINO_SwinT_OGC.py"
+        )
+        return Model(
+            model_state_dict=grounding_dino_state_dict,
+            model_config_path=grounding_dino_config.as_posix(),
+            device=device.type,
+        )
+
+    @staticmethod
+    def build_segment_anything(segment_anything_state_dict: Dict[str, torch.Tensor], device: torch.device):
+        sam = sam_model_registry["vit_h"](checkpoint=segment_anything_state_dict)
+        sam.to(device=device)
+        return SamPredictor(sam)
+
+    def detect_objects(
+        self,
+        image: np.ndarray[Any, Any],
+        prompts: List[str],
+        box_threshold: float = 0.5,
+        text_threshold: float = 0.5,
+        nms_threshold: float = 0.8,
+    ):
+        if not self.grounding_dino_model:
+            raise RuntimeError("GroundingDINO model could not load.")
+
+        detections = self.grounding_dino_model.predict_with_classes(
+            image=image, classes=prompts, box_threshold=box_threshold, text_threshold=text_threshold
+        )
+
+        nms_idx = (
+            torchvision.ops.nms(
+                torch.from_numpy(detections.xyxy), torch.from_numpy(detections.confidence), nms_threshold
+            )
+            .numpy()
+            .tolist()
+        )
+        detections.xyxy = detections.xyxy[nms_idx]
+        detections.confidence = detections.confidence[nms_idx]
+        detections.class_id = detections.class_id[nms_idx]
+
+        return detections
+
+    def segment_detections(
+        self, image: np.ndarray[Any, Any], detections: sv.Detections, prompts: List[str]
+    ) -> Dict[str, np.ndarray[Any, Any]]:
+        if not self.segment_anything_model:
+            raise RuntimeError("Segment Anything model could not be loaded")
+
+        self.segment_anything_model.set_image(image)
+        result_masks = {}
+        for box, class_id in zip(detections.xyxy, detections.class_id):
+            masks, scores, logits = self.segment_anything_model.predict(box=box, multimask_output=True)
+            index = np.argmax(scores)
+            result_masks.update({prompts[class_id]: masks[index]})
+        return result_masks
+
+    def predict(
+        self,
+        image: Image.Image,
+        prompt: str,
+        box_threshold: float = 0.5,
+        text_threshold: float = 0.5,
+        nms_threshold: float = 0.8,
+    ):
+        open_cv_image = np.array(image)
+        open_cv_image = open_cv_image[:, :, ::-1].copy()
+        prompts = prompt.split(",")
+
+        detections = self.detect_objects(open_cv_image, prompts, box_threshold, text_threshold, nms_threshold)
+        segments = self.segment_detections(open_cv_image, detections, prompts)
+
+        if len(segments) > 0:
+            combined_mask = np.zeros_like(list(segments.values())[0])
+            for mask in list(segments.values()):
+                combined_mask = np.logical_or(combined_mask, mask)
+            mask_preview = (combined_mask * 255).astype(np.uint8)
+        else:
+            mask_preview = np.zeros(open_cv_image.shape, np.uint8)
+
+        return Image.fromarray(mask_preview)

invokeai/backend/image_util/grounding_segment_anything/segment_anything/__init__.py

@@ -0,0 +1,25 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.automatic_mask_generator import (
+    SamAutomaticMaskGenerator,
+)  # noqa
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.build_sam import (  # noqa
+    build_sam,
+    build_sam_vit_b,
+    build_sam_vit_h,
+    build_sam_vit_l,
+    sam_model_registry,
+)
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.build_sam_hq import (  # noqa
+    build_sam_hq,
+    build_sam_hq_vit_b,
+    build_sam_hq_vit_h,
+    build_sam_hq_vit_l,
+    sam_hq_model_registry,
+)
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.predictor import SamPredictor  # noqa

invokeai/backend/image_util/grounding_segment_anything/segment_anything/automatic_mask_generator.py

@@ -0,0 +1,368 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Any, Dict, List, Optional, Tuple
+
+import numpy as np
+import torch
+from torchvision.ops.boxes import batched_nms, box_area  # type: ignore
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling import Sam
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.predictor import SamPredictor
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.utils.amg import (
+    MaskData,
+    area_from_rle,
+    batch_iterator,
+    batched_mask_to_box,
+    box_xyxy_to_xywh,
+    build_all_layer_point_grids,
+    calculate_stability_score,
+    coco_encode_rle,
+    generate_crop_boxes,
+    is_box_near_crop_edge,
+    mask_to_rle_pytorch,
+    remove_small_regions,
+    rle_to_mask,
+    uncrop_boxes_xyxy,
+    uncrop_masks,
+    uncrop_points,
+)
+
+
+class SamAutomaticMaskGenerator:
+    def __init__(
+        self,
+        model: Sam,
+        points_per_side: Optional[int] = 32,
+        points_per_batch: int = 64,
+        pred_iou_thresh: float = 0.88,
+        stability_score_thresh: float = 0.95,
+        stability_score_offset: float = 1.0,
+        box_nms_thresh: float = 0.7,
+        crop_n_layers: int = 0,
+        crop_nms_thresh: float = 0.7,
+        crop_overlap_ratio: float = 512 / 1500,
+        crop_n_points_downscale_factor: int = 1,
+        point_grids: Optional[List[np.ndarray]] = None,
+        min_mask_region_area: int = 0,
+        output_mode: str = "binary_mask",
+    ) -> None:
+        """
+        Using a SAM model, generates masks for the entire image.
+        Generates a grid of point prompts over the image, then filters
+        low quality and duplicate masks. The default settings are chosen
+        for SAM with a ViT-H backbone.
+
+        Arguments:
+          model (Sam): The SAM model to use for mask prediction.
+          points_per_side (int or None): The number of points to be sampled
+            along one side of the image. The total number of points is
+            points_per_side**2. If None, 'point_grids' must provide explicit
+            point sampling.
+          points_per_batch (int): Sets the number of points run simultaneously
+            by the model. Higher numbers may be faster but use more GPU memory.
+          pred_iou_thresh (float): A filtering threshold in [0,1], using the
+            model's predicted mask quality.
+          stability_score_thresh (float): A filtering threshold in [0,1], using
+            the stability of the mask under changes to the cutoff used to binarize
+            the model's mask predictions.
+          stability_score_offset (float): The amount to shift the cutoff when
+            calculated the stability score.
+          box_nms_thresh (float): The box IoU cutoff used by non-maximal
+            suppression to filter duplicate masks.
+          crops_n_layers (int): If >0, mask prediction will be run again on
+            crops of the image. Sets the number of layers to run, where each
+            layer has 2**i_layer number of image crops.
+          crops_nms_thresh (float): The box IoU cutoff used by non-maximal
+            suppression to filter duplicate masks between different crops.
+          crop_overlap_ratio (float): Sets the degree to which crops overlap.
+            In the first crop layer, crops will overlap by this fraction of
+            the image length. Later layers with more crops scale down this overlap.
+          crop_n_points_downscale_factor (int): The number of points-per-side
+            sampled in layer n is scaled down by crop_n_points_downscale_factor**n.
+          point_grids (list(np.ndarray) or None): A list over explicit grids
+            of points used for sampling, normalized to [0,1]. The nth grid in the
+            list is used in the nth crop layer. Exclusive with points_per_side.
+          min_mask_region_area (int): If >0, postprocessing will be applied
+            to remove disconnected regions and holes in masks with area smaller
+            than min_mask_region_area. Requires opencv.
+          output_mode (str): The form masks are returned in. Can be 'binary_mask',
+            'uncompressed_rle', or 'coco_rle'. 'coco_rle' requires pycocotools.
+            For large resolutions, 'binary_mask' may consume large amounts of
+            memory.
+        """
+
+        assert (points_per_side is None) != (
+            point_grids is None
+        ), "Exactly one of points_per_side or point_grid must be provided."
+        if points_per_side is not None:
+            self.point_grids = build_all_layer_point_grids(
+                points_per_side,
+                crop_n_layers,
+                crop_n_points_downscale_factor,
+            )
+        elif point_grids is not None:
+            self.point_grids = point_grids
+        else:
+            raise ValueError("Can't have both points_per_side and point_grid be None.")
+
+        assert output_mode in [
+            "binary_mask",
+            "uncompressed_rle",
+            "coco_rle",
+        ], f"Unknown output_mode {output_mode}."
+        if output_mode == "coco_rle":
+            from pycocotools import mask as mask_utils  # type: ignore # noqa: F401
+
+        if min_mask_region_area > 0:
+            import cv2  # type: ignore # noqa: F401
+
+        self.predictor = SamPredictor(model)
+        self.points_per_batch = points_per_batch
+        self.pred_iou_thresh = pred_iou_thresh
+        self.stability_score_thresh = stability_score_thresh
+        self.stability_score_offset = stability_score_offset
+        self.box_nms_thresh = box_nms_thresh
+        self.crop_n_layers = crop_n_layers
+        self.crop_nms_thresh = crop_nms_thresh
+        self.crop_overlap_ratio = crop_overlap_ratio
+        self.crop_n_points_downscale_factor = crop_n_points_downscale_factor
+        self.min_mask_region_area = min_mask_region_area
+        self.output_mode = output_mode
+
+    @torch.no_grad()
+    def generate(self, image: np.ndarray) -> List[Dict[str, Any]]:
+        """
+        Generates masks for the given image.
+
+        Arguments:
+          image (np.ndarray): The image to generate masks for, in HWC uint8 format.
+
+        Returns:
+           list(dict(str, any)): A list over records for masks. Each record is
+             a dict containing the following keys:
+               segmentation (dict(str, any) or np.ndarray): The mask. If
+                 output_mode='binary_mask', is an array of shape HW. Otherwise,
+                 is a dictionary containing the RLE.
+               bbox (list(float)): The box around the mask, in XYWH format.
+               area (int): The area in pixels of the mask.
+               predicted_iou (float): The model's own prediction of the mask's
+                 quality. This is filtered by the pred_iou_thresh parameter.
+               point_coords (list(list(float))): The point coordinates input
+                 to the model to generate this mask.
+               stability_score (float): A measure of the mask's quality. This
+                 is filtered on using the stability_score_thresh parameter.
+               crop_box (list(float)): The crop of the image used to generate
+                 the mask, given in XYWH format.
+        """
+
+        # Generate masks
+        mask_data = self._generate_masks(image)
+
+        # Filter small disconnected regions and holes in masks
+        if self.min_mask_region_area > 0:
+            mask_data = self.postprocess_small_regions(
+                mask_data,
+                self.min_mask_region_area,
+                max(self.box_nms_thresh, self.crop_nms_thresh),
+            )
+
+        # Encode masks
+        if self.output_mode == "coco_rle":
+            mask_data["segmentations"] = [coco_encode_rle(rle) for rle in mask_data["rles"]]
+        elif self.output_mode == "binary_mask":
+            mask_data["segmentations"] = [rle_to_mask(rle) for rle in mask_data["rles"]]
+        else:
+            mask_data["segmentations"] = mask_data["rles"]
+
+        # Write mask records
+        curr_anns = []
+        for idx in range(len(mask_data["segmentations"])):
+            ann = {
+                "segmentation": mask_data["segmentations"][idx],
+                "area": area_from_rle(mask_data["rles"][idx]),
+                "bbox": box_xyxy_to_xywh(mask_data["boxes"][idx]).tolist(),
+                "predicted_iou": mask_data["iou_preds"][idx].item(),
+                "point_coords": [mask_data["points"][idx].tolist()],
+                "stability_score": mask_data["stability_score"][idx].item(),
+                "crop_box": box_xyxy_to_xywh(mask_data["crop_boxes"][idx]).tolist(),
+            }
+            curr_anns.append(ann)
+
+        return curr_anns
+
+    def _generate_masks(self, image: np.ndarray) -> MaskData:
+        orig_size = image.shape[:2]
+        crop_boxes, layer_idxs = generate_crop_boxes(orig_size, self.crop_n_layers, self.crop_overlap_ratio)
+
+        # Iterate over image crops
+        data = MaskData()
+        for crop_box, layer_idx in zip(crop_boxes, layer_idxs):
+            crop_data = self._process_crop(image, crop_box, layer_idx, orig_size)
+            data.cat(crop_data)
+
+        # Remove duplicate masks between crops
+        if len(crop_boxes) > 1:
+            # Prefer masks from smaller crops
+            scores = 1 / box_area(data["crop_boxes"])
+            scores = scores.to(data["boxes"].device)
+            keep_by_nms = batched_nms(
+                data["boxes"].float(),
+                scores,
+                torch.zeros(len(data["boxes"])),  # categories
+                iou_threshold=self.crop_nms_thresh,
+            )
+            data.filter(keep_by_nms)
+
+        data.to_numpy()
+        return data
+
+    def _process_crop(
+        self,
+        image: np.ndarray,
+        crop_box: List[int],
+        crop_layer_idx: int,
+        orig_size: Tuple[int, ...],
+    ) -> MaskData:
+        # Crop the image and calculate embeddings
+        x0, y0, x1, y1 = crop_box
+        cropped_im = image[y0:y1, x0:x1, :]
+        cropped_im_size = cropped_im.shape[:2]
+        self.predictor.set_image(cropped_im)
+
+        # Get points for this crop
+        points_scale = np.array(cropped_im_size)[None, ::-1]
+        points_for_image = self.point_grids[crop_layer_idx] * points_scale
+
+        # Generate masks for this crop in batches
+        data = MaskData()
+        for (points,) in batch_iterator(self.points_per_batch, points_for_image):
+            batch_data = self._process_batch(points, cropped_im_size, crop_box, orig_size)
+            data.cat(batch_data)
+            del batch_data
+        self.predictor.reset_image()
+
+        # Remove duplicates within this crop.
+        keep_by_nms = batched_nms(
+            data["boxes"].float(),
+            data["iou_preds"],
+            torch.zeros(len(data["boxes"])),  # categories
+            iou_threshold=self.box_nms_thresh,
+        )
+        data.filter(keep_by_nms)
+
+        # Return to the original image frame
+        data["boxes"] = uncrop_boxes_xyxy(data["boxes"], crop_box)
+        data["points"] = uncrop_points(data["points"], crop_box)
+        data["crop_boxes"] = torch.tensor([crop_box for _ in range(len(data["rles"]))])
+
+        return data
+
+    def _process_batch(
+        self,
+        points: np.ndarray,
+        im_size: Tuple[int, ...],
+        crop_box: List[int],
+        orig_size: Tuple[int, ...],
+    ) -> MaskData:
+        orig_h, orig_w = orig_size
+
+        # Run model on this batch
+        transformed_points = self.predictor.transform.apply_coords(points, im_size)
+        in_points = torch.as_tensor(transformed_points, device=self.predictor.device)
+        in_labels = torch.ones(in_points.shape[0], dtype=torch.int, device=in_points.device)
+        masks, iou_preds, _ = self.predictor.predict_torch(
+            in_points[:, None, :],
+            in_labels[:, None],
+            multimask_output=True,
+            return_logits=True,
+        )
+
+        # Serialize predictions and store in MaskData
+        data = MaskData(
+            masks=masks.flatten(0, 1),
+            iou_preds=iou_preds.flatten(0, 1),
+            points=torch.as_tensor(points.repeat(masks.shape[1], axis=0)),
+        )
+        del masks
+
+        # Filter by predicted IoU
+        if self.pred_iou_thresh > 0.0:
+            keep_mask = data["iou_preds"] > self.pred_iou_thresh
+            data.filter(keep_mask)
+
+        # Calculate stability score
+        data["stability_score"] = calculate_stability_score(
+            data["masks"], self.predictor.model.mask_threshold, self.stability_score_offset
+        )
+        if self.stability_score_thresh > 0.0:
+            keep_mask = data["stability_score"] >= self.stability_score_thresh
+            data.filter(keep_mask)
+
+        # Threshold masks and calculate boxes
+        data["masks"] = data["masks"] > self.predictor.model.mask_threshold
+        data["boxes"] = batched_mask_to_box(data["masks"])
+
+        # Filter boxes that touch crop boundaries
+        keep_mask = ~is_box_near_crop_edge(data["boxes"], crop_box, [0, 0, orig_w, orig_h])
+        if not torch.all(keep_mask):
+            data.filter(keep_mask)
+
+        # Compress to RLE
+        data["masks"] = uncrop_masks(data["masks"], crop_box, orig_h, orig_w)
+        data["rles"] = mask_to_rle_pytorch(data["masks"])
+        del data["masks"]
+
+        return data
+
+    @staticmethod
+    def postprocess_small_regions(mask_data: MaskData, min_area: int, nms_thresh: float) -> MaskData:
+        """
+        Removes small disconnected regions and holes in masks, then reruns
+        box NMS to remove any new duplicates.
+
+        Edits mask_data in place.
+
+        Requires open-cv as a dependency.
+        """
+        if len(mask_data["rles"]) == 0:
+            return mask_data
+
+        # Filter small disconnected regions and holes
+        new_masks = []
+        scores = []
+        for rle in mask_data["rles"]:
+            mask = rle_to_mask(rle)
+
+            mask, changed = remove_small_regions(mask, min_area, mode="holes")
+            unchanged = not changed
+            mask, changed = remove_small_regions(mask, min_area, mode="islands")
+            unchanged = unchanged and not changed
+
+            new_masks.append(torch.as_tensor(mask).unsqueeze(0))
+            # Give score=0 to changed masks and score=1 to unchanged masks
+            # so NMS will prefer ones that didn't need postprocessing
+            scores.append(float(unchanged))
+
+        # Recalculate boxes and remove any new duplicates
+        masks = torch.cat(new_masks, dim=0)
+        boxes = batched_mask_to_box(masks)
+        keep_by_nms = batched_nms(
+            boxes.float(),
+            torch.as_tensor(scores),
+            torch.zeros(len(boxes)),  # categories
+            iou_threshold=nms_thresh,
+        )
+
+        # Only recalculate RLEs for masks that have changed
+        for i_mask in keep_by_nms:
+            if scores[i_mask] == 0.0:
+                mask_torch = masks[i_mask].unsqueeze(0)
+                mask_data["rles"][i_mask] = mask_to_rle_pytorch(mask_torch)[0]
+                mask_data["boxes"][i_mask] = boxes[i_mask]  # update res directly
+        mask_data.filter(keep_by_nms)
+
+        return mask_data

invokeai/backend/image_util/grounding_segment_anything/segment_anything/build_sam.py

@@ -0,0 +1,111 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from functools import partial
+
+import torch
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling import (
+    ImageEncoderViT,
+    MaskDecoder,
+    PromptEncoder,
+    Sam,
+    TwoWayTransformer,
+)
+
+
+def build_sam_vit_h(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=1280,
+        encoder_depth=32,
+        encoder_num_heads=16,
+        encoder_global_attn_indexes=[7, 15, 23, 31],
+        checkpoint=checkpoint,
+    )
+
+
+build_sam = build_sam_vit_h
+
+
+def build_sam_vit_l(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=1024,
+        encoder_depth=24,
+        encoder_num_heads=16,
+        encoder_global_attn_indexes=[5, 11, 17, 23],
+        checkpoint=checkpoint,
+    )
+
+
+def build_sam_vit_b(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=768,
+        encoder_depth=12,
+        encoder_num_heads=12,
+        encoder_global_attn_indexes=[2, 5, 8, 11],
+        checkpoint=checkpoint,
+    )
+
+
+sam_model_registry = {
+    "default": build_sam,
+    "vit_h": build_sam,
+    "vit_l": build_sam_vit_l,
+    "vit_b": build_sam_vit_b,
+}
+
+
+def _build_sam(
+    encoder_embed_dim,
+    encoder_depth,
+    encoder_num_heads,
+    encoder_global_attn_indexes,
+    checkpoint=None,
+):
+    prompt_embed_dim = 256
+    image_size = 1024
+    vit_patch_size = 16
+    image_embedding_size = image_size // vit_patch_size
+    sam = Sam(
+        image_encoder=ImageEncoderViT(
+            depth=encoder_depth,
+            embed_dim=encoder_embed_dim,
+            img_size=image_size,
+            mlp_ratio=4,
+            norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
+            num_heads=encoder_num_heads,
+            patch_size=vit_patch_size,
+            qkv_bias=True,
+            use_rel_pos=True,
+            global_attn_indexes=encoder_global_attn_indexes,
+            window_size=14,
+            out_chans=prompt_embed_dim,
+        ),
+        prompt_encoder=PromptEncoder(
+            embed_dim=prompt_embed_dim,
+            image_embedding_size=(image_embedding_size, image_embedding_size),
+            input_image_size=(image_size, image_size),
+            mask_in_chans=16,
+        ),
+        mask_decoder=MaskDecoder(
+            num_multimask_outputs=3,
+            transformer=TwoWayTransformer(
+                depth=2,
+                embedding_dim=prompt_embed_dim,
+                mlp_dim=2048,
+                num_heads=8,
+            ),
+            transformer_dim=prompt_embed_dim,
+            iou_head_depth=3,
+            iou_head_hidden_dim=256,
+        ),
+        pixel_mean=[123.675, 116.28, 103.53],
+        pixel_std=[58.395, 57.12, 57.375],
+    )
+    sam.eval()
+    if checkpoint is not None:
+        sam.load_state_dict(checkpoint)
+    return sam

invokeai/backend/image_util/grounding_segment_anything/segment_anything/build_sam_hq.py

