ip_adapter/ip_adapter_faceid.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_faceid import LoRAAttnProcessor, LoRAIPAttnProcessor


class MLPProjModel(torch.nn.Module):
    """SD model with image prompt"""
    def __init__(self, cross_attention_dim=768, id_embeddings_dim=512, num_tokens=4):
        super().__init__()
        
        self.cross_attention_dim = cross_attention_dim
        self.num_tokens = num_tokens
        
        self.proj = torch.nn.Sequential(
            torch.nn.Linear(id_embeddings_dim, id_embeddings_dim*2),
            torch.nn.GELU(),
            torch.nn.Linear(id_embeddings_dim*2, cross_attention_dim*num_tokens),
        )
        self.norm = torch.nn.LayerNorm(cross_attention_dim)
        
    def forward(self, id_embeds):
        x = self.proj(id_embeds)
        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
        x = self.norm(x)
        return x


class IPAdapterFaceID:
    def __init__(self, sd_pipe, ip_ckpt, device, lora_rank=128, num_tokens=4):
        self.device = device
        self.ip_ckpt = ip_ckpt
        self.lora_rank = lora_rank
        self.num_tokens = num_tokens

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

        # image proj model
        self.image_proj_model = self.init_proj()

        self.load_ip_adapter()

    def init_proj(self):
        image_proj_model = MLPProjModel(
            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
            id_embeddings_dim=512,
            num_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] = LoRAAttnProcessor(
                    hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, rank=self.lora_rank,
                ).to(self.device, dtype=torch.float16)
            else:
                attn_procs[name] = LoRAIPAttnProcessor(
                    hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, scale=1.0, rank=self.lora_rank, num_tokens=self.num_tokens,
                ).to(self.device, dtype=torch.float16)
        unet.set_attn_processor(attn_procs)

    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"])

    @torch.inference_mode()
    def get_image_embeds(self, faceid_embeds):
        
        faceid_embeds = faceid_embeds.to(self.device, dtype=torch.float16)
        image_prompt_embeds = self.image_proj_model(faceid_embeds)
        uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(faceid_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, LoRAIPAttnProcessor):
                attn_processor.scale = scale

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

       
        num_prompts = faceid_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(faceid_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 = torch.Generator(self.device).manual_seed(seed) if seed is not None else None
        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