Chenyang Xu

We introduce AniMuse, a two-stage framework for text-driven animal mesh animation directly from raw meshes, without predefined skeletons, joint names, or manual rigging. At its core are Semantic Gaussian Bones (SGBs), a compact skeleton-free deformation representation decoded from a globally shared learnable query book and trained through explicit linear blend skinning with topology-aware mask-gated weights. The shared query book yields stable cross-instance bone slots, providing a mesh-native control space for text-conditioned generation and SGB-slot motion inpainting. A DiT-based diffusion model generates per-bone SE(3) trajectories from text and geometric bone latents, while allowing users to clamp selected SGB slots and inpaint the remaining full-body motion. On DeformingThings4D, our rig reduces bidirectional CD-L1 by 39% over the best neural skeleton baseline, and a forward-only variant achieves the lowest CD-L2 overall. On the out-of-domain AnimalML3D benchmark, AniMuse improves overall motion quality over skeleton-based and vertex-based baselines.

We reformulate 3D assembly as a joint problem combining assembly and generation. These two processes are mutually reinforcing: assembly provides part-level structural priors for generation, while generation injects holistic shape context that resolves ambiguities in assembly. We design a physically based simulation pipeline to model the fossilization process of bones and implement FractureBEM for modeling brittle fracture behavior.

We present GARF, a generalizable flow matching-based 3D reassembly method trained on 1.9 Million fractures, enabling precise real-world fragment pose alignment. Our method generalizes across diverse fracture patterns without category-specific priors.