Think Before You Move: Latent Motion Reasoning for Text-to-Motion Generation

Yijie Qian^1* · Juncheng Wang^2* · Yuxiang Feng¹ · Chao Xu³ · Wang Lu⁴ · Yang Liu^3,5 ·
Baigui Sun³ · Yiqiang Chen⁴ · Yong Liu^1† · Shujun Wang<sup>2†

1</sup>Zhejiang University ²The Hong Kong Polytechnic University ³IROOTECH TECHNOLOGY & Wolf 1069 b Lab, Sany Group
⁴Institute of Computing Technology, Chinese Academy of Sciences ⁵King's College London

🔥 News

• [2025-01] Inference code and pre-trained checkpoints released! 🎉
• [2025-01] Paper submitted to arXiv.

📌 TODO List

• Release inference code and checkpoints for T2M-GPT baseline
• Release training code and evaluation scripts for T2M-GPT baseline
• Release complete code and checkpoints for MotionStreamer baseline
• Release processed training datasets and full documentation

🎯 Overview

We propose Latent Motion Reasoning (LMR), a novel framework that reformulates text-to-motion generation as a hierarchical "Think-then-Act" process. Unlike existing methods that directly translate text to motion (System 1), LMR introduces an intermediate reasoning phase (System 2) that operates in a learned, motion-aligned latent space.

Key Contributions:

• Concept: We identify the Semantic-Kinematic Impedance Mismatch in T2M and propose Latent System 2 Reasoning as a solution
• Analysis: We reveal knowledge capacity and density properties in motion tokenization
• Method: We propose a Dual-Granularity Tokenizer that decouples motion planning from execution
• Performance: 71% and 64% FID reduction on HumanML3D and KIT-ML datasets for T2M-GPT baseline

🔧 Installation

Python Virtual Environment

conda create -n lmr python=3.10
conda activate lmr

Install PyTorch

pip install torch torchvision torchaudio

Install CLIP

pip install git+https://github.com/openai/CLIP.git

Install Other Dependencies

pip install -r requirements.txt

Hugging Face Mirror (Optional)

If Hugging Face is not directly accessible, you can use the HF-mirror:

pip install -U huggingface_hub
export HF_ENDPOINT=https://hf-mirror.com

📥 Model Checkpoints

Download the pre-trained model checkpoints:

hf download qianyijie/lmr --local-dir ./checkpoints/t2m/

The checkpoint structure should be:

./checkpoints/t2m
├── dual_tokenizer
│   ├── checkpoint.pth
│   ├── run.log
│   └── train_config.json
└── GPT
    ├── checkpoint.pth
    └── train_config.json

🚀 Demo Inference

Basic Text-to-Motion Generation

python generate.py --trans_checkpoint checkpoints/t2m/GPT/checkpoint.pth --text "a person walks forward and then turns around" --output_dir ./outputs

Using Shell Script

bash run_generation.sh

Generation Options

# Specify custom text
python generate.py --trans_checkpoint checkpoints/t2m/GPT/checkpoint.pth --text "your text description here"

# Adjust generation parameters
python generate.py \
  --trans_checkpoint checkpoints/t2m/GPT/checkpoint.pth \
  --text "a person jumps twice" \
  --num_samples 5 \
  --seed 42 \
  --guidance_scale 2.0

Output Format

Generated motions will be saved in the specified output directory:

• Motion files (.npy format)

📊 Main Results

HumanML3D Dataset

Method	R-Precision (Top-1) ↑	FID ↓	MM-Dist ↓	Diversity →
T2M-GPT	0.492	0.141	3.121	9.761
MoMask	0.521	0.045	2.958	-
BAMM	0.525	0.055	2.919	9.717
LMR (Ours)	0.537	0.040	2.895	9.668

KIT-ML Dataset

Method	R-Precision (Top-1) ↑	FID ↓	MM-Dist ↓	Diversity →
T2M-GPT	0.416	0.514	3.007	10.921
MoMask	0.433	0.204	2.779	-
BAMM	0.438	0.183	2.723	11.008
LMR (Ours)	0.483	0.181	2.636	11.032

🏗️ Framework Architecture

LMR consists of two main components:

1. Dual-Granularity Tokenizer: Disentangles motion into:

   • Reasoning Latent (compressed, semantically rich): For planning global topology
   • Execution Latent (high-frequency): For preserving physical fidelity

2. LMR-Generator: Hierarchical generation process:

   • Phase I (Reasoning): Autoregressively generates reasoning tokens
   • Phase II (Execution): Generates execution tokens conditioned on reasoning tokens

📖 Citation

If you find our work helpful, please consider citing:

@misc{qian2025thinkmovelatentmotion,
      title={Think Before You Move: Latent Motion Reasoning for Text-to-Motion Generation}, 
      author={Yijie Qian and Juncheng Wang and Yuxiang Feng and Chao Xu and Wang Lu and Yang Liu and Baigui Sun and Yiqiang Chen and Yong Liu and Shujun Wang},
      year={2025},
      eprint={2512.24100},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.24100}, 
}

🌹 Acknowledgement

This repository builds upon the following awesome projects:

• T2M-GPT
• HumanML3D
• MotionStreamer
• MoMask
• CLIP

📄 License

This project is released under the MIT License. The code is for academic purposes only.

⭐ Star History