Classifier-free diffusion guidance is a method that allows diffusion models to trade off sample quality and diversity without requiring an external classifier. Unlike traditional classifier guidance, which uses the gradient of an image classifier to improve sample quality, classifier-free guidance jointly trains a conditional and an unconditional diffusion model. By combining the score estimates from both models, it achieves a similar trade-off between sample quality and diversity as classifier guidance. The method involves training a conditional diffusion model and an unconditional diffusion model together. During sampling, the score estimates from both models are combined with a mixing weight to control the trade-off between sample quality and diversity. This approach eliminates the need for an external classifier, simplifying the training pipeline and reducing computational overhead. Classifier-free guidance has been shown to produce high-quality samples with a balance between fidelity and diversity. Experiments on ImageNet demonstrate that it can achieve a similar trade-off between FID and Inception scores as classifier guidance, with the best results achieved at higher guidance strengths. The method is also more efficient in terms of training and sampling, as it avoids the need to train an additional classifier. The key advantage of classifier-free guidance is its simplicity and efficiency. It requires only a minor modification to the training and sampling processes, making it easier to implement and more scalable. Additionally, it avoids the limitations of classifier guidance, such as the need for a pre-trained classifier and the potential inconsistency of classifiers derived from Bayes' rule. Overall, classifier-free guidance provides a powerful alternative to traditional classifier guidance, enabling high-quality sample generation without the need for an external classifier. This method has the potential to be widely applicable in various generative modeling tasks, offering a more efficient and flexible approach to balancing sample quality and diversity.Classifier-free diffusion guidance is a method that allows diffusion models to trade off sample quality and diversity without requiring an external classifier. Unlike traditional classifier guidance, which uses the gradient of an image classifier to improve sample quality, classifier-free guidance jointly trains a conditional and an unconditional diffusion model. By combining the score estimates from both models, it achieves a similar trade-off between sample quality and diversity as classifier guidance. The method involves training a conditional diffusion model and an unconditional diffusion model together. During sampling, the score estimates from both models are combined with a mixing weight to control the trade-off between sample quality and diversity. This approach eliminates the need for an external classifier, simplifying the training pipeline and reducing computational overhead. Classifier-free guidance has been shown to produce high-quality samples with a balance between fidelity and diversity. Experiments on ImageNet demonstrate that it can achieve a similar trade-off between FID and Inception scores as classifier guidance, with the best results achieved at higher guidance strengths. The method is also more efficient in terms of training and sampling, as it avoids the need to train an additional classifier. The key advantage of classifier-free guidance is its simplicity and efficiency. It requires only a minor modification to the training and sampling processes, making it easier to implement and more scalable. Additionally, it avoids the limitations of classifier guidance, such as the need for a pre-trained classifier and the potential inconsistency of classifiers derived from Bayes' rule. Overall, classifier-free guidance provides a powerful alternative to traditional classifier guidance, enabling high-quality sample generation without the need for an external classifier. This method has the potential to be widely applicable in various generative modeling tasks, offering a more efficient and flexible approach to balancing sample quality and diversity.