run_sample.py

#!/usr/bin/env python
"""
VBench I2V 采样脚本

用法:
    python run_sample.py --model_name my_model --resolution 16-9
"""

import os
import argparse

# 设置模型缓存路径
os.environ['VBENCH_CACHE_DIR'] = './vbench2_beta_i2v/pretrained_models'
os.environ['HF_HOME'] = './vbench2_beta_i2v/pretrained_models/huggingface'
os.environ['TORCH_HOME'] = './vbench2_beta_i2v/pretrained_models/torch'

import torch
from PIL import Image
from samplers import I2VSampler


class MySampler(I2VSampler):
    """I2V 采样器 - 实现 sample 方法"""

    def setup(self):
        # TODO: 加载你的模型
        pass

    def sample(self, image_path: str, prompt: str, index: int) -> torch.Tensor:
        # TODO: 替换为你的模型采样代码

        # 标准视频分辨率 (16:9 -> 576x320)
        resolution_map = {
            "1-1": (512, 512),
            "8-5": (512, 320),
            "7-4": (448, 256),
            "16-9": (576, 320),
        }
        W, H = resolution_map[self.resolution]

        # 当前是噪声占位
        T, C = 16, 3
        video = torch.randint(0, 256, (T, H, W, C), dtype=torch.uint8)
        return video


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="VBench I2V 采样")
    parser.add_argument("--model_name", type=str, required=True, help="模型名称")
    parser.add_argument("--resolution", type=str, default="16-9",
                        choices=["1-1", "8-5", "7-4", "16-9"], help="分辨率")
    parser.add_argument("--seed", type=int, default=42, help="随机种子")
    parser.add_argument("--fps", type=int, default=8, help="视频帧率")
    args = parser.parse_args()

    sampler = MySampler(
        model_name=args.model_name,
        resolution=args.resolution,
        seed=args.seed,
        fps=args.fps,
    )
    sampler.run()