PhyProbe

Physical consistency violation scorer for AI-generated videos.

Project page · Paper · Code

Higher score → more severe physical consistency violations (0.0 = no violation detected, 1.0 = strong violation).

Files

File	Contents
pecore_paper.pt	Quality-head weights (EMA) + model config, ~4 MB

The checkpoint holds only the trained quality head. The frozen PE-Core-G14-448 backbone is downloaded separately (~9 GB) through perception_models.

Usage

pip install phyprobe   # Python >= 3.10
pip install git+https://github.com/facebookresearch/perception_models.git

import phyprobe

model = phyprobe.PhyProbe()                     # uses the weights bundled in the package
model = phyprobe.PhyProbe(download=True)        # or fetch this checkpoint from the Hub
score = model.score("video.mp4")                # 48 frames sampled across the whole video
score = model.score("video.mp4", clip_sec=3.0)  # first 3 s only (the training window)

Model details

• Backbone: PE-Core-G14-448 (frozen, bf16)
• Head: per-frame features pooled over time (mean / std / max) → MLP 3840 → 256 → 128 → 1
• Input: 48 frames resampled to 16 fps, letterboxed to 480×480
• Training: pairwise Bradley–Terry + Huber regression, 10k steps

Citation

NeurIPS 2026:

@inproceedings{ku2026phyprobe,
  title   = {PhyProbe: Rethinking Physical Consistency Evaluation in Generated Videos},
  author  = {Ku, Max and Fan, Jiaojiao and Hao, Zekun and Ferroni, Francesco and Wang, Heng and Chen, Wenhu and Liu, Ming-Yu and Chattopadhyay, Prithvijit},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year    = {2026},
  eprint  = {2609.38377},
  archivePrefix = {arXiv},
  url     = {https://arxiv.org/abs/2609.38377}
}

arXiv:

@misc{ku2026phyproberethinkingphysicalconsistency,
      title={PhyProbe: Rethinking Physical Consistency Evaluation in Generated Videos}, 
      author={Max Ku and Jiaojiao Fan and Zekun Hao and Francesco Ferroni and Heng Wang and Wenhu Chen and Ming-Yu Liu and Prithvijit Chattopadhyay},
      year={2026},
      eprint={2609.38377},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.38377}, 
}