# SDXS: Real-Time One-Step Latent Diffusion Models with Image Conditions

**Authors:** Yuda Song, Zehao Sun, Xuanwu Yin  
**arXiv:** [2403.16627](https://arxiv.org/abs/2403.16627)  
**Date:** 2024-03-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.16627v1/

## Key points

- Authors Yuda Song, Zehao Sun, and Xuanwu Yin discuss advancements in diffusion models for image generation
- Introduction of a dual approach focusing on model miniaturization and reducing sampling steps to decrease latency
- Leveraging knowledge distillation techniques to streamline U-Net and image decoder architectures
- Innovative one-step DM training technique using feature matching and score distillation
- Showcasing how these advancements enable real-time image generation with reduced latency, expanding possibilities for applications like image manipulation and synthesis

## Summary

, , , , 


In their paper titled "SDXS: Real-Time One-Step Latent Diffusion Models with Image Conditions," authors Yuda Song, Zehao Sun, and Xuanwu Yin discuss the recent advancements in diffusion models that have propelled them to the forefront of image generation. The authors introduce a novel dual approach that focuses on model miniaturization and reducing the number of sampling steps to significantly decrease model latency. This methodology leverages knowledge distillation techniques to streamline the U-Net and image decoder architectures, as well as introduces an innovative one-step DM training technique that utilizes feature matching and score distillation. The study showcases how advancements in diffusion models coupled with innovative methodologies can lead to real-time image generation with reduced latency, opening up new possibilities for various applications such as image manipulation and synthesis.