# MobileDiffusion: Subsecond Text-to-Image Generation on Mobile Devices

**Authors:** Yang Zhao, Yanwu Xu, Zhisheng Xiao, Tingbo Hou  
**arXiv:** [2311.16567](https://arxiv.org/abs/2311.16567)  
**Date:** 2023-11-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.16567v1/

## Key points

- MobileDiffusion introduced by researchers Zhao, Xu, Xiao, and Hou
- Extensive optimizations in architecture and sampling techniques
- Reduced redundancy and enhanced computational efficiency without compromising image quality
- Employed distillation and diffusion-GAN finetuning techniques for 8-step and 1-step inference processes
- Achieved sub-second inference speed for generating high-quality $512\times512$ images on mobile devices
- Overcomes limitations in deploying text-to-image models on mobile platforms
- Establishes MobileDiffusion as a state-of-the-art solution for efficient text-to-image generation

## Summary

In the realm of text-to-image generation, the deployment of large-scale models on mobile devices has been hindered by their significant model size and slow inference speed. To address this challenge, a team of researchers led by Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou have introduced MobileDiffusion. This innovative text-to-image diffusion model is the result of extensive optimizations in both architecture and sampling techniques. Through meticulous examination of model architecture design, the team successfully reduced redundancy and enhanced computational efficiency while minimizing the model's parameter count. This was achieved without compromising on high image generation quality. Additionally, distillation and diffusion-GAN finetuning techniques were employed to enable 8-step and 1-step inference processes respectively. Empirical studies encompassing both quantitative and qualitative analyses showcased the effectiveness of these proposed techniques. Notably, MobileDiffusion achieved an impressive sub-second inference speed for generating high-quality $512\times512$ images on mobile devices—setting a new benchmark in the field. This groundbreaking work not only overcomes existing limitations in deploying text-to-image models on mobile platforms but also establishes MobileDiffusion as a state-of-the-art solution for efficient text-to-image generation. The contributions made by Zhao et al. pave the way for further advancements in mobile-based image synthesis technologies.