# Human Motion Diffusion Model

**Authors:** Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, Amit H. Bermano  
**arXiv:** [2209.14916](https://arxiv.org/abs/2209.14916)  
**Date:** 2022-09-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.14916v2/

## Key points

- Natural and expressive human motion generation is challenging in computer animation due to diversity, perceptual sensitivity, and difficulty in accurately describing it.
- Current generative solutions are low-quality or limited in expressiveness.
- Diffusion models have shown remarkable generative capabilities in other domains and are promising candidates for human motion due to their many-to-many nature.
- Motion Diffusion Model (MDM) is introduced as a carefully adapted classifier-free diffusion-based generative model for the human motion domain.
- MDM is transformer-based, combining insights from motion generation literature.
- MDM predicts the sample rather than the noise in each diffusion step, facilitating the use of established geometric losses on the locations and velocities of the motion such as foot contact loss.
- MDM achieves state-of-the-art quality in several motion generation tasks while requiring only about three days of training on a lightweight resource.
- In text-to-motion tasks, MDM generates coherent motions that achieve state-of-the-art results on leading benchmarks for text-to-motion and action-to-motion.
- Human evaluators prefer MDM generated motions over real motions 42% of the time.
- In action to motion tasks, MDM outperforms state of the art models designed specifically for this task on common benchmarks.
- MDM also demonstrates completion and editing by adapting diffusion image inpainting to set a motion prefix and suffix and using our model to fill in gaps under textual conditions while maintaining semantic input.
- By performing inpainting in joints space rather than temporally, MDM also demonstrates semantic editing of specific body parts without changing others.
- Overall, Motion Diffusion Model introduces a promising solution to generate high quality diverse human motions with controllability while being lightweight and efficient compared to existing methods.

## Summary

The generation of natural and expressive human motion is a challenging task in computer animation due to the diversity of possible motions, human perceptual sensitivity, and the difficulty of accurately describing it. Current generative solutions are either low-quality or limited in expressiveness. However, diffusion models have shown remarkable generative capabilities in other domains and are promising candidates for human motion due to their many-to-many nature. The Motion Diffusion Model (MDM) is introduced as a carefully adapted classifier-free diffusion-based generative model for the human motion domain. MDM is transformer-based, combining insights from motion generation literature. A notable design choice is the prediction of the sample rather than the noise in each diffusion step, facilitating the use of established geometric losses on the locations and velocities of the motion such as foot contact loss. MDM's generic approach enables different modes of conditioning and different generation tasks. The MDM framework achieves state-of-the-art quality in several motion generation tasks while requiring only about three days of training on a lightweight resource. In text-to-motion tasks, MDM generates coherent motions that achieve state-of-the-art results on leading benchmarks for text-to-motion and action-to-motion. In addition, human evaluators prefer MDM generated motions over real motions 42% of the time. In action to motion tasks, MDM outperforms state of the art models designed specifically for this task on common benchmarks. MDM also demonstrates completion and editing by adapting diffusion image inpainting to set a motion prefix and suffix and using our model to fill in gaps under textual conditions while maintaining semantic input. By performing inpainting in joints space rather than temporally, MDM also demonstrates semantic editing of specific body parts without changing others. Overall, Motion Diffusion Model introduces a promising solution to generate high quality diverse human motions with controllability while being lightweight and efficient compared to existing methods. The code can be found at https://github.com/GuyTevet/motiondiffusionmodel .