@@ -0,0 +1,126 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from functools import partial
+
+import torch
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling import (
+    ImageEncoderViT,
+    MaskDecoderHQ,
+    PromptEncoder,
+    Sam,
+    TwoWayTransformer,
+)
+
+
+def build_sam_hq_vit_h(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=1280,
+        encoder_depth=32,
+        encoder_num_heads=16,
+        encoder_global_attn_indexes=[7, 15, 23, 31],
+        checkpoint=checkpoint,
+    )
+
+
+build_sam_hq = build_sam_hq_vit_h
+
+
+def build_sam_hq_vit_l(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=1024,
+        encoder_depth=24,
+        encoder_num_heads=16,
+        encoder_global_attn_indexes=[5, 11, 17, 23],
+        checkpoint=checkpoint,
+    )
+
+
+def build_sam_hq_vit_b(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=768,
+        encoder_depth=12,
+        encoder_num_heads=12,
+        encoder_global_attn_indexes=[2, 5, 8, 11],
+        checkpoint=checkpoint,
+    )
+
+
+sam_hq_model_registry = {
+    "default": build_sam_hq_vit_h,
+    "vit_h": build_sam_hq_vit_h,
+    "vit_l": build_sam_hq_vit_l,
+    "vit_b": build_sam_hq_vit_b,
+}
+
+
+def _build_sam(
+    encoder_embed_dim,
+    encoder_depth,
+    encoder_num_heads,
+    encoder_global_attn_indexes,
+    checkpoint=None,
+):
+    prompt_embed_dim = 256
+    image_size = 1024
+    vit_patch_size = 16
+    image_embedding_size = image_size // vit_patch_size
+    sam = Sam(
+        image_encoder=ImageEncoderViT(
+            depth=encoder_depth,
+            embed_dim=encoder_embed_dim,
+            img_size=image_size,
+            mlp_ratio=4,
+            norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
+            num_heads=encoder_num_heads,
+            patch_size=vit_patch_size,
+            qkv_bias=True,
+            use_rel_pos=True,
+            global_attn_indexes=encoder_global_attn_indexes,
+            window_size=14,
+            out_chans=prompt_embed_dim,
+        ),
+        prompt_encoder=PromptEncoder(
+            embed_dim=prompt_embed_dim,
+            image_embedding_size=(image_embedding_size, image_embedding_size),
+            input_image_size=(image_size, image_size),
+            mask_in_chans=16,
+        ),
+        mask_decoder=MaskDecoderHQ(
+            num_multimask_outputs=3,
+            transformer=TwoWayTransformer(
+                depth=2,
+                embedding_dim=prompt_embed_dim,
+                mlp_dim=2048,
+                num_heads=8,
+            ),
+            transformer_dim=prompt_embed_dim,
+            iou_head_depth=3,
+            iou_head_hidden_dim=256,
+            vit_dim=encoder_embed_dim,
+        ),
+        pixel_mean=[123.675, 116.28, 103.53],
+        pixel_std=[58.395, 57.12, 57.375],
+    )
+    # sam.eval()
+    if checkpoint is not None:
+        with open(checkpoint, "rb") as f:
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+            state_dict = torch.load(f, map_location=device)
+        info = sam.load_state_dict(state_dict, strict=False)
+        print(info)
+    for n, p in sam.named_parameters():
+        if (
+            "hf_token" not in n
+            and "hf_mlp" not in n
+            and "compress_vit_feat" not in n
+            and "embedding_encoder" not in n
+            and "embedding_maskfeature" not in n
+        ):
+            p.requires_grad = False
+
+    return sam

invokeai/backend/image_util/grounding_segment_anything/segment_anything/modeling/__init__.py

@@ -0,0 +1,20 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.image_encoder import (
+    ImageEncoderViT,
+)
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.mask_decoder import MaskDecoder
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.mask_decoder_hq import (
+    MaskDecoderHQ,
+)
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.prompt_encoder import (
+    PromptEncoder,
+)
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.sam import Sam
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.transformer import (
+    TwoWayTransformer,
+)

invokeai/backend/image_util/grounding_segment_anything/segment_anything/modeling/common.py

@@ -0,0 +1,43 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Type
+
+import torch
+import torch.nn as nn
+
+
+class MLPBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        mlp_dim: int,
+        act: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        super().__init__()
+        self.lin1 = nn.Linear(embedding_dim, mlp_dim)
+        self.lin2 = nn.Linear(mlp_dim, embedding_dim)
+        self.act = act()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.lin2(self.act(self.lin1(x)))
+
+
+# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
+# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119  # noqa
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x

invokeai/backend/image_util/grounding_segment_anything/segment_anything/modeling/image_encoder.py

@@ -0,0 +1,395 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Optional, Tuple, Type
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.common import (
+    LayerNorm2d,
+    MLPBlock,
+)
+
+
+# This class and its supporting functions below lightly adapted from the ViTDet backbone available at: https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/vit.py # noqa
+class ImageEncoderViT(nn.Module):
+    def __init__(
+        self,
+        img_size: int = 1024,
+        patch_size: int = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        out_chans: int = 256,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_abs_pos: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        global_attn_indexes: Tuple[int, ...] = (),
+    ) -> None:
+        """
+        Args:
+            img_size (int): Input image size.
+            patch_size (int): Patch size.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+            depth (int): Depth of ViT.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_abs_pos (bool): If True, use absolute positional embeddings.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks.
+            global_attn_indexes (list): Indexes for blocks using global attention.
+        """
+        super().__init__()
+        self.img_size = img_size
+
+        self.patch_embed = PatchEmbed(
+            kernel_size=(patch_size, patch_size),
+            stride=(patch_size, patch_size),
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+        )
+
+        self.pos_embed: Optional[nn.Parameter] = None
+        if use_abs_pos:
+            # Initialize absolute positional embedding with pretrain image size.
+            self.pos_embed = nn.Parameter(torch.zeros(1, img_size // patch_size, img_size // patch_size, embed_dim))
+
+        self.blocks = nn.ModuleList()
+        for i in range(depth):
+            block = Block(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                use_rel_pos=use_rel_pos,
+                rel_pos_zero_init=rel_pos_zero_init,
+                window_size=window_size if i not in global_attn_indexes else 0,
+                input_size=(img_size // patch_size, img_size // patch_size),
+            )
+            self.blocks.append(block)
+
+        self.neck = nn.Sequential(
+            nn.Conv2d(
+                embed_dim,
+                out_chans,
+                kernel_size=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+            nn.Conv2d(
+                out_chans,
+                out_chans,
+                kernel_size=3,
+                padding=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.patch_embed(x)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+
+        interm_embeddings = []
+        for blk in self.blocks:
+            x = blk(x)
+            if blk.window_size == 0:
+                interm_embeddings.append(x)
+
+        x = self.neck(x.permute(0, 3, 1, 2))
+
+        return x, interm_embeddings
+
+
+class Block(nn.Module):
+    """Transformer blocks with support of window attention and residual propagation blocks"""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks. If it equals 0, then
+                use global attention.
+            input_size (tuple(int, int) or None): Input resolution for calculating the relative
+                positional parameter size.
+        """
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            input_size=input_size if window_size == 0 else (window_size, window_size),
+        )
+
+        self.norm2 = norm_layer(dim)
+        self.mlp = MLPBlock(embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer)
+
+        self.window_size = window_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shortcut = x
+        x = self.norm1(x)
+        # Window partition
+        if self.window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, self.window_size)
+
+        x = self.attn(x)
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, self.window_size, pad_hw, (H, W))
+
+        x = shortcut + x
+        x = x + self.mlp(self.norm2(x))
+
+        return x
+
+
+class Attention(nn.Module):
+    """Multi-head Attention block with relative position embeddings."""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads.
+            qkv_bias (bool):  If True, add a learnable bias to query, key, value.
+            rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            input_size (tuple(int, int) or None): Input resolution for calculating the relative
+                positional parameter size.
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.proj = nn.Linear(dim, dim)
+
+        self.use_rel_pos = use_rel_pos
+        if self.use_rel_pos:
+            assert input_size is not None, "Input size must be provided if using relative positional encoding."
+            # initialize relative positional embeddings
+            self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
+            self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, H, W, _ = x.shape
+        # qkv with shape (3, B, nHead, H * W, C)
+        qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        # q, k, v with shape (B * nHead, H * W, C)
+        q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)
+
+        attn = (q * self.scale) @ k.transpose(-2, -1)
+
+        if self.use_rel_pos:
+            attn = add_decomposed_rel_pos(attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W))
+
+        attn = attn.softmax(dim=-1)
+        x = (attn @ v).view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1)
+        x = self.proj(x)
+
+        return x
+
+
+def window_partition(x: torch.Tensor, window_size: int) -> Tuple[torch.Tensor, Tuple[int, int]]:
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(
+    windows: torch.Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
+) -> torch.Tensor:
+    """
+    Window unpartition into original sequences and removing padding.
+    Args:
+        windows (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
+    """
+    Get relative positional embeddings according to the relative positions of
+        query and key sizes.
+    Args:
+        q_size (int): size of query q.
+        k_size (int): size of key k.
+        rel_pos (Tensor): relative position embeddings (L, C).
+
+    Returns:
+        Extracted positional embeddings according to relative positions.
+    """
+    max_rel_dist = int(2 * max(q_size, k_size) - 1)
+    # Interpolate rel pos if needed.
+    if rel_pos.shape[0] != max_rel_dist:
+        # Interpolate rel pos.
+        rel_pos_resized = F.interpolate(
+            rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
+            size=max_rel_dist,
+            mode="linear",
+        )
+        rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
+    else:
+        rel_pos_resized = rel_pos
+
+    # Scale the coords with short length if shapes for q and k are different.
+    q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
+    k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
+    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
+
+    return rel_pos_resized[relative_coords.long()]
+
+
+def add_decomposed_rel_pos(
+    attn: torch.Tensor,
+    q: torch.Tensor,
+    rel_pos_h: torch.Tensor,
+    rel_pos_w: torch.Tensor,
+    q_size: Tuple[int, int],
+    k_size: Tuple[int, int],
+) -> torch.Tensor:
+    """
+    Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
+    https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py   # noqa B950
+    Args:
+        attn (Tensor): attention map.
+        q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
+        rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
+        rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
+        q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
+        k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
+
+    Returns:
+        attn (Tensor): attention map with added relative positional embeddings.
+    """
+    q_h, q_w = q_size
+    k_h, k_w = k_size
+    Rh = get_rel_pos(q_h, k_h, rel_pos_h)
+    Rw = get_rel_pos(q_w, k_w, rel_pos_w)
+
+    B, _, dim = q.shape
+    r_q = q.reshape(B, q_h, q_w, dim)
+    rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
+    rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
+
+    attn = (attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]).view(
+        B, q_h * q_w, k_h * k_w
+    )
+
+    return attn
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, int] = (16, 16),
+        stride: Tuple[int, int] = (16, 16),
+        padding: Tuple[int, int] = (0, 0),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ) -> None:
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x

invokeai/backend/image_util/grounding_segment_anything/segment_anything/modeling/mask_decoder.py

@@ -0,0 +1,171 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List, Tuple, Type
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.common import LayerNorm2d
+
+
+class MaskDecoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        transformer_dim: int,
+        transformer: nn.Module,
+        num_multimask_outputs: int = 3,
+        activation: Type[nn.Module] = nn.GELU,
+        iou_head_depth: int = 3,
+        iou_head_hidden_dim: int = 256,
+    ) -> None:
+        """
+        Predicts masks given an image and prompt embeddings, using a
+        transformer architecture.
+
+        Arguments:
+          transformer_dim (int): the channel dimension of the transformer
+          transformer (nn.Module): the transformer used to predict masks
+          num_multimask_outputs (int): the number of masks to predict
+            when disambiguating masks
+          activation (nn.Module): the type of activation to use when
+            upscaling masks
+          iou_head_depth (int): the depth of the MLP used to predict
+            mask quality
+          iou_head_hidden_dim (int): the hidden dimension of the MLP
+            used to predict mask quality
+        """
+        super().__init__()
+        self.transformer_dim = transformer_dim
+        self.transformer = transformer
+
+        self.num_multimask_outputs = num_multimask_outputs
+
+        self.iou_token = nn.Embedding(1, transformer_dim)
+        self.num_mask_tokens = num_multimask_outputs + 1
+        self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
+
+        self.output_upscaling = nn.Sequential(
+            nn.ConvTranspose2d(transformer_dim, transformer_dim // 4, kernel_size=2, stride=2),
+            LayerNorm2d(transformer_dim // 4),
+            activation(),
+            nn.ConvTranspose2d(transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2),
+            activation(),
+        )
+        self.output_hypernetworks_mlps = nn.ModuleList(
+            [MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3) for i in range(self.num_mask_tokens)]
+        )
+
+        self.iou_prediction_head = MLP(transformer_dim, iou_head_hidden_dim, self.num_mask_tokens, iou_head_depth)
+
+    def forward(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        multimask_output: bool,
+        hq_token_only: bool,
+        interm_embeddings: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Predict masks given image and prompt embeddings.
+
+        Arguments:
+          image_embeddings (torch.Tensor): the embeddings from the image encoder
+          image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
+          sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
+          dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
+          multimask_output (bool): Whether to return multiple masks or a single
+            mask.
+
+        Returns:
+          torch.Tensor: batched predicted masks
+          torch.Tensor: batched predictions of mask quality
+        """
+        masks, iou_pred = self.predict_masks(
+            image_embeddings=image_embeddings,
+            image_pe=image_pe,
+            sparse_prompt_embeddings=sparse_prompt_embeddings,
+            dense_prompt_embeddings=dense_prompt_embeddings,
+        )
+
+        # Select the correct mask or masks for output
+        if multimask_output:
+            mask_slice = slice(1, None)
+        else:
+            mask_slice = slice(0, 1)
+        masks = masks[:, mask_slice, :, :]
+        iou_pred = iou_pred[:, mask_slice]
+
+        # Prepare output
+        return masks, iou_pred
+
+    def predict_masks(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Predicts masks. See 'forward' for more details."""
+        # Concatenate output tokens
+        output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight], dim=0)
+        output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.size(0), -1, -1)
+        tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
+
+        # Expand per-image data in batch direction to be per-mask
+        src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
+        src = src + dense_prompt_embeddings
+        pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
+        b, c, h, w = src.shape
+
+        # Run the transformer
+        hs, src = self.transformer(src, pos_src, tokens)
+        iou_token_out = hs[:, 0, :]
+        mask_tokens_out = hs[:, 1 : (1 + self.num_mask_tokens), :]
+
+        # Upscale mask embeddings and predict masks using the mask tokens
+        src = src.transpose(1, 2).view(b, c, h, w)
+        upscaled_embedding = self.output_upscaling(src)
+        hyper_in_list: List[torch.Tensor] = []
+        for i in range(self.num_mask_tokens):
+            hyper_in_list.append(self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]))
+        hyper_in = torch.stack(hyper_in_list, dim=1)
+        b, c, h, w = upscaled_embedding.shape
+        masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)
+
+        # Generate mask quality predictions
+        iou_pred = self.iou_prediction_head(iou_token_out)
+
+        return masks, iou_pred
+
+
+# Lightly adapted from
+# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
+class MLP(nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+        num_layers: int,
+        sigmoid_output: bool = False,
+    ) -> None:
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
+        self.sigmoid_output = sigmoid_output
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
+        if self.sigmoid_output:
+            x = F.sigmoid(x)
+        return x

invokeai/backend/image_util/grounding_segment_anything/segment_anything/modeling/mask_decoder_hq.py

@@ -0,0 +1,233 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# Modified by HQ-SAM team
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List, Tuple, Type
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.common import LayerNorm2d
+
+
+class MaskDecoderHQ(nn.Module):
+    def __init__(
+        self,
+        *,
+        transformer_dim: int,
+        transformer: nn.Module,
+        num_multimask_outputs: int = 3,
+        activation: Type[nn.Module] = nn.GELU,
+        iou_head_depth: int = 3,
+        iou_head_hidden_dim: int = 256,
+        vit_dim: int = 1024,
+    ) -> None:
+        """
+        Predicts masks given an image and prompt embeddings, using a
+        transformer architecture.
+
+        Arguments:
+          transformer_dim (int): the channel dimension of the transformer
+          transformer (nn.Module): the transformer used to predict masks
+          num_multimask_outputs (int): the number of masks to predict
+            when disambiguating masks
+          activation (nn.Module): the type of activation to use when
+            upscaling masks
+          iou_head_depth (int): the depth of the MLP used to predict
+            mask quality
+          iou_head_hidden_dim (int): the hidden dimension of the MLP
+            used to predict mask quality
+        """
+        super().__init__()
+        self.transformer_dim = transformer_dim
+        self.transformer = transformer
+
+        self.num_multimask_outputs = num_multimask_outputs
+
+        self.iou_token = nn.Embedding(1, transformer_dim)
+        self.num_mask_tokens = num_multimask_outputs + 1
+        self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
+
+        self.output_upscaling = nn.Sequential(
+            nn.ConvTranspose2d(transformer_dim, transformer_dim // 4, kernel_size=2, stride=2),
+            LayerNorm2d(transformer_dim // 4),
+            activation(),
+            nn.ConvTranspose2d(transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2),
+            activation(),
+        )
+        self.output_hypernetworks_mlps = nn.ModuleList(
+            [MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3) for i in range(self.num_mask_tokens)]
+        )
+
+        self.iou_prediction_head = MLP(transformer_dim, iou_head_hidden_dim, self.num_mask_tokens, iou_head_depth)
+
+        # HQ-SAM parameters
+        self.hf_token = nn.Embedding(1, transformer_dim)  # HQ-Ouptput-Token
+        self.hf_mlp = MLP(
+            transformer_dim, transformer_dim, transformer_dim // 8, 3
+        )  # corresponding new MLP layer for HQ-Ouptput-Token
+        self.num_mask_tokens = self.num_mask_tokens + 1
+
+        # three conv fusion layers for obtaining HQ-Feature
+        self.compress_vit_feat = nn.Sequential(
+            nn.ConvTranspose2d(vit_dim, transformer_dim, kernel_size=2, stride=2),
+            LayerNorm2d(transformer_dim),
+            nn.GELU(),
+            nn.ConvTranspose2d(transformer_dim, transformer_dim // 8, kernel_size=2, stride=2),
+        )
+
+        self.embedding_encoder = nn.Sequential(
+            nn.ConvTranspose2d(transformer_dim, transformer_dim // 4, kernel_size=2, stride=2),
+            LayerNorm2d(transformer_dim // 4),
+            nn.GELU(),
+            nn.ConvTranspose2d(transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2),
+        )
+        self.embedding_maskfeature = nn.Sequential(
+            nn.Conv2d(transformer_dim // 8, transformer_dim // 4, 3, 1, 1),
+            LayerNorm2d(transformer_dim // 4),
+            nn.GELU(),
+            nn.Conv2d(transformer_dim // 4, transformer_dim // 8, 3, 1, 1),
+        )
+
+    def forward(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        multimask_output: bool,
+        hq_token_only: bool,
+        interm_embeddings: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Predict masks given image and prompt embeddings.
+
+        Arguments:
+          image_embeddings (torch.Tensor): the embeddings from the ViT image encoder
+          image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
+          sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
+          dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
+          multimask_output (bool): Whether to return multiple masks or a single
+            mask.
+
+        Returns:
+          torch.Tensor: batched predicted masks
+          torch.Tensor: batched predictions of mask quality
+        """
+        vit_features = interm_embeddings[0].permute(
+            0, 3, 1, 2
+        )  # early-layer ViT feature, after 1st global attention block in ViT
+        hq_features = self.embedding_encoder(image_embeddings) + self.compress_vit_feat(vit_features)
+
+        masks, iou_pred = self.predict_masks(
+            image_embeddings=image_embeddings,
+            image_pe=image_pe,
+            sparse_prompt_embeddings=sparse_prompt_embeddings,
+            dense_prompt_embeddings=dense_prompt_embeddings,
+            hq_features=hq_features,
+        )
+
+        # Select the correct mask or masks for output
+        if multimask_output:
+            # mask with highest score
+            mask_slice = slice(1, self.num_mask_tokens - 1)
+            iou_pred = iou_pred[:, mask_slice]
+            iou_pred, max_iou_idx = torch.max(iou_pred, dim=1)
+            iou_pred = iou_pred.unsqueeze(1)
+            masks_multi = masks[:, mask_slice, :, :]
+            masks_sam = masks_multi[torch.arange(masks_multi.size(0)), max_iou_idx].unsqueeze(1)
+        else:
+            # singale mask output, default
+            mask_slice = slice(0, 1)
+            iou_pred = iou_pred[:, mask_slice]
+            masks_sam = masks[:, mask_slice]
+
+        masks_hq = masks[:, slice(self.num_mask_tokens - 1, self.num_mask_tokens)]
+        if hq_token_only:
+            masks = masks_hq
+        else:
+            masks = masks_sam + masks_hq
+        # Prepare output
+        return masks, iou_pred
+
+    def predict_masks(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        hq_features: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Predicts masks. See 'forward' for more details."""
+        # Concatenate output tokens
+        output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight, self.hf_token.weight], dim=0)
+        output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.size(0), -1, -1)
+        tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
+
+        # Expand per-image data in batch direction to be per-mask
+        src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
+        src = src + dense_prompt_embeddings
+        pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
+        b, c, h, w = src.shape
+
+        # Run the transformer
+        hs, src = self.transformer(src, pos_src, tokens)
+        iou_token_out = hs[:, 0, :]
+        mask_tokens_out = hs[:, 1 : (1 + self.num_mask_tokens), :]
+
+        # Upscale mask embeddings and predict masks using the mask tokens
+        src = src.transpose(1, 2).view(b, c, h, w)
+
+        upscaled_embedding_sam = self.output_upscaling(src)
+        upscaled_embedding_hq = self.embedding_maskfeature(upscaled_embedding_sam) + hq_features.repeat(b, 1, 1, 1)
+
+        hyper_in_list: List[torch.Tensor] = []
+        for i in range(self.num_mask_tokens):
+            if i < self.num_mask_tokens - 1:
+                hyper_in_list.append(self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]))
+            else:
+                hyper_in_list.append(self.hf_mlp(mask_tokens_out[:, i, :]))
+
+        hyper_in = torch.stack(hyper_in_list, dim=1)
+        b, c, h, w = upscaled_embedding_sam.shape
+
+        masks_sam = (hyper_in[:, : self.num_mask_tokens - 1] @ upscaled_embedding_sam.view(b, c, h * w)).view(
+            b, -1, h, w
+        )
+        masks_sam_hq = (hyper_in[:, self.num_mask_tokens - 1 :] @ upscaled_embedding_hq.view(b, c, h * w)).view(
+            b, -1, h, w
+        )
+        masks = torch.cat([masks_sam, masks_sam_hq], dim=1)
+        # Generate mask quality predictions
+        iou_pred = self.iou_prediction_head(iou_token_out)
+
+        return masks, iou_pred
+
+
+# Lightly adapted from
+# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
+class MLP(nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+        num_layers: int,
+        sigmoid_output: bool = False,
+    ) -> None:
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
+        self.sigmoid_output = sigmoid_output
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
+        if self.sigmoid_output:
+            x = F.sigmoid(x)
+        return x

invokeai/backend/image_util/grounding_segment_anything/segment_anything/modeling/prompt_encoder.py

