TailorRig

TailorRig is Tencent LIGHTSPEED's garment-aware humanoid auto-rigging framework and companion data: it keeps a user-specified body skeleton template while generating controllable, structured auxiliary bones for clothing and accessories. Paper: https://arxiv.org/abs/2611.02902 ; code: https://github.com/Tencent/TailorRig ; dataset: https://huggingface.co/datasets/tencent/TailorRig .

Human-Mix / TailorRig dataset on HF (~39,073 skinned humanoid .npz samples, ~34 GB):

• subsets rignet 715, humanrig 11,419, articulation_xlv2 23,452, vroid 3,487 (in-house via VRoid Studio);
• shared schema for mixed training; body_joint_mask separates body vs clothing bones;
• targets auto-rigging, auto-skinning, joint-position regression, etc.

Method pitch: semantic labels → auxiliary bone chains plus hierarchical chain-to-bone skinning to cut deformation artifacts and weight bleeding.

Why it matters

Generic auto-riggers often ignore skirts, cloaks and other garment DOFs. TailorRig treats clothing motion as first-class and ships a large public mesh corpus.

License caveat

Dataset uses the TailorRig License (Tencent, 2026): academic research only — no commercial/production use; keep notices on redistribution. Read LICENSE before use.

Limits

• HF Dataset Viewer blankness is expected (npz-inside-zip);
• Normalize coordinates before mixing subsets;
• Paper numbers and code maturity need independent checks.

Who it's for

• Teams working on character rigging, skinning, digital-human / game asset pipelines;
• Academic projects needing large humanoid sets with auxiliary bones (non-commercial license);
• People following Tencent LIGHTSPEED 3D / rigging work.

Getting started

1. Read arXiv 2611.02902 and https://github.com/Tencent/TailorRig ;
2. Pull subset zips via hf_hub_download("tencent/TailorRig", ...);
3. Rebuild samples with numpy + COO skinning weights per the card;
4. Filter body_joint_mask=False clothing bones before topology learning.