Organizations: LIGHTSPEED, Tencent, China · Xi’an Jiaotong University, China · Peking University, China · LIGHTSPEED, Tencent, United States · Tsinghua University, China
Humanoid rigging for modern characters with complex garments remains challenging due to the need for dense auxiliary skeletons and precise skinning control. Existing automatic methods primarily recover rigs from geometry, providing limited artist control over auxiliary articulation and often lacking explicit semantic or kinematic-chain structure for dense garment bones, leading to instability and weight bleeding. We present TailorRig, a framework that reformulates garment-aware humanoid rigging as a structured, controllable full-rig generation problem. This formulation shifts the goal from automatic rig recovery to artist-driven rig authoring, where both body-template compatibility and auxiliary articulation are explicit design choices. TailorRig exposes two explicit axes of artist control within a unified framework: a user-specified body skeleton template that preserves body-rig compatibility, and artist-provided semantic labels that drive an autoregressive generator to synthesize requested category-specific auxiliary bone chains. To robustly assign skinning weights under dense skeletal configurations, we further introduce a hierarchical chain-to-bone formulation: vertices are first associated to a kinematic chain, and bone weights are then refined within that chain, reducing weight bleeding between spatially adjacent but semantically unrelated structures. We evaluate TailorRig on a diverse set of characters and introduce Human-Mix, a large-scale dataset of over 38,000 humanoid models curated for garment-aware rigging. Experiments show that TailorRig improves skeleton accuracy, deformation stability, controllability, and practical usability over existing methods. Project page: https://xxxx0601.github.io/trig/ ; code: https://github.com/Tencent/TailorRig ; dataset: https://huggingface.co/datasets/tencent/TailorRig
Figures & tables
Figure 1. TailorRig turns artist controls into animation-ready garment-aware rigs. TailorRig accepts any body template and any subset of auxiliary category labels ( left ). Different label choices yield distinct full rigs ( middle , with chosen labels marked under each result), all driving stable animation ( right ). 3D characters courtesy of Adobe Mixamo ( Adobe Inc., 2021 ) . The teaser image presents a three-stage pipeline for character rigging.
Figure 2. Overview of TailorRig. Given a mesh, a body skeleton template, and auxiliary category labels, TailorRig produces an animation-ready rig in two stages. (1) Controllable skeleton generation : a Topology-Aware body decoder (TAMB) aligns the template to the mesh, and a Label-Aware autoregressive decoder synthesizes category-specific auxiliary chains, both use the same hierarchical skeleton-encoder architecture with stage-specific parameters. (2) Hierarchical chain-to-bone skinning : vertices are first assigned to kinematic chains and then refined into per-bone weights, suppressing cross-chain weight bleeding. Pipeline diagram of TailorRig showing two stages: controllable skeleton generation that aligns a body template and synthesizes auxiliary bone chains, followed by hierarchical chain-to-bone skinning weight prediction.
Source
#Raw
#Curated
#Aux.
Articulation-XL-v2 ( Song et al., 2025b )
48,635
22,533
4,060
HumanRig ( Chu et al., 2025 ) †
11,434
11,434
0
VRoid ∗ ( Isozaki et al., 2021 )
3,404
3,404
3,337
ModelResource ( Xu et al., 2020 )
2,703
717
427
Total (Human-Mix)
– ‡
38,088
7,824
Table 1. Source composition and curation statistics of Human-Mix . #Raw: characters collected before curation; #Curated: characters retained after curation; #Aux.: characters containing at least one auxiliary bone category. Table listing the source composition and curation statistics of the Human-Mix dataset.
Figure 3. Auxiliary-class distribution in Human-Mix. Bars report per-class counts across the 10 auxiliary classes among the 7,824 characters containing at least one auxiliary bone; the Body group (11 body kinematic-chain classes) covers all 38,088 characters and is reported in the legend. Bar chart showing the per-class distribution of the ten auxiliary bone categories in the Human-Mix dataset, alongside eleven body kinematic-chain classes that cover all characters.
Body Skeleton Alignment
Auxiliary Bone Generation
Skinning Weight
Method
J2J ↓
J2B ↓
B2B ↓
J2J ↓
J2B ↓
B2B ↓
L1↓
Def- L1↓
MIA ( Guo et al., 2025b )
0.065
0.053
0.031
–
–
–
14.43
1.56
TARig ( Ma and Zhang, 2023 )
0.061
0.045
0.034
–
–
–
28.25
2.68
NBS ( Li et al., 2021 )
0.141
0.121
0.111
–
–
–
62.86
28.52
RigNet ( Xu et al., 2020 )
0.147
0.111
0.100
0.141
0.127
0.118
8.56
2.14
MagicArti. ( Song et al., 2025b )
0.064
0.046
0.031
0.202
0.209
0.197
– ∥
– ∥
Table 2. Quantitative comparison across all rigging tasks in TailorRig. Values in Skinning Weight are scaled by 103 . Table of quantitative comparisons with baseline methods across all rigging tasks, reporting skeleton and skinning metrics.
