UniMate targets animation across varied 3D skeletons without retraining

UniMate is presented as a unified diffusion transformer that generates animation for a rigged 3D asset based on a text prompt. Its stated scope is arbitrary skeletons, without a separate training run for each skeleton or optimization during inference. The description contrasts this with approaches that need skeleton-specific preparation.

The paper identifies a bottleneck between automatic rigging and usable motion: animation-ready 3D assets are becoming available at scale, while generating motion to drive them remains difficult. It characterizes existing learned animators as constrained by skeleton topology. Some use category-specific templates, while others require fine-tuning for each skeleton and reference motions during inference. For AI builders working with varied rigged assets, the proposed setup matters because it aims to avoid those skeleton-specific requirements.

The stated inputs are a rigged 3D asset and a text prompt. To use the described setup, a builder supplies those inputs; the source does not identify additional input requirements. The model uses a graph-aware attention bias, spectral RoPE based on the graph Laplacian, and a global topological conditioner pooled from the asset’s rest pose. These are the three components named in the release description.

The release also includes UniML3D, a dataset of 13,006 canonicalized motion sequences paired with text. Its listed coverage includes bipeds, quadrupeds, birds, marine animals, insects, snakes, and articulated rigid objects. That range describes a dataset spanning different kinds of bodies and articulated structures, rather than only one character category. The source presents the dataset alongside the model as part of the work.

The description establishes the intended inputs, the model’s stated approach to varied skeletons, and the dataset’s size and coverage. It does not provide benchmark figures or report comparative animation performance. It also does not state animation speed or specify which asset formats are supported. Builders can therefore use the description to understand the proposed model design and data scope, but this source does not establish how well the approach performs in a particular production workflow.