InvokeAI/invokeai/backend/ip_adapter/ip_adapter.py

# copied from https://github.com/tencent-ailab/IP-Adapter (Apache License 2.0)
#   and modified as needed

from contextlib import contextmanager
from typing import Optional

import torch
from diffusers.models import UNet2DConditionModel

# FIXME: Getting errors when trying to use PyTorch 2.0 versions of IPAttnProcessor and AttnProcessor
#   so for now falling back to the default versions
# from .utils import is_torch2_available
# if is_torch2_available:
#     from .attention_processor import IPAttnProcessor2_0 as IPAttnProcessor, AttnProcessor2_0 as AttnProcessor
# else:
#     from .attention_processor import IPAttnProcessor, AttnProcessor
from PIL import Image
from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection

from .attention_processor import AttnProcessor, IPAttnProcessor
from .resampler import Resampler


class ImageProjModel(torch.nn.Module):
    """Image Projection Model"""

    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
        super().__init__()

        self.cross_attention_dim = cross_attention_dim
        self.clip_extra_context_tokens = clip_extra_context_tokens
        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
        self.norm = torch.nn.LayerNorm(cross_attention_dim)

    def forward(self, image_embeds):
        embeds = image_embeds
        clip_extra_context_tokens = self.proj(embeds).reshape(
            -1, self.clip_extra_context_tokens, self.cross_attention_dim
        )
        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
        return clip_extra_context_tokens


class IPAdapter:
    """IP-Adapter: https://arxiv.org/pdf/2308.06721.pdf"""

    def __init__(
        self,
        ip_adapter_ckpt_path: str,
        device: torch.device,
        dtype: torch.dtype = torch.float16,
        num_tokens: int = 4,
    ):
        self.device = device
        self.dtype = dtype

        self._ip_adapter_ckpt_path = ip_adapter_ckpt_path
        self._num_tokens = num_tokens

        self._clip_image_processor = CLIPImageProcessor()

        # Fields to be initialized later in initialize().
        self._unet = None
        self._image_proj_model = None
        self._attn_processors = None

        self._state_dict = torch.load(self._ip_adapter_ckpt_path, map_location="cpu")

    def is_initialized(self):
        return self._unet is not None and self._image_proj_model is not None and self._attn_processors is not None

    def initialize(self, unet: UNet2DConditionModel, image_encoder: CLIPVisionModelWithProjection):
        """Finish the model initialization process.

        HACK: This is separate from __init__ for compatibility with the model manager. The full initialization requires
        access to the UNet model to be patched, which can not easily be passed to __init__ by the model manager.

        Args:
            unet (UNet2DConditionModel): The UNet whose attention blocks will be patched by this IP-Adapter.
        """
        if self.is_initialized():
            raise Exception("IPAdapter has already been initialized.")

        self._unet = unet
        # TODO(ryand): Eliminate the need to pass the image_encoder to initialize(). It should be possible to infer the
        # necessary information from the state_dict.
        self._image_proj_model = self._init_image_proj_model(image_encoder)
        self._attn_processors = self._prepare_attention_processors()

        # Copy the weights from the _state_dict into the models.
        self._image_proj_model.load_state_dict(self._state_dict["image_proj"])
        ip_layers = torch.nn.ModuleList(self._attn_processors.values())
        ip_layers.load_state_dict(self._state_dict["ip_adapter"])

        self._state_dict = None

    def to(self, device: torch.device, dtype: Optional[torch.dtype] = None):
        self.device = device
        if dtype is not None:
            self.dtype = dtype

        for model in [self._image_proj_model, self._attn_processors]:
            # If this is called before initialize(), then some models will still be None. We just update the non-None
            # models.
            if model is not None:
                model.to(device=self.device, dtype=self.dtype)

    def _init_image_proj_model(self, image_encoder: CLIPVisionModelWithProjection):
        image_proj_model = ImageProjModel(
            cross_attention_dim=self._unet.config.cross_attention_dim,
            clip_embeddings_dim=image_encoder.config.projection_dim,
            clip_extra_context_tokens=self._num_tokens,
        ).to(self.device, dtype=self.dtype)
        return image_proj_model

