Motius

An open framework for motion research

Generate.
Train. Evaluate.

Different methods.
A shared research workflow.

Run released motion models, train HYMotion, and compare results with explicit evaluation protocols. Connect their outputs across motion formats, skeletons, and characters.

Explore models Get started
01 Model inference 02 HYMotion training 03 Evaluation & benchmarks 04 Conversion & retargeting

01 / Models & tasks

Start with the task.
Choose the method.

Explore established and recent methods through task-oriented pipelines, with model cards, checkpoints, and rendered outputs.

Text → motionHYMotion
Motion from a description.

Input: “a person jumps with legs open while clapping with hands over head simultaneously.” HYMotion output, rendered as an SMPL mesh.

Model card ↗
Music → danceBailando
A different input. Another task.

AIST++ waacking music conditions Bailando's generated dance. This mesh preview is silent; audio-synchronized comparisons are available in the benchmark.

Benchmark ↗

Selected integrations, not the full catalog. Inference and training support are documented separately.

All models & tasks

02 / Inference & training

From a checkpoint
to your own experiment.

Spend less time reconciling separate repositories. Use documented task APIs for inference and explicit data contracts for training.

Run a released model

A task-oriented inference API.

Load a packaged checkpoint and generate motion. The HYMotion Lite example below produces a six-second sequence at its native 30 fps.

from motius import Pipeline

pipe = Pipeline.from_pretrained(
    "ZeyuLing/Motius-HYMotion-T2M-1.0-Lite",
    device="cuda",
)
out = pipe.infer_text_to_motion(
    ["a person practices tai chi"],
    num_frames=[180],
)
motion = out["latent"]

Requires model weights and a compatible GPU environment. The example is independent of the preview above.

Inference guide ↗

Train & fine-tune HYMotion

Your data. A repeatable recipe.

Use the native trainer with public configurations, distributed execution, checkpoints, and full-state resume.

# Linux / Bash · replace dataset paths
MOTIUS_DATA_ROOT=/path/to/hymotion201 \
MOTIUS_TRAIN_MANIFEST=train.json \
MOTIUS_MOTION_STATS=/path/to/stats \
bash tools/dist_train.sh \
  configs/hymotion_t2m/train_hymotion_t2m.py 8 \
  --work-dir outputs/training/hymotion_t2m \
  --auto-resume

Prepare HY-Motion-201 arrays, matching normalization statistics, and cached Qwen3/CLIP text features. This launch uses eight GPUs.

Data preparation & training ↗

03 / Evaluation & benchmarks

Compare methods.
Keep the protocol explicit.

A model name is not an evaluation setting. Keep the dataset split, representation, evaluator checkpoint, and scored predictions traceable.

Choose the evaluation space.

HumanML3D Official HumanML3D-263 · 20 fps MotionStreamer MotionStreamer-272 · 30 fps Universal TMR Motius-trained TMR reproduction · SMPL-22 joints · 30 fps

Universal TMR is trained on HYMotion Data SFT and the single-person MotionHub training union. It is not an official TMR checkpoint. Compare methods within a fixed protocol, not scores across different embedding spaces.

All evaluators & physical diagnostics ↗

Inspect measured results.

Open the leaderboard, then inspect the generated examples. Publication status distinguishes complete comparisons from metric-only settings.

Text-to-motionHumanML3D Sequential generationBABEL Temporal completionHumanML3D Music-to-danceAIST++
All benchmarks, protocols & reports ↗

04 / Motion conversion & character tools

Take the output further.

Bridge model representations, retarget motion to different skeletons, and export animated characters. A connected toolkit alongside the research workflow.

Motion representationsHumanML3D · SMPL · SOMA · ARDY · G1
Keep the motion. Change the representation.

The same source clip across skeletons, body models, and Unitree G1. Joint-only and cross-skeleton conversions may be approximate.

Conversion routes ↗
Character retargetingSMPL → Mixamo
One motion, four characters.

Amy, Maria, Michelle, and Remy from Adobe Mixamo, alongside the source skeleton and SMPL mesh.

FBX export ↗
Automatic riggingMesh → animated character
From an unrigged mesh to motion.

Three downloaded characters with textures and generated skeletons, driven by the same motion.

AutoRig guide ↗

AutoRig uses Make-It-Animatable, followed by Motius normalization and retargeting. Character credits, licenses & downloads.

Make it your workflow

Start your
next experiment.

Clone the framework, install the core, and choose a released model. Checkpoints and body-model assets are downloaded separately.

Read the quickstart
Install from source
git lfs install
git clone https://github.com/ZeyuLing/Motius.git
cd Motius
git lfs pull
python -m pip install -e ".[dev]"

Requires Python 3.10+ and Git LFS.