stzhao/sekai_video

Sekai-Game videos (nanovvm)

The game subsets of Sekai (Li et al., Sekai: A Video Dataset towards World Exploration, arXiv:2506.15675), repackaged for video-model training with clip manifests. Source: Lixsp11/Sekai at revision 3cf08273b01fa2f6c7019f523c9383caf0583ffc (sekai-game-drone.zip, sekai-game-walking.zip.part_aa/_ab, SHA-256 checked against the source's LFS records).

License: non-commercial research use only, under the original Sekai license reproduced verbatim in LICENSE (its notice, conditions and disclaimer apply to this redistribution). No commercial use. Sekai-Real (YouTube) videos are not included.

• 932 drone clips (2.59 h, 300 frames each) and 1,618 walking clips (26.97 h, 1,800 frames each)
• HEVC, 1920x1080, 30 fps. The mp4 and camera npz files are byte-for-byte the members of the source zips (CRC-checked); nothing was re-encoded or cut. Every video decodes to exactly the frame range in its file name (<source>_<start>_<end>.mp4, end - start frames).

Layout

data/data.json                       one entry per video
data/<subset>-NNN.zip                ~10 GB per archive (stored); members <subset>/<clip>.mp4
camera/<subset>.zip                  members <subset>/<clip>.npz (normalized intrinsics + per-frame extrinsics)
manifests/<subset>/<setting>/        clip manifests (manifest.json + meta.yaml)
LICENSE                              the original Sekai license

<subset> is sekai-game-drone or sekai-game-walking. data.json fields: id, video_path, split (the subset), category (scene), camera_path, caption, location, scene, crowd_density, weather, time_of_day, frames, fps, width, height (annotations from the source train/<subset>.csv).

hf download stzhao/sekai_video --repo-type dataset --local-dir sekai_video
cd sekai_video/data && for z in *.zip; do unzip -q "$z"; done
cd ../camera && for z in *.zip; do unzip -q "$z"; done

After extraction each data/<subset>/ directory is a video root; manifest paths are relative to it.

Clip manifests

Every video is cut from frame 0 into non-overlapping windows of 1 + (rgb_seq_len - 1) * frame_interval frames; the incomplete tail is dropped. Clip counts:

Setting	sekai-game-drone	sekai-game-walking
cond1gen16_interval3	5,592	58,248
cond1gen15_interval3	5,592	63,102
cond1gen15_interval1	16,776	181,216

cond1gen16_interval3: 17 frames, interval 3 (span 49); cond1gen15_interval3: 16 frames, interval 3 (span 46); cond1gen15_interval1: 16 consecutive frames. Single split (train), as in the source.

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}
}