@@ -0,0 +1,212 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Any, Optional, Tuple, Type
+
+import numpy as np
+import torch
+from torch import nn
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.common import LayerNorm2d
+
+
+class PromptEncoder(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        image_embedding_size: Tuple[int, int],
+        input_image_size: Tuple[int, int],
+        mask_in_chans: int,
+        activation: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        """
+        Encodes prompts for input to SAM's mask decoder.
+
+        Arguments:
+          embed_dim (int): The prompts' embedding dimension
+          image_embedding_size (tuple(int, int)): The spatial size of the
+            image embedding, as (H, W).
+          input_image_size (int): The padded size of the image as input
+            to the image encoder, as (H, W).
+          mask_in_chans (int): The number of hidden channels used for
+            encoding input masks.
+          activation (nn.Module): The activation to use when encoding
+            input masks.
+        """
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.input_image_size = input_image_size
+        self.image_embedding_size = image_embedding_size
+        self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
+
+        self.num_point_embeddings: int = 4  # pos/neg point + 2 box corners
+        point_embeddings = [nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)]
+        self.point_embeddings = nn.ModuleList(point_embeddings)
+        self.not_a_point_embed = nn.Embedding(1, embed_dim)
+
+        self.mask_input_size = (4 * image_embedding_size[0], 4 * image_embedding_size[1])
+        self.mask_downscaling = nn.Sequential(
+            nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans // 4),
+            activation(),
+            nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans),
+            activation(),
+            nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
+        )
+        self.no_mask_embed = nn.Embedding(1, embed_dim)
+
+    def get_dense_pe(self) -> torch.Tensor:
+        """
+        Returns the positional encoding used to encode point prompts,
+        applied to a dense set of points the shape of the image encoding.
+
+        Returns:
+          torch.Tensor: Positional encoding with shape
+            1x(embed_dim)x(embedding_h)x(embedding_w)
+        """
+        return self.pe_layer(self.image_embedding_size).unsqueeze(0)
+
+    def _embed_points(
+        self,
+        points: torch.Tensor,
+        labels: torch.Tensor,
+        pad: bool,
+    ) -> torch.Tensor:
+        """Embeds point prompts."""
+        points = points + 0.5  # Shift to center of pixel
+        if pad:
+            padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
+            padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
+            points = torch.cat([points, padding_point], dim=1)
+            labels = torch.cat([labels, padding_label], dim=1)
+        point_embedding = self.pe_layer.forward_with_coords(points, self.input_image_size)
+        point_embedding[labels == -1] = 0.0
+        point_embedding[labels == -1] += self.not_a_point_embed.weight
+        point_embedding[labels == 0] += self.point_embeddings[0].weight
+        point_embedding[labels == 1] += self.point_embeddings[1].weight
+        return point_embedding
+
+    def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
+        """Embeds box prompts."""
+        boxes = boxes + 0.5  # Shift to center of pixel
+        coords = boxes.reshape(-1, 2, 2)
+        corner_embedding = self.pe_layer.forward_with_coords(coords, self.input_image_size)
+        corner_embedding[:, 0, :] += self.point_embeddings[2].weight
+        corner_embedding[:, 1, :] += self.point_embeddings[3].weight
+        return corner_embedding
+
+    def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
+        """Embeds mask inputs."""
+        mask_embedding = self.mask_downscaling(masks)
+        return mask_embedding
+
+    def _get_batch_size(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> int:
+        """
+        Gets the batch size of the output given the batch size of the input prompts.
+        """
+        if points is not None:
+            return points[0].shape[0]
+        elif boxes is not None:
+            return boxes.shape[0]
+        elif masks is not None:
+            return masks.shape[0]
+        else:
+            return 1
+
+    def _get_device(self) -> torch.device:
+        return self.point_embeddings[0].weight.device
+
+    def forward(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Embeds different types of prompts, returning both sparse and dense
+        embeddings.
+
+        Arguments:
+          points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
+            and labels to embed.
+          boxes (torch.Tensor or none): boxes to embed
+          masks (torch.Tensor or none): masks to embed
+
+        Returns:
+          torch.Tensor: sparse embeddings for the points and boxes, with shape
+            BxNx(embed_dim), where N is determined by the number of input points
+            and boxes.
+          torch.Tensor: dense embeddings for the masks, in the shape
+            Bx(embed_dim)x(embed_H)x(embed_W)
+        """
+        bs = self._get_batch_size(points, boxes, masks)
+        sparse_embeddings = torch.empty((bs, 0, self.embed_dim), device=self._get_device())
+        if points is not None:
+            coords, labels = points
+            point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
+            sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
+        if boxes is not None:
+            box_embeddings = self._embed_boxes(boxes)
+            sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
+
+        if masks is not None:
+            dense_embeddings = self._embed_masks(masks)
+        else:
+            dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
+                bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
+            )
+
+        return sparse_embeddings, dense_embeddings
+
+
+class PositionEmbeddingRandom(nn.Module):
+    """
+    Positional encoding using random spatial frequencies.
+    """
+
+    def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
+        super().__init__()
+        if scale is None or scale <= 0.0:
+            scale = 1.0
+        self.register_buffer(
+            "positional_encoding_gaussian_matrix",
+            scale * torch.randn((2, num_pos_feats)),
+        )
+
+    def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
+        """Positionally encode points that are normalized to [0,1]."""
+        # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
+        coords = 2 * coords - 1
+        coords = coords @ self.positional_encoding_gaussian_matrix
+        coords = 2 * np.pi * coords
+        # outputs d_1 x ... x d_n x C shape
+        return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
+
+    def forward(self, size: Tuple[int, int]) -> torch.Tensor:
+        """Generate positional encoding for a grid of the specified size."""
+        h, w = size
+        device: Any = self.positional_encoding_gaussian_matrix.device
+        grid = torch.ones((h, w), device=device, dtype=torch.float32)
+        y_embed = grid.cumsum(dim=0) - 0.5
+        x_embed = grid.cumsum(dim=1) - 0.5
+        y_embed = y_embed / h
+        x_embed = x_embed / w
+
+        pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
+        return pe.permute(2, 0, 1)  # C x H x W
+
+    def forward_with_coords(self, coords_input: torch.Tensor, image_size: Tuple[int, int]) -> torch.Tensor:
+        """Positionally encode points that are not normalized to [0,1]."""
+        coords = coords_input.clone()
+        coords[:, :, 0] = coords[:, :, 0] / image_size[1]
+        coords[:, :, 1] = coords[:, :, 1] / image_size[0]
+        return self._pe_encoding(coords.to(torch.float))  # B x N x C

invokeai/backend/image_util/grounding_segment_anything/segment_anything/modeling/sam.py

@@ -0,0 +1,178 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Any, Dict, List, Tuple
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.image_encoder import (
+    ImageEncoderViT,
+)
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.mask_decoder import MaskDecoder
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.prompt_encoder import (
+    PromptEncoder,
+)
+
+
+class Sam(nn.Module):
+    mask_threshold: float = 0.0
+    image_format: str = "RGB"
+
+    def __init__(
+        self,
+        image_encoder: ImageEncoderViT,
+        prompt_encoder: PromptEncoder,
+        mask_decoder: MaskDecoder,
+        pixel_mean: List[float] = [123.675, 116.28, 103.53],
+        pixel_std: List[float] = [58.395, 57.12, 57.375],
+    ) -> None:
+        """
+        SAM predicts object masks from an image and input prompts.
+
+        Arguments:
+          image_encoder (ImageEncoderViT): The backbone used to encode the
+            image into image embeddings that allow for efficient mask prediction.
+          prompt_encoder (PromptEncoder): Encodes various types of input prompts.
+          mask_decoder (MaskDecoder): Predicts masks from the image embeddings
+            and encoded prompts.
+          pixel_mean (list(float)): Mean values for normalizing pixels in the input image.
+          pixel_std (list(float)): Std values for normalizing pixels in the input image.
+        """
+        super().__init__()
+        self.image_encoder = image_encoder
+        self.prompt_encoder = prompt_encoder
+        self.mask_decoder = mask_decoder
+        self.register_buffer("pixel_mean", torch.Tensor(pixel_mean).view(-1, 1, 1), False)
+        self.register_buffer("pixel_std", torch.Tensor(pixel_std).view(-1, 1, 1), False)
+
+    @property
+    def device(self) -> Any:
+        return self.pixel_mean.device
+
+    @torch.no_grad()
+    def forward(
+        self,
+        batched_input: List[Dict[str, Any]],
+        multimask_output: bool,
+    ) -> List[Dict[str, torch.Tensor]]:
+        """
+        Predicts masks end-to-end from provided images and prompts.
+        If prompts are not known in advance, using SamPredictor is
+        recommended over calling the model directly.
+
+        Arguments:
+          batched_input (list(dict)): A list over input images, each a
+            dictionary with the following keys. A prompt key can be
+            excluded if it is not present.
+              'image': The image as a torch tensor in 3xHxW format,
+                already transformed for input to the model.
+              'original_size': (tuple(int, int)) The original size of
+                the image before transformation, as (H, W).
+              'point_coords': (torch.Tensor) Batched point prompts for
+                this image, with shape BxNx2. Already transformed to the
+                input frame of the model.
+              'point_labels': (torch.Tensor) Batched labels for point prompts,
+                with shape BxN.
+              'boxes': (torch.Tensor) Batched box inputs, with shape Bx4.
+                Already transformed to the input frame of the model.
+              'mask_inputs': (torch.Tensor) Batched mask inputs to the model,
+                in the form Bx1xHxW.
+          multimask_output (bool): Whether the model should predict multiple
+            disambiguating masks, or return a single mask.
+
+        Returns:
+          (list(dict)): A list over input images, where each element is
+            as dictionary with the following keys.
+              'masks': (torch.Tensor) Batched binary mask predictions,
+                with shape BxCxHxW, where B is the number of input promts,
+                C is determiend by multimask_output, and (H, W) is the
+                original size of the image.
+              'iou_predictions': (torch.Tensor) The model's predictions
+                of mask quality, in shape BxC.
+              'low_res_logits': (torch.Tensor) Low resolution logits with
+                shape BxCxHxW, where H=W=256. Can be passed as mask input
+                to subsequent iterations of prediction.
+        """
+        input_images = torch.stack([self.preprocess(x["image"]) for x in batched_input], dim=0)
+        image_embeddings = self.image_encoder(input_images)
+
+        outputs = []
+        for image_record, curr_embedding in zip(batched_input, image_embeddings, strict=False):
+            if "point_coords" in image_record:
+                points = (image_record["point_coords"], image_record["point_labels"])
+            else:
+                points = None
+            sparse_embeddings, dense_embeddings = self.prompt_encoder(
+                points=points,
+                boxes=image_record.get("boxes", None),
+                masks=image_record.get("mask_inputs", None),
+            )
+            low_res_masks, iou_predictions = self.mask_decoder(
+                image_embeddings=curr_embedding.unsqueeze(0),
+                image_pe=self.prompt_encoder.get_dense_pe(),
+                sparse_prompt_embeddings=sparse_embeddings,
+                dense_prompt_embeddings=dense_embeddings,
+                multimask_output=multimask_output,
+            )
+            masks = self.postprocess_masks(
+                low_res_masks,
+                input_size=image_record["image"].shape[-2:],
+                original_size=image_record["original_size"],
+            )
+            masks = masks > self.mask_threshold
+            outputs.append(
+                {
+                    "masks": masks,
+                    "iou_predictions": iou_predictions,
+                    "low_res_logits": low_res_masks,
+                }
+            )
+        return outputs
+
+    def postprocess_masks(
+        self,
+        masks: torch.Tensor,
+        input_size: Tuple[int, ...],
+        original_size: Tuple[int, ...],
+    ) -> torch.Tensor:
+        """
+        Remove padding and upscale masks to the original image size.
+
+        Arguments:
+          masks (torch.Tensor): Batched masks from the mask_decoder,
+            in BxCxHxW format.
+          input_size (tuple(int, int)): The size of the image input to the
+            model, in (H, W) format. Used to remove padding.
+          original_size (tuple(int, int)): The original size of the image
+            before resizing for input to the model, in (H, W) format.
+
+        Returns:
+          (torch.Tensor): Batched masks in BxCxHxW format, where (H, W)
+            is given by original_size.
+        """
+        masks = F.interpolate(
+            masks,
+            (self.image_encoder.img_size, self.image_encoder.img_size),
+            mode="bilinear",
+            align_corners=False,
+        )
+        masks = masks[..., : input_size[0], : input_size[1]]
+        masks = F.interpolate(masks, original_size, mode="bilinear", align_corners=False)
+        return masks
+
+    def preprocess(self, x: torch.Tensor) -> torch.Tensor:
+        """Normalize pixel values and pad to a square input."""
+        # Normalize colors
+        x = (x - self.pixel_mean) / self.pixel_std
+
+        # Pad
+        h, w = x.shape[-2:]
+        padh = self.image_encoder.img_size - h
+        padw = self.image_encoder.img_size - w
+        x = F.pad(x, (0, padw, 0, padh))
+        return x

invokeai/backend/image_util/grounding_segment_anything/segment_anything/modeling/transformer.py

@@ -0,0 +1,232 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from typing import Tuple, Type
+
+import torch
+from torch import Tensor, nn
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling.common import MLPBlock
+
+
+class TwoWayTransformer(nn.Module):
+    def __init__(
+        self,
+        depth: int,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+    ) -> None:
+        """
+        A transformer decoder that attends to an input image using
+        queries whose positional embedding is supplied.
+
+        Args:
+          depth (int): number of layers in the transformer
+          embedding_dim (int): the channel dimension for the input embeddings
+          num_heads (int): the number of heads for multihead attention. Must
+            divide embedding_dim
+          mlp_dim (int): the channel dimension internal to the MLP block
+          activation (nn.Module): the activation to use in the MLP block
+        """
+        super().__init__()
+        self.depth = depth
+        self.embedding_dim = embedding_dim
+        self.num_heads = num_heads
+        self.mlp_dim = mlp_dim
+        self.layers = nn.ModuleList()
+
+        for i in range(depth):
+            self.layers.append(
+                TwoWayAttentionBlock(
+                    embedding_dim=embedding_dim,
+                    num_heads=num_heads,
+                    mlp_dim=mlp_dim,
+                    activation=activation,
+                    attention_downsample_rate=attention_downsample_rate,
+                    skip_first_layer_pe=(i == 0),
+                )
+            )
+
+        self.final_attn_token_to_image = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)
+        self.norm_final_attn = nn.LayerNorm(embedding_dim)
+
+    def forward(
+        self,
+        image_embedding: Tensor,
+        image_pe: Tensor,
+        point_embedding: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+          image_embedding (torch.Tensor): image to attend to. Should be shape
+            B x embedding_dim x h x w for any h and w.
+          image_pe (torch.Tensor): the positional encoding to add to the image. Must
+            have the same shape as image_embedding.
+          point_embedding (torch.Tensor): the embedding to add to the query points.
+            Must have shape B x N_points x embedding_dim for any N_points.
+
+        Returns:
+          torch.Tensor: the processed point_embedding
+          torch.Tensor: the processed image_embedding
+        """
+        # BxCxHxW -> BxHWxC == B x N_image_tokens x C
+        bs, c, h, w = image_embedding.shape
+        image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
+        image_pe = image_pe.flatten(2).permute(0, 2, 1)
+
+        # Prepare queries
+        queries = point_embedding
+        keys = image_embedding
+
+        # Apply transformer blocks and final layernorm
+        for layer in self.layers:
+            queries, keys = layer(
+                queries=queries,
+                keys=keys,
+                query_pe=point_embedding,
+                key_pe=image_pe,
+            )
+
+        # Apply the final attenion layer from the points to the image
+        q = queries + point_embedding
+        k = keys + image_pe
+        attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm_final_attn(queries)
+
+        return queries, keys
+
+
+class TwoWayAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int = 2048,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+        skip_first_layer_pe: bool = False,
+    ) -> None:
+        """
+        A transformer block with four layers: (1) self-attention of sparse
+        inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
+        block on sparse inputs, and (4) cross attention of dense inputs to sparse
+        inputs.
+
+        Arguments:
+          embedding_dim (int): the channel dimension of the embeddings
+          num_heads (int): the number of heads in the attention layers
+          mlp_dim (int): the hidden dimension of the mlp block
+          activation (nn.Module): the activation of the mlp block
+          skip_first_layer_pe (bool): skip the PE on the first layer
+        """
+        super().__init__()
+        self.self_attn = Attention(embedding_dim, num_heads)
+        self.norm1 = nn.LayerNorm(embedding_dim)
+
+        self.cross_attn_token_to_image = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)
+        self.norm2 = nn.LayerNorm(embedding_dim)
+
+        self.mlp = MLPBlock(embedding_dim, mlp_dim, activation)
+        self.norm3 = nn.LayerNorm(embedding_dim)
+
+        self.norm4 = nn.LayerNorm(embedding_dim)
+        self.cross_attn_image_to_token = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)
+
+        self.skip_first_layer_pe = skip_first_layer_pe
+
+    def forward(self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor) -> Tuple[Tensor, Tensor]:
+        # Self attention block
+        if self.skip_first_layer_pe:
+            queries = self.self_attn(q=queries, k=queries, v=queries)
+        else:
+            q = queries + query_pe
+            attn_out = self.self_attn(q=q, k=q, v=queries)
+            queries = queries + attn_out
+        queries = self.norm1(queries)
+
+        # Cross attention block, tokens attending to image embedding
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm2(queries)
+
+        # MLP block
+        mlp_out = self.mlp(queries)
+        queries = queries + mlp_out
+        queries = self.norm3(queries)
+
+        # Cross attention block, image embedding attending to tokens
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
+        keys = keys + attn_out
+        keys = self.norm4(keys)
+
+        return queries, keys
+
+
+class Attention(nn.Module):
+    """
+    An attention layer that allows for downscaling the size of the embedding
+    after projection to queries, keys, and values.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        downsample_rate: int = 1,
+    ) -> None:
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.internal_dim = embedding_dim // downsample_rate
+        self.num_heads = num_heads
+        assert self.internal_dim % num_heads == 0, "num_heads must divide embedding_dim."
+
+        self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.k_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.v_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
+
+    def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
+        b, n, c = x.shape
+        x = x.reshape(b, n, num_heads, c // num_heads)
+        return x.transpose(1, 2)  # B x N_heads x N_tokens x C_per_head
+
+    def _recombine_heads(self, x: Tensor) -> Tensor:
+        b, n_heads, n_tokens, c_per_head = x.shape
+        x = x.transpose(1, 2)
+        return x.reshape(b, n_tokens, n_heads * c_per_head)  # B x N_tokens x C
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
+        # Input projections
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+
+        # Separate into heads
+        q = self._separate_heads(q, self.num_heads)
+        k = self._separate_heads(k, self.num_heads)
+        v = self._separate_heads(v, self.num_heads)
+
+        # Attention
+        _, _, _, c_per_head = q.shape
+        attn = q @ k.permute(0, 1, 3, 2)  # B x N_heads x N_tokens x N_tokens
+        attn = attn / math.sqrt(c_per_head)
+        attn = torch.softmax(attn, dim=-1)
+
+        # Get output
+        out = attn @ v
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out

invokeai/backend/image_util/grounding_segment_anything/segment_anything/predictor.py

@@ -0,0 +1,271 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.modeling import Sam
+from invokeai.backend.image_util.grounding_segment_anything.segment_anything.utils.transforms import ResizeLongestSide
+
+
+class SamPredictor:
+    def __init__(
+        self,
+        sam_model: Sam,
+    ) -> None:
+        """
+        Uses SAM to calculate the image embedding for an image, and then
+        allow repeated, efficient mask prediction given prompts.
+
+        Arguments:
+          sam_model (Sam): The model to use for mask prediction.
+        """
+        super().__init__()
+        self.model = sam_model
+        self.transform = ResizeLongestSide(sam_model.image_encoder.img_size)
+        self.reset_image()
+
+    def set_image(
+        self,
+        image: np.ndarray,
+        image_format: str = "RGB",
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method.
+
+        Arguments:
+          image (np.ndarray): The image for calculating masks. Expects an
+            image in HWC uint8 format, with pixel values in [0, 255].
+          image_format (str): The color format of the image, in ['RGB', 'BGR'].
+        """
+        assert image_format in [
+            "RGB",
+            "BGR",
+        ], f"image_format must be in ['RGB', 'BGR'], is {image_format}."
+        # import pdb;pdb.set_trace()
+        if image_format != self.model.image_format:
+            image = image[..., ::-1]
+
+        # Transform the image to the form expected by the model
+        # import pdb;pdb.set_trace()
+        input_image = self.transform.apply_image(image)
+        input_image_torch = torch.as_tensor(input_image, device=self.device)
+        input_image_torch = input_image_torch.permute(2, 0, 1).contiguous()[None, :, :, :]
+
+        self.set_torch_image(input_image_torch, image.shape[:2])
+
+    @torch.no_grad()
+    def set_torch_image(
+        self,
+        transformed_image: torch.Tensor,
+        original_image_size: Tuple[int, ...],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method. Expects the input
+        image to be already transformed to the format expected by the model.
+
+        Arguments:
+          transformed_image (torch.Tensor): The input image, with shape
+            1x3xHxW, which has been transformed with ResizeLongestSide.
+          original_image_size (tuple(int, int)): The size of the image
+            before transformation, in (H, W) format.
+        """
+        assert (
+            len(transformed_image.shape) == 4
+            and transformed_image.shape[1] == 3
+            and max(*transformed_image.shape[2:]) == self.model.image_encoder.img_size
+        ), f"set_torch_image input must be BCHW with long side {self.model.image_encoder.img_size}."
+        self.reset_image()
+
+        self.original_size = original_image_size
+        self.input_size = tuple(transformed_image.shape[-2:])
+        input_image = self.model.preprocess(transformed_image)
+        self.features, self.interm_features = self.model.image_encoder(input_image)
+        self.is_image_set = True
+
+    def predict(
+        self,
+        point_coords: Optional[np.ndarray] = None,
+        point_labels: Optional[np.ndarray] = None,
+        box: Optional[np.ndarray] = None,
+        mask_input: Optional[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        hq_token_only: bool = False,
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+
+        Arguments:
+          point_coords (np.ndarray or None): A Nx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (np.ndarray or None): A length N array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          box (np.ndarray or None): A length 4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form 1xHxW, where
+            for SAM, H=W=256.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (np.ndarray): The output masks in CxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (np.ndarray): An array of length C containing the model's
+            predictions for the quality of each mask.
+          (np.ndarray): An array of shape CxHxW, where C is the number
+            of masks and H=W=256. These low resolution logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self.is_image_set:
+            raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
+
+        # Transform input prompts
+        coords_torch, labels_torch, box_torch, mask_input_torch = None, None, None, None
+        if point_coords is not None:
+            assert point_labels is not None, "point_labels must be supplied if point_coords is supplied."
+            point_coords = self.transform.apply_coords(point_coords, self.original_size)
+            coords_torch = torch.as_tensor(point_coords, dtype=torch.float, device=self.device)
+            labels_torch = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
+            coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]
+        if box is not None:
+            box = self.transform.apply_boxes(box, self.original_size)
+            box_torch = torch.as_tensor(box, dtype=torch.float, device=self.device)
+            box_torch = box_torch[None, :]
+        if mask_input is not None:
+            mask_input_torch = torch.as_tensor(mask_input, dtype=torch.float, device=self.device)
+            mask_input_torch = mask_input_torch[None, :, :, :]
+
+        masks, iou_predictions, low_res_masks = self.predict_torch(
+            coords_torch,
+            labels_torch,
+            box_torch,
+            mask_input_torch,
+            multimask_output,
+            return_logits=return_logits,
+            hq_token_only=hq_token_only,
+        )
+
+        masks_np = masks[0].detach().cpu().numpy()
+        iou_predictions_np = iou_predictions[0].detach().cpu().numpy()
+        low_res_masks_np = low_res_masks[0].detach().cpu().numpy()
+        return masks_np, iou_predictions_np, low_res_masks_np
+
+    @torch.no_grad()
+    def predict_torch(
+        self,
+        point_coords: Optional[torch.Tensor],
+        point_labels: Optional[torch.Tensor],
+        boxes: Optional[torch.Tensor] = None,
+        mask_input: Optional[torch.Tensor] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        hq_token_only: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+        Input prompts are batched torch tensors and are expected to already be
+        transformed to the input frame using ResizeLongestSide.
+
+        Arguments:
+          point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (torch.Tensor or None): A BxN array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          boxes (np.ndarray or None): A Bx4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form Bx1xHxW, where
+            for SAM, H=W=256. Masks returned by a previous iteration of the
+            predict method do not need further transformation.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (torch.Tensor): The output masks in BxCxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (torch.Tensor): An array of shape BxC containing the model's
+            predictions for the quality of each mask.
+          (torch.Tensor): An array of shape BxCxHxW, where C is the number
+            of masks and H=W=256. These low res logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self.is_image_set:
+            raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
+
+        if point_coords is not None:
+            points = (point_coords, point_labels)
+        else:
+            points = None
+
+        # Embed prompts
+        sparse_embeddings, dense_embeddings = self.model.prompt_encoder(
+            points=points,
+            boxes=boxes,
+            masks=mask_input,
+        )
+
+        # Predict masks
+        low_res_masks, iou_predictions = self.model.mask_decoder(
+            image_embeddings=self.features,
+            image_pe=self.model.prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            hq_token_only=hq_token_only,
+            interm_embeddings=self.interm_features,
+        )
+
+        # Upscale the masks to the original image resolution
+        masks = self.model.postprocess_masks(low_res_masks, self.input_size, self.original_size)
+
+        if not return_logits:
+            masks = masks > self.model.mask_threshold
+
+        return masks, iou_predictions, low_res_masks
+
+    def get_image_embedding(self) -> torch.Tensor:
+        """
+        Returns the image embeddings for the currently set image, with
+        shape 1xCxHxW, where C is the embedding dimension and (H,W) are
+        the embedding spatial dimension of SAM (typically C=256, H=W=64).
+        """
+        if not self.is_image_set:
+            raise RuntimeError("An image must be set with .set_image(...) to generate an embedding.")
+        assert self.features is not None, "Features must exist if an image has been set."
+        return self.features
+
+    @property
+    def device(self) -> torch.device:
+        return self.model.device
+
+    def reset_image(self) -> None:
+        """Resets the currently set image."""
+        self.is_image_set = False
+        self.features = None
+        self.orig_h = None
+        self.orig_w = None
+        self.input_h = None
+        self.input_w = None

