PanFlow Dataset

The PanFlow dataset supports the research presented in the paper PanFlow: Decoupled Motion Control for Panoramic Video Generation.

PanFlow is a novel framework for controllable 360° panoramic video generation that decouples motion input into two interpretable components: rotation flow and derotated flow. This dataset is a large-scale, motion-rich panoramic video dataset with frame-level pose and optical flow annotations, curated to enable precise motion control, produce loop-consistent panoramas, and support applications such as motion transfer and panoramic video editing.

Paper: https://huggingface.co/papers/2512.00832 Code: https://github.com/chengzhag/PanFlow Video Overview: https://www.youtube.com/watch?v=sFTWwlHjNtg

By conditioning diffusion on spherical-warped motion noise, PanFlow enables precise motion control, produces loop-consistent panoramas, and supports applications such as motion transfer:

and panoramic video editing:

Dataset Structure and Details

The PanFlow dataset provides camera pose annotations for 300k clips. It also includes pre-generated latent and noise cache for a filtered subset to speed up training.

The underlying video data is derived from the 360-1M dataset, which consists of YouTube videos licensed under CC BY 4.0.

Citation

If you use the PanFlow dataset in your research, please cite the original paper:

@inproceedings{zhang2025panflow,
  title={PanFlow: Decoupled Motion Control for Panoramic Video Generation},
  author={Zhang, Cheng and Liang, Hanwen and Chen, Donny Y and Wu, Qianyi and Plataniotis, Konstantinos N and Gambardella, Camilo Cruz and Cai, Jianfei},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}