Anates-Labs-Research/Physics-IQ-Verified

Physics-IQ Verified Dataset

This repository hosts the Physics-IQ Verified benchmark data for evaluating physical understanding in generative video models.

Physics-IQ Verified is derived from the original Physics-IQ benchmark dataset.

Original Physics-IQ

• Paper: Do generative video models understand physical principles?
• Repository: Code | Dataset in Google Cloud

Physics-IQ Verified (Recommended)

• Paper: Physics-IQ Verified
• Repository: Code | Dataset: Here in this repo :)

We gratefully acknowledge the original Physics-IQ authors for creating and releasing the benchmark.

Relation to the original Physics-IQ dataset

Physics-IQ Verified is not a plain copy of the original Physics-IQ dataset. It is a modified and verified benchmark variant designed to improve the robustness, traceability, and interpretability of the evaluation.

Compared to the original Physics-IQ data, Physics-IQ Verified introduces targeted changes including:

• refined prompts for ambiguous, incomplete, temporally imprecise, vague, or factually incorrect text descriptions,
• verified modifications to ground-truth videos and masks where spurious metric activations or visual artifacts could affect evaluation,
• a sample-level scoring setup aligned with the Physics-IQ Verified evaluation pipeline.

What was refined?

Physics-IQ Verified refines the original benchmark along three axes: prompt quality, artifact removal, and score aggregation.

Across the 198 take-1 benchmark samples, Physics-IQ Verified modifies or refines 57.6% of samples. In particular:

• 34.8% of prompts were refined because the original text description was ambiguous, incomplete, temporally imprecise, vague, or factually incorrect.
• 29.8% of videos contained artifacts or spurious metric activations that were not part of the physical phenomenon being evaluated.
• 20 videos contained both prompt issues and artifact-related modifications.
• At the frame level, 66.2% of frames contain active physical motion or metric activation, and 27.1% of those active frames were modified to remove artifact-driven activations.

Download

Install the Hugging Face CLI:

pip install -U huggingface_hub

Download the dataset into the folder name expected by the Physics-IQ benchmark code:

hf download Anates-Labs-Research/Physics-IQ-Verified \
  --repo-type dataset \
  --local-dir physics-IQ-benchmark-verified

After download, the local structure should be:

physics-IQ-benchmark-verified/
|-- full-videos/
|-- split-videos/
|-- switch-frames/
`-- video-masks/

Place physics-IQ-benchmark-verified/ in your working directory before running the benchmark.

Usage

Use this dataset with the Physics-IQ Verified benchmark code:

GitHub: Physics-IQ Verified

Physics-IQ Verified is the recommended benchmark variant.

Citation

Please cite both the original Physics-IQ benchmark and Physics-IQ Verified when using this dataset.

Original Physics-IQ

@article{motamed2026physics,
  title={Do generative video models understand physical principles?},
  author={Saman Motamed and Laura Culp and Kevin Swersky and Priyank Jaini and Robert Geirhos},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={948--958},
  year={2026}
}

Physics-IQ Verified

@article{radsch2026verified,
  author  = {Rädsch, Tim and Asano, Yuki M. and Kuehne, Hilde and Bauer, Stefan and Jaini, Priyank and Geirhos, Robert and Lüth, Carsten T.},
  title   = {Physics-IQ Verified},
  journal = {arXiv preprint arXiv:2606.18943},
  year    = {2026},
}

License

CC BY 4.0

Brought to you with love from the Anates Labs team.