← Back to feed
2026-09-04visiondatacommunity code

UniMate: One Unified Model to Animate Diverse Skeletons

Linzhan Mou, Jiahui Lei, Zhiyang Dou, Chenyue Cai, Chaoyue Song, Adam Finkelstein, Szymon Rusinkiewicz

PDF preview unavailable
Read on arXiv →

Key claim

UniMate enables efficient motion synthesis for diverse skeletons.

In plain English

Imagine you're creating a video game and need to animate characters with different body structures. Traditionally, animators rely on specific templates for each character type, which can be time-consuming and limits creativity. This approach often leads to issues like needing to fine-tune animations for each unique skeleton or using reference motions that may not fit well, which is what's called template rigidity. As a result, animators face bottlenecks when trying to produce diverse and dynamic movements efficiently.

To address these challenges, the authors developed UniMate, a unified foundation model that generates motion for any skeleton based on a 3D asset and a simple text prompt. The model leverages a topology-aware diffusion transformer that incorporates skeletal structure into its attention mechanisms, allowing it to understand and generate movements that are contextually appropriate for various body types. This means animators can create animations without the need for extensive retraining or optimization for each new character. Compared to previous methods, UniMate not only enhances the quality and efficiency of motion generation but also supports innovative features like zero-shot transfer across different skeleton types and text-guided editing, making it a significant step forward for creators in the animation space.

Novelty
8.5/10

UniMate introduces a novel approach to motion synthesis that integrates skeletal topology into a unified model.

Reliability
8.0/10

The evaluation against state-of-the-art baselines demonstrates solid performance improvements across multiple metrics.

GitHub1 repo
Friedrich-M/UniMateCommunity