UniMate Checkpoints

Pretrained checkpoints for UniMate: One Unified Model to Animate Diverse Skeletons (SIGGRAPH Asia 2026), a single text-conditioned flow-matching model that generates motion for arbitrary skeletons: animals, humanoids and rigged objects.

Project Page · Paper · Video · Code · Datasets · Interactive Demo

Models

Model	Training data	Architecture	Params	Steps	Status
unimate_uniml3d_f60_preview	UniML3D (Truebones ZOO, Mixamo, Objaverse-XL); 60-frame clips at 30 fps; skeletons with 5 to 60 joints	Graph attention with AdaLN text conditioning; 10 layers, width 512	74.1M	120k	Preview

unimate_uniml3d_f60_preview

A preview model, trained for about one day (23 hours on 6 NVIDIA H100 GPUs) with the released configuration configs/uniml3d_60frames_graph_adaln.json. It is intended as a reproducible reference point for the public code rather than as the final model.

Checkpoints are saved every 10k steps; checkpoint_step_120000.pt is the recommended one. Each checkpoint stores the model weights, their exponential moving average (EMA), and the optimizer and scheduler state, so training can be resumed from it. Inference uses the EMA weights.

Data version

The UniML3D dataset will continue to be updated. The models currently in this repository were trained on the current release of the data (September 2026); models trained on later data releases will be published here with their data version noted.

Roadmap

This repository will be updated as new models are released:

• a full UniMate model on UniML3D;
• a separate model for each dataset (Truebones ZOO, Mixamo, Objaverse-XL).

Repository layout

<model>/
  config.json                  resolved training configuration, read by inference
  dataset_stats.npy            feature normalization statistics
  checkpoints/
    checkpoint_step_<N>.pt     model, EMA, optimizer and scheduler state
  logs/                        TensorBoard training curves
  samples/
    step_<NNNNNN>/             motions generated at each checkpoint (step_000000: before training)
      <object_type>-<i>_fk.mp4   joints from forward kinematics of the predicted rotations
      <object_type>-<i>_ric.mp4  predicted joint positions

Usage

Set up the environment and the dataset features as described in the code repository. Sampling reads the target skeleton from dataset/features/<dataset>/, so the features the model was trained on must be present.

Download a model (here the recommended checkpoint only):

hf download Linzhan/UniMate \
    --include "unimate_uniml3d_f60_preview/*.json" "unimate_uniml3d_f60_preview/*.npy" \
              "unimate_uniml3d_f60_preview/checkpoints/checkpoint_step_120000.pt" \
    --local-dir outputs

Generate motion from text (the latest checkpoint in the folder is used):

python -m unimate.inference.sample \
    --exp_dir outputs/unimate_uniml3d_f60_preview \
    --test_cases_json test_cases.json \
    --num_repetitions 3

Resume training from a checkpoint:

bash scripts/run_train.sh configs/uniml3d_60frames_graph_adaln.json -- \
    --resume outputs/unimate_uniml3d_f60_preview/checkpoints/checkpoint_step_120000.pt

Reproduce the preview model:

accelerate launch --num_processes 6 -m unimate.training.train \
    --config configs/uniml3d_60frames_graph_adaln.json

License

The UniMate code is released under the MIT License. The training data remain governed by the licenses of their original sources: Adobe's Mixamo terms of use, the per-object licenses of Objaverse-XL, and the commercial license of the Truebones ZOO pack. Please review these terms before using the models.

Citation

@article{mou2026unimate,
  title   = {UniMate: One Unified Model to Animate Diverse Skeletons},
  author  = {Mou, Linzhan and Lei, Jiahui and Dou, Zhiyang and Cai, Chenyue and Song, Chaoyue and Finkelstein, Adam and Rusinkiewicz, Szymon},
  journal = {arXiv preprint arXiv:2609.05415},
  year    = {2026}
}