Skip to main content
Back

TailorRig Solves the Hardest Part of Rigging Stylized Game Characters

A SIGGRAPH Asia 2026 paper reframes garment-aware rigging as controllable generation, released alongside a 38,000-plus humanoid dataset.

  • TailorRig, a SIGGRAPH Asia 2026 paper, reframes garment-aware rigging as controllable generation. arXiv
  • Artists supply a body skeleton template plus semantic labels like skirt, hair, or cape.
  • An autoregressive decoder synthesizes category-specific auxiliary bone chains for each label.
  • A hierarchical chain-to-bone skinning scheme cuts weight bleeding between adjacent garment and body parts.
  • Human-Mix dataset of 38,088 rigged humanoids with a 21-class bone taxonomy is released on Hugging Face.
  • Outperforms Pinocchio and UniRig on garment-rich characters, with cleaner skeletons and less body-cloth coupling.

Rigging a stylized character with a flowing skirt, cape, long hair, and dangling accessories requires auxiliary bone chains, carefully isolated skinning weights, and a hierarchy that matches the studio’s animation pipeline. The TailorRig paper, accepted for SIGGRAPH Asia 2026 Conference Papers, proposes a learned system that generates those structures from artist controls. Its authors have also published the Human-Mix training dataset on Hugging Face.

TailorRig matters because geometry-driven auto-rigging often struggles with dense, flexible parts. The model accepts a body skeleton template and semantic categories such as skirt, hair, sleeve, cape, and accessory, then generates the requested auxiliary chains and skinning weights around that specification.

The pipeline in five lines

Question TailorRig’s approach
What goes in? A humanoid mesh, a body skeleton template, and artist-selected semantic categories.
What comes out? A full skeleton with category-specific auxiliary chains and skinning weights.
What stays controlled? The body template preserves compatibility with the intended animation rig.
How are weights assigned? Vertices are associated with a kinematic chain before weights are refined across bones in that chain.
What is public? The paper and Human-Mix dataset are available; the linked repository does not yet provide inference code or a model checkpoint.

Why garments break geometry-first rigs

Geometry-first auto-rigging systems infer a skeleton from the mesh. That approach works best when body shape provides a clear signal for major joints such as hips, knees, shoulders, and elbows. Skirts, capes, hair strands, and hanging accessories introduce many plausible bone layouts that geometry alone cannot resolve.

Garment-heavy characters also place unrelated surfaces close together. A skirt may sit beside both thighs, while a cape may overlap the spine and arms. A proximity-driven system can assign skirt vertices to a leg, merge a cape chain into the torso, or produce auxiliary bones without useful semantic grouping. Animation then exposes weight bleeding, unwanted body-garment coupling, and interpenetration.