Hannah motion model
The text-to-motion model behind Hannah, Vanth Labs' local AI assistant with a voice, a 3D body and hands. Given the sentence Hannah is about to say (plus its spoken duration and an emotion), it produces SMPL-X body motion (55 joints, axis-angle, 30 fps) so that every sentence gets its own body language instead of a looping idle animation.
Two stages, both in this repo:
| File | Stage | Size |
|---|---|---|
vae/latest.pt |
Part-wise motion VAE (latent space) | 174 MB |
flow/latest.pt |
Latent flow-matching DiT, conditioned on word-level T5 embeddings, emotion and a motion prefix | 213 MB |
Code, training scripts and the serving module: github.com/Vanth-Labs/motion-model
(pip install "motionlab[serve] @ git+https://github.com/Vanth-Labs/motion-model.git@v0.1.1",
then python -m motionlab.serve; the server downloads these weights on first start).
License and provenance
Trained on BEAT2 (co-speech mocap, CC BY-NC 4.0)
and evaluated with SMPL-X. These weights inherit that license: non-commercial use only.
The text encoder is a frozen flan-t5-base, pulled from the Hub at runtime.
Made by Vanth Labs, Lima, Peru. Questions: hello@vanthlabs.org