TEDWB1k-preview

⚠️ This is the public preview of initialneil/TEDWB1k. The full dataset (1,431 TED-talk speakers, ~120 GB) is hosted at the gated repo above. This preview repo exists so the HuggingFace Dataset Viewer can render the per-subject thumbnails and tracking grids without going through the gated EULA flow.

If you want the full training data (frames + mattes + per-frame SMPL-X / FLAME tracking for all 1,431 subjects), go to the gated main repo, accept the agreement, and use load_tedwb1k.py.

What's in this preview

This repo is schema-identical to the gated main repo, but the heavy per-subject data (frames.tar, mattes.tar, tracking pickles) is included only for the 12 subjects in train_subset_x12 as a working sample. For the other 1,419 subjects, only the metadata and QC visualizations are present (so the viewer table still lists all 1,431 entries).

What	Subjects	Size
Per-split parquets with embedded source-frame thumbnails	1,431	~210 MB
metadata/previews/<id>.jpg (1024×1024 source frames)	1,431	~210 MB
metadata/ehm/<id>.jpg (full-res SMPL-X overlay grids)	1,431	~17.6 GB
metadata/flame/<id>.jpg (full-res FLAME overlay grids)	1,431	~8.0 GB
metadata/base/<id>.jpg (full-res PIXIE+Sapiens grids)	1,431	~5.2 GB
Per-subject heavy data (frames.tar + mattes.tar + tracking)	12	~540 MB
Total preview		~32 GB

The HF Dataset Viewer above renders 5 tabs (train, train_subset_x1, train_subset_x12, train_val, test) with one row per subject, the per-subject frame and shot counts, and a thumbnail of the first source frame. Each thumbnail is the actual shots_images/<id>/<first_shot>/000000.jpg that the tracker consumed.

Quick start

If you only want to play with one of the 12 sample subjects (no agreement required):

pip install huggingface_hub

python -c "
from huggingface_hub import snapshot_download
snapshot_download(
    'initialneil/TEDWB1k-preview',
    repo_type='dataset',
    allow_patterns='subjects/-2Dj9M71JAc/*',
    local_dir='./tedwb1k_x1',
)
"

That gives you tracking pickles + frames.tar + mattes.tar for one sample subject in a few seconds. To turn it into the 5-file bundle that HolisticAvatar's TrackedData expects, use the same load_tedwb1k.py from the main repo:

wget https://huggingface.co/datasets/initialneil/TEDWB1k/raw/main/load_tedwb1k.py
python load_tedwb1k.py --split train_subset_x1 --out ./tedwb1k_x1 \
    --repo_id initialneil/TEDWB1k-preview

For the full 1,361-subject training set, request access at the gated main repo.

Per-subject visualizations

Each of the 1,431 subjects has 4 standalone visualization files under metadata/:

• metadata/previews/<id>.jpg — clean 1024×1024 source frame (the first frame of the first shot). Embedded in the parquet preview column too.
• metadata/ehm/<id>.jpg — full-resolution SMPL-X overlay grid from the final tracking stage (large vertical contact sheet, ~13 MB).
• metadata/flame/<id>.jpg — FLAME overlay grid from the intermediate face-fitting stage (~6 MB).
• metadata/base/<id>.jpg — stage-1 PIXIE+Sapiens overlay grid (~4 MB).

You can fetch any single subject's visualization with one hf_hub_download call:

from huggingface_hub import hf_hub_download
path = hf_hub_download(
    'initialneil/TEDWB1k-preview',
    'metadata/ehm/05jJodDVJRQ.jpg',
    repo_type='dataset',
)

Splits

Same as the main repo:

Split	Subjects	Notes
train_subset_x1	1	tiny single-subject overfit (⊂ train)
train_subset_x12	12	small overfit (⊂ train) — the only subjects with downloadable heavy data in this preview
train_val	20	training monitor (⊂ train)
test	70	identity-disjoint evaluation
train	1,361	full training pool
total	1,431

train (1,361) and test (70) are identity-disjoint and together cover all 1,431 subjects. train_subset_x1, train_subset_x12, and train_val are subsets of train.

Why two repos?

HuggingFace's Dataset Viewer cannot render tabs/thumbnails for gated datasets — the worker that computes split names runs without a user identity and can't satisfy the gating EULA. The full TEDWB1k is gated for TED-content compliance, so to keep the viewer working we mirror the metadata + a 12-subject sample to this public preview repo.

For full discussion see this thread: https://discuss.huggingface.co/t/after-gated-user-access-was-enabled-the-huggingface-not-showing-dataset-viewer/157333

License

CC-BY-NC-ND 4.0 — same as the main repo. Non-commercial research use only. Attribution required. No derivatives — you may not distribute modified or remixed versions of this dataset.

The tracking parameters, JPG frames, and mattes are all derived works of TED talk videos that are themselves CC-BY-NC-ND on ted.com. This dataset matches the upstream license to remain compatible with TED's source restrictions.

Links

• Full gated dataset: https://huggingface.co/datasets/initialneil/TEDWB1k
• Tracking pipeline: https://github.com/initialneil/HolisticTracker
• HolisticAvatar (downstream model): https://github.com/initialneil/HolisticAvatar