PhyRubric: Rubric-Grounded Diagnostic Evaluation of Physical Commonsense in Text-to-Video Generation
Across ten text-to-video models, the authors find physical commonsense remains far from solved.
PhyRubric builds a caption-specific checklist before viewing a video, then scores individual semantic and physical requirements using timestamped evidence. Its benchmark covers 34 categories across five physical domains, identifying which principles a video violates. The team also fine-tuned an open-source judge on 3,788 videos. They report closer agreement with human judgments than existing open-source judges, matching the proprietary Gemini 4 Argon.
HF Daily Papers' note
PhyRubric builds a caption-specific checklist before viewing a video, then scores individual semantic and physical requirements using timestamped evidence. Its benchmark covers 34 categories across five physical domains, identifying which principles a video violates. The team also fine-tuned an open-source judge on 3,788 videos. They report closer agreement with human judgments than existing open-source judges, matching the proprietary Gemini 4 Argon.
HF Daily Papers' note
score 5