VideoPhy-2
by UCLA and Google Research (Hritik Bansal, Clark Peng, Yonatan Bitton, Roman Goldenberg, Aditya Grover, Kai-Wei Chang)
What is VideoPhy-2? VideoPhy-2 is an image & vision eval made by UCLA and Google Research (Hritik Bansal, Clark Peng, Yonatan Bitton, Roman Goldenberg, Aditya Grover, Kai-Wei Chang). An action-centric physical-commonsense benchmark for text-to-video models. It curates 200 diverse actions and detailed prompts; human raters score semantic adherence, physical commonsense and specific physical-rule violations. The official page on 14 Nov 2026 reported that the best listed open model, Wan2.1-14B, reached 21.9% joint performance on the hard subset. ICLR 2026 (arXiv:2503.06800). An automatic evaluator, VideoPhy2-eval, is also released. The earlier VideoPhy paper (arXiv:2406.03520, ICLR 2025) and MIT-licensed code are at github.com/Hritikbansal/videophy.
Editor note
An action-centric physical-commonsense benchmark for text-to-video models. It curates 200 diverse actions and detailed prompts; human raters score semantic adherence, physical commonsense and specific physical-rule violations. The official page on 14 Nov 2026 reported that the best listed open model, Wan2.1-14B, reached 21.9% joint performance on the hard subset. ICLR 2026 (arXiv:2503.06800). An automatic evaluator, VideoPhy2-eval, is also released. The earlier VideoPhy paper (arXiv:2406.03520, ICLR 2025) and MIT-licensed code are at github.com/Hritikbansal/videophy.
- What it tests
- Semantic adherence plus physical commonsense of action videos, including a hard subset and rule-level violations.
- Original source
- videophy2.github.io/
- Last updated
- ICLR 2026; VideoPhy repo LICENSE re-read 14 Nov 2026 (MIT)
- License
- MIT (Hritikbansal/videophy LICENSE, re-read 14 Nov 2026)
Related image & vision evals
Discussion
No one has discussed VideoPhy-2 yet. Have you run it? Share what it caught, what it missed or how you set it up.