invokeai/backend/image_util/grounding_segment_anything/segment_anything/utils/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

invokeai/backend/image_util/grounding_segment_anything/segment_anything/utils/amg.py

@@ -0,0 +1,330 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from copy import deepcopy
+from itertools import product
+from typing import Any, Dict, Generator, ItemsView, List, Tuple
+
+import numpy as np
+import torch
+
+
+class MaskData:
+    """
+    A structure for storing masks and their related data in batched format.
+    Implements basic filtering and concatenation.
+    """
+
+    def __init__(self, **kwargs) -> None:
+        for v in kwargs.values():
+            assert isinstance(
+                v, (list, np.ndarray, torch.Tensor)
+            ), "MaskData only supports list, numpy arrays, and torch tensors."
+        self._stats = dict(**kwargs)
+
+    def __setitem__(self, key: str, item: Any) -> None:
+        assert isinstance(
+            item, (list, np.ndarray, torch.Tensor)
+        ), "MaskData only supports list, numpy arrays, and torch tensors."
+        self._stats[key] = item
+
+    def __delitem__(self, key: str) -> None:
+        del self._stats[key]
+
+    def __getitem__(self, key: str) -> Any:
+        return self._stats[key]
+
+    def items(self) -> ItemsView[str, Any]:
+        return self._stats.items()
+
+    def filter(self, keep: torch.Tensor) -> None:
+        for k, v in self._stats.items():
+            if v is None:
+                self._stats[k] = None
+            elif isinstance(v, torch.Tensor):
+                self._stats[k] = v[torch.as_tensor(keep, device=v.device)]
+            elif isinstance(v, np.ndarray):
+                self._stats[k] = v[keep.detach().cpu().numpy()]
+            elif isinstance(v, list) and keep.dtype == torch.bool:
+                self._stats[k] = [a for i, a in enumerate(v) if keep[i]]
+            elif isinstance(v, list):
+                self._stats[k] = [v[i] for i in keep]
+            else:
+                raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
+
+    def cat(self, new_stats: "MaskData") -> None:
+        for k, v in new_stats.items():
+            if k not in self._stats or self._stats[k] is None:
+                self._stats[k] = deepcopy(v)
+            elif isinstance(v, torch.Tensor):
+                self._stats[k] = torch.cat([self._stats[k], v], dim=0)
+            elif isinstance(v, np.ndarray):
+                self._stats[k] = np.concatenate([self._stats[k], v], axis=0)
+            elif isinstance(v, list):
+                self._stats[k] = self._stats[k] + deepcopy(v)
+            else:
+                raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
+
+    def to_numpy(self) -> None:
+        for k, v in self._stats.items():
+            if isinstance(v, torch.Tensor):
+                self._stats[k] = v.detach().cpu().numpy()
+
+
+def is_box_near_crop_edge(
+    boxes: torch.Tensor, crop_box: List[int], orig_box: List[int], atol: float = 20.0
+) -> torch.Tensor:
+    """Filter masks at the edge of a crop, but not at the edge of the original image."""
+    crop_box_torch = torch.as_tensor(crop_box, dtype=torch.float, device=boxes.device)
+    orig_box_torch = torch.as_tensor(orig_box, dtype=torch.float, device=boxes.device)
+    boxes = uncrop_boxes_xyxy(boxes, crop_box).float()
+    near_crop_edge = torch.isclose(boxes, crop_box_torch[None, :], atol=atol, rtol=0)
+    near_image_edge = torch.isclose(boxes, orig_box_torch[None, :], atol=atol, rtol=0)
+    near_crop_edge = torch.logical_and(near_crop_edge, ~near_image_edge)
+    return torch.any(near_crop_edge, dim=1)
+
+
+def box_xyxy_to_xywh(box_xyxy: torch.Tensor) -> torch.Tensor:
+    box_xywh = deepcopy(box_xyxy)
+    box_xywh[2] = box_xywh[2] - box_xywh[0]
+    box_xywh[3] = box_xywh[3] - box_xywh[1]
+    return box_xywh
+
+
+def batch_iterator(batch_size: int, *args) -> Generator[List[Any], None, None]:
+    assert len(args) > 0 and all(
+        len(a) == len(args[0]) for a in args
+    ), "Batched iteration must have inputs of all the same size."
+    n_batches = len(args[0]) // batch_size + int(len(args[0]) % batch_size != 0)
+    for b in range(n_batches):
+        yield [arg[b * batch_size : (b + 1) * batch_size] for arg in args]
+
+
+def mask_to_rle_pytorch(tensor: torch.Tensor) -> List[Dict[str, Any]]:
+    """
+    Encodes masks to an uncompressed RLE, in the format expected by
+    pycoco tools.
+    """
+    # Put in fortran order and flatten h,w
+    b, h, w = tensor.shape
+    tensor = tensor.permute(0, 2, 1).flatten(1)
+
+    # Compute change indices
+    diff = tensor[:, 1:] ^ tensor[:, :-1]
+    change_indices = diff.nonzero()
+
+    # Encode run length
+    out = []
+    for i in range(b):
+        cur_idxs = change_indices[change_indices[:, 0] == i, 1]
+        cur_idxs = torch.cat(
+            [
+                torch.tensor([0], dtype=cur_idxs.dtype, device=cur_idxs.device),
+                cur_idxs + 1,
+                torch.tensor([h * w], dtype=cur_idxs.dtype, device=cur_idxs.device),
+            ]
+        )
+        btw_idxs = cur_idxs[1:] - cur_idxs[:-1]
+        counts = [] if tensor[i, 0] == 0 else [0]
+        counts.extend(btw_idxs.detach().cpu().tolist())
+        out.append({"size": [h, w], "counts": counts})
+    return out
+
+
+def rle_to_mask(rle: Dict[str, Any]) -> np.ndarray:
+    """Compute a binary mask from an uncompressed RLE."""
+    h, w = rle["size"]
+    mask = np.empty(h * w, dtype=bool)
+    idx = 0
+    parity = False
+    for count in rle["counts"]:
+        mask[idx : idx + count] = parity
+        idx += count
+        parity ^= True
+    mask = mask.reshape(w, h)
+    return mask.transpose()  # Put in C order
+
+
+def area_from_rle(rle: Dict[str, Any]) -> int:
+    return sum(rle["counts"][1::2])
+
+
+def calculate_stability_score(masks: torch.Tensor, mask_threshold: float, threshold_offset: float) -> torch.Tensor:
+    """
+    Computes the stability score for a batch of masks. The stability
+    score is the IoU between the binary masks obtained by thresholding
+    the predicted mask logits at high and low values.
+    """
+    # One mask is always contained inside the other.
+    # Save memory by preventing unnecesary cast to torch.int64
+    intersections = (masks > (mask_threshold + threshold_offset)).sum(-1, dtype=torch.int16).sum(-1, dtype=torch.int32)
+    unions = (masks > (mask_threshold - threshold_offset)).sum(-1, dtype=torch.int16).sum(-1, dtype=torch.int32)
+    return intersections / unions
+
+
+def build_point_grid(n_per_side: int) -> np.ndarray:
+    """Generates a 2D grid of points evenly spaced in [0,1]x[0,1]."""
+    offset = 1 / (2 * n_per_side)
+    points_one_side = np.linspace(offset, 1 - offset, n_per_side)
+    points_x = np.tile(points_one_side[None, :], (n_per_side, 1))
+    points_y = np.tile(points_one_side[:, None], (1, n_per_side))
+    points = np.stack([points_x, points_y], axis=-1).reshape(-1, 2)
+    return points
+
+
+def build_all_layer_point_grids(n_per_side: int, n_layers: int, scale_per_layer: int) -> List[np.ndarray]:
+    """Generates point grids for all crop layers."""
+    points_by_layer = []
+    for i in range(n_layers + 1):
+        n_points = int(n_per_side / (scale_per_layer**i))
+        points_by_layer.append(build_point_grid(n_points))
+    return points_by_layer
+
+
+def generate_crop_boxes(
+    im_size: Tuple[int, ...], n_layers: int, overlap_ratio: float
+) -> Tuple[List[List[int]], List[int]]:
+    """
+    Generates a list of crop boxes of different sizes. Each layer
+    has (2**i)**2 boxes for the ith layer.
+    """
+    crop_boxes, layer_idxs = [], []
+    im_h, im_w = im_size
+    short_side = min(im_h, im_w)
+
+    # Original image
+    crop_boxes.append([0, 0, im_w, im_h])
+    layer_idxs.append(0)
+
+    def crop_len(orig_len, n_crops, overlap):
+        return int(math.ceil((overlap * (n_crops - 1) + orig_len) / n_crops))
+
+    for i_layer in range(n_layers):
+        n_crops_per_side = 2 ** (i_layer + 1)
+        overlap = int(overlap_ratio * short_side * (2 / n_crops_per_side))
+
+        crop_w = crop_len(im_w, n_crops_per_side, overlap)
+        crop_h = crop_len(im_h, n_crops_per_side, overlap)
+
+        crop_box_x0 = [int((crop_w - overlap) * i) for i in range(n_crops_per_side)]
+        crop_box_y0 = [int((crop_h - overlap) * i) for i in range(n_crops_per_side)]
+
+        # Crops in XYWH format
+        for x0, y0 in product(crop_box_x0, crop_box_y0):
+            box = [x0, y0, min(x0 + crop_w, im_w), min(y0 + crop_h, im_h)]
+            crop_boxes.append(box)
+            layer_idxs.append(i_layer + 1)
+
+    return crop_boxes, layer_idxs
+
+
+def uncrop_boxes_xyxy(boxes: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
+    x0, y0, _, _ = crop_box
+    offset = torch.tensor([[x0, y0, x0, y0]], device=boxes.device)
+    # Check if boxes has a channel dimension
+    if len(boxes.shape) == 3:
+        offset = offset.unsqueeze(1)
+    return boxes + offset
+
+
+def uncrop_points(points: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
+    x0, y0, _, _ = crop_box
+    offset = torch.tensor([[x0, y0]], device=points.device)
+    # Check if points has a channel dimension
+    if len(points.shape) == 3:
+        offset = offset.unsqueeze(1)
+    return points + offset
+
+
+def uncrop_masks(masks: torch.Tensor, crop_box: List[int], orig_h: int, orig_w: int) -> torch.Tensor:
+    x0, y0, x1, y1 = crop_box
+    if x0 == 0 and y0 == 0 and x1 == orig_w and y1 == orig_h:
+        return masks
+    # Coordinate transform masks
+    pad_x, pad_y = orig_w - (x1 - x0), orig_h - (y1 - y0)
+    pad = (x0, pad_x - x0, y0, pad_y - y0)
+    return torch.nn.functional.pad(masks, pad, value=0)
+
+
+def remove_small_regions(mask: np.ndarray, area_thresh: float, mode: str) -> Tuple[np.ndarray, bool]:
+    """
+    Removes small disconnected regions and holes in a mask. Returns the
+    mask and an indicator of if the mask has been modified.
+    """
+    import cv2  # type: ignore
+
+    assert mode in ["holes", "islands"]
+    correct_holes = mode == "holes"
+    working_mask = (correct_holes ^ mask).astype(np.uint8)
+    n_labels, regions, stats, _ = cv2.connectedComponentsWithStats(working_mask, 8)
+    sizes = stats[:, -1][1:]  # Row 0 is background label
+    small_regions = [i + 1 for i, s in enumerate(sizes) if s < area_thresh]
+    if len(small_regions) == 0:
+        return mask, False
+    fill_labels = [0] + small_regions
+    if not correct_holes:
+        fill_labels = [i for i in range(n_labels) if i not in fill_labels]
+        # If every region is below threshold, keep largest
+        if len(fill_labels) == 0:
+            fill_labels = [int(np.argmax(sizes)) + 1]
+    mask = np.isin(regions, fill_labels)
+    return mask, True
+
+
+def coco_encode_rle(uncompressed_rle: Dict[str, Any]) -> Dict[str, Any]:
+    from pycocotools import mask as mask_utils  # type: ignore
+
+    h, w = uncompressed_rle["size"]
+    rle = mask_utils.frPyObjects(uncompressed_rle, h, w)
+    rle["counts"] = rle["counts"].decode("utf-8")  # Necessary to serialize with json
+    return rle
+
+
+def batched_mask_to_box(masks: torch.Tensor) -> torch.Tensor:
+    """
+    Calculates boxes in XYXY format around masks. Return [0,0,0,0] for
+    an empty mask. For input shape C1xC2x...xHxW, the output shape is C1xC2x...x4.
+    """
+    # torch.max below raises an error on empty inputs, just skip in this case
+    if torch.numel(masks) == 0:
+        return torch.zeros(*masks.shape[:-2], 4, device=masks.device)
+
+    # Normalize shape to CxHxW
+    shape = masks.shape
+    h, w = shape[-2:]
+    if len(shape) > 2:
+        masks = masks.flatten(0, -3)
+    else:
+        masks = masks.unsqueeze(0)
+
+    # Get top and bottom edges
+    in_height, _ = torch.max(masks, dim=-1)
+    in_height_coords = in_height * torch.arange(h, device=in_height.device)[None, :]
+    bottom_edges, _ = torch.max(in_height_coords, dim=-1)
+    in_height_coords = in_height_coords + h * (~in_height)
+    top_edges, _ = torch.min(in_height_coords, dim=-1)
+
+    # Get left and right edges
+    in_width, _ = torch.max(masks, dim=-2)
+    in_width_coords = in_width * torch.arange(w, device=in_width.device)[None, :]
+    right_edges, _ = torch.max(in_width_coords, dim=-1)
+    in_width_coords = in_width_coords + w * (~in_width)
+    left_edges, _ = torch.min(in_width_coords, dim=-1)
+
+    # If the mask is empty the right edge will be to the left of the left edge.
+    # Replace these boxes with [0, 0, 0, 0]
+    empty_filter = (right_edges < left_edges) | (bottom_edges < top_edges)
+    out = torch.stack([left_edges, top_edges, right_edges, bottom_edges], dim=-1)
+    out = out * (~empty_filter).unsqueeze(-1)
+
+    # Return to original shape
+    if len(shape) > 2:
+        out = out.reshape(*shape[:-2], 4)
+    else:
+        out = out[0]
+
+    return out

invokeai/backend/image_util/grounding_segment_anything/segment_anything/utils/transforms.py

@@ -0,0 +1,92 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from copy import deepcopy
+from typing import Tuple
+
+import numpy as np
+import torch
+from torch.nn import functional as F
+from torchvision.transforms.functional import resize, to_pil_image  # type: ignore
+
+
+class ResizeLongestSide:
+    """
+    Resizes images to longest side 'target_length', as well as provides
+    methods for resizing coordinates and boxes. Provides methods for
+    transforming both numpy array and batched torch tensors.
+    """
+
+    def __init__(self, target_length: int) -> None:
+        self.target_length = target_length
+
+    def apply_image(self, image: np.ndarray) -> np.ndarray:
+        """
+        Expects a numpy array with shape HxWxC in uint8 format.
+        """
+        target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
+        return np.array(resize(to_pil_image(image), target_size))
+
+    def apply_coords(self, coords: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
+        """
+        Expects a numpy array of length 2 in the final dimension. Requires the
+        original image size in (H, W) format.
+        """
+        old_h, old_w = original_size
+        new_h, new_w = self.get_preprocess_shape(original_size[0], original_size[1], self.target_length)
+        coords = deepcopy(coords).astype(float)
+        coords[..., 0] = coords[..., 0] * (new_w / old_w)
+        coords[..., 1] = coords[..., 1] * (new_h / old_h)
+        return coords
+
+    def apply_boxes(self, boxes: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
+        """
+        Expects a numpy array shape Bx4. Requires the original image size
+        in (H, W) format.
+        """
+        boxes = self.apply_coords(boxes.reshape(-1, 2, 2), original_size)
+        return boxes.reshape(-1, 4)
+
+    def apply_image_torch(self, image: torch.Tensor) -> torch.Tensor:
+        """
+        Expects batched images with shape BxCxHxW and float format. This
+        transformation may not exactly match apply_image. apply_image is
+        the transformation expected by the model.
+        """
+        # Expects an image in BCHW format. May not exactly match apply_image.
+        target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
+        return F.interpolate(image, target_size, mode="bilinear", align_corners=False, antialias=True)
+
+    def apply_coords_torch(self, coords: torch.Tensor, original_size: Tuple[int, ...]) -> torch.Tensor:
+        """
+        Expects a torch tensor with length 2 in the last dimension. Requires the
+        original image size in (H, W) format.
+        """
+        old_h, old_w = original_size
+        new_h, new_w = self.get_preprocess_shape(original_size[0], original_size[1], self.target_length)
+        coords = deepcopy(coords).to(torch.float)
+        coords[..., 0] = coords[..., 0] * (new_w / old_w)
+        coords[..., 1] = coords[..., 1] * (new_h / old_h)
+        return coords
+
+    def apply_boxes_torch(self, boxes: torch.Tensor, original_size: Tuple[int, ...]) -> torch.Tensor:
+        """
+        Expects a torch tensor with shape Bx4. Requires the original image
+        size in (H, W) format.
+        """
+        boxes = self.apply_coords_torch(boxes.reshape(-1, 2, 2), original_size)
+        return boxes.reshape(-1, 4)
+
+    @staticmethod
+    def get_preprocess_shape(oldh: int, oldw: int, long_side_length: int) -> Tuple[int, int]:
+        """
+        Compute the output size given input size and target long side length.
+        """
+        scale = long_side_length * 1.0 / max(oldh, oldw)
+        newh, neww = oldh * scale, oldw * scale
+        neww = int(neww + 0.5)
+        newh = int(newh + 0.5)
+        return (newh, neww)

invokeai/backend/image_util/segment_anything/mask_refinement.py

@@ -1,50 +0,0 @@
-# This file contains utilities for Grounded-SAM mask refinement based on:
-# https://github.com/NielsRogge/Transformers-Tutorials/blob/a39f33ac1557b02ebfb191ea7753e332b5ca933f/Grounding%20DINO/GroundingDINO_with_Segment_Anything.ipynb
-
-
-import cv2
-import numpy as np
-import numpy.typing as npt
-
-
-def mask_to_polygon(mask: npt.NDArray[np.uint8]) -> list[tuple[int, int]]:
-    """Convert a binary mask to a polygon.
-
-    Returns:
-        list[list[int]]: List of (x, y) coordinates representing the vertices of the polygon.
-    """
-    # Find contours in the binary mask.
-    contours, _ = cv2.findContours(mask.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
-
-    # Find the contour with the largest area.
-    largest_contour = max(contours, key=cv2.contourArea)
-
-    # Extract the vertices of the contour.
-    polygon = largest_contour.reshape(-1, 2).tolist()
-
-    return polygon
-
-
-def polygon_to_mask(
-    polygon: list[tuple[int, int]], image_shape: tuple[int, int], fill_value: int = 1
-) -> npt.NDArray[np.uint8]:
-    """Convert a polygon to a segmentation mask.
-
-    Args:
-        polygon (list): List of (x, y) coordinates representing the vertices of the polygon.
-        image_shape (tuple): Shape of the image (height, width) for the mask.
-        fill_value (int): Value to fill the polygon with.
-
-    Returns:
-        np.ndarray: Segmentation mask with the polygon filled (with value 255).
-    """
-    # Create an empty mask.
-    mask = np.zeros(image_shape, dtype=np.uint8)
-
-    # Convert polygon to an array of points.
-    pts = np.array(polygon, dtype=np.int32)
-
-    # Fill the polygon with white color (255).
-    cv2.fillPoly(mask, [pts], color=(fill_value,))
-
-    return mask

