ip_adapter_instantstyle/ip_adapter.py

import os
import glob
from typing import List

import torch
import torch.nn as nn
from diffusers import StableDiffusionPipeline
from diffusers.pipelines.controlnet import MultiControlNetModel
from PIL import Image
from safetensors import safe_open
from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection

from .utils import is_torch2_available, get_generator

L = 4


def pos_encode(x, L):
    pos_encode = []

    for freq in range(L):
        pos_encode.append(torch.cos(2**freq * torch.pi * x))
        pos_encode.append(torch.sin(2**freq * torch.pi * x))
    pos_encode = torch.cat(pos_encode, dim=1)
    return pos_encode


if is_torch2_available():
    from .attention_processor import (
        AttnProcessor2_0 as AttnProcessor,
    )
    from .attention_processor import (
        CNAttnProcessor2_0 as CNAttnProcessor,
    )
    from .attention_processor import (
        IPAttnProcessor2_0 as IPAttnProcessor,
    )
else:
    from .attention_processor import AttnProcessor, CNAttnProcessor, IPAttnProcessor
from .resampler import Resampler


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

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

        self.generator = None
        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 MLPProjModel(torch.nn.Module):
    """SD model with image prompt"""

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

        self.proj = torch.nn.Sequential(
            torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim),
            torch.nn.GELU(),
            torch.nn.Linear(clip_embeddings_dim, cross_attention_dim),
            torch.nn.LayerNorm(cross_attention_dim),
        )

    def forward(self, image_embeds):
        clip_extra_context_tokens = self.proj(image_embeds)
        return clip_extra_context_tokens


class IPAdapter:
    def __init__(
        self,
        sd_pipe,
        image_encoder_path,
        ip_ckpt,
        device,
        num_tokens=4,
        target_blocks=["block"],
    ):
        self.device = device
        self.image_encoder_path = image_encoder_path
        self.ip_ckpt = ip_ckpt
        self.num_tokens = num_tokens
        self.target_blocks = target_blocks

        self.pipe = sd_pipe.to(self.device)
        self.set_ip_adapter()

        # load image encoder
        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(
            self.image_encoder_path
        ).to(self.device, dtype=torch.float16)
        self.clip_image_processor = CLIPImageProcessor()
        # image proj model
        self.image_proj_model = self.init_proj()

        self.load_ip_adapter()

    def init_proj(self):
        image_proj_model = ImageProjModel(
            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
            clip_embeddings_dim=self.image_encoder.config.projection_dim,
            clip_extra_context_tokens=self.num_tokens,
        ).to(self.device, dtype=torch.float16)
        return image_proj_model

    def set_ip_adapter(self):
        unet = self.pipe.unet
        attn_procs = {}
        for name in unet.attn_processors.keys():
            cross_attention_dim = (
                None
                if name.endswith("attn1.processor")
                else unet.config.cross_attention_dim
            )
            if name.startswith("mid_block"):
                hidden_size = unet.config.block_out_channels[-1]
            elif name.startswith("up_blocks"):
                block_id = int(name[len("up_blocks.")])
                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
            elif name.startswith("down_blocks"):
                block_id = int(name[len("down_blocks.")])
                hidden_size = unet.config.block_out_channels[block_id]
            if cross_attention_dim is None:
                attn_procs[name] = AttnProcessor()
            else:
                selected = False
                for block_name in self.target_blocks:
                    if block_name in name:
                        selected = True
                        break
                if selected:
                    attn_procs[name] = IPAttnProcessor(
                        hidden_size=hidden_size,
                        cross_attention_dim=cross_attention_dim,
                        scale=1.0,
                        num_tokens=self.num_tokens,
                    ).to(self.device, dtype=torch.float16)
                else:
                    attn_procs[name] = IPAttnProcessor(
                        hidden_size=hidden_size,
                        cross_attention_dim=cross_attention_dim,
                        scale=1.0,
                        num_tokens=self.num_tokens,
                        skip=True,
                    ).to(self.device, dtype=torch.float16)
        unet.set_attn_processor(attn_procs)
        if hasattr(self.pipe, "controlnet"):
            if isinstance(self.pipe.controlnet, MultiControlNetModel):
                for controlnet in self.pipe.controlnet.nets:
                    controlnet.set_attn_processor(
                        CNAttnProcessor(num_tokens=self.num_tokens)
                    )
            else:
                self.pipe.controlnet.set_attn_processor(
                    CNAttnProcessor(num_tokens=self.num_tokens)
                )

