PhyRubric: rubric-based benchmark diagnoses physical commonsense in text-to-video
Huiyi Chen · hf · 2026-11-12
UIC researchers introduce PhyRubric, a rubric-grounded diagnostic benchmark and judge for physical commonsense in text-to-video generation. Motivated by T2V models being positioned as world simulators, it replaces subjective single-score ratings with structured per-principle diagnosis (semantic adherence, physical law compliance). Results show current models still struggle with physical commonsense, and the judge aligns closely with human judgments. Code and paper are open-sourced.
More from Multimodal
- MPEG Diffusion renders subjects hi-res and backgrounds low-res for up to 4.8× faster video generation — xxunhuang · 2026-11-12
- fal plugin brings image & video generation into ChatGPT, producing a 60-second ad — OpenAIDevs · 2026-11-12
- Runway adds Seedance 2.5 with 4K video generation support — runwayml · 2026-11-12
- Recraft upgrades vector output quality, ships Vectorizer Pro on all plans — aziz4ai · 2026-11-12
- Fable 5.2 demo: stunning video of the bacterial flagellar motor — Dr_Singularity · 2026-11-12
- MiniMax H3 Image vs MOSS Image: 16MP in 6 steps against 2MP in 50 steps — zmarcoz2 · 2026-11-12