The paper introduces classifier-free guidance, a method to enhance sample quality in diffusion models without the need for an additional classifier. Classifier guidance, proposed by Dhariwal and Nichol, combines the score estimate of a diffusion model with the gradient of an image classifier to achieve a trade-off between sample quality and diversity. However, this approach requires training a separate classifier, which can be complex and limited by the quality of the classifier. Classifier-free guidance jointly trains both a conditional and an unconditional diffusion model, then combines their score estimates to achieve the same trade-off. This method does not require a classifier and can be implemented with a single neural network, making it simpler and more flexible. The authors demonstrate that classifier-free guidance can produce high-fidelity samples similar to those obtained with classifier guidance, while also avoiding the potential issues associated with classifier gradients. Experiments on ImageNet show that classifier-free guidance effectively trades off Inception scores (IS) and Fréchet Inception Distance (FID) scores, achieving competitive results and sometimes outperforming previous methods. The method is particularly effective at moderate levels of guidance strength, where both IS and FID scores are optimized. The paper also discusses the implications of classifier-free guidance and suggests future directions for improving sample diversity and performance in various applications.The paper introduces classifier-free guidance, a method to enhance sample quality in diffusion models without the need for an additional classifier. Classifier guidance, proposed by Dhariwal and Nichol, combines the score estimate of a diffusion model with the gradient of an image classifier to achieve a trade-off between sample quality and diversity. However, this approach requires training a separate classifier, which can be complex and limited by the quality of the classifier. Classifier-free guidance jointly trains both a conditional and an unconditional diffusion model, then combines their score estimates to achieve the same trade-off. This method does not require a classifier and can be implemented with a single neural network, making it simpler and more flexible. The authors demonstrate that classifier-free guidance can produce high-fidelity samples similar to those obtained with classifier guidance, while also avoiding the potential issues associated with classifier gradients. Experiments on ImageNet show that classifier-free guidance effectively trades off Inception scores (IS) and Fréchet Inception Distance (FID) scores, achieving competitive results and sometimes outperforming previous methods. The method is particularly effective at moderate levels of guidance strength, where both IS and FID scores are optimized. The paper also discusses the implications of classifier-free guidance and suggests future directions for improving sample diversity and performance in various applications.