# Simple diffusion: End-to-end diffusion for high resolution images

**Authors:** Emiel Hoogeboom, Jonathan Heek, Tim Salimans  
**arXiv:** [2301.11093](https://arxiv.org/abs/2301.11093)  
**Date:** 2023-01-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.11093v2/

## Key points

- The paper addresses the challenge of applying diffusion models to high resolution images
- Existing approaches focus on lower dimensional spaces or use cascades with multiple super-resolution levels
- The authors aim to improve denoising diffusion for high resolution images while keeping the model simple
- Four main findings are presented: adjusting noise schedule, scaling only a specific part of the architecture, adding dropout at specific locations, and using downsampling as an effective strategy
- By combining these techniques, state-of-the-art results on image generation are achieved without using sampling modifiers on ImageNet
- The paper provides valuable insights and practical recommendations for training denoising diffusion models for high resolution images

## Summary

The paper titled "Simple diffusion: End-to-end diffusion for high resolution images" addresses the challenge of applying diffusion models in pixel space to high resolution images. Existing approaches either focus on diffusion in lower dimensional spaces or use cascades with multiple super-resolution levels of generation. However, these approaches introduce additional complexity to the diffusion framework. In this study, the authors aim to improve denoising diffusion for high resolution images while keeping the model as simple as possible. They investigate how to train a standard denoising diffusion model on high resolution images and achieve performance comparable to alternate approaches. The paper presents four main findings. Firstly, the noise schedule should be adjusted specifically for high resolution images. Secondly, it is sufficient to scale only a particular part of the architecture instead of scaling the entire model. Thirdly, dropout should be added at specific locations in the architecture to enhance performance. Lastly, downsampling proves to be an effective strategy for avoiding high resolution feature maps. By combining these simple yet effective techniques, the authors achieve state-of-the-art results on image generation among diffusion models without using sampling modifiers on ImageNet. Overall, this paper provides valuable insights into training denoising diffusion models for high resolution images and offers practical recommendations for improving their performance while maintaining simplicity in the model architecture.