invokeai/backend/image_util/segment_anything/segment_anything_pipeline.py

@@ -1,53 +0,0 @@
-from typing import Optional
-
-import torch
-from PIL import Image
-from transformers.models.sam import SamModel
-from transformers.models.sam.processing_sam import SamProcessor
-
-from invokeai.backend.raw_model import RawModel
-
-
-class SegmentAnythingPipeline(RawModel):
-    """A wrapper class for the transformers SAM model and processor that makes it compatible with the model manager."""
-
-    def __init__(self, sam_model: SamModel, sam_processor: SamProcessor):
-        self._sam_model = sam_model
-        self._sam_processor = sam_processor
-
-    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None):
-        # HACK(ryand): The SAM pipeline does not work on MPS devices. We only allow it to be moved to CPU or CUDA.
-        if device is not None and device.type not in {"cpu", "cuda"}:
-            device = None
-        self._sam_model.to(device=device, dtype=dtype)
-
-    def calc_size(self) -> int:
-        # HACK(ryand): Fix the circular import issue.
-        from invokeai.backend.model_manager.load.model_util import calc_module_size
-
-        return calc_module_size(self._sam_model)
-
-    def segment(self, image: Image.Image, bounding_boxes: list[list[int]]) -> torch.Tensor:
-        """Run the SAM model.
-
-        Args:
-            image (Image.Image): The image to segment.
-            bounding_boxes (list[list[int]]): The bounding box prompts. Each bounding box is in the format
-                [xmin, ymin, xmax, ymax].
-
-        Returns:
-            torch.Tensor: The segmentation masks. dtype: torch.bool. shape: [num_masks, channels, height, width].
-        """
-        # Add batch dimension of 1 to the bounding boxes.
-        boxes = [bounding_boxes]
-        inputs = self._sam_processor(images=image, input_boxes=boxes, return_tensors="pt").to(self._sam_model.device)
-        outputs = self._sam_model(**inputs)
-        masks = self._sam_processor.post_process_masks(
-            masks=outputs.pred_masks,
-            original_sizes=inputs.original_sizes,
-            reshaped_input_sizes=inputs.reshaped_input_sizes,
-        )
-
-        # There should be only one batch.
-        assert len(masks) == 1
-        return masks[0]

invokeai/backend/load_flux_model.py

@@ -1,129 +0,0 @@
-import json
-import os
-import time
-from pathlib import Path
-from typing import Union
-
-import torch
-from diffusers.models.model_loading_utils import load_state_dict
-from diffusers.models.transformers.transformer_flux import FluxTransformer2DModel
-from diffusers.utils import (
-    CONFIG_NAME,
-    SAFE_WEIGHTS_INDEX_NAME,
-    SAFETENSORS_WEIGHTS_NAME,
-    _get_checkpoint_shard_files,
-    is_accelerate_available,
-)
-from optimum.quanto import qfloat8
-from optimum.quanto.models import QuantizedDiffusersModel
-from optimum.quanto.models.shared_dict import ShardedStateDict
-
-from invokeai.backend.requantize import requantize
-
-
-class QuantizedFluxTransformer2DModel(QuantizedDiffusersModel):
-    base_class = FluxTransformer2DModel
-
-    @classmethod
-    def from_pretrained(cls, model_name_or_path: Union[str, os.PathLike]):
-        if cls.base_class is None:
-            raise ValueError("The `base_class` attribute needs to be configured.")
-
-        if not is_accelerate_available():
-            raise ValueError("Reloading a quantized diffusers model requires the accelerate library.")
-        from accelerate import init_empty_weights
-
-        if os.path.isdir(model_name_or_path):
-            # Look for a quantization map
-            qmap_path = os.path.join(model_name_or_path, cls._qmap_name())
-            if not os.path.exists(qmap_path):
-                raise ValueError(f"No quantization map found in {model_name_or_path}: is this a quantized model ?")
-
-            # Look for original model config file.
-            model_config_path = os.path.join(model_name_or_path, CONFIG_NAME)
-            if not os.path.exists(model_config_path):
-                raise ValueError(f"{CONFIG_NAME} not found in {model_name_or_path}.")
-
-            with open(qmap_path, "r", encoding="utf-8") as f:
-                qmap = json.load(f)
-
-            with open(model_config_path, "r", encoding="utf-8") as f:
-                original_model_cls_name = json.load(f)["_class_name"]
-            configured_cls_name = cls.base_class.__name__
-            if configured_cls_name != original_model_cls_name:
-                raise ValueError(
-                    f"Configured base class ({configured_cls_name}) differs from what was derived from the provided configuration ({original_model_cls_name})."
-                )
-
-            # Create an empty model
-            config = cls.base_class.load_config(model_name_or_path)
-            with init_empty_weights():
-                model = cls.base_class.from_config(config)
-
-            # Look for the index of a sharded checkpoint
-            checkpoint_file = os.path.join(model_name_or_path, SAFE_WEIGHTS_INDEX_NAME)
-            if os.path.exists(checkpoint_file):
-                # Convert the checkpoint path to a list of shards
-                _, sharded_metadata = _get_checkpoint_shard_files(model_name_or_path, checkpoint_file)
-                # Create a mapping for the sharded safetensor files
-                state_dict = ShardedStateDict(model_name_or_path, sharded_metadata["weight_map"])
-            else:
-                # Look for a single checkpoint file
-                checkpoint_file = os.path.join(model_name_or_path, SAFETENSORS_WEIGHTS_NAME)
-                if not os.path.exists(checkpoint_file):
-                    raise ValueError(f"No safetensor weights found in {model_name_or_path}.")
-                # Get state_dict from model checkpoint
-                state_dict = load_state_dict(checkpoint_file)
-
-            # Requantize and load quantized weights from state_dict
-            requantize(model, state_dict=state_dict, quantization_map=qmap)
-            model.eval()
-            return cls(model)
-        else:
-            raise NotImplementedError("Reloading quantized models directly from the hub is not supported yet.")
-
-
-def load_flux_transformer(path: Path) -> FluxTransformer2DModel:
-    # model = FluxTransformer2DModel.from_pretrained(path, local_files_only=True, torch_dtype=torch.bfloat16)
-    model_8bit_path = path / "quantized"
-    if model_8bit_path.exists():
-        # The quantized model exists, load it.
-        # TODO(ryand): The requantize(...) operation in from_pretrained(...) is very slow. This seems like
-        # something that we should be able to make much faster.
-        q_model = QuantizedFluxTransformer2DModel.from_pretrained(model_8bit_path)
-
-        # Access the underlying wrapped model.
-        # We access the wrapped model, even though it is private, because it simplifies the type checking by
-        # always returning a FluxTransformer2DModel from this function.
-        model = q_model._wrapped
-    else:
-        # The quantized model does not exist yet, quantize and save it.
-        # TODO(ryand): Loading in float16 and then quantizing seems to result in NaNs. In order to run this on
-        # GPUs that don't support bfloat16, we would need to host the quantized model instead of generating it
-        # here.
-        model = FluxTransformer2DModel.from_pretrained(path, local_files_only=True, torch_dtype=torch.bfloat16)
-        assert isinstance(model, FluxTransformer2DModel)
-
-        q_model = QuantizedFluxTransformer2DModel.quantize(model, weights=qfloat8)
-
-        model_8bit_path.mkdir(parents=True, exist_ok=True)
-        q_model.save_pretrained(model_8bit_path)
-
-        # (See earlier comment about accessing the wrapped model.)
-        model = q_model._wrapped
-
-    assert isinstance(model, FluxTransformer2DModel)
-    return model
-
-
-def main():
-    start = time.time()
-    model = load_flux_transformer(
-        Path("/data/invokeai/models/.download_cache/black-forest-labs_flux.1-schnell/FLUX.1-schnell/transformer/")
-    )
-    print(f"Time to load: {time.time() - start}s")
-    print("hi")
-
-
-if __name__ == "__main__":
-    main()

invokeai/backend/lora.py

@@ -3,13 +3,12 @@
 
 import bisect
 from pathlib import Path
-from typing import Dict, List, Optional, Set, Tuple, Union
+from typing import Dict, List, Optional, Tuple, Union
 
 import torch
 from safetensors.torch import load_file
 from typing_extensions import Self
 
-import invokeai.backend.util.logging as logger
 from invokeai.backend.model_manager import BaseModelType
 from invokeai.backend.raw_model import RawModel
 
@@ -47,19 +46,9 @@ class LoRALayerBase:
         self.rank = None  # set in layer implementation
         self.layer_key = layer_key
 
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+    def get_weight(self, orig_weight: Optional[torch.Tensor]) -> torch.Tensor:
         raise NotImplementedError()
 
-    def get_bias(self, orig_bias: torch.Tensor) -> Optional[torch.Tensor]:
-        return self.bias
-
-    def get_parameters(self, orig_module: torch.nn.Module) -> Dict[str, torch.Tensor]:
-        params = {"weight": self.get_weight(orig_module.weight)}
-        bias = self.get_bias(orig_module.bias)
-        if bias is not None:
-            params["bias"] = bias
-        return params
-
     def calc_size(self) -> int:
         model_size = 0
         for val in [self.bias]:
@@ -71,17 +60,6 @@ class LoRALayerBase:
         if self.bias is not None:
             self.bias = self.bias.to(device=device, dtype=dtype)
 
-    def check_keys(self, values: Dict[str, torch.Tensor], known_keys: Set[str]):
-        """Log a warning if values contains unhandled keys."""
-        # {"alpha", "bias_indices", "bias_values", "bias_size"} are hard-coded, because they are handled by
-        # `LoRALayerBase`. Sub-classes should provide the known_keys that they handled.
-        all_known_keys = known_keys | {"alpha", "bias_indices", "bias_values", "bias_size"}
-        unknown_keys = set(values.keys()) - all_known_keys
-        if unknown_keys:
-            logger.warning(
-                f"Unexpected keys found in LoRA/LyCORIS layer, model might work incorrectly! Keys: {unknown_keys}"
-            )
-
 
 # TODO: find and debug lora/locon with bias
 class LoRALayer(LoRALayerBase):
@@ -98,19 +76,14 @@ class LoRALayer(LoRALayerBase):
 
         self.up = values["lora_up.weight"]
         self.down = values["lora_down.weight"]
-        self.mid = values.get("lora_mid.weight", None)
+        if "lora_mid.weight" in values:
+            self.mid: Optional[torch.Tensor] = values["lora_mid.weight"]
+        else:
+            self.mid = None
 
         self.rank = self.down.shape[0]
-        self.check_keys(
-            values,
-            {
-                "lora_up.weight",
-                "lora_down.weight",
-                "lora_mid.weight",
-            },
-        )
 
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+    def get_weight(self, orig_weight: Optional[torch.Tensor]) -> torch.Tensor:
         if self.mid is not None:
             up = self.up.reshape(self.up.shape[0], self.up.shape[1])
             down = self.down.reshape(self.down.shape[0], self.down.shape[1])
@@ -152,23 +125,20 @@ class LoHALayer(LoRALayerBase):
         self.w1_b = values["hada_w1_b"]
         self.w2_a = values["hada_w2_a"]
         self.w2_b = values["hada_w2_b"]
-        self.t1 = values.get("hada_t1", None)
-        self.t2 = values.get("hada_t2", None)
+
+        if "hada_t1" in values:
+            self.t1: Optional[torch.Tensor] = values["hada_t1"]
+        else:
+            self.t1 = None
+
+        if "hada_t2" in values:
+            self.t2: Optional[torch.Tensor] = values["hada_t2"]
+        else:
+            self.t2 = None
 
         self.rank = self.w1_b.shape[0]
-        self.check_keys(
-            values,
-            {
-                "hada_w1_a",
-                "hada_w1_b",
-                "hada_w2_a",
-                "hada_w2_b",
-                "hada_t1",
-                "hada_t2",
-            },
-        )
 
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+    def get_weight(self, orig_weight: Optional[torch.Tensor]) -> torch.Tensor:
         if self.t1 is None:
             weight: torch.Tensor = (self.w1_a @ self.w1_b) * (self.w2_a @ self.w2_b)
 
@@ -216,39 +186,37 @@ class LoKRLayer(LoRALayerBase):
     ):
         super().__init__(layer_key, values)
 
-        self.w1 = values.get("lokr_w1", None)
-        if self.w1 is None:
+        if "lokr_w1" in values:
+            self.w1: Optional[torch.Tensor] = values["lokr_w1"]
+            self.w1_a = None
+            self.w1_b = None
+        else:
+            self.w1 = None
             self.w1_a = values["lokr_w1_a"]
             self.w1_b = values["lokr_w1_b"]
 
-        self.w2 = values.get("lokr_w2", None)
-        if self.w2 is None:
+        if "lokr_w2" in values:
+            self.w2: Optional[torch.Tensor] = values["lokr_w2"]
+            self.w2_a = None
+            self.w2_b = None
+        else:
+            self.w2 = None
             self.w2_a = values["lokr_w2_a"]
             self.w2_b = values["lokr_w2_b"]
 
-        self.t2 = values.get("lokr_t2", None)
+        if "lokr_t2" in values:
+            self.t2: Optional[torch.Tensor] = values["lokr_t2"]
+        else:
+            self.t2 = None
 
-        if self.w1_b is not None:
-            self.rank = self.w1_b.shape[0]
-        elif self.w2_b is not None:
-            self.rank = self.w2_b.shape[0]
+        if "lokr_w1_b" in values:
+            self.rank = values["lokr_w1_b"].shape[0]
+        elif "lokr_w2_b" in values:
+            self.rank = values["lokr_w2_b"].shape[0]
         else:
             self.rank = None  # unscaled
 
-        self.check_keys(
-            values,
-            {
-                "lokr_w1",
-                "lokr_w1_a",
-                "lokr_w1_b",
-                "lokr_w2",
-                "lokr_w2_a",
-                "lokr_w2_b",
-                "lokr_t2",
-            },
-        )
-
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+    def get_weight(self, orig_weight: Optional[torch.Tensor]) -> torch.Tensor:
         w1: Optional[torch.Tensor] = self.w1
         if w1 is None:
             assert self.w1_a is not None
@@ -304,9 +272,7 @@ class LoKRLayer(LoRALayerBase):
 
 
 class FullLayer(LoRALayerBase):
-    # bias handled in LoRALayerBase(calc_size, to)
     # weight: torch.Tensor
-    # bias: Optional[torch.Tensor]
 
     def __init__(
         self,
@@ -316,12 +282,15 @@ class FullLayer(LoRALayerBase):
         super().__init__(layer_key, values)
 
         self.weight = values["diff"]
-        self.bias = values.get("diff_b", None)
+
+        if len(values.keys()) > 1:
+            _keys = list(values.keys())
+            _keys.remove("diff")
+            raise NotImplementedError(f"Unexpected keys in lora diff layer: {_keys}")
 
         self.rank = None  # unscaled
-        self.check_keys(values, {"diff", "diff_b"})
 
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+    def get_weight(self, orig_weight: Optional[torch.Tensor]) -> torch.Tensor:
         return self.weight
 
     def calc_size(self) -> int:
@@ -350,9 +319,8 @@ class IA3Layer(LoRALayerBase):
         self.on_input = values["on_input"]
 
         self.rank = None  # unscaled
-        self.check_keys(values, {"weight", "on_input"})
 
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+    def get_weight(self, orig_weight: Optional[torch.Tensor]) -> torch.Tensor:
         weight = self.weight
         if not self.on_input:
             weight = weight.reshape(-1, 1)
@@ -490,19 +458,16 @@ class LoRAModelRaw(RawModel):  # (torch.nn.Module):
             state_dict = cls._convert_sdxl_keys_to_diffusers_format(state_dict)
 
         for layer_key, values in state_dict.items():
-            # Detect layers according to LyCORIS detection logic(`weight_list_det`)
-            # https://github.com/KohakuBlueleaf/LyCORIS/tree/8ad8000efb79e2b879054da8c9356e6143591bad/lycoris/modules
-
             # lora and locon
-            if "lora_up.weight" in values:
+            if "lora_down.weight" in values:
                 layer: AnyLoRALayer = LoRALayer(layer_key, values)
 
             # loha
-            elif "hada_w1_a" in values:
+            elif "hada_w1_b" in values:
                 layer = LoHALayer(layer_key, values)
 
             # lokr
-            elif "lokr_w1" in values or "lokr_w1_a" in values:
+            elif "lokr_w1_b" in values or "lokr_w1" in values:
                 layer = LoKRLayer(layer_key, values)
 
             # diff
@@ -510,7 +475,7 @@ class LoRAModelRaw(RawModel):  # (torch.nn.Module):
                 layer = FullLayer(layer_key, values)
 
             # ia3
-            elif "on_input" in values:
+            elif "weight" in values and "on_input" in values:
                 layer = IA3Layer(layer_key, values)
 
             else:

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

@@ -98,9 +98,6 @@ class StableDiffusionDiffusersModel(GenericDiffusersLoader):
                 ModelVariantType.Normal: StableDiffusionXLPipeline,
                 ModelVariantType.Inpaint: StableDiffusionXLInpaintPipeline,
             },
-            BaseModelType.StableDiffusionXLRefiner: {
-                ModelVariantType.Normal: StableDiffusionXLPipeline,
-            },
         }
         assert isinstance(config, MainCheckpointConfig)
         try:

invokeai/backend/model_manager/load/model_util.py

@@ -11,8 +11,6 @@ from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.schedulers.scheduling_utils import SchedulerMixin
 from transformers import CLIPTokenizer
 
-from invokeai.backend.image_util.grounding_dino.grounding_dino_pipeline import GroundingDinoPipeline
-from invokeai.backend.image_util.segment_anything.segment_anything_pipeline import SegmentAnythingPipeline
 from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
 from invokeai.backend.lora import LoRAModelRaw
 from invokeai.backend.model_manager.config import AnyModel
@@ -36,17 +34,7 @@ def calc_model_size_by_data(logger: logging.Logger, model: AnyModel) -> int:
     elif isinstance(model, CLIPTokenizer):
         # TODO(ryand): Accurately calculate the tokenizer's size. It's small enough that it shouldn't matter for now.
         return 0
-    elif isinstance(
-        model,
-        (
-            TextualInversionModelRaw,
-            IPAdapter,
-            LoRAModelRaw,
-            SpandrelImageToImageModel,
-            GroundingDinoPipeline,
-            SegmentAnythingPipeline,
-        ),
-    ):
+    elif isinstance(model, (TextualInversionModelRaw, IPAdapter, LoRAModelRaw, SpandrelImageToImageModel)):
         return model.calc_size()
     else:
         # TODO(ryand): Promote this from a log to an exception once we are confident that we are handling all of the

