UniMate: One Unified Model to Animate Diverse Skeletons

While automated rigging can generate 3D assets at scale, synthesizing motion across diverse skeleton topologies has remained bottlenecked by template constraints and per-skeleton fine-tuning. UniMate addresses this with a topology-aware diffusion transformer that integrates graph-Laplacian spectral rotary embeddings and graph-biased attention. Trained on UniML3D, a dataset of 13,006 sequences covering bipedal, quadrupedal, marine, and serpentine forms, it delivers text-driven motion across arbitrary rigs without test-time optimization.