VEditBench: Holistic Benchmark for Text-Guided Video Editing
Keywords: Benchmark, Generative Models, Video Editing
Abstract: Video editing usually requires substantial human expertise and effort. However, recent advances in generative models have democratized this process, enabling video edits to be made using simple textual instructions. Despite this progress, the absence of a standardized and comprehensive benchmark has made it difficult to compare different methods within a common framework. To address this gap, we introduce VEditBench, a comprehensive benchmark for text-guided video editing (TGVE). VEditBench offers several key features: (1) 420 real-world videos spanning diverse categories and durations, including 300 short videos (2-4 seconds) and 120 longer videos (10-20 seconds); (2) 6 editing tasks that capture a broad range of practical editing challenges: object insertion, object removal, object swap, scene replacement, motion change, and style translation; (3) 9 evaluation dimensions to assess the semantic fidelity and visual quality of edits. We evaluate ten state-of-the-art video editing models using VEditBench, offering an in-depth analysis of their performance across metrics, tasks, and models. We hope VEditBench will provide valuable insights to the community and serve as the standard benchmark for TGVE models following its open-sourcing.
Supplementary Material: zip
Primary Area: datasets and benchmarks
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8243
Loading