[AAAI 25] VE-Bench: Subjective-Aligned Benchmark Suite for Text-Driven Video Editing Quality Assessment
(* Corresponding author)
from MMCAL, Peking University
TL;DR: VE-Bench is an evaluation suite for text-driven video editing, consisting of a quality assessment model to provide a human-aligned metric for edited videos, and a database containing rich video-prompt pairs and the corresponding human scores.
VE-Bench DB contains a rich collection of source videos, including real-world videos, AIGC videos, and CG videos, covering various aspects such as people, objects, animals, and landscapes. It also includes a variety of editing instructions across different categories, including semantic editing like addition, removal, replacement, etc., as well as structural changes in size, shape, etc., and stylizations such as color, texture, etc. Additionally, it features editing results based on different video editing models. We conducted a subjective experiment involving 24 participants from diverse backgrounds, resulting in 28,080 score samples. We further trained VE-Bench QA model based on this data. The left image below shows the box plot of average scores obtained by each model during the subjective experiment, while the right image illustrates the scores for each model across different types of prompts.
Left: Average score distributions of 8 editing methods. Right: Performance on different types of prompts from previous video-editing methods.
Download all models from google drive and put them into ckpts
.
python infer.py --single_test --src_path ${path_to_source_video} --dst_path ${path_to_dst_video} --prompt ${editing_prompt}
# Run on example videos
# python infer.py --single_test --src_path "./data/src/00433tokenflow_baby_gaze.mp4" --dst_path "./data/edited/00433tokenflow_baby_gaze.mp4" --prompt "A black-haired boy is turning his head"
python infer.py --data_path ${path_to_data_folder} --label_path ${path_to_prompt_txt_file}
Part of the code is developed based on DOVER and BLIP. We would like to thank the authors for their contributions to the community.
If your have any comments or questions, feel free to contact sunshk@stu.pku.edu.cn.
@article{sun2024bench,
title={VE-Bench: Subjective-Aligned Benchmark Suite for Text-Driven Video Editing Quality Assessment},
author={Sun, Shangkun and Liang, Xiaoyu and Fan, Songlin and Gao, Wenxu and Gao, Wei},
journal={arXiv preprint arXiv:2408.11481},
year={2024}
}