SegMo: Segment-aligned Text to 3D Human Motion Generation (WACV 2026)

Code repository for the paper "SegMo: Segment-aligned Text to 3D Human Motion Generation".

Contents

1. Install
2. Run
3. Reference

Install

This code respository is based on MoMask, please refer to this repository for the installation of the environment.

Specifically, you will need to download the following data before running the code:

• HumanML3D dataset.
• KIT-ML dataset processed in HumanML3D format.
• Glove data used for the word vectorizer.
• Evaluators from MoMask.

Run

Train the Residual VQ-VAE

python train_rvq.py --exp-name vq_name

You can skip this step using the trained models provided by MoMask.

Train the Mask Transformer

python train_mtrans.py --exp-name mtrans_name --vq-name vq_name

Train the Residual Transformer

python train_rtrans.py --exp-name rtrans_name --vq-name vq_name --mtrans-name mtrans_name

Evaluation

python eval.py --vq-name vq_name --mtrans-name mtrans_name --rtrans-name rtrans_name

Reference

Our code is based on MoMask. Thanks to these authors for their great work.