PhyRubric-Bench

📄 Paper: PhyRubric: Rubric-Grounded Diagnostic Evaluation of Physical Commonsense in Text-to-Video Generation

The evaluation videos of PhyRubric, a rubric-grounded diagnostic framework for physical commonsense in text-to-video generation.

245 captions spanning 34 categories across five physical domains (Mechanics, Heat, Optics, Electricity, Acoustics), each with a caption-specific rubric written before any video was seen, rendered by ten video generators: 2,433 clips in total, every one labelled item by item with timestamped evidence and scored for semantic adherence (SA) and physical commonsense (PC) on 1-5.

<generator>/eval/<ID>.mp4        2,433 clips, 5-6 s
annotations/eval/<generator>.json   the gold labels (copy of the GitHub repository; captions and rubrics live there)

Generators: CogVideoX1.5-5B, Cosmos-Predict2.5-2B, Cosmos-Predict2.5-14B, HunyuanVideo-1.5 (480p), LTX-Video-2B, LTX-2.3-22B, Wan2.2-T2V-A14B, Wan2.2-TI2V-5B, Seedance 2.5 and Veo 3.1.

Usage

Code, the scoring rule, PhyRubric-Judge and the evaluation scripts are at https://github.com/chenyil6/phyRubric. Download this dataset into the repository so that the relative video paths in the annotations resolve:

git clone https://github.com/chenyil6/phyRubric && cd phyRubric
hf download kittyhy/PhyRubric-Bench --repo-type dataset --local-dir PhyRubric-Bench

The training split is PhyRubric-Train; the judge weights are PhyRubric-Judge.

The videos were generated by the listed models from our captions and are released for research use; each generator's own licence applies to its outputs.