Classifier-free diffusion model guidance

Classifier-free guidance in diffusion models controls image generation, balancing creativity and predictability. It conditions the model on text or class embeddings, with a guidance scale determining prompt adherence. Higher scales increase fidelity but reduce diversity. This method improves upon earlier approaches by eliminating additional networks and allowing generalization to unseen classes. It trains conditional and unconditional models jointly, combining their scores to balance sample quality and diversity during inference.