UniMate: One Unified Model to Animate Diverse Skeletons

UniMate is a topology-aware diffusion transformer that generates text-driven articulated motion for arbitrary rigged 3D skeletons without test-time optimization or per-skeleton retraining, trained on a 13,006-sequence dataset.

UniMate is a unified foundation model that animates arbitrary rigged 3D skeletons from a rigged asset and a text prompt, with no test-time optimization and no per-skeleton retraining or fine-tuning. Its architecture is a topology-aware diffusion transformer combining graph-aware attention bias built from joint relations and geodesic distances, a spectral rotary position embedding that generalizes RoPE to arbitrary kinematic trees via the graph Laplacian, and a topological conditioner. It was trained on UniML3D, a curated dataset of 13,006 motion sequences spanning bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid-object skeleton topologies. According to both reports, UniMate outperforms baselines (the arXiv report specifies state-of-the-art baselines) and supports zero-shot cross-topology transfer, in-betweening, and text-guided editing; the Hugging Face report additionally lists motion expansion as a supported capability. The sources agree on the model name, approach, and the 13,006-sequence dataset count; they differ only in minor wording, with the topological conditioner described as 'global' in one report and 'rest-pose' in the other, and the no-retraining claim phrased as 'per-skeleton retraining' versus 'per-skeleton fine-tuning'.

• UniMate animates arbitrary rigged 3D skeletons from an asset and a text prompt with no test-time optimization and no per-skeleton retraining/fine-tuning.
• Architecture: topology-aware diffusion transformer combining graph-aware attention bias (joint relations and geodesic distances), a spectral rotary position embedding generalizing RoPE to kinematic trees via the graph Laplacian, and a…
• Trained on UniML3D: 13,006 motion sequences spanning bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid-object skeleton topologies.
• Reported to outperform baselines; the arXiv report (2026-09-04) describes this as outperforming state-of-the-art baselines.
• Supports zero-shot cross-topology transfer, in-betweening, and text-guided editing; motion expansion is listed as a supported capability only in the Hugging Face report (2026-09-03).
• Source discrepancy: the topological conditioner is called a 'global topological conditioner' in the Hugging Face report and a 'rest-pose topological conditioner' in the arXiv report.