Classifier-Free Guidance
A technique that improves conditional generation quality by interpolating between conditional and unconditional predictions.
Overview
Classifier-free guidance (CFG) is a technique for improving the quality and adherence of conditional generation in diffusion models. During training, the conditioning signal (e.g., text prompt) is randomly dropped some percentage of the time. At inference, the model makes both conditional and unconditional predictions, and the final output extrapolates away from the unconditional prediction toward the conditional one.
Key Details
The guidance scale parameter controls the strength of conditioning — higher values produce images more closely matching the prompt but with less diversity. CFG has become standard in virtually all text-to-image systems (Stable Diffusion, DALL-E 3, Midjourney) and is key to their ability to follow text prompts faithfully.