This is the implementation of ShapeMove, a framework for generating body-shape-aware human motion from text. ShapeMove combines a quantized VAE with continuous shape conditioning and a pretrained language model to synthesize realistic, shape-aligned motions from natural language descriptions.

⚙️ Environment Setup

conda create -n shapemove python=3.10
conda activate shapemove
pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

⌛️ Download Data and Base Models

bash scripts/download_models.sh

This step will download our pretrained ShapeMove model trained with the AMASS dataset, the flan-t5-base language model, and the SMPL neutral model for visualization.

📐 Inference Model

bash scripts/demo.sh

The output motion and shape beta will be saved under outputs.

🎞️ Render Motion

Blender setup

Follow the setup steps in TEMOS.

After installing blender and required packages in the python environment of blender, run the following command to ensure installation:

blender --background --version

This should return Blender 2.93.18.

Render Meshes with Blender

# generate mesh with given beta and motion .npy file
python -m utils.mesh --dir [path/to/inference/output/folder]
# generate image from blender (with obj/ply file)
blender --background -noaudio --python utils/blender_render.py -- --mode=video --dir [path/to/mesh/folder]
# gather generated image and make video
python utils/visualization.py --dir [path/to/mesh/folder]

Citations

Shape My Moves (this work)

@article{shapemove,
    author    = {Liao, Ting-Hsuan and Zhou, Yi and Shen, Yu and Huang, Chun-Hao Paul and Mitra, Saayan and Huang, Jia-Bin and Bhattacharya, Uttaran},
    title     = {Shape My Moves: Text-Driven Shape-Aware Synthesis of Human Motions},
    journal   = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2025}
}

SMPL

Please follow the official instructions to cite SMPL.

AMASS

Please follow the official instructions to cite AMASS.

Licenses

Acknowledgments

Some great resources we benefit from: MotionGPT, T2M-GPT and text-to-motion.