A Comprehensive Pipeline for Large-Scale 3D Humanoid Rigging Dataset Creation

Automated 3D character rigging is vital for accelerating content creation in animation, games, and virtual environments. However, existing auto-rigging methods rely on scarce and inconsistent datasets, which are expensive to build because of the need for manual annotations and often reflect artist-specific rigging styles. These limitations hinder scalability and generalization. To address this issue, we propose a fully automated, text-driven rigging pipeline that eliminates the need for manual input. Our approach first generates a high-quality T-pose character image by using a text-to-image diffusion model. A Zero123-based novel view synthesis is then used to create multiview images of the same character. A 2D pose-estimation network extracts skeletal keypoints from each view, while a multiview image-to-mesh model reconstructs a detailed 3D mesh. Finally, a heat-diffusion-based skinning algorithm is employed to bind the 3D skeleton to the mesh, producing an animation-ready, deformable character. This pipeline enables scalable and consistent rigging across various character types, without relying on handcrafted rigs or artist-specific datasets. Our method offers a data-efficient solution for 3D character rigging, and its effectiveness is demonstrated through experiments involving diverse styles and poses. It provides a promising direction for democratizing rigging in data-constrained scenarios.