VBench-2.0
by Shanghai AI Laboratory, NTU S-Lab, Sun Yat-Sen University and CUHK (Dian Zheng, Ziqi Huang and collaborators)
What is VBench-2.0? VBench-2.0 is an image & vision eval made by Shanghai AI Laboratory, NTU S-Lab, Sun Yat-Sen University and CUHK (Dian Zheng, Ziqi Huang and collaborators). A later suite on the VBench line that scores generated videos on intrinsic faithfulness rather than only visual polish. It uses five categories — Human Fidelity, Controllability, Creativity, Physics and Commonsense — broken into 18 fine-grained dimensions, mixing generalist VLMs/LLMs with specialist anomaly detectors and human-preference checks. Paper: arXiv:2503.21755 (March 2025). Code lives in the Vchitect/VBench repository. The official Space names a Total Score, but the results CSV we requested from the project's Hugging Face dataset returned HTTP 401 on 14 Nov 2026.
Editor note
A later suite on the VBench line that scores generated videos on intrinsic faithfulness rather than only visual polish. It uses five categories — Human Fidelity, Controllability, Creativity, Physics and Commonsense — broken into 18 fine-grained dimensions, mixing generalist VLMs/LLMs with specialist anomaly detectors and human-preference checks. Paper: arXiv:2503.21755 (March 2025). Code lives in the Vchitect/VBench repository. The official Space names a Total Score, but the results CSV we requested from the project's Hugging Face dataset returned HTTP 401 on 14 Nov 2026.
- What it tests
- Whether generated video follows anatomy, control, creativity, physics and commonsense, not just looks coherent.
- Original source
- vchitect.github.io/VBench-2.0-project/
- Last updated
- Project page dated Mar 2025; parent repo last push 2026-08-21; LICENSE re-read 14 Nov 2026
- License
- Apache-2.0 (parent Vchitect/VBench repo LICENSE, re-read 14 Nov 2026)
Related image & vision evals
Discussion
No one has discussed VBench-2.0 yet. Have you run it? Share what it caught, what it missed or how you set it up.