PhyRubric paper finds Seedance 2.5 physically faithful in 44.2% of text-to-video cases
4.8/10Significance
A paper posted on November 11, 2026 introduces PhyRubric, a benchmark of 245 captions rendered into 2,433 videos by ten models. The authors report Seedance 2.5 at 44.20% and Veo 3.1 at 36.20%.
What changed
The best open-source model, HunyuanVideo-1.5-8B, scored 14.93%, and the authors say 25.0% of all videos never show the captioned event.
Why it matters
It gives a measured view of how often video generators get basic physics right.
What remains uncertain
The authors call it an analysis tool, not a definitive standard. Part of the paper’s judge results could not be checked.
What to watch
Watch for other labs reporting scores on PhyRubric.
Source
Why this ranks here
A clear benchmark result, but a single new paper.
- Impact4
- Reach5
- Novelty6
- Commitment4
- Coverage quality5
- Future potential5