RefVIE-Bench

Project Page | Paper | GitHub

RefVIE-Bench is a comprehensive evaluation benchmark introduced in the paper Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance. It is specifically designed to assess instruction-reference-following capabilities in video editing models, featuring source videos, reference images (for both subjects and backgrounds), and natural language instructions.

Sample Usage

To run inference on this benchmark using the official Kiwi-Edit framework, you can use the following command:

python infer.py \
  --ckpt_path path_to_ckpt \
  --bench refvie \
  --max_frame 81 \
  --max_pixels 921600 \
  --save_dir ./infer_results/exp_name/

Directory layout

• refvie_bench.yaml: Configuration file containing mapping for instructions, reference images, and source videos.
• ref_images/background/: Reference images used for background-guided editing.
• ref_images/subjects/: Reference images used for subject-guided editing.
• source_videos/: The original video sequences.

Included media

• Total referenced media files: 86
• Reference images: 54
  • Background images: 8
  • Subject images: 46
• Source videos: 32

Notes

• File paths in refvie_bench.yaml are preserved relative to this release directory.

Citation

If you use our code in your work, please cite our paper:

@misc{kiwiedit,
      title={Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance}, 
      author={Yiqi Lin and Guoqiang Liang and Ziyun Zeng and Zechen Bai and Yanzhe Chen and Mike Zheng Shou},
      year={2026},
      eprint={2603.02175},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.02175}, 
}