# T2M-GPT: Generating Human Motion from Textual Descriptions with discrete Representations

**Authors:** Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen  
**arXiv:** [2301.06052](https://arxiv.org/abs/2301.06052)  
**Date:** 2023-01-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.06052v1/

## Key points

- Study titled "T2M-GPT: Generating Human Motion from Textual Descriptions with discrete Representations"
- Authors: Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen
- Introduces a conditional generative framework combining VQ-VAE and GPT for human motion synthesis
- Demonstrates efficacy through analyses on the HumanML3D dataset
- Identifies dataset size as a critical factor affecting performance
- Emphasizes the importance of dataset size in achieving high-quality results
- Provides valuable insights to the field of human motion synthesis
- Suggests avenues for future research in enhancing generative frameworks

## Summary

, , , , 

In the study titled "T2M-GPT: Generating Human Motion from Textual Descriptions with discrete Representations," authors Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen delve into a comprehensive exploration of a conditional generative framework for human motion synthesis. The T2M-GPT model combines Vector Quantised-Variational AutoEncoder (VQ-VAE) and Generative Pre-trained Transformer (GPT) to generate high-quality human motion from textual descriptions. Through detailed analyses on the HumanML3D dataset, the researchers demonstrate the efficacy of their approach and identify dataset size as a limiting factor affecting its performance. This study not only contributes valuable insights to the field but also highlights avenues for future research in enhancing generative frameworks for human motion synthesis. <break>

<break>

<break>

Keywords: , , , , <kd>Human motion synthesis</kbd>

In their study "T2M-GPT: Generating Human Motion from Textual Descriptions with discrete Representations," authors Jianrong Zhang et al. explore a conditional generative framework that combines Vector Quantised-Variational AutoEncoder (VQ-VAE) and Generative Pre-trained Transformer (GPT) to synthesize human motion from text. They showcase the effectiveness of their approach through detailed analyses on the HumanML3D dataset, highlighting the importance of dataset size in achieving high-quality results. This study not only provides valuable insights but also suggests avenues for future research in improving generative frameworks for human motion synthesis.