app.py

#!/usr/bin/env python3
"""
Zero123++ Gradio Demo
用于 Hugging Face Spaces 部署
"""

import gradio as gr
import torch
from PIL import Image
from diffusers import DiffusionPipeline, EulerAncestralDiscreteScheduler
import os

# 全局变量存储 pipeline
pipeline = None

def load_model():
    """加载模型"""
    global pipeline

    if pipeline is not None:
        return

    print("正在加载模型...")

    # 检查 CUDA 可用性
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    dtype = torch.float16 if torch.cuda.is_available() else torch.float32

    # 加载 pipeline
    pipeline = DiffusionPipeline.from_pretrained(
        "sudo-ai/zero123plus-v1.1",
        custom_pipeline="sudo-ai/zero123plus-pipeline",
        torch_dtype=dtype
    )

    # 设置调度器
    pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(
        pipeline.scheduler.config,
        timestep_spacing='trailing'
    )

    pipeline.to(device)
    print(f"✓ 模型加载完成 (设备: {device})")

def process_image(input_image, remove_bg=False):
    """
    处理输入图像,生成多视图

    输入:
        input_image: PIL Image
        remove_bg: 是否移除背景 (暂未实现)

    输出:
        result_image: PIL Image (多视图合成图)
    """
    if input_image is None:
        return None

    # 确保模型已加载
    load_model()

    try:
        # 预处理图像 - 转为正方形
        img = input_image

        # 如果不是正方形,裁剪为正方形
        if img.size[0] != img.size[1]:
            size = min(img.size)
            img = img.crop((
                (img.size[0] - size) // 2,
                (img.size[1] - size) // 2,
                (img.size[0] + size) // 2,
                (img.size[1] + size) // 2
            ))

        # 调整到推荐尺寸
        target_size = 320
        if img.size[0] != target_size:
            img = img.resize((target_size, target_size), Image.LANCZOS)

        # 运行推理
        print("正在生成多视图...")
        result = pipeline(img).images[0]

        return result

    except Exception as e:
        print(f"错误: {e}")
        raise gr.Error(f"处理失败: {str(e)}")

# 创建 Gradio 界面
def create_demo():
    with gr.Blocks(title="Zero123++ Demo") as demo:
        gr.Markdown("""
        # Zero123++ 多视图生成

        将单张图片转换为 6 个不同角度的视图

        **输入要求:**
        - 建议使用正方形图片
        - 推荐分辨率 >= 320x320
        - 脚本会自动裁剪和调整非正方形图片

        **输出说明:**
        - 生成 6 个视图 (2行 x 3列)
        - 方位角: 30°, 90°, 150°, 210°, 270°, 330°
        - 仰角: 30°, -20°, 30°, -20°, 30°, -20°
        """)

        with gr.Row():
            with gr.Column():
                input_image = gr.Image(
                    label="输入图片",
                    type="pil",
                    height=400
                )

                # remove_bg = gr.Checkbox(
                #     label="移除背景 (实验性)",
                #     value=False
                # )

                generate_btn = gr.Button("生成多视图", variant="primary")

            with gr.Column():
                output_image = gr.Image(
                    label="多视图输出",
                    type="pil",
                    height=400
                )

        gr.Examples(
            examples=[
                ["examples/example1.png"],
                ["examples/example2.png"],
            ],
            inputs=input_image,
            label="示例图片"
        )

        gr.Markdown("""
        ### 技术说明
        - 模型: [sudo-ai/zero123plus-v1.1](https://huggingface.co/sudo-ai/zero123plus-v1.1)
        - 首次运行需要加载模型,可能需要 20-30 秒
        - 生成时间约 30-60 秒 (取决于硬件)

        ### 引用
        ```bibtex
        @misc{shi2023zero123plus,
            title={Zero123++: a Single Image to Consistent Multi-view Diffusion Base Model},
            author={Ruoxi Shi and Hansheng Chen and Zhuoyang Zhang and Minghua Liu and Chao Xu and Xinyue Wei and Linghao Chen and Chong Zeng and Hao Su},
            year={2023},
            eprint={2310.15110},
            archivePrefix={arXiv},
            primaryClass={cs.CV}
        }
        ```
        """)

        # 绑定事件
        generate_btn.click(
            fn=process_image,
            inputs=[input_image],
            outputs=output_image
        )

    return demo

if __name__ == "__main__":
    # 预加载模型
    load_model()

    # 启动 demo
    demo = create_demo()
    demo.queue()
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False
    )