invokeai/backend/model_manager/starter_models.py

@@ -187,171 +187,164 @@ STARTER_MODELS: list[StarterModel] = [
     # endregion
     # region ControlNet
     StarterModel(
-        name="QRCode Monster v2 (SD1.5)",
+        name="QRCode Monster",
         base=BaseModelType.StableDiffusion1,
-        source="monster-labs/control_v1p_sd15_qrcode_monster::v2",
-        description="ControlNet model that generates scannable creative QR codes",
-        type=ModelType.ControlNet,
-    ),
-    StarterModel(
-        name="QRCode Monster (SDXL)",
-        base=BaseModelType.StableDiffusionXL,
-        source="monster-labs/control_v1p_sdxl_qrcode_monster",
-        description="ControlNet model that generates scannable creative QR codes",
+        source="monster-labs/control_v1p_sd15_qrcode_monster",
+        description="Controlnet model that generates scannable creative QR codes",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="canny",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11p_sd15_canny",
-        description="ControlNet weights trained on sd-1.5 with canny conditioning.",
+        description="Controlnet weights trained on sd-1.5 with canny conditioning.",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="inpaint",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11p_sd15_inpaint",
-        description="ControlNet weights trained on sd-1.5 with canny conditioning, inpaint version",
+        description="Controlnet weights trained on sd-1.5 with canny conditioning, inpaint version",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="mlsd",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11p_sd15_mlsd",
-        description="ControlNet weights trained on sd-1.5 with canny conditioning, MLSD version",
+        description="Controlnet weights trained on sd-1.5 with canny conditioning, MLSD version",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="depth",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11f1p_sd15_depth",
-        description="ControlNet weights trained on sd-1.5 with depth conditioning",
+        description="Controlnet weights trained on sd-1.5 with depth conditioning",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="normal_bae",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11p_sd15_normalbae",
-        description="ControlNet weights trained on sd-1.5 with normalbae image conditioning",
+        description="Controlnet weights trained on sd-1.5 with normalbae image conditioning",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="seg",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11p_sd15_seg",
-        description="ControlNet weights trained on sd-1.5 with seg image conditioning",
+        description="Controlnet weights trained on sd-1.5 with seg image conditioning",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="lineart",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11p_sd15_lineart",
-        description="ControlNet weights trained on sd-1.5 with lineart image conditioning",
+        description="Controlnet weights trained on sd-1.5 with lineart image conditioning",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="lineart_anime",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11p_sd15s2_lineart_anime",
-        description="ControlNet weights trained on sd-1.5 with anime image conditioning",
+        description="Controlnet weights trained on sd-1.5 with anime image conditioning",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="openpose",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11p_sd15_openpose",
-        description="ControlNet weights trained on sd-1.5 with openpose image conditioning",
+        description="Controlnet weights trained on sd-1.5 with openpose image conditioning",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="scribble",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11p_sd15_scribble",
-        description="ControlNet weights trained on sd-1.5 with scribble image conditioning",
+        description="Controlnet weights trained on sd-1.5 with scribble image conditioning",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="softedge",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11p_sd15_softedge",
-        description="ControlNet weights trained on sd-1.5 with soft edge conditioning",
+        description="Controlnet weights trained on sd-1.5 with soft edge conditioning",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="shuffle",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11e_sd15_shuffle",
-        description="ControlNet weights trained on sd-1.5 with shuffle image conditioning",
+        description="Controlnet weights trained on sd-1.5 with shuffle image conditioning",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="tile",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11f1e_sd15_tile",
-        description="ControlNet weights trained on sd-1.5 with tiled image conditioning",
+        description="Controlnet weights trained on sd-1.5 with tiled image conditioning",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="ip2p",
         base=BaseModelType.StableDiffusion1,
         source="lllyasviel/control_v11e_sd15_ip2p",
-        description="ControlNet weights trained on sd-1.5 with ip2p conditioning.",
+        description="Controlnet weights trained on sd-1.5 with ip2p conditioning.",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="canny-sdxl",
         base=BaseModelType.StableDiffusionXL,
-        source="xinsir/controlNet-canny-sdxl-1.0",
-        description="ControlNet weights trained on sdxl-1.0 with canny conditioning, by Xinsir.",
+        source="xinsir/controlnet-canny-sdxl-1.0",
+        description="Controlnet weights trained on sdxl-1.0 with canny conditioning, by Xinsir.",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="depth-sdxl",
         base=BaseModelType.StableDiffusionXL,
-        source="diffusers/controlNet-depth-sdxl-1.0",
-        description="ControlNet weights trained on sdxl-1.0 with depth conditioning.",
+        source="diffusers/controlnet-depth-sdxl-1.0",
+        description="Controlnet weights trained on sdxl-1.0 with depth conditioning.",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="softedge-dexined-sdxl",
         base=BaseModelType.StableDiffusionXL,
-        source="SargeZT/controlNet-sd-xl-1.0-softedge-dexined",
-        description="ControlNet weights trained on sdxl-1.0 with dexined soft edge preprocessing.",
+        source="SargeZT/controlnet-sd-xl-1.0-softedge-dexined",
+        description="Controlnet weights trained on sdxl-1.0 with dexined soft edge preprocessing.",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="depth-16bit-zoe-sdxl",
         base=BaseModelType.StableDiffusionXL,
-        source="SargeZT/controlNet-sd-xl-1.0-depth-16bit-zoe",
-        description="ControlNet weights trained on sdxl-1.0 with Zoe's preprocessor (16 bits).",
+        source="SargeZT/controlnet-sd-xl-1.0-depth-16bit-zoe",
+        description="Controlnet weights trained on sdxl-1.0 with Zoe's preprocessor (16 bits).",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="depth-zoe-sdxl",
         base=BaseModelType.StableDiffusionXL,
-        source="diffusers/controlNet-zoe-depth-sdxl-1.0",
-        description="ControlNet weights trained on sdxl-1.0 with Zoe's preprocessor (32 bits).",
+        source="diffusers/controlnet-zoe-depth-sdxl-1.0",
+        description="Controlnet weights trained on sdxl-1.0 with Zoe's preprocessor (32 bits).",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="openpose-sdxl",
         base=BaseModelType.StableDiffusionXL,
-        source="xinsir/controlNet-openpose-sdxl-1.0",
-        description="ControlNet weights trained on sdxl-1.0 compatible with the DWPose processor by Xinsir.",
+        source="xinsir/controlnet-openpose-sdxl-1.0",
+        description="Controlnet weights trained on sdxl-1.0 compatible with the DWPose processor by Xinsir.",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="scribble-sdxl",
         base=BaseModelType.StableDiffusionXL,
-        source="xinsir/controlNet-scribble-sdxl-1.0",
-        description="ControlNet weights trained on sdxl-1.0 compatible with various lineart processors and black/white sketches by Xinsir.",
+        source="xinsir/controlnet-scribble-sdxl-1.0",
+        description="Controlnet weights trained on sdxl-1.0 compatible with various lineart processors and black/white sketches by Xinsir.",
         type=ModelType.ControlNet,
     ),
     StarterModel(
         name="tile-sdxl",
         base=BaseModelType.StableDiffusionXL,
-        source="xinsir/controlNet-tile-sdxl-1.0",
-        description="ControlNet weights trained on sdxl-1.0 with tiled image conditioning",
+        source="xinsir/controlnet-tile-sdxl-1.0",
+        description="Controlnet weights trained on sdxl-1.0 with tiled image conditioning",
         type=ModelType.ControlNet,
     ),
     # endregion

invokeai/backend/model_manager/util/select_hf_files.py

@@ -54,7 +54,6 @@ def filter_files(
                 "lora_weights.safetensors",
                 "weights.pb",
                 "onnx_data",
-                "spiece.model", # Added for `black-forest-labs/FLUX.1-schnell`.
             )
         ):
             paths.append(file)
@@ -63,7 +62,7 @@ def filter_files(
         # downloading random checkpoints that might also be in the repo. However there is no guarantee
         # that a checkpoint doesn't contain "model" in its name, and no guarantee that future diffusers models
         # will adhere to this naming convention, so this is an area to be careful of.
-        elif re.search(r"model.*\.(safetensors|bin|onnx|xml|pth|pt|ckpt|msgpack)$", file.name):
+        elif re.search(r"model(\.[^.]+)?\.(safetensors|bin|onnx|xml|pth|pt|ckpt|msgpack)$", file.name):
             paths.append(file)
 
     # limit search to subfolder if requested
@@ -98,9 +97,7 @@ def _filter_by_variant(files: List[Path], variant: ModelRepoVariant) -> Set[Path
             if variant == ModelRepoVariant.Flax:
                 result.add(path)
 
-        # Note: '.model' was added to support:
-        # https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/768d12a373ed5cc9ef9a9dea7504dc09fcc14842/tokenizer_2/spiece.model
-        elif path.suffix in [".json", ".txt", ".model"]:
+        elif path.suffix in [".json", ".txt"]:
             result.add(path)
 
         elif variant in [
@@ -143,23 +140,6 @@ def _filter_by_variant(files: List[Path], variant: ModelRepoVariant) -> Set[Path
             continue
 
     for candidate_list in subfolder_weights.values():
-        # Check if at least one of the files has the explicit fp16 variant.
-        at_least_one_fp16 = False
-        for candidate in candidate_list:
-            if len(candidate.path.suffixes) == 2 and candidate.path.suffixes[0] == ".fp16":
-                at_least_one_fp16 = True
-                break
-
-        if not at_least_one_fp16:
-            # If none of the candidates in this candidate_list have the explicit fp16 variant label, then this
-            # candidate_list probably doesn't adhere to the variant naming convention that we expected. In this case,
-            # we'll simply keep all the candidates. An example of a model that hits this case is
-            # `black-forest-labs/FLUX.1-schnell` (as of commit 012d2fd).
-            for candidate in candidate_list:
-                result.add(candidate.path)
-
-        # The candidate_list seems to have the expected variant naming convention. We'll select the highest scoring
-        # candidate.
         highest_score_candidate = max(candidate_list, key=lambda candidate: candidate.score)
         if highest_score_candidate:
             result.add(highest_score_candidate.path)

invokeai/backend/model_patcher.py

@@ -17,9 +17,8 @@ from invokeai.backend.lora import LoRAModelRaw
 from invokeai.backend.model_manager import AnyModel
 from invokeai.backend.model_manager.load.optimizations import skip_torch_weight_init
 from invokeai.backend.onnx.onnx_runtime import IAIOnnxRuntimeModel
-from invokeai.backend.stable_diffusion.extensions.lora import LoRAExt
 from invokeai.backend.textual_inversion import TextualInversionManager, TextualInversionModelRaw
-from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
+from invokeai.backend.util.devices import TorchDevice
 
 """
 loras = [
@@ -86,13 +85,13 @@ class ModelPatcher:
         cls,
         unet: UNet2DConditionModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
-        cached_weights: Optional[Dict[str, torch.Tensor]] = None,
+        model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
     ) -> Generator[None, None, None]:
         with cls.apply_lora(
             unet,
             loras=loras,
             prefix="lora_unet_",
-            cached_weights=cached_weights,
+            model_state_dict=model_state_dict,
         ):
             yield
 
@@ -102,9 +101,9 @@ class ModelPatcher:
         cls,
         text_encoder: CLIPTextModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
-        cached_weights: Optional[Dict[str, torch.Tensor]] = None,
+        model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
     ) -> Generator[None, None, None]:
-        with cls.apply_lora(text_encoder, loras=loras, prefix="lora_te_", cached_weights=cached_weights):
+        with cls.apply_lora(text_encoder, loras=loras, prefix="lora_te_", model_state_dict=model_state_dict):
             yield
 
     @classmethod
@@ -114,7 +113,7 @@ class ModelPatcher:
         model: AnyModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
         prefix: str,
-        cached_weights: Optional[Dict[str, torch.Tensor]] = None,
+        model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
     ) -> Generator[None, None, None]:
         """
         Apply one or more LoRAs to a model.
@@ -122,26 +121,66 @@ class ModelPatcher:
         :param model: The model to patch.
         :param loras: An iterator that returns the LoRA to patch in and its patch weight.
         :param prefix: A string prefix that precedes keys used in the LoRAs weight layers.
-        :cached_weights: Read-only copy of the model's state dict in CPU, for unpatching purposes.
+        :model_state_dict: Read-only copy of the model's state dict in CPU, for unpatching purposes.
         """
-        original_weights = OriginalWeightsStorage(cached_weights)
+        original_weights = {}
         try:
-            for lora_model, lora_weight in loras:
-                LoRAExt.patch_model(
-                    model=model,
-                    prefix=prefix,
-                    lora=lora_model,
-                    lora_weight=lora_weight,
-                    original_weights=original_weights,
-                )
-                del lora_model
+            with torch.no_grad():
+                for lora, lora_weight in loras:
+                    # assert lora.device.type == "cpu"
+                    for layer_key, layer in lora.layers.items():
+                        if not layer_key.startswith(prefix):
+                            continue
 
-            yield
+                        # TODO(ryand): A non-negligible amount of time is currently spent resolving LoRA keys. This
+                        # should be improved in the following ways:
+                        # 1. The key mapping could be more-efficiently pre-computed. This would save time every time a
+                        #    LoRA model is applied.
+                        # 2. From an API perspective, there's no reason that the `ModelPatcher` should be aware of the
+                        #    intricacies of Stable Diffusion key resolution. It should just expect the input LoRA
+                        #    weights to have valid keys.
+                        assert isinstance(model, torch.nn.Module)
+                        module_key, module = cls._resolve_lora_key(model, layer_key, prefix)
+
+                        # All of the LoRA weight calculations will be done on the same device as the module weight.
+                        # (Performance will be best if this is a CUDA device.)
+                        device = module.weight.device
+                        dtype = module.weight.dtype
+
+                        if module_key not in original_weights:
+                            if model_state_dict is not None:  # we were provided with the CPU copy of the state dict
+                                original_weights[module_key] = model_state_dict[module_key + ".weight"]
+                            else:
+                                original_weights[module_key] = module.weight.detach().to(device="cpu", copy=True)
+
+                        layer_scale = layer.alpha / layer.rank if (layer.alpha and layer.rank) else 1.0
+
+                        # We intentionally move to the target device first, then cast. Experimentally, this was found to
+                        # be significantly faster for 16-bit CPU tensors being moved to a CUDA device than doing the
+                        # same thing in a single call to '.to(...)'.
+                        layer.to(device=device)
+                        layer.to(dtype=torch.float32)
+                        # TODO(ryand): Using torch.autocast(...) over explicit casting may offer a speed benefit on CUDA
+                        # devices here. Experimentally, it was found to be very slow on CPU. More investigation needed.
+                        layer_weight = layer.get_weight(module.weight) * (lora_weight * layer_scale)
+                        layer.to(device=TorchDevice.CPU_DEVICE)
+
+                        assert isinstance(layer_weight, torch.Tensor)  # mypy thinks layer_weight is a float|Any ??!
+                        if module.weight.shape != layer_weight.shape:
+                            # TODO: debug on lycoris
+                            assert hasattr(layer_weight, "reshape")
+                            layer_weight = layer_weight.reshape(module.weight.shape)
+
+                        assert isinstance(layer_weight, torch.Tensor)  # mypy thinks layer_weight is a float|Any ??!
+                        module.weight += layer_weight.to(dtype=dtype)
+
+            yield  # wait for context manager exit
 
         finally:
+            assert hasattr(model, "get_submodule")  # mypy not picking up fact that torch.nn.Module has get_submodule()
             with torch.no_grad():
-                for param_key, weight in original_weights.get_changed_weights():
-                    model.get_parameter(param_key).copy_(weight)
+                for module_key, weight in original_weights.items():
+                    model.get_submodule(module_key).weight.copy_(weight)
 
     @classmethod
     @contextmanager

invokeai/backend/quantization/fast_quantized_diffusion_model.py

@@ -1,77 +0,0 @@
-import json
-import os
-from typing import Union
-
-from diffusers.models.model_loading_utils import load_state_dict
-from diffusers.utils import (
-    CONFIG_NAME,
-    SAFE_WEIGHTS_INDEX_NAME,
-    SAFETENSORS_WEIGHTS_NAME,
-    _get_checkpoint_shard_files,
-    is_accelerate_available,
-)
-from optimum.quanto.models import QuantizedDiffusersModel
-from optimum.quanto.models.shared_dict import ShardedStateDict
-
-from invokeai.backend.requantize import requantize
-
-
-class FastQuantizedDiffusersModel(QuantizedDiffusersModel):
-    @classmethod
-    def from_pretrained(cls, model_name_or_path: Union[str, os.PathLike]):
-        """We override the `from_pretrained()` method in order to use our custom `requantize()` implementation."""
-        if cls.base_class is None:
-            raise ValueError("The `base_class` attribute needs to be configured.")
-
-        if not is_accelerate_available():
-            raise ValueError("Reloading a quantized diffusers model requires the accelerate library.")
-        from accelerate import init_empty_weights
-
-        if os.path.isdir(model_name_or_path):
-            # Look for a quantization map
-            qmap_path = os.path.join(model_name_or_path, cls._qmap_name())
-            if not os.path.exists(qmap_path):
-                raise ValueError(f"No quantization map found in {model_name_or_path}: is this a quantized model ?")
-
-            # Look for original model config file.
-            model_config_path = os.path.join(model_name_or_path, CONFIG_NAME)
-            if not os.path.exists(model_config_path):
-                raise ValueError(f"{CONFIG_NAME} not found in {model_name_or_path}.")
-
-            with open(qmap_path, "r", encoding="utf-8") as f:
-                qmap = json.load(f)
-
-            with open(model_config_path, "r", encoding="utf-8") as f:
-                original_model_cls_name = json.load(f)["_class_name"]
-            configured_cls_name = cls.base_class.__name__
-            if configured_cls_name != original_model_cls_name:
-                raise ValueError(
-                    f"Configured base class ({configured_cls_name}) differs from what was derived from the provided configuration ({original_model_cls_name})."
-                )
-
-            # Create an empty model
-            config = cls.base_class.load_config(model_name_or_path)
-            with init_empty_weights():
-                model = cls.base_class.from_config(config)
-
-            # Look for the index of a sharded checkpoint
-            checkpoint_file = os.path.join(model_name_or_path, SAFE_WEIGHTS_INDEX_NAME)
-            if os.path.exists(checkpoint_file):
-                # Convert the checkpoint path to a list of shards
-                _, sharded_metadata = _get_checkpoint_shard_files(model_name_or_path, checkpoint_file)
-                # Create a mapping for the sharded safetensor files
-                state_dict = ShardedStateDict(model_name_or_path, sharded_metadata["weight_map"])
-            else:
-                # Look for a single checkpoint file
-                checkpoint_file = os.path.join(model_name_or_path, SAFETENSORS_WEIGHTS_NAME)
-                if not os.path.exists(checkpoint_file):
-                    raise ValueError(f"No safetensor weights found in {model_name_or_path}.")
-                # Get state_dict from model checkpoint
-                state_dict = load_state_dict(checkpoint_file)
-
-            # Requantize and load quantized weights from state_dict
-            requantize(model, state_dict=state_dict, quantization_map=qmap)
-            model.eval()
-            return cls(model)
-        else:
-            raise NotImplementedError("Reloading quantized models directly from the hub is not supported yet.")

invokeai/backend/quantization/fast_quantized_transformers_model.py

@@ -1,61 +0,0 @@
-import json
-import os
-from typing import Union
-
-from optimum.quanto.models import QuantizedTransformersModel
-from optimum.quanto.models.shared_dict import ShardedStateDict
-from transformers import AutoConfig
-from transformers.modeling_utils import get_checkpoint_shard_files, load_state_dict
-from transformers.utils import SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_NAME, is_accelerate_available
-
-from invokeai.backend.requantize import requantize
-
-
-class FastQuantizedTransformersModel(QuantizedTransformersModel):
-    @classmethod
-    def from_pretrained(cls, model_name_or_path: Union[str, os.PathLike]):
-        """We override the `from_pretrained()` method in order to use our custom `requantize()` implementation."""
-        if cls.auto_class is None:
-            raise ValueError(
-                "Quantized models cannot be reloaded using {cls}: use a specialized quantized class such as QuantizedModelForCausalLM instead."
-            )
-        if not is_accelerate_available():
-            raise ValueError("Reloading a quantized transformers model requires the accelerate library.")
-        from accelerate import init_empty_weights
-
-        if os.path.isdir(model_name_or_path):
-            # Look for a quantization map
-            qmap_path = os.path.join(model_name_or_path, cls._qmap_name())
-            if not os.path.exists(qmap_path):
-                raise ValueError(f"No quantization map found in {model_name_or_path}: is this a quantized model ?")
-            with open(qmap_path, "r", encoding="utf-8") as f:
-                qmap = json.load(f)
-            # Create an empty model
-            config = AutoConfig.from_pretrained(model_name_or_path)
-            with init_empty_weights():
-                model = cls.auto_class.from_config(config)
-            # Look for the index of a sharded checkpoint
-            checkpoint_file = os.path.join(model_name_or_path, SAFE_WEIGHTS_INDEX_NAME)
-            if os.path.exists(checkpoint_file):
-                # Convert the checkpoint path to a list of shards
-                checkpoint_file, sharded_metadata = get_checkpoint_shard_files(model_name_or_path, checkpoint_file)
-                # Create a mapping for the sharded safetensor files
-                state_dict = ShardedStateDict(model_name_or_path, sharded_metadata["weight_map"])
-            else:
-                # Look for a single checkpoint file
-                checkpoint_file = os.path.join(model_name_or_path, SAFE_WEIGHTS_NAME)
-                if not os.path.exists(checkpoint_file):
-                    raise ValueError(f"No safetensor weights found in {model_name_or_path}.")
-                # Get state_dict from model checkpoint
-                state_dict = load_state_dict(checkpoint_file)
-            # Requantize and load quantized weights from state_dict
-            requantize(model, state_dict=state_dict, quantization_map=qmap)
-            if getattr(model.config, "tie_word_embeddings", True):
-                # Tie output weight embeddings to input weight embeddings
-                # Note that if they were quantized they would NOT be tied
-                model.tie_weights()
-            # Set model in evaluation mode as it is done in transformers
-            model.eval()
-            return cls(model)
-        else:
-            raise NotImplementedError("Reloading quantized models directly from the hub is not supported yet.")

invokeai/backend/requantize.py

@@ -1,53 +0,0 @@
-from typing import Any, Dict
-
-import torch
-from optimum.quanto.quantize import _quantize_submodule
-
-# def custom_freeze(model: torch.nn.Module):
-#     for name, m in model.named_modules():
-#         if isinstance(m, QModuleMixin):
-#             m.weight =
-#             m.freeze()
-
-
-def requantize(
-    model: torch.nn.Module,
-    state_dict: Dict[str, Any],
-    quantization_map: Dict[str, Dict[str, str]],
-    device: torch.device = None,
-):
-    if device is None:
-        device = next(model.parameters()).device
-        if device.type == "meta":
-            device = torch.device("cpu")
-
-    # Quantize the model with parameters from the quantization map
-    for name, m in model.named_modules():
-        qconfig = quantization_map.get(name, None)
-        if qconfig is not None:
-            weights = qconfig["weights"]
-            if weights == "none":
-                weights = None
-            activations = qconfig["activations"]
-            if activations == "none":
-                activations = None
-            _quantize_submodule(model, name, m, weights=weights, activations=activations)
-
-    # Move model parameters and buffers to CPU before materializing quantized weights
-    for name, m in model.named_modules():
-
-        def move_tensor(t, device):
-            if t.device.type == "meta":
-                return torch.empty_like(t, device=device)
-            return t.to(device)
-
-        for name, param in m.named_parameters(recurse=False):
-            setattr(m, name, torch.nn.Parameter(move_tensor(param, "cpu")))
-        for name, param in m.named_buffers(recurse=False):
-            setattr(m, name, move_tensor(param, "cpu"))
-    # Freeze model and move to target device
-    # freeze(model)
-    # model.to(device)
-
-    # Load the quantized model weights
-    model.load_state_dict(state_dict, strict=False)

invokeai/backend/stable_diffusion/__init__.py

@@ -7,9 +7,11 @@ from invokeai.backend.stable_diffusion.diffusers_pipeline import (  # noqa: F401
     StableDiffusionGeneratorPipeline,
 )
 from invokeai.backend.stable_diffusion.diffusion import InvokeAIDiffuserComponent  # noqa: F401
+from invokeai.backend.stable_diffusion.seamless import set_seamless  # noqa: F401
 
 __all__ = [
     "PipelineIntermediateState",
     "StableDiffusionGeneratorPipeline",
     "InvokeAIDiffuserComponent",
+    "set_seamless",
 ]

invokeai/backend/stable_diffusion/extensions/base.py

@@ -2,14 +2,14 @@ from __future__ import annotations
 
 from contextlib import contextmanager
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Callable, Dict, List
+from typing import TYPE_CHECKING, Callable, Dict, List, Optional
 
+import torch
 from diffusers import UNet2DConditionModel
 
 if TYPE_CHECKING:
     from invokeai.backend.stable_diffusion.denoise_context import DenoiseContext
     from invokeai.backend.stable_diffusion.extension_callback_type import ExtensionCallbackType
-    from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
 
 
 @dataclass
@@ -56,17 +56,5 @@ class ExtensionBase:
         yield None
 
     @contextmanager
-    def patch_unet(self, unet: UNet2DConditionModel, original_weights: OriginalWeightsStorage):
-        """A context manager for applying patches to the UNet model. The context manager's lifetime spans the entire
-        diffusion process. Weight unpatching is handled upstream, and is achieved by saving unchanged weights by
-        `original_weights.save` function. Note that this enables some performance optimization by avoiding redundant
-        operations. All other patches (e.g. changes to tensor shapes, function monkey-patches, etc.) should be unpatched
-        by this context manager.
-
-        Args:
-            unet (UNet2DConditionModel): The UNet model on execution device to patch.
-            original_weights (OriginalWeightsStorage): A storage with copy of the model's original weights in CPU, for
-                unpatching purposes. Extension should save tensor which being modified in this storage, also extensions
-                can access original weights values.
-        """
-        yield
+    def patch_unet(self, unet: UNet2DConditionModel, cached_weights: Optional[Dict[str, torch.Tensor]] = None):
+        yield None

invokeai/backend/stable_diffusion/extensions/freeu.py

@@ -1,15 +1,15 @@
 from __future__ import annotations
 
 from contextlib import contextmanager
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Dict, Optional
 
