Hugging Face researchers have unveiled UniMate, a unified foundation model that can animate arbitrary skeletons from a rigged 3D asset and a text prompt, eliminating the need for per-skeleton retraining or test-time optimization. The work addresses a key bottleneck in automatic rigging: while generating animation-ready 3D assets has scaled, producing the motion to drive them remains constrained by topology-specific models.
UniMate is built on a topology-aware diffusion transformer that integrates skeletal structure into the attention mechanism through three innovations: a graph-aware attention bias derived from pairwise joint relations and geodesic distances, a spectral rotary position embedding that generalizes RoPE to arbitrary kinematic trees via the graph Laplacian, and a global topological conditioner attention-pooled from the rest-pose skeleton.
To train the model, the team curated UniML3D, a dataset of 13,006 motion sequences spanning bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid objects, all unified through canonicalization and paired with text descriptions. This diversity enables UniMate to support zero-shot cross-topology transfer, in-betweening, expansion, and text-guided editing.
According to the paper, UniMate outperforms state-of-the-art baselines in quality, generalization, and efficiency. The project page and code are publicly available, and the paper has been recommended by the Semantic Scholar API alongside related works on neural motion blending and rigging.