T2V-CompBench
by Kaiyue Sun and collaborators (CVPR 2025)
What is T2V-CompBench? T2V-CompBench is an image & vision eval made by Kaiyue Sun and collaborators (CVPR 2025). A compositional text-to-video benchmark (CVPR 2025) for whether generated video follows multi-object, attribute and relation bindings in the prompt. Official code: github.com/KaiyueSun98/T2V-CompBench. Still cited as a composition baseline; the GitHub repo had no LICENSE file when checked on 14 Nov 2026, so the licence field is left empty.
Editor note
A compositional text-to-video benchmark (CVPR 2025) for whether generated video follows multi-object, attribute and relation bindings in the prompt. Official code: github.com/KaiyueSun98/T2V-CompBench. Still cited as a composition baseline; the GitHub repo had no LICENSE file when checked on 14 Nov 2026, so the licence field is left empty.
- What it tests
- Compositional prompt following in text-to-video generation.
- Original source
- t2v-compbench-2025.github.io/
- Last updated
- Repo last push 2025-10-25 (GitHub API, earlier read 14 Nov 2026); LICENSE still absent 14 Nov 2026
- License
- Not checked yet — see source
Related image & vision evals
Discussion
No one has discussed T2V-CompBench yet. Have you run it? Share what it caught, what it missed or how you set it up.