HumanRig: Learning Automatic Rigging for Humanoid Characters Using a Large‑Scale Dataset

HumanRig is the first large‑scale dataset specifically designed for automatic rigging of 3D humanoid characters. It contains 11,434 high‑quality AI‑generated T‑pose meshes that share a unified skeleton topology, making them directly usable in mainstream animation engines. The dataset also provides per‑vertex skinning weight matrices, 3D joint positions, frontal images with camera parameters, and 2D joint locations.

The lack of comprehensive, standardized datasets has been a major bottleneck for research on automated 3D character binding. Existing collections such as RigNet (≈1.7K meshes) suffer from limited scale, inconsistent skeleton topologies, and incomplete joint annotations, while datasets based on SMPL focus only on realistic human bodies and do not generalize to diverse character types.

Based on HumanRig, the authors propose a novel data‑driven automatic rigging framework. The pipeline consists of three core modules: (1) a Prior‑Guided Skeleton Estimator (PGSE) that projects 2D skeletal priors from frontal images into 3D space to obtain a coarse skeleton; (2) a U‑shaped Point Transformer that encodes mesh geometry without relying on edge information, improving robustness on complex AI‑generated meshes; and (3) a Mesh‑Skeleton Mutual Attention Network (MSMAN) that fuses mesh and skeleton features in a high‑level semantic space, enabling joint optimization of skeleton construction and skinning weight estimation.

The overall loss combines a mean‑squared‑error term for skeleton prediction and a KL‑divergence term for skinning weight estimation.

Extensive experiments demonstrate that the proposed method significantly outperforms prior approaches such as RigNet and NBS on both skeleton creation and skinning quality. Quantitative metrics and visual comparisons show superior joint accuracy, mesh deformation quality, and robustness across diverse body proportions, clothing, and accessories. The HumanRig dataset itself also exhibits clear advantages in scale, diversity, and skeleton consistency compared with existing benchmarks.

Key contributions are: (1) the HumanRig dataset, providing the largest and most diverse collection of humanoid meshes for rigging research; (2) the PGSE module that leverages 2D priors to simplify 3D skeleton learning; (3) the U‑shaped Point Transformer for effective mesh encoding; and (4) the MSMAN attention mechanism that jointly optimizes skeleton and skinning.

Future directions include fully automated character generation from textual or sketch inputs and end‑to‑end frameworks that produce both rigged models and animated sequences.