README.md

metadata

pretty_name: VideoVAE Datasets
task_categories:
  - video-generation

VideoVAE-Datasets

Quality-filtered real-world videos for training video VAE decoders. Final training set: 84,335 videos (see training-metadata/).

training-metadata/     final qualified + deduplicated video list, curation results  <- start here
  stage_0/             pilot set: every OpenViD-HD video in part_48..part_55, unfiltered (local videos only)
OpenViD-HD/
  videos/{min132,min121-131,min81-120}/part-NNNN/shard-NNN.tar
  metadata/quality/part_N.json, index.json
  metadata/captions/OpenVidHD.csv, source.json   original OpenVidHD captions (join on video file name)
  latents/wan2.1/chunks-33frames/   Wan2.1 VAE latents (33-frame sections) + manifests/ + README.md
UltraVideo/
  videos/{min132,min121-131,min81-120}/clips_short_960/shard-NNN.tar
  videos/{min132,min121-131,min81-120}/clips_short_1920/shard-NNN.tar   1920x1080 versions, quality-passed only
  metadata/quality/clips_short_960.json, clips_short_1920.json, index.json
Miradata/
  videos/min132/miradata_v1/shard-NNN.tar
  metadata/quality/miradata_v1.json, index.json

Dataset	Source	Qualified	Final (after dedup)
OpenViD-HD	nkp37/OpenVid-1M, OpenVidHD parts	78,583	72,982
UltraVideo	APRIL-AIGC/UltraVideo, clips_short_960	11,143	9,698
MiraData	MiraData v1, real-world clips	1,868	1,655

• videos/ folders are frame-count ranges (original MP4 bytes, no trimming or re-encoding). Video TARs contain every qualified video; MiraData TARs also contain videos that failed quality. Select training videos with training-metadata/train.jsonl.
• metadata/quality/*.json hold per-video CPU quality measurements and gate results for every source video, passed or not. OpenViD-HD quality covers part_1, part_2 and part_10..part_60; quality-passed videos of part_48..part_55 are included in the TARs.
• UltraVideo clips_short_1920 holds the same clips as clips_short_960 at 1920x1080, quality-checked with the same gates; it is not part of the final training set. Passed rows in clips_short_1920.json carry hf_archive / hf_archive_member.
• All paths in metadata are relative to the dataset root (e.g. OpenViD-HD/videos/part_48/x.mp4 after extracting TARs as <root>/<Dataset>/<archive member>).
• Download everything except latents: hf download mingyang-wu/VideoVAE-Datasets --repo-type dataset --exclude "*/latents/*" --local-dir VideoVAE-Datasets

evaluation/

Reconstruction evaluation sets, one TAR each. Extract with tar -xf <archive> -C <root> to get <root>/<Dataset>/...; checksums are in evaluation/SHA256SUMS.

Archive	Contents	Source
Inter4K.tar	1,000 Inter4K MP4s at Inter4K/60fps/UHD/<id>.mp4, original files unmodified (3840×2160, ~60 fps, ~5 s)	Inter4K, Alexandros Stergiou and Ronald Poppe; fetched from the Hugging Face mirror Shilin-LU/Inter4k
WebVid10M-val-1k.tar	1,000 WebVid-10M validation videos	WebVid-10M
Large-Motion-100-recreated.tar	100 high-motion clips (80 WebVid + 20 Inter4K) chosen by optical flow; a recreation, not the original unreleased split	built in this project from the two sets above
MPI-Sintel.tar	Official MPI-Sintel frames with 5 training scenes held out as a custom validation split	MPI-Sintel

Inter4K license. Inter4K is © 2021 Alexandros Stergiou and is licensed under CC BY-NC-SA 4.0 (see the original licence.txt): non-commercial use only, with attribution, and adaptations under the same license. The files here are redistributed unmodified. The same terms apply to the 20 Inter4K clips in Large-Motion-100-recreated.tar. Cite:

@article{stergiou2021adapool,
  title   = {AdaPool: Exponential Adaptive Pooling for Information-Retaining Downsampling},
  author  = {Stergiou, Alexandros and Poppe, Ronald},
  journal = {arXiv preprint},
  year    = {2021}
}

The other evaluation sets keep the terms of their original sources.