VarenWan

VarenWan contains synthetic equine turn-around sequences for animal appearance and avatar reconstruction, covering horses, zebras, and donkeys.

Category	Sequences
Horse	9,003
Zebra	3,788
Donkey	4,973
Total	17,764

Each stored sequence contains 49 RGB frames, 49 foreground masks, and a NumPy .npz file with VAREN parameters. The manuscript describes 48-frame sequences; this upload preserves the actual 49-frame files without dropping or resampling frames.

Contents

The dataset is distributed as 139 tar shards, usually containing 128 sequences each; the final shard contains 100 sequences. Sharding preserves the original sequence directory structure while avoiding millions of separate repository files.

data/
  varenwan-00000.tar
  ...
  varenwan-00138.tar
metadata/sequences.jsonl
dataset_info.json
upload_manifest.json
upload_status.json
restore.py

After extraction:

VarenWan/
  horse_00001/
    images/0000.jpg ... 0048.jpg
    masks/0000.png ... 0048.png
    horse_00001.npz
  donkey_donkey_00000/
    images/
    masks/
    donkey_donkey_00000.npz
  zebra_zebra_00000/
    images/
    masks/
    zebra_zebra_00000.npz

Typical annotation fields are pose, global_orient, shape, transl, and tail_scale. Two horse sequences also contain auxiliary preprocessing outputs; these files are preserved in their original relative locations.

metadata/sequences.jsonl maps every sequence to its species, frame count, original file count, byte size, and shard. upload_manifest.json records shard sizes and SHA-256 checksums.

Dataset viewer

The dataset viewer displays the sequence index from metadata/sequences.jsonl: one row per sequence, including species, frame and mask counts, file count, byte size, and archive shard. The sequences split is the complete index, not a train/validation/test partition.

The TAR files preserve the original sequence folders and are download archives, not WebDataset-format samples. Use the restoration instructions below to access the RGB frames, masks, and VAREN parameters.

Download and restore

With Python 3.12 or later, install the Hugging Face client, download restore.py from this repository, and run:

pip install huggingface_hub
python restore.py --repo-id luoxue-star/VarenWan --output-dir ./VarenWan

For a private repository, authenticate first with a token that can read it. The restoration script downloads, verifies, and extracts one shard at a time, removing temporary archives after extraction. Original RGB, mask, and annotation file contents are preserved. Hidden operating-system caches and transfer checkpoints are excluded.

Upload completeness

During the initial upload, shards may appear progressively. A complete release has status: complete in upload_status.json, 139 shard files, and 17,764 indexed sequences. Check that status before downloading the full release.