    def load_ip_adapter(self):
        if os.path.splitext(self.ip_ckpt)[-1] == ".safetensors":
            state_dict = {"image_proj": {}, "ip_adapter": {}}
            with safe_open(self.ip_ckpt, framework="pt", device="cpu") as f:
                for key in f.keys():
                    if key.startswith("image_proj."):
                        state_dict["image_proj"][key.replace("image_proj.", "")] = (
                            f.get_tensor(key)
                        )
                    elif key.startswith("ip_adapter."):
                        state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = (
                            f.get_tensor(key)
                        )
        else:
            state_dict = torch.load(self.ip_ckpt, map_location="cpu")
        self.image_proj_model.load_state_dict(state_dict["image_proj"])
        ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
        ip_layers.load_state_dict(state_dict["ip_adapter"], strict=False)

    @torch.inference_mode()
    def get_image_embeds(
        self, pil_image=None, clip_image_embeds=None, content_prompt_embeds=None
    ):
        if pil_image is not None:
            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 = self.image_encoder(
                clip_image.to(self.device, dtype=torch.float16)
            ).image_embeds
        else:
            clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)

        if content_prompt_embeds is not None:
            print(clip_image_embeds.shape)
            print(content_prompt_embeds.shape)
            clip_image_embeds = clip_image_embeds - content_prompt_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

    @torch.inference_mode()
    def generate_image_edit_dir(
        self,
        pil_image=None,
        content_prompt_embeds=None,
        edit_mlps: dict[torch.nn.Module, float] = None,
    ):
        print("Combining multiple MLPs!")
        if pil_image is not None:
            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 = self.image_encoder(
                clip_image.to(self.device, dtype=torch.float16)
            ).image_embeds
            pred_editing_dirs = [
                net(
                    clip_image_embeds,
                    torch.Tensor([strength]).to(self.device, dtype=torch.float16),
                )
                for net, strength in edit_mlps.items()
            ]

        clip_image_embeds = clip_image_embeds + sum(pred_editing_dirs)

        if content_prompt_embeds is not None:
            clip_image_embeds = clip_image_embeds - content_prompt_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

    @torch.inference_mode()
    def get_image_edit_dir(
        self,
        start_image=None,
        pil_image=None,
        pil_image2=None,
        content_prompt_embeds=None,
        edit_strength=1.0,
    ):
        print("Blending Two Materials!")
        if pil_image is not None:
            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 = self.image_encoder(
                clip_image.to(self.device, dtype=torch.float16)
            ).image_embeds

        if pil_image2 is not None:
            if isinstance(pil_image2, Image.Image):
                pil_image2 = [pil_image2]
            clip_image2 = self.clip_image_processor(
                images=pil_image2, return_tensors="pt"
            ).pixel_values
            clip_image_embeds2 = self.image_encoder(
                clip_image2.to(self.device, dtype=torch.float16)
            ).image_embeds

        if start_image is not None:
            if isinstance(start_image, Image.Image):
                start_image = [start_image]
            clip_image_start = self.clip_image_processor(
                images=start_image, return_tensors="pt"
            ).pixel_values
            clip_image_embeds_start = self.image_encoder(
                clip_image_start.to(self.device, dtype=torch.float16)
            ).image_embeds

        if content_prompt_embeds is not None:
            clip_image_embeds = clip_image_embeds - content_prompt_embeds
            clip_image_embeds2 = clip_image_embeds2 - content_prompt_embeds

        # clip_image_embeds += edit_strength * (clip_image_embeds2 - clip_image_embeds)
        clip_image_embeds = clip_image_embeds_start + edit_strength * (
            clip_image_embeds2 - clip_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):
        for attn_processor in self.pipe.unet.attn_processors.values():
            if isinstance(attn_processor, IPAttnProcessor):
                attn_processor.scale = scale

    def set_scale(self, scale):
        for attn_processor in self.pipe.unet.attn_processors.values():
            if isinstance(attn_processor, IPAttnProcessor):
                attn_processor.scale = scale

    def generate(
        self,
        pil_image=None,
        clip_image_embeds=None,
        prompt=None,
        negative_prompt=None,
        scale=1.0,
        num_samples=4,
        seed=None,
        guidance_scale=7.5,
        num_inference_steps=30,
        neg_content_emb=None,
        **kwargs,
    ):
        self.set_scale(scale)

        if pil_image is not None:
            num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
        else:
            num_prompts = clip_image_embeds.size(0)

        if prompt is None:
            prompt = "best quality, high quality"
        if negative_prompt is None:
            negative_prompt = (
                "monochrome, lowres, bad anatomy, worst quality, low quality"
            )

        if not isinstance(prompt, List):
            prompt = [prompt] * num_prompts
        if not isinstance(negative_prompt, List):
            negative_prompt = [negative_prompt] * num_prompts

