app.py

import gradio as gr
from gradio_image_annotation import image_annotator
from diffusers import EulerDiscreteScheduler
import torch
import os
import random

from migc.migc_pipeline import StableDiffusionMIGCPipeline, MIGCProcessor, AttentionStore
from migc.migc_utils import seed_everything, load_migc
from huggingface_hub import hf_hub_download

# 下载模型文件
migc_ckpt_path = hf_hub_download(repo_id="limuloo1999/MIGC", filename="MIGC_SD14.ckpt")
RV_path = hf_hub_download(repo_id="SG161222/Realistic_Vision_V6.0_B1_noVAE", filename="Realistic_Vision_V6.0_NV_B1.safetensors")
anime_path = hf_hub_download(repo_id="ckpt/cetus-mix", filename="cetusMix_v4.safetensors")

# -------- 风格切换器类 --------
class StyleSwitcher:
    def __init__(self):
        self.pipe = None
        self.attn_store = AttentionStore()
        self.styles = {
            "realistic": RV_path,
            "anime": anime_path
        }
        self.current_style = None

    def load_model(self, style):
        if style == self.current_style:
            return self.pipe

        if self.pipe:
            del self.pipe
            torch.cuda.empty_cache()
            print(f"[Info] Switched from {self.current_style} to {style}.")

        model_path = self.styles[style]
        print(f"[Info] Loading {style} model...")

        self.pipe = StableDiffusionMIGCPipeline.from_single_file(
            model_path,
            torch_dtype=torch.float32
        )
        self.pipe.safety_checker = None
        self.pipe.attention_store = self.attn_store
        load_migc(self.pipe.unet, self.attn_store, migc_ckpt_path, attn_processor=MIGCProcessor)
        self.pipe = self.pipe.to("cuda" if torch.cuda.is_available() else "cpu")
        self.pipe.scheduler = EulerDiscreteScheduler.from_config(self.pipe.scheduler.config)

        self.current_style = style
        return self.pipe

style_switcher = StyleSwitcher()

# ⬇️ 新增函数：返回随机 seed
def generate_random_seed():
    return random.randint(0, 2**32 - 1)

# 生成函数
def get_boxes_json(annotations, seed_value, edit_mode, style_selection):
    seed_everything(seed_value)

    pipe = style_switcher.load_model(style_selection)

    image = annotations["image"]
    width = image.shape[1]
    height = image.shape[0]
    boxes = annotations["boxes"]
    prompt_final = [[]]
    bboxes = [[]]
    for box in boxes:
        box["xmin"] /= width
        box["xmax"] /= width
        box["ymin"] /= height
        box["ymax"] /= height
        prompt_final[0].append(box["label"])
        bboxes[0].append([box["xmin"], box["ymin"], box["xmax"], box["ymax"]])

    prompt = ", ".join(prompt_final[0])
    prompt_final[0].insert(0, prompt)

    negative_prompt = 'worst quality, low quality, bad anatomy, watermark, text, blurry'
    output_image = pipe(prompt_final, bboxes, num_inference_steps=30, guidance_scale=7.5,
                        MIGCsteps=15, aug_phase_with_and=False, negative_prompt=negative_prompt,
                        sa_preserve=True, use_sa_preserve=edit_mode).images[0]
    return output_image

# 示例标注图
example_annotation = {
    "image": os.path.join(os.path.dirname(__file__), "background.png"),
    "boxes": [],
}

# ------------- Gradio UI -------------
with gr.Blocks() as demo:
    with gr.Tab("DreamRenderer", id="DreamRenderer"):
        with gr.Row():
            with gr.Column(scale=1):
                annotator = image_annotator(example_annotation, height=512, width=512)
            with gr.Column(scale=1):
                generated_image = gr.Image(label="Generated Image", height=512, width=512)
                seed_input = gr.Number(label="Seed (Optional)", precision=0)
                seed_random_btn = gr.Button("🎲 Random Seed")
                edit_mode_toggle = gr.Checkbox(label="Edit Mode")
                style_selector = gr.Radio(choices=["realistic", "anime"], label="风格选择", value="realistic")

        button_get = gr.Button("生成图像")
        button_get.click(
            fn=get_boxes_json,
            inputs=[annotator, seed_input, edit_mode_toggle, style_selector],
            outputs=generated_image
        )

        seed_random_btn.click(fn=generate_random_seed, inputs=[], outputs=seed_input)

if __name__ == "__main__":
    demo.launch(share=True)