# Diffusion Models Beat GANs on Image Synthesis

**Authors:** Prafulla Dhariwal, Alex Nichol  
**arXiv:** [2105.05233](https://arxiv.org/abs/2105.05233)  
**Date:** 2021-05-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2105.05233v1/

## Key points

- Authors Prafulla Dhariwal and Alex Nichol present groundbreaking advancement in generative models
- Diffusion models surpass current state-of-the-art generative models in image sample quality
- Meticulous exploration of model architectures through ablations enhances unconditional image synthesis significantly
- Introduction of a novel approach for conditional image synthesis by incorporating classifier guidance
- Impressive results with achieved Frechet Inception Distance (FID) scores: 2.97 at 128 x 128, 4.59 at 256 x 256, and 7.72 at 512 x 512 resolution on ImageNet
- Scores outperform current state-of-the-art generative models and match BigGAN-deep performance with just 25 forward passes per sample
- Integration of classifier guidance with upsampling diffusion models further improves FID to 3.85 on ImageNet at 512 x 512 resolution
- Code publicly available for reproducibility and further research: https://github.com/openai/guided-diffusion

## Summary

In their paper titled "Diffusion Models Beat GANs on Image Synthesis," authors Prafulla Dhariwal and Alex Nichol present a groundbreaking advancement in the field of generative models. They demonstrate that diffusion models can achieve image sample quality surpassing the current state-of-the-art generative models. Through a meticulous exploration of model architectures via a series of ablations, they enhance unconditional image synthesis significantly. Additionally, the authors introduce a novel approach for conditional image synthesis by incorporating classifier guidance. This method allows for the optimization of sample quality while efficiently balancing diversity and leveraging gradients from a classifier. The results are impressive, with an achieved Frechet Inception Distance (FID) of 2.97 on ImageNet at 128 x 128 resolution, 4.59 at 256 x 256 resolution, and 7.72 at 512 x 512 resolution. These scores outperform current state-of-the-art generative models and even match BigGAN-deep performance with just 25 forward passes per sample, showcasing superior distribution coverage. Furthermore, the integration of classifier guidance with upsampling diffusion models yields even better results, further improving FID to 3.85 on ImageNet at 512 x 512 resolution. The authors have made their code publicly available at https://github.com/openai/guided-diffusion for reproducibility and further research in this cutting-edge area of study. Overall, this research not only pushes the boundaries of image synthesis quality but also introduces innovative techniques that can potentially revolutionize the field of generative modeling.