A topology-aware diffusion Transformer encodes joint relationships, graph distances, spectral positions, and the rest-pose skeleton. Training uses UniML3D, a collection of 13,006 text-paired motion sequences. During sampling, selected motion tokens can be held fixed to support editing, interpolation, or continuation.
UniMate is useful for animation research and workflows with heterogeneous character rigs. Training and inference code and dataset resources are public, but pretrained checkpoints and some reproducibility assets are marked as forthcoming. Some source motion assets are commercially licensed and cannot be redistributed with the dataset.
