OmniHuman Dataset

OmniHuman is a large-scale, human-centric audio-visual dataset for video understanding and generation. Each sample provides a multimodal annotation (sample_json): id-aware video captions, structured subject descriptions, speech transcripts (with timing, language, and emotion), audio captions, and basic video attributes (fps, duration, resolution). Per-frame body/hand tracking (tracking_npz) is also included.

Original videos are not redistributed. Instead, every sample includes its YouTube source_url together with clip_start_sec / clip_end_sec, so you can locate and download the exact source clip yourself.

What's included

All assets are stored as tar shards under archives/, each with an index CSV:

Asset	Archive	Description
sample_json	sample_json_part_*.tar.gz + sample_json_index.csv	Per-sample audio-visual annotation: captions, subjects, speech, audio
metadata	metadata_part_*.tar.gz + metadata_index.csv	JSONL index files for scanning and loading samples
tracking_npz	tracking_npz_part_*.tar + tracking_npz_index.csv	Per-frame SMPL/MANO body & hand tracking (.npz)

omnihuman/
├── README.md
├── scripts/        # extraction & utility scripts
└── archives/
    ├── sample_json_index.csv
    ├── sample_json_part_*.tar.gz
    ├── metadata_index.csv
    ├── metadata_part_*.tar.gz
    ├── tracking_npz_index.csv
    └── tracking_npz_part_*.tar

The train/ and test/ directories are reconstructed by extracting the archives.

sample_json content

sample_json/xxx.json is the per-sample audio-visual annotation. It typically contains:

1. Source — source_url (YouTube link), clip_start_sec, clip_end_sec.
2. Captions — English/Chinese captions, REF-linked variants (double-person case), replacement-text variants, and audio captions.
3. Speech — speaker language, transcript text, emotion, on/offscreen flags.
4. Subjects (output) — appearance, action, expression, position, subject type, main-subject flag.
5. Audio / video metadata — background audio fields, fps, duration, resolution.
6. Person & identity — person_id, matched identity (e.g. REF_1), face_id, frame span, audio alignment, blur/quality statistics.
7. Quality/consistency signals — e.g. semantic_consistency.

Download from Hugging Face

You need the full archives/ contents on disk before extraction.

Option A: git lfs

git lfs install
git clone https://huggingface.co/datasets/<HF_DATASET_ID>
cd <HF_DATASET_ID>
git lfs pull

Option B: huggingface-cli

pip install -U "huggingface_hub[cli]"
huggingface-cli login
huggingface-cli download <HF_DATASET_ID> --repo-type dataset --local-dir . --local-dir-use-symlinks False

Extract from archives

Run from the repo root (the directory containing archives/ and scripts/):

for asset in sample_json metadata tracking_npz; do
  python scripts/extract_asset_from_archives.py --repo-root . --asset "$asset" --all
done

Add --skip-existing to resume after an interruption.

Shards can also be unpacked with plain tar (member paths match the relpath column of the index):

tar xzf archives/sample_json_part_00000.tar.gz
tar xf  archives/tracking_npz_part_00000.tar

Dataset layout (after extraction)

The repo root contains train/ and test/ splits, each divided into single/ (single-person) and double/ (double-person) subsets:

omnihuman/
├── archives/
├── train/
│   ├── single/
│   │   ├── sample_json/
│   │   ├── metadata/
│   │   └── tracking_npz/
│   └── double/
│       └── ...
└── test/
    └── ...

Folder	Description
sample_json/	Per-sample audio-visual annotation (captions, subjects, speech, audio)
metadata/	JSONL index files for scanning and loading
tracking_npz/	Tracking .npz. For double/ samples both persons are in the same file.

Naming and sharding

• Sample name is derived from the original clip stem.
• Duplicate basenames are disambiguated with __dupXXXX.
• Metadata files are named <split>_<subset>-NNNNN.jsonl (e.g. train_single-00000.jsonl).
• Each shard contains up to 2000 samples.

Index CSV format

Each archives/<asset>_index.csv has columns:

relpath,archive,member,size_bytes

• relpath: repo-relative path restored on extraction.
• archive: repo-relative tar shard path.
• member: tar member path (same as relpath).
• size_bytes: original (uncompressed) file size.