MoRAE DiT-XL
Consolidated inference checkpoint for MoRAE: Flow-Friendly Self-Supervised Latents for Text-to-Motion Generation.
Authors: Yifei Zhu, Mingyi Shi, Yangyang Cai, Miao Cheng, Yoshifumi Kitamura, and Taku Komura.
The checkpoint contains exactly the components needed for text-to-motion inference:
- OpenAI CLIP ViT-B/32 text tower
- MoRAE DiT-XL EMA generator
- MoRAE feature decoder and essential-67 motion decoder
- latent and HumanML3D normalization statistics
It excludes all optimizer state, scheduler state, non-EMA generator weights, the JEPA encoder, the FAE encoder, evaluators, and training data.
Usage
git clone https://github.com/zhuyifeiabcd1/MoRAE.git
cd MoRAE
pip install -e .
python generate.py \
--text "A person walks forward, raises their right arm, then walks backward." \
--seconds 6 \
--output outputs/example
The code downloads morae_dit_xl.safetensors from this repository automatically.
Evaluation
The released HumanML3D checkpoint uses classifier-free guidance 8.5 and 24 Euler time points. Under the essential-dimension evaluator it obtains FID 0.0885 +/- 0.0037 over 20 repeats.
Output representation
The generator outputs HumanML3D essential-67 motion at 20 FPS. The public code also recovers 22-joint positions and renders a stick-figure MP4 or GIF.
License and attribution
The MoRAE release code is MIT licensed. The checkpoint includes the text-side parameters of the MIT-licensed OpenAI CLIP ViT-B/32 model. See the GitHub repository for full third-party notices.