Classifier-Free Guidance in Conditional Generative Models
An inference technique for conditional diffusion and flow-based generators that trades diversity for fidelity to the condition by extrapolating between a conditional and an unconditional prediction, using a single model trained with the condition randomly dropped. Covers the guidance scale, the combined conditional/unconditional training scheme, and its effect on sample quality.
This Concept is waiting for its first lesson!
An inference technique for conditional diffusion and flow-based generators that trades diversity for fidelity to the condition by extrapolating between a conditional and an unconditional prediction, using a single model trained with the condition randomly dropped. Covers the guidance scale, the combined conditional/unconditional training scheme, and its effect on sample quality.
Are you a teacher? Sign in to start contributing.
Sign In