# Make-An-Animation: Large-Scale Text-conditional 3D Human Motion Generation

**Authors:** Samaneh Azadi, Akbar Shah, Thomas Hayes, Devi Parikh, Sonal Gupta  
**arXiv:** [2305.09662](https://arxiv.org/abs/2305.09662)  
**Date:** 2023-05-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.09662v1/

## Key points

- Make-An-Animation is a text-conditioned human motion generation model
- Developed to improve the quality of generated motions in animation and robotics applications
- Diffusion models have improved the quality of generated motions, but existing approaches are limited by small-scale motion capture data
- Make-An-Animation has been trained on large-scale image-text datasets in two stages:
- First stage: trained on curated dataset of (text, static pseudo-pose) pairs extracted from image-text datasets
- Second stage: fine-tuned on motion capture data by adding additional layers to model the temporal dimension.

## Summary

Make-An-Animation is a text-conditioned human motion generation model that has been developed to improve the quality of generated motions in applications spanning animation and robotics. The recent application of diffusion models for motion generation has enabled significant improvements in the quality of generated motions. However, existing approaches are limited by their reliance on relatively small-scale motion capture data, leading to poor performance on more diverse, in-the-wild prompts. To address this limitation, Make-An-Animation has been trained on large-scale image-text datasets in two stages. In the first stage, the model is trained on a curated dataset of (text, static pseudo-pose) pairs extracted from image-text datasets. In the second stage, it is fine-tuned on motion capture data by adding additional layers to model the temporal dimension.