UniMate can animate any 3D skeleton from text

UniMate is an open-source animation foundation model that brings practically any rigged 3D asset to life directly from text. Most learned animators remain stubbornly locked to human bipedal templates, failing the moment you hand them a quadruped, a serpentine creature, or an articulated mechanical prop.

UniMate changes that dynamic. You hand it an already-rigged 3D mesh alongside a text prompt, and it synthesizes coherent joint motion in real time. It requires zero fine-tuning on the target skeleton and no test-time optimization.

What it does

The model handles an extraordinarily diverse range of skeletal topologies, spanning bipeds, quadrupeds, birds, marine life, insects, serpents, and multi-joint rigid objects. In demonstrations, it comfortably drives everything from Garfield and Gundam mechs to animated flowers and satellites.

Beyond basic text-to-motion generation, UniMate supports three distinct zero-shot editing workflows straight out of the box, requiring no auxiliary networks or retraining.

You can perform spatial motion editing by locking specific joints, such as keeping a head turn fixed while regenerating the legs to walk forward. It also performs motion in-betweening given start and end poses, as well as sequential expansion to stitch together multi-stage movements smoothly.

Under the hood and open-source availability

Instead of treating skeleton joints as static sequential vectors, UniMate relies on a topology-aware diffusion transformer. It incorporates skeletal trees directly into attention via graph-aware relational biases, rest-pose topological conditioning, and a spectral rotary position embedding generalized through the graph Laplacian.

To train the system, the researchers curated UniML3D, a massive dataset featuring 13,769 motion clips across 7,430 distinct skeletons sourced from Mixamo, Objaverse-XL, and Truebones. The training pipeline utilizes flow matching with geodesic rotation and smoothness losses.

Accepted to SIGGRAPH Asia 2026, the project is completely open source. The team has released the inference code, training pipelines, dataset processing scripts, and preview model checkpoints on Hugging Face. A preprocessing utility handles standard GLB, glTF, and FBX files directly.

Why it matters

For indie game developers, technical artists, and 3D animators, UniMate is a massive time saver. Rigging non-human creatures is relatively accessible today, but keyframing believable cycles for monsters, quadrupeds, or mechanical gadgets remains an agonizing bottleneck.

We think UniMate's biggest achievement is proving that motion models do not need bespoke architectures for every creature class. While complex multi-contact interactions can still produce glitches on unconventional rigs, it establishes a compelling foundation for general-purpose 3D character animation that is well worth spinning up locally.