Efficient-Large-Model/SANA-WM-example-training-dataset

SANA-WM Example Training Dataset

This repository contains the Sekai-Game training subset and LTX2 VAE latent cache used by the SANA-WM stage-1 example training configs.

License and Terms

This dataset is redistributed for non-commercial research use only under the same terms as the official Sekai project. See LICENSE and NOTICE.md.

The raw videos, camera trajectories, annotations, and derived VAE latent cache are all subject to the Sekai non-commercial research-only terms. Do not use this dataset, or derivatives of it, for commercial purposes unless you have obtained separate rights from the original rights holders.

Source

This is a redistributed/repacked SANA-WM training subset derived from the official Sekai project:

• Project/code: https://github.com/Lixsp11/sekai-codebase
• Official dataset hub: https://huggingface.co/datasets/Lixsp11/Sekai-Project
• Paper: https://arxiv.org/abs/2506.15675

This repository is not the official Sekai release.

Contents

data/
  sekai_game_train_961frames_16fps_ovl640/
    sekai_game_train_00000000.zip
    sekai_game_train_00000000_camera.npz
    sekai_game_train_00000000_*.json
    sekai_game_train_00000000_*.jsonl
  vae_cache/
    LTX2VAE_diffusers_704x1280/
      sekai_game_train_961frames_16fps_ovl640/
        sekai_game_train_00000000.zip

The raw zip contains 1,618 MP4 clips and 1,618 JSON sidecars. The VAE cache zip contains 1,618 NPZ latent files. See data_manifest.json for byte sizes and entry counts.

Use with SANA-WM Stage-1 Training

Download the dataset into the root of an NVlabs-Sana checkout so that the data/... paths line up with the example configs:

cd NVlabs-Sana
hf download Efficient-Large-Model/SANA-WM-example-training-dataset \
  --repo-type dataset \
  --local-dir .

Then run a SANA-WM stage-1 config that expects:

data:
  data_dir:
    sekai_game: data/sekai_game_train_961frames_16fps_ovl640
  vae_cache_dir: data/vae_cache/LTX2VAE_diffusers_704x1280/sekai_game_train_961frames_16fps_ovl640

If you download the dataset elsewhere, update those two config paths accordingly.

Citation

@article{li2025sekai,
  title={Sekai: A Video Dataset towards World Exploration},
  author={Zhen Li and Chuanhao Li and Xiaofeng Mao and Shaoheng Lin and Ming Li and Shitian Zhao and Zhaopan Xu and Xinyue Li and Yukang Feng and Jianwen Sun and Zizhen Li and Fanrui Zhang and Jiaxin Ai and Zhixiang Wang and Yuwei Wu and Tong He and Jiangmiao Pang and Yu Qiao and Yunde Jia and Kaipeng Zhang},
  journal={arXiv preprint arXiv:2506.15675},
  year={2025}
}