UniMate

UniMate

Generates motion for skeletons of any shape, from animals to humanoids to rigged objects, from a text description. Suits animation researchers and technical artists, under a non-commercial license.

Made by Linzhan on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: UniMate turns a text prompt into motion for a skeleton of any shape, whether an animal, a humanoid or a rigged object. It is built on flow matching. This repo holds its pretrained checkpoints. The paper behind it is titled One Unified Model to Animate Diverse Skeletons.

Who it suits: Animation researchers and technical artists who want to turn a written motion description into movement for a rigged character or creature.

Industry | Who uses it | How they use it | Example
Animation | Technical artists | Generate motion for a rigged character from text | Make a quadruped trot forward
Games | Prototype teams | Test movement ideas on varied skeletons | Animate an eagle or a shark asset
Research | Motion researchers | Study one model across many skeleton types | Compare checkpoints on the same prompt

If you need to use it commercially, you can skip it, since the license is non-commercial.

Where to use it:

• The UniMate code repository, in its own Python environment
• Rigged GLB, glTF or FBX assets after the card's preprocessing step
• Windows, macOS or Linux, for downloading the files

Use case | Who | What they get back
Text to motion | Animators | 60 frames at 30 fps of joint motion
Animate a mesh | Technical artists | One GLB or FBX per motion
Mixamo or Truebones rigs | Researchers | Dedicated checkpoints for those skeleton sets

How to set it up:

• Python with pip
• The separate UniMate code repository and its environment
• The checkpoint you want, recommended is unimate_uniml3d_f60_v3

1. Install the Hugging Face hub package.
2. Download the repo, or only the model folder you need.
3. From the code repository's root, run the generation script on an asset.

pip install -U huggingface_hub
hf download Linzhan/UniMate

How to use it: From the code repository's root, download one model, generate motion for an example asset with a text prompt, then drive the asset's mesh with the motions.

hf download Linzhan/UniMate --repo-type model --local-dir outputs \
    --include "unimate_uniml3d_f60_v3/*.json" --include "unimate_uniml3d_f60_v3/*.npy" \
    --include "unimate_uniml3d_f60_v3/checkpoints/checkpoint_step_150000.pt"

python -m unimate.inference.sample --exp_dir outputs/unimate_uniml3d_f60_v3 --asset assets/examples/unitree_go2 \
    --prompt "An object trots forward." "An object rears up on its hind legs." --num_repetitions 3 \
    --output_dir outputs/samples/custom

bash scripts/run_animate_motion.sh outputs/samples/custom

Tool | What it does
unimate.inference.sample | Generates motion from text for an asset
scripts/run_animate_motion.sh | Drives each asset's mesh with its motions
data_process.rig_preprocess | Labels the joints of your own rigged file

Example: You run the sample command on the example quadruped asset with the prompts that it trots forward and that it rears up. You get three repetitions of each, and the last script turns each motion into a mesh file.

For newcomers and experts: If you are new, run the three commands on the shipped example asset first. Then try your own rig through the preprocessing step, which stops so you can check the joint labels.

The recommended model is a transformer denoiser with graph attention and text conditioning through AdaLN, trained with flow matching on skeletons of 5 to 70 joints. Text comes from a frozen flan-t5-base encoder, downloaded on first use. Output is position, rotation relative to the T-pose and velocity for each joint.

A text prompt and a skeleton go in
flan-t5-base encodes the text
The denoiser runs flow matching
60 frames of joint motion come out
A script drives the mesh

Good to know:

• The Mixamo checkpoint works for one rig only, the 22-joint Mixamo skeleton
• The Truebones model covers 73 of 74 species, all but Dragon
• License: CC BY-NC 4.0, so non-commercial use only
• Cost to run: none named, you supply the hardware
• Data it reaches: your prompts and rigged assets, and the files downloaded from Hugging Face
• Not for: commercial use, under the non-commercial license

Title: Denoiser parameters in millions
unimate_uniml3d_f60_v3: 74.1 million
unimate_uniml3d_f60_v3_cross_attn: 66.2 million
unimate_mixamo_f60_v2: 47.8 million
unimate_truebones_f60_v2: 47.8 million

On Hugging Face, 13 November 2026

• Library: pytorch
• Downloads in the last 30 days: 533
• Likes: 97
• License: cc-by-nc-4.0
• Languages: en
• Quantized versions: 1
• Fine-tunes: 3
• Tags: pytorch, motion-generation, text-to-motion, character-animation, skeletal-animation, flow-matching, unimate

Links

• Model: https://huggingface.co/Linzhan/UniMate
• Files: https://huggingface.co/Linzhan/UniMate/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Linzhan/UniMate

Then follow the usage in its model card. Nothing in this kit runs by itself.