        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
            pil_image=pil_image,
            clip_image_embeds=clip_image_embeds,
            content_prompt_embeds=neg_content_emb,
        )
        bs_embed, seq_len, _ = image_prompt_embeds.shape
        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
        image_prompt_embeds = image_prompt_embeds.view(
            bs_embed * num_samples, seq_len, -1
        )
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
            1, num_samples, 1
        )
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
            bs_embed * num_samples, seq_len, -1
        )

        with torch.inference_mode():
            prompt_embeds_, negative_prompt_embeds_ = self.pipe.encode_prompt(
                prompt,
                device=self.device,
                num_images_per_prompt=num_samples,
                do_classifier_free_guidance=True,
                negative_prompt=negative_prompt,
            )
            prompt_embeds = torch.cat([prompt_embeds_, image_prompt_embeds], dim=1)
            negative_prompt_embeds = torch.cat(
                [negative_prompt_embeds_, uncond_image_prompt_embeds], dim=1
            )

        generator = get_generator(seed, self.device)

        images = self.pipe(
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps,
            generator=generator,
            **kwargs,
        ).images

        return images


class IPAdapterXL(IPAdapter):
    """SDXL"""

    def generate(
        self,
        pil_image,
        prompt=None,
        negative_prompt=None,
        scale=1.0,
        num_samples=4,
        seed=None,
        num_inference_steps=30,
        neg_content_emb=None,
        neg_content_prompt=None,
        neg_content_scale=1.0,
        clip_strength=1.0,
        **kwargs,
    ):
        self.set_scale(scale)

        num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)

        if prompt is None:
            prompt = "best quality, high quality"
        if negative_prompt is None:
            negative_prompt = (
                "monochrome, lowres, bad anatomy, worst quality, low quality"
            )

        if not isinstance(prompt, List):
            prompt = [prompt] * num_prompts
        if not isinstance(negative_prompt, List):
            negative_prompt = [negative_prompt] * num_prompts

        if neg_content_emb is None:
            if neg_content_prompt is not None:
                with torch.inference_mode():
                    (
                        prompt_embeds_,  # torch.Size([1, 77, 2048])
                        negative_prompt_embeds_,
                        pooled_prompt_embeds_,  # torch.Size([1, 1280])
                        negative_pooled_prompt_embeds_,
                    ) = self.pipe.encode_prompt(
                        neg_content_prompt,
                        num_images_per_prompt=num_samples,
                        do_classifier_free_guidance=True,
                        negative_prompt=negative_prompt,
                    )
                    pooled_prompt_embeds_ *= neg_content_scale
            else:
                pooled_prompt_embeds_ = neg_content_emb
        else:
            pooled_prompt_embeds_ = None

        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
            pil_image, content_prompt_embeds=pooled_prompt_embeds_
        )
        bs_embed, seq_len, _ = image_prompt_embeds.shape
        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
        image_prompt_embeds = image_prompt_embeds.view(
            bs_embed * num_samples, seq_len, -1
        )
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
            1, num_samples, 1
        )
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
            bs_embed * num_samples, seq_len, -1
        )
        print("CLIP Strength is {}".format(clip_strength))
        image_prompt_embeds *= clip_strength
        uncond_image_prompt_embeds *= clip_strength

        with torch.inference_mode():
            (
                prompt_embeds,
                negative_prompt_embeds,
                pooled_prompt_embeds,
                negative_pooled_prompt_embeds,
            ) = self.pipe.encode_prompt(
                prompt,
                num_images_per_prompt=num_samples,
                do_classifier_free_guidance=True,
                negative_prompt=negative_prompt,
            )
            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
            negative_prompt_embeds = torch.cat(
                [negative_prompt_embeds, uncond_image_prompt_embeds], dim=1
            )

        self.generator = get_generator(seed, self.device)

