SeMoCo: Semantic Motion Tokens with Body-Structured Flow Matching
Introduction
SeMoCo couples a semantically supervised motion codec with body-structured flow matching for text-to-motion generation. An autoregressive model predicts a semantic code sequence from text, and a conditional flow generates continuous kinematic features for the root, torso, and limbs. The semantic codes and kinematic features are decoded together into motion.
This repository contains the pretrained HumanML3D and Ω-MotionVerse models for the main results. Each directory has the tokenizer, semantic model, six-part flow, and the normalization statistics used with those weights.
Model Files
| File | Contents |
|---|---|
humanml3d/tokenizer.pt |
HumanML3D codec |
humanml3d/semantic_ar.pt |
HumanML3D semantic model |
humanml3d/flow.pt |
HumanML3D six-part flow |
humanml3d/kin_stats.npz, humanml3d/feature_stats.npz |
Latent and Guo263 normalization |
omega/tokenizer.pt |
Ω-MotionVerse codec |
omega/semantic_ar.pt |
Ω-MotionVerse semantic model |
omega/flow.pt |
Ω-MotionVerse six-part flow |
omega/kin_stats.npz, omega/feature_stats.npz |
Latent and UMR499 normalization |
Evaluation
| Dataset | FID | R@1 | R@2 | R@3 | MM-Dist |
|---|---|---|---|---|---|
| HumanML3D | 0.0327 | 0.5561 | 0.7473 | 0.8365 | 2.750 |
| Ω-MotionVerse | 0.01238 | 0.7325 | 0.8143 | 0.8446 | 0.963 |
HumanML3D reports the mean of 20 repeats with the checkpoint held fixed. Ω-MotionVerse is one evaluation of the full test set.
Usage
hf download poisonousID/SeMoCo --local-dir checkpoints
Then generate with the configs from the code repository:
semoco-generate --config configs/humanml3d_inference.yaml \
--inputs data/prompts.npz --output runs/motion.npz
Use configs/omega_inference.yaml for Ω-MotionVerse.
License
Apache-2.0.
Citation
@article{semoco,
title = {SeMoCo: Semantic Motion Tokens with Body-Structured Flow Matching},
author = {OMEGA-team},
journal = {arXiv preprint arXiv:2608.24334},
year = {2026}
}