+import torch
 from diffusers import UNet2DConditionModel
 
 from invokeai.backend.stable_diffusion.extensions.base import ExtensionBase
 
 if TYPE_CHECKING:
     from invokeai.app.shared.models import FreeUConfig
-    from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
 
 
 class FreeUExt(ExtensionBase):
@@ -21,7 +21,7 @@ class FreeUExt(ExtensionBase):
         self._freeu_config = freeu_config
 
     @contextmanager
-    def patch_unet(self, unet: UNet2DConditionModel, original_weights: OriginalWeightsStorage):
+    def patch_unet(self, unet: UNet2DConditionModel, cached_weights: Optional[Dict[str, torch.Tensor]] = None):
         unet.enable_freeu(
             b1=self._freeu_config.b1,
             b2=self._freeu_config.b2,

invokeai/backend/stable_diffusion/extensions/inpaint.py

@@ -1,120 +0,0 @@
-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):
-    """An extension for inpainting with non-inpainting models. See `InpaintModelExt` for inpainting with inpainting
-    models.
-    """
-
-    def __init__(
-        self,
-        mask: torch.Tensor,
-        is_gradient_mask: bool,
-    ):
-        """Initialize InpaintExt.
-        Args:
-            mask (torch.Tensor): The inpainting mask. Shape: (1, 1, latent_height, latent_width). Values are
-                expected to be in the range [0, 1]. A value of 1 means that the corresponding 'pixel' should not be
-                inpainted.
-            is_gradient_mask (bool): If True, mask is interpreted as a gradient mask meaning that the mask values range
-                from 0 to 1. If False, mask is interpreted as binary mask meaning that the mask values are either 0 or
-                1.
-        """
-        super().__init__()
-        self._mask = mask
-        self._is_gradient_mask = is_gradient_mask
-
-        # Noise, which used to noisify unmasked part of image
-        # if noise provided to context, then it will be used
-        # if no noise provided, then noise will be generated based on seed
-        self._noise: Optional[torch.Tensor] = None
-
-    @staticmethod
-    def _is_normal_model(unet: UNet2DConditionModel):
-        """Checks if the provided UNet belongs to a regular model.
-        The `in_channels` of a UNet vary depending on model type:
-        - normal - 4
-        - depth - 5
-        - inpaint - 9
-        """
-        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:
-            threshold = (t.item()) / ctx.scheduler.config.num_train_timesteps
-            mask_bool = mask < 1 - threshold
-            masked_input = torch.where(mask_bool, latents, mask_latents)
-        else:
-            masked_input = torch.lerp(latents, mask_latents.to(dtype=latents.dtype), 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 ValueError(
-                "InpaintExt should be used only on normal (non-inpainting) models. This could be caused by an "
-                "inpainting model that was incorrectly marked as a non-inpainting model. In some cases, this can be "
-                "fixed by removing and re-adding the model (so that it gets re-probed)."
-            )
-
-        self._mask = self._mask.to(device=ctx.latents.device, dtype=ctx.latents.dtype)
-
-        self._noise = ctx.inputs.noise
-        # '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 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)
-
-    # Use negative order to make extensions with default order work with patched latents
-    @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: redo this with preview events rewrite
-    # Use negative order to make extensions with default order work with patched latents
-    @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)
-
-    # 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 < 1, ctx.latents, ctx.inputs.orig_latents)
-        else:
-            ctx.latents = torch.lerp(ctx.latents, ctx.inputs.orig_latents, self._mask)

invokeai/backend/stable_diffusion/extensions/inpaint_model.py

@@ -1,88 +0,0 @@
-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):
-    """An extension for inpainting with inpainting models. See `InpaintExt` for inpainting with non-inpainting
-    models.
-    """
-
-    def __init__(
-        self,
-        mask: Optional[torch.Tensor],
-        masked_latents: Optional[torch.Tensor],
-        is_gradient_mask: bool,
-    ):
-        """Initialize InpaintModelExt.
-        Args:
-            mask (Optional[torch.Tensor]): The inpainting mask. Shape: (1, 1, latent_height, latent_width). Values are
-                expected to be in the range [0, 1]. A value of 1 means that the corresponding 'pixel' should not be
-                inpainted.
-            masked_latents (Optional[torch.Tensor]): Latents of initial image, with masked out by black color inpainted area.
-                If mask provided, then too should be provided. Shape: (1, 1, latent_height, latent_width)
-            is_gradient_mask (bool): If True, mask is interpreted as a gradient mask meaning that the mask values range
-                from 0 to 1. If False, mask is interpreted as binary mask meaning that the mask values are either 0 or
-                1.
-        """
-        super().__init__()
-        if mask is not None and masked_latents is None:
-            raise ValueError("Source image required for inpaint mask when inpaint model used!")
-
-        # Inverse mask, because inpaint models treat mask as: 0 - remain same, 1 - inpaint
-        self._mask = None
-        if mask is not None:
-            self._mask = 1 - mask
-        self._masked_latents = masked_latents
-        self._is_gradient_mask = is_gradient_mask
-
-    @staticmethod
-    def _is_inpaint_model(unet: UNet2DConditionModel):
-        """Checks if the provided UNet belongs to a regular model.
-        The `in_channels` of a UNet vary depending on model type:
-        - normal - 4
-        - depth - 5
-        - inpaint - 9
-        """
-        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 ValueError("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)
-
-    # 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,
-        )
-
-    # 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)

invokeai/backend/stable_diffusion/extensions/lora.py

@@ -1,137 +0,0 @@
-from __future__ import annotations
-
-from contextlib import contextmanager
-from typing import TYPE_CHECKING, Tuple
-
-import torch
-from diffusers import UNet2DConditionModel
-
-from invokeai.backend.stable_diffusion.extensions.base import ExtensionBase
-from invokeai.backend.util.devices import TorchDevice
-
-if TYPE_CHECKING:
-    from invokeai.app.invocations.model import ModelIdentifierField
-    from invokeai.app.services.shared.invocation_context import InvocationContext
-    from invokeai.backend.lora import LoRAModelRaw
-    from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
-
-
-class LoRAExt(ExtensionBase):
-    def __init__(
-        self,
-        node_context: InvocationContext,
-        model_id: ModelIdentifierField,
-        weight: float,
-    ):
-        super().__init__()
-        self._node_context = node_context
-        self._model_id = model_id
-        self._weight = weight
-
-    @contextmanager
-    def patch_unet(self, unet: UNet2DConditionModel, original_weights: OriginalWeightsStorage):
-        lora_model = self._node_context.models.load(self._model_id).model
-        self.patch_model(
-            model=unet,
-            prefix="lora_unet_",
-            lora=lora_model,
-            lora_weight=self._weight,
-            original_weights=original_weights,
-        )
-        del lora_model
-
-        yield
-
-    @classmethod
-    @torch.no_grad()
-    def patch_model(
-        cls,
-        model: torch.nn.Module,
-        prefix: str,
-        lora: LoRAModelRaw,
-        lora_weight: float,
-        original_weights: OriginalWeightsStorage,
-    ):
-        """
-        Apply one or more LoRAs to a model.
-        :param model: The model to patch.
-        :param lora: LoRA model to patch in.
-        :param lora_weight: LoRA patch weight.
-        :param prefix: A string prefix that precedes keys used in the LoRAs weight layers.
-        :param original_weights: Storage with original weights, filled by weights which lora patches, used for unpatching.
-        """
-
-        if lora_weight == 0:
-            return
-
-        # assert lora.device.type == "cpu"
-        for layer_key, layer in lora.layers.items():
-            if not layer_key.startswith(prefix):
-                continue
-
-            # TODO(ryand): A non-negligible amount of time is currently spent resolving LoRA keys. This
-            # should be improved in the following ways:
-            # 1. The key mapping could be more-efficiently pre-computed. This would save time every time a
-            #    LoRA model is applied.
-            # 2. From an API perspective, there's no reason that the `ModelPatcher` should be aware of the
-            #    intricacies of Stable Diffusion key resolution. It should just expect the input LoRA
-            #    weights to have valid keys.
-            assert isinstance(model, torch.nn.Module)
-            module_key, module = cls._resolve_lora_key(model, layer_key, prefix)
-
-            # All of the LoRA weight calculations will be done on the same device as the module weight.
-            # (Performance will be best if this is a CUDA device.)
-            device = module.weight.device
-            dtype = module.weight.dtype
-
-            layer_scale = layer.alpha / layer.rank if (layer.alpha and layer.rank) else 1.0
-
-            # We intentionally move to the target device first, then cast. Experimentally, this was found to
-            # be significantly faster for 16-bit CPU tensors being moved to a CUDA device than doing the
-            # same thing in a single call to '.to(...)'.
-            layer.to(device=device)
-            layer.to(dtype=torch.float32)
-
-            # TODO(ryand): Using torch.autocast(...) over explicit casting may offer a speed benefit on CUDA
-            # devices here. Experimentally, it was found to be very slow on CPU. More investigation needed.
-            for param_name, lora_param_weight in layer.get_parameters(module).items():
-                param_key = module_key + "." + param_name
-                module_param = module.get_parameter(param_name)
-
-                # save original weight
-                original_weights.save(param_key, module_param)
-
-                if module_param.shape != lora_param_weight.shape:
-                    # TODO: debug on lycoris
-                    lora_param_weight = lora_param_weight.reshape(module_param.shape)
-
-                lora_param_weight *= lora_weight * layer_scale
-                module_param += lora_param_weight.to(dtype=dtype)
-
-            layer.to(device=TorchDevice.CPU_DEVICE)
-
-    @staticmethod
-    def _resolve_lora_key(model: torch.nn.Module, lora_key: str, prefix: str) -> Tuple[str, torch.nn.Module]:
-        assert "." not in lora_key
-
-        if not lora_key.startswith(prefix):
-            raise Exception(f"lora_key with invalid prefix: {lora_key}, {prefix}")
-
-        module = model
-        module_key = ""
-        key_parts = lora_key[len(prefix) :].split("_")
-
-        submodule_name = key_parts.pop(0)
-
-        while len(key_parts) > 0:
-            try:
-                module = module.get_submodule(submodule_name)
-                module_key += "." + submodule_name
-                submodule_name = key_parts.pop(0)
-            except Exception:
-                submodule_name += "_" + key_parts.pop(0)
-
-        module = module.get_submodule(submodule_name)
-        module_key = (module_key + "." + submodule_name).lstrip(".")
-
-        return (module_key, module)

invokeai/backend/stable_diffusion/extensions/seamless.py

@@ -1,71 +0,0 @@
-from __future__ import annotations
-
-from contextlib import contextmanager
-from typing import Callable, Dict, List, Optional, Tuple
-
-import torch
-import torch.nn as nn
-from diffusers import UNet2DConditionModel
-from diffusers.models.lora import LoRACompatibleConv
-
-from invokeai.backend.stable_diffusion.extensions.base import ExtensionBase
-
-
-class SeamlessExt(ExtensionBase):
-    def __init__(
-        self,
-        seamless_axes: List[str],
-    ):
-        super().__init__()
-        self._seamless_axes = seamless_axes
-
-    @contextmanager
-    def patch_unet(self, unet: UNet2DConditionModel, cached_weights: Optional[Dict[str, torch.Tensor]] = None):
-        with self.static_patch_model(
-            model=unet,
-            seamless_axes=self._seamless_axes,
-        ):
-            yield
-
-    @staticmethod
-    @contextmanager
-    def static_patch_model(
-        model: torch.nn.Module,
-        seamless_axes: List[str],
-    ):
-        if not seamless_axes:
-            yield
-            return
-
-        x_mode = "circular" if "x" in seamless_axes else "constant"
-        y_mode = "circular" if "y" in seamless_axes else "constant"
-
-        # override conv_forward
-        # https://github.com/huggingface/diffusers/issues/556#issuecomment-1993287019
-        def _conv_forward_asymmetric(
-            self, input: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None
-        ):
-            self.paddingX = (self._reversed_padding_repeated_twice[0], self._reversed_padding_repeated_twice[1], 0, 0)
-            self.paddingY = (0, 0, self._reversed_padding_repeated_twice[2], self._reversed_padding_repeated_twice[3])
-            working = torch.nn.functional.pad(input, self.paddingX, mode=x_mode)
-            working = torch.nn.functional.pad(working, self.paddingY, mode=y_mode)
-            return torch.nn.functional.conv2d(
-                working, weight, bias, self.stride, torch.nn.modules.utils._pair(0), self.dilation, self.groups
-            )
-
-        original_layers: List[Tuple[nn.Conv2d, Callable]] = []
-        try:
-            for layer in model.modules():
-                if not isinstance(layer, torch.nn.Conv2d):
-                    continue
-
-                if isinstance(layer, LoRACompatibleConv) and layer.lora_layer is None:
-                    layer.lora_layer = lambda *x: 0
-                original_layers.append((layer, layer._conv_forward))
-                layer._conv_forward = _conv_forward_asymmetric.__get__(layer, torch.nn.Conv2d)
-
-            yield
-
-        finally:
-            for layer, orig_conv_forward in original_layers:
-                layer._conv_forward = orig_conv_forward

invokeai/backend/stable_diffusion/extensions/t2i_adapter.py

@@ -1,120 +0,0 @@
-from __future__ import annotations
-
-import math
-from typing import TYPE_CHECKING, List, Optional, Union
-
-import torch
-from diffusers import T2IAdapter
-from PIL.Image import Image
-
-from invokeai.app.util.controlnet_utils import prepare_control_image
-from invokeai.backend.model_manager import BaseModelType
-from invokeai.backend.stable_diffusion.diffusion.conditioning_data import ConditioningMode
-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.app.invocations.model import ModelIdentifierField
-    from invokeai.app.services.shared.invocation_context import InvocationContext
-    from invokeai.app.util.controlnet_utils import CONTROLNET_RESIZE_VALUES
-    from invokeai.backend.stable_diffusion.denoise_context import DenoiseContext
-
-
-class T2IAdapterExt(ExtensionBase):
-    def __init__(
-        self,
-        node_context: InvocationContext,
-        model_id: ModelIdentifierField,
-        image: Image,
-        weight: Union[float, List[float]],
-        begin_step_percent: float,
-        end_step_percent: float,
-        resize_mode: CONTROLNET_RESIZE_VALUES,
-    ):
-        super().__init__()
-        self._node_context = node_context
-        self._model_id = model_id
-        self._image = image
-        self._weight = weight
-        self._resize_mode = resize_mode
-        self._begin_step_percent = begin_step_percent
-        self._end_step_percent = end_step_percent
-
-        self._adapter_state: Optional[List[torch.Tensor]] = None
-
-        # The max_unet_downscale is the maximum amount that the UNet model downscales the latent image internally.
-        model_config = self._node_context.models.get_config(self._model_id.key)
-        if model_config.base == BaseModelType.StableDiffusion1:
-            self._max_unet_downscale = 8
-        elif model_config.base == BaseModelType.StableDiffusionXL:
-            self._max_unet_downscale = 4
-        else:
-            raise ValueError(f"Unexpected T2I-Adapter base model type: '{model_config.base}'.")
-
-    @callback(ExtensionCallbackType.SETUP)
-    def setup(self, ctx: DenoiseContext):
-        t2i_model: T2IAdapter
-        with self._node_context.models.load(self._model_id) as t2i_model:
-            _, _, latents_height, latents_width = ctx.inputs.orig_latents.shape
-
-            self._adapter_state = self._run_model(
-                model=t2i_model,
-                image=self._image,
-                latents_height=latents_height,
-                latents_width=latents_width,
-            )
-
-    def _run_model(
-        self,
-        model: T2IAdapter,
-        image: Image,
-        latents_height: int,
-        latents_width: int,
-    ):
-        # Resize the T2I-Adapter input image.
-        # We select the resize dimensions so that after the T2I-Adapter's total_downscale_factor is applied, the
-        # result will match the latent image's dimensions after max_unet_downscale is applied.
-        input_height = latents_height // self._max_unet_downscale * model.total_downscale_factor
-        input_width = latents_width // self._max_unet_downscale * model.total_downscale_factor
-
-        # Note: We have hard-coded `do_classifier_free_guidance=False`. This is because we only want to prepare
-        # a single image. If CFG is enabled, we will duplicate the resultant tensor after applying the
-        # T2I-Adapter model.
-        #
-        # Note: We re-use the `prepare_control_image(...)` from ControlNet for T2I-Adapter, because it has many
-        # of the same requirements (e.g. preserving binary masks during resize).
-        t2i_image = prepare_control_image(
-            image=image,
-            do_classifier_free_guidance=False,
-            width=input_width,
-            height=input_height,
-            num_channels=model.config["in_channels"],
-            device=model.device,
-            dtype=model.dtype,
-            resize_mode=self._resize_mode,
-        )
-
-        return model(t2i_image)
-
-    @callback(ExtensionCallbackType.PRE_UNET)
-    def pre_unet_step(self, ctx: DenoiseContext):
-        # skip if model not active in current step
-        total_steps = len(ctx.inputs.timesteps)
-        first_step = math.floor(self._begin_step_percent * total_steps)
-        last_step = math.ceil(self._end_step_percent * total_steps)
-        if ctx.step_index < first_step or ctx.step_index > last_step:
-            return
-
-        weight = self._weight
-        if isinstance(weight, list):
-            weight = weight[ctx.step_index]
-
-        adapter_state = self._adapter_state
-        if ctx.conditioning_mode == ConditioningMode.Both:
-            adapter_state = [torch.cat([v] * 2) for v in adapter_state]
-
-        if ctx.unet_kwargs.down_intrablock_additional_residuals is None:
-            ctx.unet_kwargs.down_intrablock_additional_residuals = [v * weight for v in adapter_state]
-        else:
-            for i, value in enumerate(adapter_state):
-                ctx.unet_kwargs.down_intrablock_additional_residuals[i] += value * weight

invokeai/backend/stable_diffusion/extensions_manager.py

@@ -7,7 +7,6 @@ import torch
 from diffusers import UNet2DConditionModel
 
 from invokeai.app.services.session_processor.session_processor_common import CanceledException
-from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
 
 if TYPE_CHECKING:
     from invokeai.backend.stable_diffusion.denoise_context import DenoiseContext
@@ -68,15 +67,9 @@ class ExtensionsManager:
         if self._is_canceled and self._is_canceled():
             raise CanceledException
 
-        original_weights = OriginalWeightsStorage(cached_weights)
-        try:
-            with ExitStack() as exit_stack:
-                for ext in self._extensions:
-                    exit_stack.enter_context(ext.patch_unet(unet, original_weights))
+        # TODO: create weight patch logic in PR with extension which uses it
+        with ExitStack() as exit_stack:
+            for ext in self._extensions:
+                exit_stack.enter_context(ext.patch_unet(unet, cached_weights))
 
-                yield None
-
-        finally:
-            with torch.no_grad():
-                for param_key, weight in original_weights.get_changed_weights():
-                    unet.get_parameter(param_key).copy_(weight)
+            yield None

invokeai/backend/stable_diffusion/schedulers/schedulers.py

@@ -20,14 +20,10 @@ from diffusers import (
 )
 from diffusers.schedulers.scheduling_utils import SchedulerMixin
 
