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.

This model is derived from Tencent HY-MOTION 1.0 and is released under the Tencent HY-MOTION 1.0 Community License Agreement (see LICENSE).

TACD · HY-Motion-Lite, 8 steps

Text-to-motion in 8 sampling steps, distilled from HY-Motion 1.0.

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

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

💻 Code: https://github.com/vkgo/TACD

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
Trained only from the teacher on the model's own samples, with no motion data

Usage

pip install torch "transformers>=4.51" safetensors einops numpy ftfy

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 with the right hand"],
                     duration=4.0, seed=0)   # up to 12 s, 30 fps
out.joints   # (1, 120, 22, 3) joint positions in metres
out.rot6d    # (1, 120, 22, 6) local joint rotations
out.transl   # (1, 120, 3)     root translation

The text encoders (Qwen/Qwen3-0.6B, openai/clip-vit-large-patch14) are downloaded on first use; for local copies, call model.load_text_encoders(llm_path=..., clip_path=...).

GitHub repository

github.com/vkgo/TACD has the code around this model:

humanoid/ drives a Unitree G1 from text: a resident server generates with this model, retargets with UMR and streams the motion to NVIDIA's SONIC controller; type a prompt, press Enter
tools/fresh_env_test.py checks the model in a new environment

git clone https://github.com/vkgo/TACD && cd TACD
python tools/fresh_env_test.py --repo weijinhuang/TACD-HY-Motion-Lite --out check.npz

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

License

Released under the Tencent HY-MOTION 1.0 Community License Agreement (LICENSE). 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}
}

Downloads last month: 48

Safetensors

Model size

0.5B params

Tensor type

F32

Model tree for weijinhuang/TACD-HY-Motion-Lite

Base model

tencent/HY-Motion-1.0

Finetuned

(1)

this model

Paper for weijinhuang/TACD-HY-Motion-Lite

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

Paper • 2610.02867 • Published 20 days ago • 4