Generative 3D Rigging
Learning skeletons, skinning, and semantic correspondence for controllable 3D assets.
Generated 3D assets become substantially more useful when they can move. My work in this direction studies how to infer controllable structure—joints, skeletal connectivity, skinning, and semantic correspondence—from diverse 3D geometry.
ARMO unifies joint and connectivity prediction in an autoregressive skeleton-generation framework and introduces OmniRig, a multi-category dataset with 79,499 rigged meshes. DRiVE combines 3D Gaussian character generation with diffusion-based joint prediction and the AnimeRig dataset to animate complex clothing and hair. SRIF uses diffusion image morphing and flow estimation to recover dense semantic shape correspondence.
