PhysicTran38K

We provide 38K video-based dataset for physics-aware image editing, by casting editing as physical state transitions. To know the details of the dataset creation pipeline, please refer to Section 3.2 of [1].

Dataset Download and Loading

huggingface-cli download diffusion-cot/PhysicTran38K --local-dir /path/to/data_path

Then please refer to our training code for data loading.

✍️ Citation

If you find our work inspiring or use our codebase in your research, please consider giving a star ⭐ and a citation~

@article{zhao2026statics,
  title={From Statics to Dynamics: Physics-Aware Image Editing with Latent Transition Priors},
  author={Zhao, Liangbing and Zhuo, Le and Paul, Sayak and Li, Hongsheng and Elhoseiny, Mohamed},
  journal={arXiv preprint arXiv:2602.21778},
  year={2026}
}