Body Skeleton Alignment
Auxiliary Bone Generation
Skinning Weight
Variant
J2J ↓
J2B ↓
B2B ↓
J2J ↓
J2B ↓
B2B ↓
L1↓
Def- L1↓
Ours w/o Lsem
0.029
0.018
0.015
0.120
0.130
0.131
11.50
0.83
Ours w/o Hier. Encoder
0.026
0.015
0.013
0.096
0.106
0.102
18.60
1.88
Ours w/o TAMB Fusion
0.030
0.018
0.015
–
–
–
–
–
Ours w/o Lproj
0.031
0.020
0.017
–
–
–
–
–
Ours w/o Label-Aware Attn
–
–
–
0.107
0.126
0.125
–
–
Table 3. Ablation study on body skeleton alignment, auxiliary bone generation, and skinning quality. Values in Skinning Quality are scaled by 103 . Table of ablation results on body skeleton alignment, auxiliary bone generation, and skinning quality.
Figure 4. Body-template control. Different body templates control different rigs: a template without finger joints (left) yields a finger-less hand rig, while one with articulated fingers (right) yields an articulated one—producing distinct hand behaviors under the same motion. Side-by-side comparison of hand rigs generated from two body templates: a template without finger joints produces a finger-less hand rig, while one with articulated fingers produces an articulated hand, shown under the same driving motion.
Figure 5. Auxiliary-label control. With body skeleton and mesh fixed, different auxiliary labels control the generated bones and, under the same driving motion, lead to different animations across the skirt , accessory , and skirt+accessory rigs. Blue / red : body/auxiliary bones. Rendered comparisons showing how different auxiliary label selections (skirt, accessory, or both) change the generated bones and the resulting garment animation under the same motion, with body bones in blue and auxiliary bones in red.
Figure 6. Qualitative ablations on skeleton generation and skinning. Left: blue and green denote predicted and ground-truth skeletons, respectively. Removing semantic supervision causes both body and auxiliary bones to drift from their intended regions. Right: under articulated deformation, removing the chain-to-bone decomposition or semantic supervision introduces visible cross-part skinning bleeding and deformation artifacts. Two-panel qualitative ablation. Left: predicted skeletons in blue versus ground-truth skeletons in green, showing bone drift when semantic supervision is removed. Right: deformed meshes showing skinning artifacts when the chain-to-bone decomposition or semantic supervision is removed.
Method
Body J2J ↓
Aux J2J ↓
Skin L1 ↓
Qual. ↑
Repair Time ↓
Pinocchio ( Baran and Popović, 2007 )
0.142
0.274
0.010
2.23
3.48
Puppet. ( Song et al., 2025a )
0.082
0.226
0.033
2.98
2.70
UniRig ( Zhang et al., 2025 )
0.025
0.188
0.010
3.37
1.38
Ours
0.024
0.160
0.003
4.56
0.80
Table 4. In-the-wild production evaluation on 50 assets rated by 5 artists. Qual.: 1–5 Likert; Repair Time: estimated hours to production quality. Table summarizing the in-the-wild production evaluation with Likert quality ratings and estimated repair hours from five artists on fifty assets.
Figure 7. Skeleton comparison on garment-rich characters. Blue denotes body bones, and red denotes auxiliary or garment-related bones (manually colored for baselines when no explicit auxiliary labels are provided). Baselines often omit, merge, or mis-parent garment-related auxiliary bones, whereas TailorRig produces cleaner body–auxiliary separation under the specified labels and better matches the ground truth. Grid of garment-rich characters comparing skeletons from TailorRig and baseline rigging methods, with body bones in blue and auxiliary bones in red, highlighting cleaner body-auxiliary separation from TailorRig.
Figure 8. Skinning comparison under controlled auxiliary-bone rotations. All methods use the same ground-truth skeleton and are evaluated under the same auxiliary-bone rotations, isolating skinning-weight prediction from skeleton errors. Surface heat maps visualize skinning-weight error against the ground truth, with warmer colors indicating larger errors. TailorRig produces cleaner garment/accessory deformation, whereas baselines exhibit stronger body–auxiliary coupling and local artifacts. Deformation comparisons under identical auxiliary-bone rotations, with surface heat maps of skinning-weight error; TailorRig shows mostly cool colors (low error) while baselines show warmer error regions.
Figure 9. Frame-wise animation comparison after full rigging. After each method generates a complete skeleton and skinning weights, we drive all results with the same motion and render the same sampled frames. TailorRig better preserves garment motion relative to the body, closely matching the ground truth. In contrast, Pinocchio and UniRig exhibit garment–body coupling, and UniRig further shows leg–cloth interpenetration in the highlighted regions. Sampled animation frames driven by the same motion across rigging methods; TailorRig preserves garment motion close to the ground truth, while baselines show garment-body coupling and leg-cloth interpenetration.
BNRist, Department of Computer Science and Technology, Tsinghua University, Beijing, China · Zhili College, Tsinghua University, Beijing, China · VAST, Beijing, China