faceadapter/face_adapter.py

import os
from typing import List

import torch
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 .attention_processor import LoRAFaceAttnProcessor

from .utils import is_torch2_available, get_generator

if is_torch2_available():
    from .attention_processor import (
        AttnProcessor2_0 as AttnProcessor,
    )
else:
    from .attention_processor import AttnProcessor
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 FaceAdapterLora:
    def __init__(self, sd_pipe, image_encoder_path, id_ckpt, device, num_tokens=4,torch_type=torch.float32):
        self.device = device
        self.image_encoder_path = image_encoder_path
        self.id_ckpt = id_ckpt
        self.num_tokens = num_tokens
        self.torch_type = torch_type

        self.pipe = sd_pipe.to(self.device)
        self.set_face_adapter()
        # load image encoder
        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(
            self.device, dtype=self.torch_type
        )
        self.clip_image_processor = CLIPImageProcessor()
        # image proj model
        self.image_proj_model = self.init_proj()

        self.load_face_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=self.torch_type)
        return image_proj_model

    def set_face_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().to(self.device, dtype=self.torch_type)
            else:
                attn_procs[name] = LoRAFaceAttnProcessor(
                    hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, scale=1.0, rank=128, num_tokens=self.num_tokens,
                ).to(self.device, dtype=self.torch_type)
        unet.set_attn_processor(attn_procs)
    def load_face_adapter(self):
        state_dict = torch.load(self.id_ckpt, map_location="cpu")
        if 'state_dict' in state_dict:
            state_dict = state_dict['state_dict']
            image_proj_dict={}
            face_adapter_proj={}
            for k,v in state_dict.items():
                if k.startswith("module.image_proj_model"):
                    image_proj_dict[k.replace("module.image_proj_model.", "")] = state_dict[k]
                elif k.startswith("module.adapter_modules."):
                    face_adapter_proj[k.replace("module.adapter_modules.", "")] = state_dict[k]
                elif k.startswith("image_proj_model"):
                    image_proj_dict[k.replace("image_proj_model.", "")] = state_dict[k]
                elif k.startswith("adapter_modules."):
                    face_adapter_proj[k.replace("adapter_modules.", "")] = state_dict[k]
                else:
                    print("ERROR!")
                    return
            state_dict = {}
            state_dict['image_proj'] = image_proj_dict
            state_dict["face_adapter"] = face_adapter_proj
        self.image_proj_model.load_state_dict(state_dict["image_proj"])
        adapter_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
        adapter_layers.load_state_dict(state_dict["face_adapter"],strict=False)
    @torch.inference_mode()
    def get_image_embeds(self, pil_image=None, clip_image_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=self.torch_type)).image_embeds
        else:
            clip_image_embeds = clip_image_embeds.to(self.device, dtype=self.torch_type)
        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

    # This scales the face-adapter face_hidden_states (attn output). attn_processor.scale: default 1.0.
    # faceadapter/attention_processor.py:L283.
    def set_attn_scale(self, attn_scale):
        for attn_processor in self.pipe.unet.attn_processors.values():
            if isinstance(attn_processor, LoRAFaceAttnProcessor):
                attn_processor.scale = attn_scale

    def generate(
        self,
        pil_image=None,
        clip_image_embeds=None,
        prompt=None,
        negative_prompt=None,
        attn_scale=1,
        num_samples=4,
        seed=None,
        guidance_scale=7.5,
        num_inference_steps=30,
        **kwargs,
    ):
        self.set_attn_scale(attn_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
        )
        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 FaceAdapterPlusForVideoLora(FaceAdapterLora):
    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=self.torch_type)
        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=self.torch_type)
        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=None,
        init_image=None,
        init_image_strength=1.,
        clip_image_embeds=None,
        prompt=None,
        negative_prompt=None,
        adaface_embeds=None,
        adaface_scale=1.0,
        attn_scale=1.0,
        num_samples=1,
        seed=None,
        guidance_scale=4,
        num_inference_steps=30,
        adaface_anneal_steps=0,
        width=512,
        height=512,
        video_length=16,
        image_embed_scale=1,
        controlnet_images: torch.FloatTensor = None,
        controlnet_image_index: list = [0],
        **kwargs,
    ):
        self.set_attn_scale(attn_scale)
        num_prompts=1

        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
        num_prompt_img  = len(pil_image)
        total_image_prompt_embeds = 0
        for i in range(num_prompt_img):
            prompt_img = pil_image[i]
            image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
                pil_image=prompt_img, clip_image_embeds=clip_image_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)
            total_image_prompt_embeds += image_prompt_embeds
        total_image_prompt_embeds /= num_prompt_img
        image_prompt_embeds  = total_image_prompt_embeds
        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():
            # if do_classifier_free_guidance,
            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method.
            # https://github.com/huggingface/diffusers/blob/70f8d4b488f03730ae3bc11d4d707bafe153d10d/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py#L469
            prompt_embeds_, negative_prompt_embeds_ = self.pipe.encode_prompt(
                prompt,
                device=self.device,
                num_videos_per_prompt=num_samples,
                do_classifier_free_guidance=True,
                negative_prompt=negative_prompt,
            )

            if adaface_embeds is not None:
                prompt_embeds0_ = prompt_embeds_
                # self.torch_type == torch.float16. adaface_embeds is torch.float32.
                prompt_embeds_ = adaface_embeds.repeat(num_samples, 1, 1).to(dtype=self.torch_type) * adaface_scale
                # Scale down ID-Animator's face embeddings, so that they don't dominate the generation.
                # Note to balance image_prompt_embeds with uncond_image_prompt_embeds after scaling.
                image_prompt_embeds = image_prompt_embeds * image_embed_scale + uncond_image_prompt_embeds * (1 - image_embed_scale)
                # We still need uncond_image_prompt_embeds, otherwise the output is blank.
                prompt_embeds_end   = torch.cat([prompt_embeds_,  image_prompt_embeds], dim=1)
                prompt_embeds_begin = torch.cat([prompt_embeds0_, torch.zeros_like(image_prompt_embeds)], dim=1)
                prompt_embeds = (prompt_embeds_begin, prompt_embeds_end, adaface_anneal_steps)
            else:
                prompt_embeds = torch.cat([prompt_embeds_, image_prompt_embeds], dim=1)

            # 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)

        video = self.pipe(
            init_image=init_image,
            init_image_strength=init_image_strength,
            prompt = "",
            prompt_embeds = prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps,
            generator=generator,
            width = width,
            height=height,
            video_length = video_length,
            controlnet_images = controlnet_images,
            controlnet_image_index=controlnet_image_index,
            **kwargs,
        ).videos

        return video

    def generate_video_edit(
        self,
        pil_image=None,
        clip_image_embeds=None,
        prompt=None,
        negative_prompt=None,
        attn_scale=1.0,
        num_samples=1,
        seed=None,
        guidance_scale=7.5,
        num_inference_steps=30,
        width=512,
        height=512,
        video_length=16,
        video_latents=None,
        **kwargs,
    ):
        self.set_attn_scale(attn_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
        )
        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_videos_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)

        video = self.pipe.video_edit(
            prompt = "",
            prompt_embeds = prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps,
            generator=generator,
            width = width,
            height=height,
            video_length = video_length,
            latents=video_latents,
            **kwargs,
        ).videos

        return video