metadata
license: other
license_name: sekai-codebase-terms
license_link: https://github.com/Lixsp11/sekai-codebase
extra_gated_prompt: >-
This dataset contains annotations (camera poses, captions, clip manifests)
derived from the Sekai project. By requesting access you agree to comply with
the upstream sekai-codebase license terms
(https://github.com/Lixsp11/sekai-codebase) in addition to using the data for
academic, non-commercial research only.
extra_gated_fields:
Affiliation: text
I agree to the sekai-codebase terms and academic non-commercial use: checkbox
tags:
- video
- world-model
- camera-pose
- alayaworld
AlayaWorld v1.1 — Training Data (Sekai-Real-Walking-HQ annotations)
Annotations used by AlayaWorld training (18,208 clips): per-clip camera poses, captions and
manifests. Due to the hub's per-directory file limit, pose/ and caption/ ship as tarballs:
| File | Content |
|---|---|
sekai_real_hq.jsonl |
clip manifest (18,208 entries) |
caption.tar.gz |
per-clip captions (extract to caption/) |
pose.tar.gz |
per-clip camera poses (extract to pose/) |
tar -xzf caption.tar.gz && tar -xzf pose.tar.gz # restores the expected annotation layout
Raw videos are NOT included. Download the source clips via the upstream tooling: Lixsp11/sekai-codebase, then point the AlayaWorld configs at them:
paths:
video_base_dir: <your sekai video root>
annotation_base_dir: <this dataset download dir>
Training pipeline & data format: AlayaLab/AlayaWorld → docs/vigeo/README.md.