# Classifier-Free Diffusion Guidance

**Authors:** Jonathan Ho, Tim Salimans  
**arXiv:** [2207.12598](https://arxiv.org/abs/2207.12598)  
**Date:** 2022-07-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.12598v1/

## Key points

- Diffusion models are a powerful tool for generative modeling
- Achieving a balance between sample quality and diversity is a challenge with diffusion models
- Classifier guidance is a method that addresses this challenge by trading off mode coverage and sample fidelity in conditional diffusion models post-training
- Classifier-free guidance is proposed as an alternative approach to achieve guidance without using an image classifier
- The authors demonstrate that it is possible to perform guidance using only a pure generative model by jointly training both conditional and unconditional diffusion models
- The proposed method achieves comparable results on several benchmark datasets compared to traditional classifier-guided methods while eliminating the need for training an additional image classifier
- This approach could be useful in scenarios where there are limited resources or time constraints for training multiple models.

## Summary

In recent years, diffusion models have emerged as a powerful tool for generative modeling. However, one of the challenges with these models is achieving a balance between sample quality and diversity. Classifier guidance is a method that has been introduced to address this challenge by trading off mode coverage and sample fidelity in conditional diffusion models post-training. This technique combines the score estimate of a diffusion model with the gradient of an image classifier, which requires training an image classifier separate from the diffusion model. The question arises whether it is possible to achieve guidance without using a classifier. In their paper "Classifier-Free Diffusion Guidance," Jonathan Ho and Tim Salimans demonstrate that it is indeed possible to perform guidance using only a pure generative model. They propose a new approach called "classifier-free guidance," where they jointly train both conditional and unconditional diffusion models. By combining the resulting score estimates from these models, they are able to attain a trade-off between sample quality and diversity similar to that obtained using classifier guidance. The authors show that their proposed method achieves comparable results on several benchmark datasets compared to traditional classifier-guided methods while eliminating the need for training an additional image classifier. The results suggest that this approach could be useful in scenarios where there are limited resources or time constraints for training multiple models. Overall, this paper presents an innovative solution to the problem of balancing sample quality and diversity in generative modeling using diffusion models. The proposed method offers an alternative way of performing guidance without relying on external classifiers, making it more accessible and efficient for practical applications.