SCRIPT: Scalable Diffusion Policy with Multi-stage Training for Language-driven Physics-Based Humanoid Control
Paper • 2605.22894 • Published
Robotics
Diffusers
Safetensors
humanoid-control
physics-based-animation
diffusion-policy
flow-matching
text-to-motion
isaac-gym
pip install -U diffusers transformers accelerate
import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Zhanglele/SCRIPT", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
The data and the weights are derived from AMASS and HumanML3D. To access them, you must have registered at https://amass.is.tue.mpg.de and accepted the AMASS license, and you may use them for non-commercial scientific research only.
Log in or Sign Up to review the conditions and access this model content.
SCRIPT
Weights and data of SCRIPT: Scalable Diffusion Policy with Multi-stage Training for Language-driven Physics-Based Humanoid Control (SIGGRAPH Asia 2026).
Paper · Project page · Code · Video
Files
| Path | Size | Contents |
|---|---|---|
checkpoints/stage1_humanml3d_0.5B/ |
1.9 GB | Stage I policy (0.5B, step 80k, EMA weights), with its normaliser, null text and scheduler |
checkpoints/evaluators/humanml3d_paper/ |
0.3 GB | the text-motion evaluator of the paper |
checkpoints/evaluators/humanml3d_retrained/ |
0.3 GB | the evaluator retrained with the released code |
data/humanml3d/window_154+4/ |
52 GB | tracked training and validation trajectories with CLIP text features, the index of training windows, the normaliser |
data/humanml3d/evaluate_data/test.h5 |
9.1 GB | the 11,468 tracked test trajectories: the ground truth of the evaluation |
data/humanml3d/test_motions/ |
4.5 GB | the test set in 8 shards, for the standing start and for the start from the reference motion |
data/humanml3d/texts.tar.gz |
18 MB | the captions |
data/humanml3d/null_text.pt |
0.4 MB | the CLIP features of the empty caption, for classifier-free guidance in training |
Usage
Accept the conditions on this page, log in with hf auth login, and download into the root of the
code:
hf download Zhanglele/SCRIPT --local-dir . # everything (68 GB)
hf download Zhanglele/SCRIPT --include "checkpoints/*" --local-dir . # only the weights (2.5 GB)
tar -xzf data/humanml3d/texts.tar.gz -C data/humanml3d
The files are then where the code's configs and commands expect them, e.g.
bash scripts/eval_humanml3d.sh checkpoints/stage1_humanml3d_0.5B checkpoints/evaluators/humanml3d_paper --keep-copies
| Evaluator | MM_Dist ↓ | R@1 ↑ | R@2 ↑ | R@3 ↑ | FID ↓ | Diversity |
|---|---|---|---|---|---|---|
humanml3d_paper (--keep-copies) |
2.020 | 0.484 | 0.650 | 0.738 | 0.137 | 1.460 |
humanml3d_retrained |
1.784 | 0.494 | 0.661 | 0.754 | 0.127 | 1.352 |
License
The motions come from AMASS and HumanAct12, cut into captioned clips by HumanML3D and tracked in simulation with PHC; the data is subject to the AMASS license. The weights are released under CC BY-NC 4.0. Both are for non-commercial research only.
Citation
@article{zhang2026script,
title = {SCRIPT: Scalable Diffusion Policy with Multi-stage Training for
Language-driven Physics-Based Humanoid Control},
author = {Zhang, Jingyan and Liang, Han and Zhang, Ruichi and Li, Bin and
Zhang, Juze and Chen, Xin and Wang, Jingya and Xu, Lan and Yu, Jingyi},
journal = {arXiv preprint arXiv:2605.22894},
year = {2026}
}
- Downloads last month
- -
# Gated model: Login with a HF token with gated access permission hf auth login