        images = self.pipe(
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            num_inference_steps=num_inference_steps,
            generator=self.generator,
            **kwargs,
        ).images

        return images

    def generate_parametric_edits(
        self,
        pil_image,
        edit_mlps: dict[torch.nn.Module, float],
        prompt=None,
        negative_prompt=None,
        scale=1.0,
        num_samples=4,
        seed=None,
        num_inference_steps=30,
        neg_content_emb=None,
        neg_content_prompt=None,
        neg_content_scale=1.0,
        **kwargs,
    ):
        self.set_scale(scale)

        num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)

        if prompt is None:
            prompt = "best quality, high quality"
        if negative_prompt is None:
            negative_prompt = (
                "monochrome, lowres, bad anatomy, worst quality, low quality"
            )

        if not isinstance(prompt, List):
            prompt = [prompt] * num_prompts
        if not isinstance(negative_prompt, List):
            negative_prompt = [negative_prompt] * num_prompts

        if neg_content_emb is None:
            if neg_content_prompt is not None:
                with torch.inference_mode():
                    (
                        prompt_embeds_,  # torch.Size([1, 77, 2048])
                        negative_prompt_embeds_,
                        pooled_prompt_embeds_,  # torch.Size([1, 1280])
                        negative_pooled_prompt_embeds_,
                    ) = self.pipe.encode_prompt(
                        neg_content_prompt,
                        num_images_per_prompt=num_samples,
                        do_classifier_free_guidance=True,
                        negative_prompt=negative_prompt,
                    )
                    pooled_prompt_embeds_ *= neg_content_scale
            else:
                pooled_prompt_embeds_ = neg_content_emb
        else:
            pooled_prompt_embeds_ = None
        image_prompt_embeds, uncond_image_prompt_embeds = self.generate_image_edit_dir(
            pil_image, content_prompt_embeds=pooled_prompt_embeds_, edit_mlps=edit_mlps
        )

        bs_embed, seq_len, _ = image_prompt_embeds.shape
        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
        image_prompt_embeds = image_prompt_embeds.view(
            bs_embed * num_samples, seq_len, -1
        )
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
            1, num_samples, 1
        )
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
            bs_embed * num_samples, seq_len, -1
        )

        with torch.inference_mode():
            (
                prompt_embeds,
                negative_prompt_embeds,
                pooled_prompt_embeds,
                negative_pooled_prompt_embeds,
            ) = self.pipe.encode_prompt(
                prompt,
                num_images_per_prompt=num_samples,
                do_classifier_free_guidance=True,
                negative_prompt=negative_prompt,
            )
            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
            negative_prompt_embeds = torch.cat(
                [negative_prompt_embeds, uncond_image_prompt_embeds], dim=1
            )

        self.generator = get_generator(seed, self.device)

        images = self.pipe(
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            num_inference_steps=num_inference_steps,
            generator=self.generator,
            **kwargs,
        ).images

        return images

    def generate_edit(
        self,
        start_image,
        pil_image,
        pil_image2,
        prompt=None,
        negative_prompt=None,
        scale=1.0,
        num_samples=4,
        seed=None,
        num_inference_steps=30,
        neg_content_emb=None,
        neg_content_prompt=None,
        neg_content_scale=1.0,
        edit_strength=1.0,
        **kwargs,
    ):
        self.set_scale(scale)

        num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)

        if prompt is None:
            prompt = "best quality, high quality"
        if negative_prompt is None:
            negative_prompt = (
                "monochrome, lowres, bad anatomy, worst quality, low quality"
            )

        if not isinstance(prompt, List):
            prompt = [prompt] * num_prompts
        if not isinstance(negative_prompt, List):
            negative_prompt = [negative_prompt] * num_prompts

        if neg_content_emb is None:
            if neg_content_prompt is not None:
                with torch.inference_mode():
                    (
                        prompt_embeds_,  # torch.Size([1, 77, 2048])
                        negative_prompt_embeds_,
                        pooled_prompt_embeds_,  # torch.Size([1, 1280])
                        negative_pooled_prompt_embeds_,
                    ) = self.pipe.encode_prompt(
                        neg_content_prompt,
                        num_images_per_prompt=num_samples,
                        do_classifier_free_guidance=True,
                        negative_prompt=negative_prompt,
                    )
                    pooled_prompt_embeds_ *= neg_content_scale
            else:
                pooled_prompt_embeds_ = neg_content_emb
        else:
            pooled_prompt_embeds_ = None