-# TODO: add dpmpp_3s/dpmpp_3s_k when fix released
-# https://github.com/huggingface/diffusers/issues/9007
-
 SCHEDULER_NAME_VALUES = Literal[
     "ddim",
     "ddpm",
     "deis",
-    "deis_k",
     "lms",
     "lms_k",
     "pndm",
@@ -37,21 +33,16 @@ SCHEDULER_NAME_VALUES = Literal[
     "euler_k",
     "euler_a",
     "kdpm_2",
-    "kdpm_2_k",
     "kdpm_2_a",
-    "kdpm_2_a_k",
     "dpmpp_2s",
     "dpmpp_2s_k",
     "dpmpp_2m",
     "dpmpp_2m_k",
     "dpmpp_2m_sde",
     "dpmpp_2m_sde_k",
-    "dpmpp_3m",
-    "dpmpp_3m_k",
     "dpmpp_sde",
     "dpmpp_sde_k",
     "unipc",
-    "unipc_k",
     "lcm",
     "tcd",
 ]
@@ -59,8 +50,7 @@ SCHEDULER_NAME_VALUES = Literal[
 SCHEDULER_MAP: dict[SCHEDULER_NAME_VALUES, tuple[Type[SchedulerMixin], dict[str, Any]]] = {
     "ddim": (DDIMScheduler, {}),
     "ddpm": (DDPMScheduler, {}),
-    "deis": (DEISMultistepScheduler, {"use_karras_sigmas": False}),
-    "deis_k": (DEISMultistepScheduler, {"use_karras_sigmas": True}),
+    "deis": (DEISMultistepScheduler, {}),
     "lms": (LMSDiscreteScheduler, {"use_karras_sigmas": False}),
     "lms_k": (LMSDiscreteScheduler, {"use_karras_sigmas": True}),
     "pndm": (PNDMScheduler, {}),
@@ -69,28 +59,17 @@ SCHEDULER_MAP: dict[SCHEDULER_NAME_VALUES, tuple[Type[SchedulerMixin], dict[str,
     "euler": (EulerDiscreteScheduler, {"use_karras_sigmas": False}),
     "euler_k": (EulerDiscreteScheduler, {"use_karras_sigmas": True}),
     "euler_a": (EulerAncestralDiscreteScheduler, {}),
-    "kdpm_2": (KDPM2DiscreteScheduler, {"use_karras_sigmas": False}),
-    "kdpm_2_k": (KDPM2DiscreteScheduler, {"use_karras_sigmas": True}),
-    "kdpm_2_a": (KDPM2AncestralDiscreteScheduler, {"use_karras_sigmas": False}),
-    "kdpm_2_a_k": (KDPM2AncestralDiscreteScheduler, {"use_karras_sigmas": True}),
-    "dpmpp_2s": (DPMSolverSinglestepScheduler, {"use_karras_sigmas": False, "solver_order": 2}),
-    "dpmpp_2s_k": (DPMSolverSinglestepScheduler, {"use_karras_sigmas": True, "solver_order": 2}),
-    "dpmpp_2m": (DPMSolverMultistepScheduler, {"use_karras_sigmas": False, "solver_order": 2}),
-    "dpmpp_2m_k": (DPMSolverMultistepScheduler, {"use_karras_sigmas": True, "solver_order": 2}),
-    "dpmpp_2m_sde": (
-        DPMSolverMultistepScheduler,
-        {"use_karras_sigmas": False, "solver_order": 2, "algorithm_type": "sde-dpmsolver++"},
-    ),
-    "dpmpp_2m_sde_k": (
-        DPMSolverMultistepScheduler,
-        {"use_karras_sigmas": True, "solver_order": 2, "algorithm_type": "sde-dpmsolver++"},
-    ),
-    "dpmpp_3m": (DPMSolverMultistepScheduler, {"use_karras_sigmas": False, "solver_order": 3}),
-    "dpmpp_3m_k": (DPMSolverMultistepScheduler, {"use_karras_sigmas": True, "solver_order": 3}),
+    "kdpm_2": (KDPM2DiscreteScheduler, {}),
+    "kdpm_2_a": (KDPM2AncestralDiscreteScheduler, {}),
+    "dpmpp_2s": (DPMSolverSinglestepScheduler, {"use_karras_sigmas": False}),
+    "dpmpp_2s_k": (DPMSolverSinglestepScheduler, {"use_karras_sigmas": True}),
+    "dpmpp_2m": (DPMSolverMultistepScheduler, {"use_karras_sigmas": False}),
+    "dpmpp_2m_k": (DPMSolverMultistepScheduler, {"use_karras_sigmas": True}),
+    "dpmpp_2m_sde": (DPMSolverMultistepScheduler, {"use_karras_sigmas": False, "algorithm_type": "sde-dpmsolver++"}),
+    "dpmpp_2m_sde_k": (DPMSolverMultistepScheduler, {"use_karras_sigmas": True, "algorithm_type": "sde-dpmsolver++"}),
     "dpmpp_sde": (DPMSolverSDEScheduler, {"use_karras_sigmas": False, "noise_sampler_seed": 0}),
     "dpmpp_sde_k": (DPMSolverSDEScheduler, {"use_karras_sigmas": True, "noise_sampler_seed": 0}),
-    "unipc": (UniPCMultistepScheduler, {"use_karras_sigmas": False, "cpu_only": True}),
-    "unipc_k": (UniPCMultistepScheduler, {"use_karras_sigmas": True, "cpu_only": True}),
+    "unipc": (UniPCMultistepScheduler, {"cpu_only": True}),
     "lcm": (LCMScheduler, {}),
     "tcd": (TCDScheduler, {}),
 }

invokeai/backend/stable_diffusion/seamless.py

@@ -0,0 +1,51 @@
+from contextlib import contextmanager
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from diffusers.models.autoencoders.autoencoder_kl import AutoencoderKL
+from diffusers.models.autoencoders.autoencoder_tiny import AutoencoderTiny
+from diffusers.models.lora import LoRACompatibleConv
+from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
+
+
+@contextmanager
+def set_seamless(model: Union[UNet2DConditionModel, AutoencoderKL, AutoencoderTiny], seamless_axes: List[str]):
+    if not seamless_axes:
+        yield
+        return
+
+    # override conv_forward
+    # https://github.com/huggingface/diffusers/issues/556#issuecomment-1993287019
+    def _conv_forward_asymmetric(self, input: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None):
+        self.paddingX = (self._reversed_padding_repeated_twice[0], self._reversed_padding_repeated_twice[1], 0, 0)
+        self.paddingY = (0, 0, self._reversed_padding_repeated_twice[2], self._reversed_padding_repeated_twice[3])
+        working = torch.nn.functional.pad(input, self.paddingX, mode=x_mode)
+        working = torch.nn.functional.pad(working, self.paddingY, mode=y_mode)
+        return torch.nn.functional.conv2d(
+            working, weight, bias, self.stride, torch.nn.modules.utils._pair(0), self.dilation, self.groups
+        )
+
+    original_layers: List[Tuple[nn.Conv2d, Callable]] = []
+
+    try:
+        x_mode = "circular" if "x" in seamless_axes else "constant"
+        y_mode = "circular" if "y" in seamless_axes else "constant"
+
+        conv_layers: List[torch.nn.Conv2d] = []
+
+        for module in model.modules():
+            if isinstance(module, torch.nn.Conv2d):
+                conv_layers.append(module)
+
+        for layer in conv_layers:
+            if isinstance(layer, LoRACompatibleConv) and layer.lora_layer is None:
+                layer.lora_layer = lambda *x: 0
+            original_layers.append((layer, layer._conv_forward))
+            layer._conv_forward = _conv_forward_asymmetric.__get__(layer, torch.nn.Conv2d)
+
+        yield
+
+    finally:
+        for layer, orig_conv_forward in original_layers:
+            layer._conv_forward = orig_conv_forward

invokeai/backend/util/hotfixes.py

@@ -3,7 +3,7 @@ from typing import Any, Dict, List, Optional, Tuple, Union
 import diffusers
 import torch
 from diffusers.configuration_utils import ConfigMixin, register_to_config
-from diffusers.loaders.single_file_model import FromOriginalModelMixin
+from diffusers.loaders import FromOriginalControlNetMixin
 from diffusers.models.attention_processor import AttentionProcessor, AttnProcessor
 from diffusers.models.controlnet import ControlNetConditioningEmbedding, ControlNetOutput, zero_module
 from diffusers.models.embeddings import (
@@ -32,7 +32,7 @@ from invokeai.backend.util.logging import InvokeAILogger
 logger = InvokeAILogger.get_logger(__name__)
 
 
-class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlNetMixin):
     """
     A ControlNet model.
 

invokeai/backend/util/original_weights_storage.py

@@ -1,39 +0,0 @@
-from __future__ import annotations
-
-from typing import Dict, Iterator, Optional, Tuple
-
-import torch
-
-from invokeai.backend.util.devices import TorchDevice
-
-
-class OriginalWeightsStorage:
-    """A class for tracking the original weights of a model for patch/unpatch operations."""
-
-    def __init__(self, cached_weights: Optional[Dict[str, torch.Tensor]] = None):
-        # The original weights of the model.
-        self._weights: dict[str, torch.Tensor] = {}
-        # The keys of the weights that have been changed (via `save()`) during the lifetime of this instance.
-        self._changed_weights: set[str] = set()
-        if cached_weights:
-            self._weights.update(cached_weights)
-
-    def save(self, key: str, weight: torch.Tensor, copy: bool = True):
-        self._changed_weights.add(key)
-        if key in self._weights:
-            return
-
-        self._weights[key] = weight.detach().to(device=TorchDevice.CPU_DEVICE, copy=copy)
-
-    def get(self, key: str, copy: bool = False) -> Optional[torch.Tensor]:
-        weight = self._weights.get(key, None)
-        if weight is not None and copy:
-            weight = weight.clone()
-        return weight
-
-    def contains(self, key: str) -> bool:
-        return key in self._weights
-
-    def get_changed_weights(self) -> Iterator[Tuple[str, torch.Tensor]]:
-        for key in self._changed_weights:
-            yield key, self._weights[key]

invokeai/backend/vto_workflow/clipseg.py

@@ -0,0 +1,62 @@
+import torch
+from PIL import Image
+from transformers import AutoProcessor, CLIPSegForImageSegmentation, CLIPSegProcessor
+
+
+def load_clipseg_model() -> tuple[CLIPSegProcessor, CLIPSegForImageSegmentation]:
+    # Load the model.
+    clipseg_processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
+    clipseg_model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
+    return clipseg_processor, clipseg_model
+
+
+def run_clipseg(
+    images: list[Image.Image],
+    prompt: str,
+    clipseg_processor,
+    clipseg_model,
+    clipseg_temp: float,
+    device: torch.device,
+) -> list[Image.Image]:
+    """Run ClipSeg on a list of images.
+
+    Args:
+        clipseg_temp (float): Temperature applied to the CLIPSeg logits. Higher values cause the mask to be 'smoother'
+            and include more of the background. Recommended range: 0.5 to 1.0.
+    """
+
+    orig_image_sizes = [img.size for img in images]
+
+    prompts = [prompt] * len(images)
+    # TODO(ryand): Should we run the same image with and without the prompt to normalize for any bias in the model?
+    inputs = clipseg_processor(text=prompts, images=images, padding=True, return_tensors="pt")
+
+    # Move inputs and clipseg_model to the correct device and dtype.
+    inputs = {k: v.to(device=device) for k, v in inputs.items()}
+    clipseg_model = clipseg_model.to(device=device)
+
+    outputs = clipseg_model(**inputs)
+
+    logits = outputs.logits
+    if logits.ndim == 2:
+        # The model squeezes the batch dimension if it's 1, so we need to unsqueeze it.
+        logits = logits.unsqueeze(0)
+    probs = torch.nn.functional.sigmoid(logits / clipseg_temp)
+    # Normalize each mask to 0-255. Note that each mask is normalized independently.
+    probs = 255 * probs / probs.amax(dim=(1, 2), keepdim=True)
+
+    # Make mask greyscale.
+    masks: list[Image.Image] = []
+    for prob, orig_size in zip(probs, orig_image_sizes, strict=True):
+        mask = Image.fromarray(prob.cpu().detach().numpy()).convert("L")
+        mask = mask.resize(orig_size)
+        masks.append(mask)
+
+    return masks
+
+
+def select_device() -> torch.device:
+    if torch.cuda.is_available():
+        return torch.device("cuda")
+
+    return torch.device("cpu")

invokeai/backend/vto_workflow/extract_channel.py

@@ -0,0 +1,57 @@
+from enum import Enum
+
+import cv2
+import numpy as np
+import numpy.typing as npt
+
+
+class ImageChannel(Enum):
+    RGB_R = "RGB_R"
+    RGB_G = "RGB_G"
+    RGB_B = "RGB_B"
+
+    LAB_L = "LAB_L"
+    LAB_A = "LAB_A"
+    LAB_B = "LAB_B"
+
+    HSV_H = "HSV_H"
+    HSV_S = "HSV_S"
+    HSV_V = "HSV_V"
+
+
+def extract_channel(image: npt.NDArray[np.uint8], channel: ImageChannel) -> npt.NDArray[np.uint8]:
+    """Extract a channel from an image.
+
+    Args:
+        image (np.ndarray): Shape (H, W, 3) of dtype uint8.
+        channel (ImageChannel): The channel to extract.
+
+    Returns:
+        np.ndarray: Shape (H, W) of dtype uint8.
+    """
+    if channel == ImageChannel.RGB_R:
+        return image[:, :, 0]
+    elif channel == ImageChannel.RGB_G:
+        return image[:, :, 1]
+    elif channel == ImageChannel.RGB_B:
+        return image[:, :, 2]
+    elif channel == ImageChannel.LAB_L:
+        lab = cv2.cvtColor(image, cv2.COLOR_RGB2LAB)
+        return lab[:, :, 0]
+    elif channel == ImageChannel.LAB_A:
+        lab = cv2.cvtColor(image, cv2.COLOR_RGB2LAB)
+        return lab[:, :, 1]
+    elif channel == ImageChannel.LAB_B:
+        lab = cv2.cvtColor(image, cv2.COLOR_RGB2LAB)
+        return lab[:, :, 2]
+    elif channel == ImageChannel.HSV_H:
+        hsv = cv2.cvtColor(image, cv2.COLOR_RGB2HSV)
+        return hsv[:, :, 0]
+    elif channel == ImageChannel.HSV_S:
+        hsv = cv2.cvtColor(image, cv2.COLOR_RGB2HSV)
+        return hsv[:, :, 1]
+    elif channel == ImageChannel.HSV_V:
+        hsv = cv2.cvtColor(image, cv2.COLOR_RGB2HSV)
+        return hsv[:, :, 2]
+    else:
+        raise ValueError(f"Unknown channel: {channel}")

invokeai/backend/vto_workflow/overlay_pattern.py

@@ -0,0 +1,71 @@
+import numpy as np
+import torch
+from PIL import Image
+
+from invokeai.backend.vto_workflow.clipseg import load_clipseg_model, run_clipseg
+
+
+@torch.no_grad()
+def generate_dress_mask(model_image):
+    """Return a mask of the dress in the image.
+
+    Returns:
+        np.ndarray: Shape (H, W) of dtype bool. True where the dress is, False elsewhere.
+    """
+    clipseg_processor, clipseg_model = load_clipseg_model()
+
+    masks = run_clipseg(
+        images=[model_image],
+        prompt="a dress",
+        clipseg_processor=clipseg_processor,
+        clipseg_model=clipseg_model,
+        clipseg_temp=1.0,
+        device=torch.device("cuda"),
+    )
+
+    mask_np = np.array(masks[0])
+    thresh = 128
+    binary_mask = mask_np > thresh
+    return binary_mask
+
+
+def multiply_images(image_1: Image.Image, image_2: Image.Image) -> Image.Image:
+    """Multiply two images together.
+
+    Args:
+        image_1 (Image.Image): The first image.
+        image_2 (Image.Image): The second image.
+
+    Returns:
+        Image.Image: The product of the two images.
+    """
+    image_1_np = np.array(image_1, dtype=np.float32)
+    if image_1_np.ndim == 2:
+        # If the image is greyscale, add a channel dimension.
+        image_1_np = np.expand_dims(image_1_np, axis=-1)
+    image_2_np = np.array(image_2, dtype=np.float32)
+    if image_2_np.ndim == 2:
+        # If the image is greyscale, add a channel dimension.
+        image_2_np = np.expand_dims(image_2_np, axis=-1)
+    product_np = image_1_np * image_2_np // 255
+    product_np = np.clip(product_np, 0, 255).astype(np.uint8)
+    product = Image.fromarray(product_np)
+    return product
+
+
+@torch.inference_mode()
+def main():
+    # Load the model image.
+    model_image = Image.open("/home/ryan/src/InvokeAI/invokeai/backend/vto_workflow/dress.jpeg")
+
+    # Load the pattern image.
+    pattern_image = Image.open("/home/ryan/src/InvokeAI/invokeai/backend/vto_workflow/pattern1.jpg")
+
+    # Generate a mask for the dress.
+    mask = generate_dress_mask(model_image)
+
+    print("hi")
+
+
+if __name__ == "__main__":
+    main()

invokeai/backend/vto_workflow/seamless_mapping.py

@@ -0,0 +1,31 @@
+import math
+
+import numpy as np
+from PIL import Image
+
+
+def map_seamless_tiles(seamless_tile: Image.Image, target_hw: tuple[int, int], num_repeats_h: float) -> Image.Image:
+    """Map seamless tiles to a target size with a given number of repeats along the height dimension."""
+    # TODO(ryand): Add option to flip odd rows and columns if the tile is not seamless.
+    # - May also want the option to decide on a per-axis basis.
+
+    target_h, target_w = target_hw
+
+    # Calculate the height of the tile that is necessary to achieve the desired number of repeats.
+    # Take the ceiling so that the last tile overflows the target height.
+    target_tile_h = math.ceil(target_h / num_repeats_h)
+
+    # Resize the tile to the target height.
+    # Determine the target_tile_w that preserves the original aspect ratio.
+    target_tile_w = int(target_tile_h / seamless_tile.height * seamless_tile.width)
+    resized_tile = seamless_tile.resize((target_tile_w, target_tile_h))
+
+    # Repeat the tile along the height and width dimensions.
+    num_repeats_h_int = math.ceil(num_repeats_h)
+    num_repeats_w_int = math.ceil(target_w / target_tile_w)
+    seamless_tiles_np = np.array(resized_tile)
+    repeated_tiles_np = np.tile(seamless_tiles_np, (num_repeats_h_int, num_repeats_w_int, 1))
+
+    # Crop the repeated tiles to the target size.
+    output_pattern = Image.fromarray(repeated_tiles_np[:target_h, :target_w])
+    return output_pattern

invokeai/frontend/web/public/locales/de.json

@@ -91,8 +91,7 @@
         "viewingDesc": "Bilder in großer Galerie ansehen",
         "tab": "Tabulator",
         "enabled": "Aktiviert",
-        "disabled": "Ausgeschaltet",
-        "dontShowMeThese": "Zeig mir diese nicht"
+        "disabled": "Ausgeschaltet"
     },
     "gallery": {
         "galleryImageSize": "Bildgröße",
@@ -107,6 +106,7 @@
         "download": "Runterladen",
         "setCurrentImage": "Setze aktuelle Bild",
         "featuresWillReset": "Wenn Sie dieses Bild löschen, werden diese Funktionen sofort zurückgesetzt.",
+        "deleteImageBin": "Gelöschte Bilder werden an den Papierkorb Ihres Betriebssystems gesendet.",
         "unableToLoad": "Galerie kann nicht geladen werden",
         "downloadSelection": "Auswahl herunterladen",
         "currentlyInUse": "Dieses Bild wird derzeit in den folgenden Funktionen verwendet:",
@@ -628,10 +628,7 @@
         "private": "Private Ordner",
         "shared": "Geteilte Ordner",
         "archiveBoard": "Ordner archivieren",
-        "archived": "Archiviert",
-        "noBoards": "Kein {boardType}} Ordner",
-        "hideBoards": "Ordner verstecken",
-        "viewBoards": "Ordner ansehen"
+        "archived": "Archiviert"
     },
     "controlnet": {
         "showAdvanced": "Zeige Erweitert",
@@ -946,21 +943,6 @@
             "paragraphs": [
                 "Reduziert das Ausgangsbild auf die Breite und Höhe des Ausgangsbildes. Empfohlen zu aktivieren."
             ]
-        },
-        "structure": {
-            "paragraphs": [
-                "Die Struktur steuert, wie genau sich das Ausgabebild an das Layout des Originals hält. Eine niedrige Struktur erlaubt größere Änderungen, während eine hohe Struktur die ursprüngliche Komposition und das Layout strikter beibehält."
-            ]
-        },
-        "creativity": {
-            "paragraphs": [
-                "Die Kreativität bestimmt den Grad der Freiheit, die dem Modell beim Hinzufügen von Details gewährt wird. Eine niedrige Kreativität hält sich eng an das Originalbild, während eine hohe Kreativität mehr Veränderungen zulässt. Bei der Verwendung eines Prompts erhöht eine hohe Kreativität den Einfluss des Prompts."
-            ]
-        },
-        "scale": {
-            "paragraphs": [
-                "Die Skalierung steuert die Größe des Ausgabebildes und basiert auf einem Vielfachen der Auflösung des Originalbildes. So würde z. B. eine 2-fache Hochskalierung eines 1024x1024px Bildes eine 2048x2048px große Ausgabe erzeugen."
-            ]
         }
     },
     "invocationCache": {

invokeai/frontend/web/public/locales/en.json

@@ -31,8 +31,7 @@
         "deleteBoard": "Delete Board",
         "deleteBoardAndImages": "Delete Board and Images",
         "deleteBoardOnly": "Delete Board Only",
-        "deletedBoardsCannotbeRestored": "Deleted boards cannot be restored. Selecting 'Delete Board Only' will move images to an uncategorized state.",
-        "deletedPrivateBoardsCannotbeRestored": "Deleted boards cannot be restored. Selecting 'Delete Board Only' will move images to a private uncategorized state for the image's creator.",
+        "deletedBoardsCannotbeRestored": "Deleted boards cannot be restored",
         "hideBoards": "Hide Boards",
         "loading": "Loading...",
         "menuItemAutoAdd": "Auto-add to this Board",
@@ -373,6 +372,7 @@
         "dropToUpload": "$t(gallery.drop) to Upload",
         "deleteImage_one": "Delete Image",
         "deleteImage_other": "Delete {{count}} Images",
+        "deleteImageBin": "Deleted images will be sent to your operating system's Bin.",
         "deleteImagePermanent": "Deleted images cannot be restored.",
         "displayBoardSearch": "Display Board Search",
         "displaySearch": "Display Search",
@@ -1052,7 +1052,11 @@
         "remixImage": "Remix Image",
         "usePrompt": "Use Prompt",
         "useSeed": "Use Seed",
-        "width": "Width"
+        "width": "Width",
+        "isAllowedToUpscale": {
+            "useX2Model": "Image is too large to upscale with x4 model, use x2 model",
+            "tooLarge": "Image is too large to upscale, select smaller image"
+        }
     },
     "dynamicPrompts": {
         "showDynamicPrompts": "Show Dynamic Prompts",
@@ -1505,30 +1509,6 @@
         "seamlessTilingYAxis": {
             "heading": "Seamless Tiling Y Axis",
             "paragraphs": ["Seamlessly tile an image along the vertical axis."]
-        },
-        "upscaleModel": {
-            "heading": "Upscale Model",
-            "paragraphs": [
-                "The upscale model scales the image to the output size before details are added. Any supported upscale model may be used, but some are specialized for different kinds of images, like photos or line drawings."
-            ]
-        },
-        "scale": {
-            "heading": "Scale",
-            "paragraphs": [
-                "Scale controls the output image size, and is based on a multiple of the input image resolution. For example a 2x upscale on a 1024x1024 image would produce a 2048 x 2048 output."
-            ]
-        },
-        "creativity": {
-            "heading": "Creativity",
-            "paragraphs": [
-                "Creativity controls the amount of freedom granted to the model when adding details. Low creativity stays close to the original image, while high creativity allows for more change. When using a prompt, high creativity increases the influence of the prompt."
-            ]
-        },
-        "structure": {
-            "heading": "Structure",
-            "paragraphs": [
-                "Structure controls how closely the output image will keep to the layout of the original. Low structure allows major changes, while high structure strictly maintains the original composition and layout."
-            ]
         }
     },
     "unifiedCanvas": {
@@ -1673,8 +1653,6 @@
     },
     "upscaling": {
         "creativity": "Creativity",
-        "exceedsMaxSize": "Upscale settings exceed max size limit",
-        "exceedsMaxSizeDetails": "Max upscale limit is {{maxUpscaleDimension}}x{{maxUpscaleDimension}} pixels. Please try a smaller image or decrease your scale selection.",
         "structure": "Structure",
         "upscaleModel": "Upscale Model",
         "postProcessingModel": "Post-Processing Model",

invokeai/frontend/web/public/locales/es.json

@@ -88,6 +88,7 @@
         "deleteImage_one": "Eliminar Imagen",
         "deleteImage_many": "",
         "deleteImage_other": "",
+        "deleteImageBin": "Las imágenes eliminadas se enviarán a la papelera de tu sistema operativo.",
         "deleteImagePermanent": "Las imágenes eliminadas no se pueden restaurar.",
         "assets": "Activos",
         "autoAssignBoardOnClick": "Asignación automática de tableros al hacer clic"