Research & Innovation

TailorRig Generates a Working Skeleton for Characters Wearing Clothes, Not Just Characters

Accepted to SIGGRAPH Asia 2026, Tencent's auto-rigging system treats garments as part of the rig itself, generating auxiliary skeletons and hierarchical skinning weights for clothed humanoid models — and scored nearly three times better than the existing baseline in a user study.

Auto-rigging tools have gotten good at the body. TailorRig, a new system from Tencent’s LIGHTSPEED studio, Xi’an Jiaotong University, Peking University, and Tsinghua University, accepted to SIGGRAPH Asia 2026 and posted to arXiv November 5, tackles the part most of them skip: the character’s clothes. Instead of generating a skeleton for a bare mesh and leaving the garment to deform however the body’s bones happen to move it, TailorRig treats garment-aware rigging as its own generation problem from the start.

Clothes get their own bones

The system works by having an artist pick a body skeleton template and specify which auxiliary garment categories need rigs of their own — a loose jacket or a flowing skirt needs bones a tight t-shirt doesn’t. TailorRig then generates those auxiliary skeletons and assigns skinning weights hierarchically: garment-chain weights get resolved before the individual body bones underneath them, so cloth deformation and body deformation don’t fight each other during animation. It’s worth noting the system is bounded rather than universal — it supports a defined set of garment categories rather than an open-ended taxonomy of arbitrary clothing.

A dataset built to match the problem

To train and evaluate the approach, the team built Human-Mix, a dataset of more than 38,000 humanoid models — specifically constructed around clothed characters rather than adapted from a bare-mesh rigging dataset. That scale matters for a garment-aware task: most existing rigging datasets were built for unclothed bodies and simply don’t contain the kind of fabric-draping variety a system like this needs to generalize from.

How much faster it actually is

TailorRig’s own user study is specific about the gap: raters scored the system 4.56 out of 5, against 3.37 out of 5 for the existing baseline method, UniRig. The practical number may matter more than the rating — TailorRig’s output needed an average of 0.80 hours of manual repair to become production-usable, versus 1.38 hours for UniRig’s output. For a production pipeline processing many clothed characters, that’s roughly the difference between a rig an artist lightly cleans up and one that needs real reworking.