        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_edit_dir(
            start_image,
            pil_image,
            pil_image2,
            content_prompt_embeds=pooled_prompt_embeds_,
            edit_strength=edit_strength,
        )

        bs_embed, seq_len, _ = image_prompt_embeds.shape
        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
        image_prompt_embeds = image_prompt_embeds.view(
            bs_embed * num_samples, seq_len, -1
        )
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
            1, num_samples, 1
        )
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
            bs_embed * num_samples, seq_len, -1
        )

        with torch.inference_mode():
            (
                prompt_embeds,
                negative_prompt_embeds,
                pooled_prompt_embeds,
                negative_pooled_prompt_embeds,
            ) = self.pipe.encode_prompt(
                prompt,
                num_images_per_prompt=num_samples,
                do_classifier_free_guidance=True,
                negative_prompt=negative_prompt,
            )
            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
            negative_prompt_embeds = torch.cat(
                [negative_prompt_embeds, uncond_image_prompt_embeds], dim=1
            )

        self.generator = get_generator(seed, self.device)

        images = self.pipe(
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            num_inference_steps=num_inference_steps,
            generator=self.generator,
            **kwargs,
        ).images

        return images


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

    def init_proj(self):
        image_proj_model = Resampler(
            dim=self.pipe.unet.config.cross_attention_dim,
            depth=4,
            dim_head=64,
            heads=12,
            num_queries=self.num_tokens,
            embedding_dim=self.image_encoder.config.hidden_size,
            output_dim=self.pipe.unet.config.cross_attention_dim,
            ff_mult=4,
        ).to(self.device, dtype=torch.float16)
        return image_proj_model

    @torch.inference_mode()
    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
        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=torch.float16)
        clip_image_embeds = self.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 = self.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


class IPAdapterFull(IPAdapterPlus):
    """IP-Adapter with full features"""

    def init_proj(self):
        image_proj_model = MLPProjModel(
            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
            clip_embeddings_dim=self.image_encoder.config.hidden_size,
        ).to(self.device, dtype=torch.float16)
        return image_proj_model


class IPAdapterPlusXL(IPAdapter):
    """SDXL"""

    def init_proj(self):
        image_proj_model = Resampler(
            dim=1280,
            depth=4,
            dim_head=64,
            heads=20,
            num_queries=self.num_tokens,
            embedding_dim=self.image_encoder.config.hidden_size,
            output_dim=self.pipe.unet.config.cross_attention_dim,
            ff_mult=4,
        ).to(self.device, dtype=torch.float16)
        return image_proj_model

    @torch.inference_mode()
    def get_image_embeds(self, pil_image):
        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=torch.float16)
        clip_image_embeds = self.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 = self.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

    def generate(
        self,
        pil_image,
        prompt=None,
        negative_prompt=None,
        scale=1.0,
        num_samples=4,
        seed=None,
        num_inference_steps=30,
        **kwargs,
    ):
        self.set_scale(scale)

        num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)

        if prompt is None:
            prompt = "best quality, high quality"
        if negative_prompt is None:
            negative_prompt = (
                "monochrome, lowres, bad anatomy, worst quality, low quality"
            )

        if not isinstance(prompt, List):
            prompt = [prompt] * num_prompts
        if not isinstance(negative_prompt, List):
            negative_prompt = [negative_prompt] * num_prompts

        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
            pil_image
        )
        bs_embed, seq_len, _ = image_prompt_embeds.shape
        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
        image_prompt_embeds = image_prompt_embeds.view(
            bs_embed * num_samples, seq_len, -1
        )
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
            1, num_samples, 1
        )
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
            bs_embed * num_samples, seq_len, -1
        )

        with torch.inference_mode():
            (
                prompt_embeds,
                negative_prompt_embeds,
                pooled_prompt_embeds,
                negative_pooled_prompt_embeds,
            ) = self.pipe.encode_prompt(
                prompt,
                num_images_per_prompt=num_samples,
                do_classifier_free_guidance=True,
                negative_prompt=negative_prompt,
            )
            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
            negative_prompt_embeds = torch.cat(
                [negative_prompt_embeds, uncond_image_prompt_embeds], dim=1
            )

        generator = get_generator(seed, self.device)

        images = self.pipe(
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            num_inference_steps=num_inference_steps,
            generator=generator,
            **kwargs,
        ).images

        return images