TACD: Distilling Efficient Text-to-Motion Models via Terminal Amplification Control

📄 Paper: https://arxiv.org/abs/2610.02867

🌐 Project page: https://vkgo.github.io/TACD/

🤗 Model: https://huggingface.co/weijinhuang/TACD-HY-Motion-Lite

Text-to-motion in 8 sampling steps, distilled on-policy from HY-Motion 1.0 without motion data.

• 268 ms per motion and 2.62 GB peak GPU memory in bf16, vs 2077 ms and 17.63 GB for the 50-step HY-Motion 1.0 teacher: 7.7× faster, 6.7× less memory (RTX L40, batch 1)
• 465M HY-Motion-1.0-Lite denoiser with a Qwen3-0.6B text encoder

Quick start

One Python 3.12 environment covers the model, the demo and the humanoid server:

git clone https://github.com/vkgo/TACD && cd TACD
pip install --index-url https://download.pytorch.org/whl/cu129 torch==2.8.0   # or the CPU build of torch 2.8.0
pip install -r requirements.txt
hf auth login   # the model is gated: accept its licence on the model page first

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("weijinhuang/TACD-HY-Motion-Lite", trust_remote_code=True)
model = model.to("cuda").eval()   # or .to("cuda", torch.bfloat16) for bf16

out = model.generate(["a person walks forward, turns around and waves"], duration=4.0, seed=0)
out.joints   # (1, 120, 22, 3) joint positions in metres, 30 fps
out.rot6d    # (1, 120, 22, 6) local joint rotations
out.transl   # (1, 120, 3)     root translation

Visualize

Web demo. Type a sentence, set the duration, and the motion plays in a 3D viewer (drag to rotate, scroll to zoom); several seeds show several motions side by side.

python demo/app.py            # open http://localhost:7860

--share adds a temporary public link, --host 0.0.0.0 serves other machines, --device cpu runs without a GPU (about 20 s per 4-second motion on two CPU cores).

HTML file. Write motions to one HTML file and open it in a browser:

python demo/visualize.py "a person walks forward, turns around and waves" --duration 5 --seeds 0,1 --out walk.html

or, for motions generated in your own code:

import sys; sys.path.insert(0, "demo")
from visualize import save_html

out = model.generate(["a person jumps twice"], duration=4.0, seed=[0, 1])
save_html(out, "a person jumps twice", "jump.html")

The viewer loads three.js and the character model from public CDNs, so the browser needs an internet connection.

Repository layout

path	contents
hf_repo/	code, config and model card of the Hugging Face model (weights are on the Hub)
demo/	web demo and HTML export with a 3D viewer
humanoid/	text-to-humanoid application: TACD + retargeting + streaming to a G1 controller
tools/	smoke test of the Hugging Face model in a fresh environment

Results

HumanML3D test set:

	R@1 ↑	R@3 ↑	MM-Dist ↓	FID ↓
HY-Motion 1.0 teacher @ 8 steps	0.378	0.652	4.225	3.250
TACD @ 8 steps	0.479	0.782	3.192	1.041

Humanoid

humanoid/ drives a Unitree G1 from text: a resident server generates with TACD, retargets with UMR and streams the motion to NVIDIA's SONIC controller; type a prompt, press Enter. See humanoid/README.md.

License

The code in this repository is released under the Apache License 2.0, except:

• the model weights and the files in hf_repo/, which contain Tencent HY-Motion 1.0 code, are released under the Tencent HY-MOTION 1.0 Community License Agreement (hf_repo/LICENSE);
• demo/ contains code from the Tencent HY-Motion-1.0 Space, under the same Tencent agreement (demo/NOTICE);
• third-party files in humanoid/ keep their licences (humanoid/THIRD_PARTY.md).

Tencent HY-MOTION 1.0 is licensed under the Tencent HY-MOTION 1.0 Community License Agreement, Copyright © 2025 Tencent. All Rights Reserved. The trademark rights of “Tencent HY” are owned by Tencent or its affiliate. Powered by Tencent HY.

Citation

@article{huang2026tacd,
  title   = {TACD: Distilling Efficient Text-to-Motion Models via Terminal Amplification Control},
  author  = {Huang, Wei-Jin and Li, Yuan-Ming and Lin, Kun-Yu and Luo, Wang and Zhu, Yinlin and
             Yu, Yue and Ye, Shenghao and Yuan, Junbin and Hong, Fa-Ting and Zhang, Qing and Zheng, Wei-Shi},
  journal = {arXiv preprint arXiv:2610.02867},
  year    = {2026}
}