# Human Motion Diffusion Model

**Authors:** Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H. Bermano, Daniel Cohen-Or  
**arXiv:** [2209.14916](https://arxiv.org/abs/2209.14916)  
**Date:** 2022-09-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.14916v1/

## Key points

- The field of computer animation aims to achieve natural and expressive human motion generation
- Current generative solutions often lack quality and expressiveness
- The Motion Diffusion Model (MDM) is a promising solution for human motion generation
- MDM is a diffusion-based generative model tailored for human motion, incorporating insights from existing literature
- MDM supports various forms of conditioning and different generation tasks like text-to-motion and action-to-motion
- MDM achieves state-of-the-art results with user studies showing preference over real motions 42% of the time
- MDM excels in tasks such as inpainting gaps in motion sequences under textual conditions while maintaining semantic consistency, completion, and editing capabilities

## Summary

The field of computer animation strives to achieve natural and expressive human motion generation. This is a challenging task due to the complexity and diversity of possible motions, as well as human perceptual sensitivity to them. Current generative solutions often fall short in terms of quality and expressiveness. However, the Motion Diffusion Model (MDM) presents a promising solution for human motion generation. , , , , 

MDM is a classifier-free diffusion-based generative model specifically tailored for human motion. It incorporates insights from existing motion generation literature and utilizes a key design choice of predicting the sample rather than the noise in each diffusion step. This allows for the use of established geometric losses such as foot contact loss. , , , , <kd>Fidelity and diversity trade-off </ kd>

The MDM framework supports various forms of conditioning and different generation tasks including text-to-motion, action-to-motion, and unconditioned generation. It is trained in a classifier-free manner, enabling a balance between diversity and fidelity in generated motions. < kd >State-of-the-art results </ kd > , user studies indicating preference over real motions 42% of the time, outperforming existing models on benchmarks like HumanAct12 and UESTC - these are just some of the impressive achievements showcased by MDM in evaluations on benchmarks like HumanML3D and KIT. < kd >Text-to-motion tasks </ kd > , inpainting gaps in motion sequences under textual conditions while maintaining semantic consistency, completion and editing capabilities - MDM excels in all of these areas. Its ability to generate diverse and high-quality motions makes it a significant advancement in the field of computer animation. , , , , <kd>High-quality results </ kd>

In summary, the Motion Diffusion Model is a powerful and versatile approach to human motion generation that excels in quality across multiple tasks with minimal training requirements. Its ability to generate diverse and high-quality motions makes it a significant advancement in the field of computer animation.