app.py

import spaces
import gradio as gr
import torch
from diffusers import ControlNetModel, StableDiffusionXLControlNetImg2ImgPipeline, AutoencoderKL
from PIL import Image
import os
import time
from utils.dl_utils import dl_cn_model, dl_cn_config, dl_tagger_model, dl_lora_model
from utils.image_utils import resize_image_aspect_ratio, base_generation, background_removal
from utils.prompt_utils import execute_prompt, remove_color, remove_duplicates
from utils.tagger import modelLoad, analysis

path = os.getcwd()
cn_dir = f"{path}/controlnet"
tagger_dir = f"{path}/tagger"
lora_dir = f"{path}/lora"
os.makedirs(cn_dir, exist_ok=True)
os.makedirs(tagger_dir, exist_ok=True)
os.makedirs(lora_dir, exist_ok=True)

dl_cn_model(cn_dir)
dl_cn_config(cn_dir)
dl_tagger_model(tagger_dir)
dl_lora_model(lora_dir)

@spaces.GPU(duration=120)
def predict(lora_model, input_image_path, prompt, negative_prompt, controlnet_scale, load_model_fn):
    # LoRAモデルに基づきpipeを取得
    pipe = load_model_fn(lora_model)
    input_image = Image.open(input_image_path)
    base_image = base_generation(input_image.size, (255, 255, 255, 255)).convert("RGB")
    resize_image = resize_image_aspect_ratio(input_image)
    resize_base_image = resize_image_aspect_ratio(base_image)
    generator = torch.manual_seed(0)
    last_time = time.time()
    
    # プロンプト生成
    prompt = "masterpiece, best quality, monochrome, greyscale, lineart, white background, star-shaped pupils, " + prompt
    execute_tags = ["realistic", "nose", "asian"]
    prompt = execute_prompt(execute_tags, prompt)
    prompt = remove_duplicates(prompt)
    prompt = remove_color(prompt)
    print(prompt)
    
    # 画像生成
    output_image = pipe(
        image=resize_base_image,
        control_image=resize_image,
        strength=1.0,
        prompt=prompt,
        negative_prompt=negative_prompt,
        controlnet_conditioning_scale=float(controlnet_scale),
        generator=generator,
        num_inference_steps=30,
        eta=1.0,
    ).images[0]
    print(f"Time taken: {time.time() - last_time}")
    output_image = output_image.resize(input_image.size, Image.LANCZOS)
    return output_image


class Img2Img:
    def __init__(self):
        self.demo = self.layout()
        self.tagger_model = None
        self.input_image_path = None
        self.bg_removed_image = None
        self.pipe = None
        self.current_lora_model = None
        
    def load_model(self, lora_model):
        # 既に正しいpipeがロードされている場合は再利用
        if self.pipe and self.current_lora_model == lora_model:
            return self.pipe  # キャッシュされたpipeを返す

        # 新しいpipeの生成
        dtype = torch.float16
        vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=dtype)
        controlnet = ControlNetModel.from_pretrained(cn_dir, torch_dtype=dtype, use_safetensors=True)

        self.pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_pretrained(
            "cagliostrolab/animagine-xl-3.1", controlnet=controlnet, vae=vae, torch_dtype=dtype
        )
        self.pipe.enable_model_cpu_offload()

        # LoRAモデルの設定
        if lora_model == "とりにく風":
            self.pipe.load_lora_weights(lora_dir, weight_name="tori29umai_line.safetensors")             
        elif lora_model == "少女漫画風":
            self.pipe.load_lora_weights(lora_dir, weight_name="syoujomannga_line.safetensors")        
        elif lora_model == "劇画調風":
            self.pipe.load_lora_weights(lora_dir, weight_name="gekiga_line.safetensors")
        elif lora_model == "プレーン":
            pass

        # 現在のlora_modelを保存
        self.current_lora_model = lora_model
        return self.pipe

    def layout(self):
        css = """
        #intro{
            max-width: 32rem;
            text-align: center;
            margin: 0 auto;
        }
        """
        with gr.Blocks(css=css) as demo:
            with gr.Row():
                with gr.Column():
                    self.lora_model = gr.Dropdown(label="Image Style",  choices=["プレーン", "とりにく風", "少女漫画風"], value="プレーン")                    
                    self.input_image_path = gr.Image(label="Input image", type='filepath')
                    self.bg_removed_image_path = gr.Image(label="Background Removed Image", type='filepath')
                    
                    # 自動背景除去トリガー
                    self.input_image_path.change(
                        fn=self.auto_background_removal,
                        inputs=[self.input_image_path],
                        outputs=[self.bg_removed_image_path]
                    )

                    self.prompt = gr.Textbox(label="Prompt", lines=3)
                    self.negative_prompt = gr.Textbox(label="Negative prompt", lines=3, value="nose, asian, realistic, lowres, error, extra digit, fewer digits, cropped, worst quality,low quality, normal quality, jpeg artifacts, blurry")
                    prompt_analysis_button = gr.Button("Prompt analysis")
                    self.controlnet_scale = gr.Slider(minimum=0.4, maximum=1.0, value=0.55, step=0.01, label="Photo fidelity")                 
                    generate_button = gr.Button(value="Generate", variant="primary")

                with gr.Column():
                    self.output_image = gr.Image(type="pil", label="Output image")

            prompt_analysis_button.click(
                fn=self.process_prompt_analysis,
                inputs=[self.bg_removed_image_path],
                outputs=self.prompt
            )

            generate_button.click(
                fn=lambda lora_model, input_image_path, prompt, negative_prompt, controlnet_scale: 
                   predict(lora_model, input_image_path, prompt, negative_prompt, controlnet_scale, self.load_model),
                inputs=[self.lora_model, self.bg_removed_image_path, self.prompt, self.negative_prompt, self.controlnet_scale],
                outputs=self.output_image
            )
        return demo

    def auto_background_removal(self, input_image_path):
        if input_image_path is not None:
            bg_removed_image = background_removal(input_image_path)
            return bg_removed_image
        return None

img2img = Img2Img()
img2img.demo.queue()
img2img.demo.launch(share=True)