app.py

import gradio as gr
# import spaces #[uncomment to use ZeroGPU]
from gradio_image_annotation import image_annotator
from diffusers import StableDiffusionPipeline
import os
import torch
from diffusers import EulerDiscreteScheduler
from migc.migc_utils import seed_everything
from migc.migc_pipeline import StableDiffusionMIGCPipeline, MIGCProcessor, AttentionStore


from huggingface_hub import hf_hub_download

# 下载文件
migc_ckpt_path = hf_hub_download(
    repo_id="limuloo1999/MIGC",
    filename="MIGC_SD14.ckpt",
    repo_type="model"  # 也可以省略，默认就是 model
)

RV_path = hf_hub_download(
    repo_id="SG161222/Realistic_Vision_V6.0_B1_noVAE",
    filename="Realistic_Vision_V6.0_NV_B1.safetensors",
    repo_type="model"  # 也可以省略，默认就是 model
)



# Load model
# pipe = StableDiffusionMIGCPipeline.from_pretrained(
#     "rSG161222/Realistic_Vision_V6.0_B1_noVAE",
#     torch_dtype=torch.float32
# )
pipe = StableDiffusionMIGCPipeline.from_single_file(
    RV_path,
    torch_dtype=torch.float32
)
pipe.safety_checker = None
pipe.attention_store = AttentionStore()
from migc.migc_utils import load_migc
load_migc(pipe.unet , pipe.attention_store,
        migc_ckpt_path, attn_processor=MIGCProcessor)
pipe = pipe.to("cuda" if torch.cuda.is_available() else "cpu")
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)

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

# @spaces.GPU
def get_boxes_json(annotations):
    print(annotations)
    image = annotations["image"]
    width = image.shape[1]
    height = image.shape[0]
    boxes = annotations["boxes"]
    prompt_final = [[]]
    bboxes = [[]]
    for box in boxes:
        box["xmin"] = box["xmin"] / width
        box["xmax"] = box["xmax"] / width
        box["ymin"] = box["ymin"] / height
        box["ymax"] = box["ymax"] / height
        prompt_final[0].append(box["label"])
        bboxes[0].append([box["xmin"], box["ymin"], box["xmax"], box["ymax"]])
    # import pdb; pdb.set_trace()
    prompt = ", ".join(prompt_final[0])
    prompt_final[0].insert(0, prompt)
    negative_prompt = 'worst quality, low quality, bad anatomy, watermark, text, blurry'
    image = pipe(prompt_final, bboxes, num_inference_steps=30, guidance_scale=7.5, 
                    MIGCsteps=15, aug_phase_with_and=False, negative_prompt=negative_prompt).images[0]
    return image
    # return annotations["boxes"]

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)

        button_get = gr.Button("Generation")
        button_get.click(get_boxes_json, inputs=annotator, outputs=generated_image)


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