    def _prepare_attention_processors(self):
        """Creates a dict of attention processors that can later be injected into `self.unet`, and loads the IP-Adapter
        attention weights into them.
        """
        attn_procs = {}
        for name in self._unet.attn_processors.keys():
            cross_attention_dim = None if name.endswith("attn1.processor") else self._unet.config.cross_attention_dim
            if name.startswith("mid_block"):
                hidden_size = self._unet.config.block_out_channels[-1]
            elif name.startswith("up_blocks"):
                block_id = int(name[len("up_blocks.")])
                hidden_size = list(reversed(self._unet.config.block_out_channels))[block_id]
            elif name.startswith("down_blocks"):
                block_id = int(name[len("down_blocks.")])
                hidden_size = self._unet.config.block_out_channels[block_id]
            if cross_attention_dim is None:
                attn_procs[name] = AttnProcessor()
            else:
                attn_procs[name] = IPAttnProcessor(
                    hidden_size=hidden_size,
                    cross_attention_dim=cross_attention_dim,
                    scale=1.0,
                ).to(self.device, dtype=self.dtype)
        return attn_procs

    @contextmanager
    def apply_ip_adapter_attention(self):
        """A context manager that patches `self._unet` with this IP-Adapter's attention processors while it is active.

        Yields:
            None
        """
        if not self.is_initialized():
            raise Exception("Call IPAdapter.initialize() before calling IPAdapter.apply_ip_adapter_attention().")

        orig_attn_processors = self._unet.attn_processors
        # Make a (moderately-) shallow copy of the self._attn_processors dict, because set_attn_processor(...) actually
        # pops elements from the passed dict.
        ip_adapter_attn_processors = {k: v for k, v in self._attn_processors.items()}
        try:
            self._unet.set_attn_processor(ip_adapter_attn_processors)
            yield None
        finally:
            self._unet.set_attn_processor(orig_attn_processors)

    @torch.inference_mode()
    def get_image_embeds(self, pil_image, image_encoder: CLIPVisionModelWithProjection):
        if not self.is_initialized():
            raise Exception("Call IPAdapter.initialize() before calling IPAdapter.get_image_embeds().")

        if isinstance(pil_image, Image.Image):
            pil_image = [pil_image]
        clip_image = self._clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
        clip_image_embeds = image_encoder(clip_image.to(self.device, dtype=self.dtype)).image_embeds
        image_prompt_embeds = self._image_proj_model(clip_image_embeds)
        uncond_image_prompt_embeds = self._image_proj_model(torch.zeros_like(clip_image_embeds))
        return image_prompt_embeds, uncond_image_prompt_embeds

    def set_scale(self, scale):
        if not self.is_initialized():
            raise Exception("Call IPAdapter.initialize() before calling IPAdapter.set_scale().")

        for attn_processor in self._attn_processors.values():
            if isinstance(attn_processor, IPAttnProcessor):
                attn_processor.scale = scale


class IPAdapterPlus(IPAdapter):
    """IP-Adapter with fine-grained features"""

    def _init_image_proj_model(self, image_encoder: CLIPVisionModelWithProjection):
        image_proj_model = Resampler(
            dim=self._unet.config.cross_attention_dim,
            depth=4,
            dim_head=64,
            heads=12,
            num_queries=self._num_tokens,
            embedding_dim=image_encoder.config.hidden_size,
            output_dim=self._unet.config.cross_attention_dim,
            ff_mult=4,
        ).to(self.device, dtype=self.dtype)
        return image_proj_model

    @torch.inference_mode()
    def get_image_embeds(self, pil_image, image_encoder: CLIPVisionModelWithProjection):
        if not self.is_initialized():
            raise Exception("Call IPAdapter.initialize() before calling IPAdapter.get_image_embeds().")

        if isinstance(pil_image, Image.Image):
            pil_image = [pil_image]
        clip_image = self._clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
        clip_image = clip_image.to(self.device, dtype=self.dtype)
        clip_image_embeds = image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
        image_prompt_embeds = self._image_proj_model(clip_image_embeds)
        uncond_clip_image_embeds = image_encoder(torch.zeros_like(clip_image), output_hidden_states=True).hidden_states[
            -2
        ]
        uncond_image_prompt_embeds = self._image_proj_model(uncond_clip_image_embeds)
        return image_prompt_embeds, uncond_image_prompt_embeds