Research map / 3D vision / Face and shape
HumanRig: Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces HumanRig, a 11,434-mesh AI-generated T-pose humanoid rigging dataset with Mixamo skeletons, and a rigging framework (PGSE + point-transformer mesh encoder + MSMAN) that beats RigNet and NBS.
Problem
Automatic rigging of humanoid characters is limited by the lack of large, standardized rigged datasets, and GNN-based rigging methods struggle with complex AI-generated meshes.
Approach
Use 2D skeleton priors from a front-view image, back-projected through camera rays onto the mesh to get a coarse 3D skeleton, then refine joints and skinning jointly with a point-transformer mesh encoder and mutual cross-attention between skeleton and mesh features.
Key results
- On the HumanRig test set, the full model reaches CD-J2J 0.0027 and skinning precision 0.9271 with L1 0.0610, versus 0.0110 CD-J2J without PGSE.
- Training on full HumanRig instead of RigNetv1-human improves RigNetv1-human test CD-J2J from 0.0274 to 0.0071.
- Deformation error on RigNetv1-human drops to 0.00198 vs 0.00252 for RigNet, and on HumanRig to 0.00084 vs 0.00224 for GVB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on full HumanRig rather than RigNetv1-human | HumanRig rigging framework · HumanRig test | CD-J2J: 0.0163 -> 0.0027 | Helped |
| Add PGSE (ablation) | HumanRig · HumanRig test | CD-J2J: 0.0110 -> 0.0027 | Helped |
| Add PGSE (ablation) | HumanRig · HumanRig test | Prec: 0.8363 -> 0.9271 | Helped |
| Add MSMAN (ablation) | HumanRig · HumanRig test | CD-J2B: 0.0076 -> 0.0025 | Helped |
| Add MSMAN (ablation) | HumanRig · HumanRig test | Prec: 0.9214 -> 0.9271 | Helped |
| Replace Point Transformer with GraphSAGE (ablation) | HumanRig mesh encoder · HumanRig test | CD-J2J: 0.0027 -> 0.0031 | Hurt |
| Replace Point Transformer with GraphTransformer (ablation) | HumanRig mesh encoder · HumanRig test | CD-J2J: 0.0027 -> 0.0030 | Hurt |
| Replace Point Transformer with GMEdgeNet (ablation) | HumanRig mesh encoder · HumanRig test | CD-J2J: 0.0027 -> 0.0028 | Hurt |
| Train with diverse head-to-body ratios rather than only five-head samples (ablation) | HumanRig · HumanRig head-to-body-ratio test subsets | rigging performance | Helped |
Surprising findings
- A model trained on a 1,729-sample random HumanRig subset outperforms one trained on the full RigNetv1-human set on skeleton metrics within the RigNetv1-human test set.
Limitations
- Skeleton template does not model small parts such as fingers.
- Only humanoid characters are handled; extension to quadrupeds or other objects is untested.
- RigNet cannot process meshes with more than about 5K vertices, so some comparisons are missing on HumanRig.
Experimental setup
Datasets: HumanRig RigNetv1 (human subset, 1,729 meshes) SMPL
Benchmarks: HumanRig test set RigNetv1-human test set
Metrics: CD-J2J CD-J2B CD-B2B skinning precision L1 norm deformation error
Baselines: RigNet NBS Heat Diffusion GVB GraphSAGE GraphTransformer GMEdgeNet
Builds on: RigNet RTMPose Point Transformer Pinocchio
Closest papers by meaning
- Puppeteer: Rig and Animate Your 3D Models (NeurIPS 2025 · similarity 0.88)
- Make-It-Animatable: An Efficient Framework for Authoring Animation-Ready 3D Characters (CVPR 2025 · similarity 0.88)
- DRiVE: Diffusion-based Rigging Empowers Generation of Versatile and Expressive Characters (CVPR 2025 · similarity 0.88)
- VINECS: Video-based Neural Character Skinning (CVPR 2024 · similarity 0.87)
- TapMo: Shape-aware Motion Generation of Skeleton-free Characters (ICLR 2024 · similarity 0.87)
- Rigel3D: Rig-aware Latents for Animation-Ready 3D Asset Generation (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.