AnimationManager/server/python/birefnet-matting.py

"""
BiRefNet 图像背景移除工具
用于高质量抠图处理
"""
import sys
import os

# 添加 BiRefNet 目录到路径
birefnet_dir = os.path.join(os.path.dirname(__file__), 'BiRefNet')
sys.path.insert(0, birefnet_dir)

try:
    from transformers import AutoModelForImageSegmentation
    from PIL import Image
    import torch
    import numpy as np
    from torchvision import transforms
except ImportError as e:
    print(f"错误: 缺少必要的库 - {e}", file=sys.stderr)
    print("请安装: pip install transformers pillow torch numpy torchvision", file=sys.stderr)
    sys.exit(1)

def process_image(input_path, output_path):
    """
    处理单个图片文件
    """
    try:
        # 检查输入文件是否存在
        if not os.path.exists(input_path):
            print(f"错误: 输入文件不存在: {input_path}", file=sys.stderr)
            sys.exit(1)
        
        # 创建输出目录（如果不存在）
        output_dir = os.path.dirname(output_path)
        if output_dir and not os.path.exists(output_dir):
            os.makedirs(output_dir, exist_ok=True)
        
        print(f"正在加载 BiRefNet 模型...", flush=True)
        
        # 检查是否有本地模型路径
        local_model_path = os.path.join(birefnet_dir, 'model', 'BiRefNet')
        model_file = os.path.join(local_model_path, 'model.safetensors')
        config_file = os.path.join(local_model_path, 'birefnet.py')
        use_local = os.path.exists(model_file) and os.path.exists(config_file)
        
        # 加载 BiRefNet 模型
        try:
            if use_local:
                print(f"使用本地模型: {local_model_path}", flush=True)
                model = AutoModelForImageSegmentation.from_pretrained(
                    local_model_path,
                    trust_remote_code=True,
                    local_files_only=True
                )
            else:
                print(f"从 HuggingFace 下载模型...", flush=True)
                print("提示: 如果下载失败，请运行 download-birefnet-model.py 下载模型到本地", flush=True)
                model = AutoModelForImageSegmentation.from_pretrained(
                    'ZhengPeng7/BiRefNet',
                    trust_remote_code=True
                )
        except Exception as e:
            print(f"错误: 无法加载 BiRefNet 模型 - {e}", file=sys.stderr)
            if not use_local:
                print("提示: 请确保已安装 transformers 库并可以访问 HuggingFace", file=sys.stderr)
                print("或者运行以下命令下载模型到本地:", file=sys.stderr)
                print("  python download-birefnet-model.py", file=sys.stderr)
            sys.exit(1)
        
        # 设置设备
        device = 'cuda' if torch.cuda.is_available() else 'cpu'
        model = model.to(device)
        model.eval()
        
        half_precision = device == 'cuda'
        if half_precision:
            model = model.half()  # 使用 FP16 加速
        
        print(f"正在处理图片: {os.path.basename(input_path)}", flush=True)
        
        # 读取输入图像
        image_ori = Image.open(input_path).convert('RGB')
        original_size = image_ori.size
        
        # 预处理图像（使用 BiRefNet 的预处理方式）
        try:
            from image_proc import ImagePreprocessor
            resolution = (1024, 1024)  # 默认分辨率
            image_preprocessor = ImagePreprocessor(resolution=resolution)
            image_proc = image_preprocessor.proc(image_ori)
            image_proc = image_proc.unsqueeze(0)
        except ImportError:
            # 如果无法导入，使用简单的预处理
            transform = transforms.Compose([
                transforms.Resize((1024, 1024)),
                transforms.ToTensor(),
                transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
            ])
            image_proc = transform(image_ori).unsqueeze(0)
        
        image_proc = image_proc.to(device)
        if half_precision:
            image_proc = image_proc.half()
        
        # 推理
        print("正在进行VIP抠图处理...", flush=True)
        with torch.no_grad():
            preds = model(image_proc)[-1].sigmoid().cpu()
        
        # 后处理：将预测结果转换为图像
        pred = preds[0].squeeze()
        pred_pil = transforms.ToPILImage()(pred)
        
        # 调整到原始尺寸
        pred_pil = pred_pil.resize(original_size, Image.Resampling.LANCZOS)
        
        # 应用遮罩到原图
        image_rgba = image_ori.convert('RGBA')
        mask_array = np.array(pred_pil)
        
        # 创建 RGBA 图像
        rgba_array = np.array(image_rgba)
        rgba_array[:, :, 3] = mask_array
        
        # 保存结果
        output_image = Image.fromarray(rgba_array, 'RGBA')
        output_image.save(output_path, 'PNG')
        
        print("VIP抠图完成", flush=True)
        
    except Exception as error:
        print(f"错误: {error}", file=sys.stderr)
        import traceback
        traceback.print_exc()
        sys.exit(1)

def main():
    if len(sys.argv) < 3:
        print("错误: 需要提供输入和输出路径", file=sys.stderr)
        print("用法: python birefnet-matting.py <输入文件> <输出文件>", file=sys.stderr)
        sys.exit(1)
    
    input_path = sys.argv[1]
    output_path = sys.argv[2]
    
    process_image(input_path, output_path)

if __name__ == '__main__':
    main()