A technique used in diffusion models to control how closely the generated image follows the text prompt. It works by interpolating between the predictions of a conditional model (prompted) and an unconditional model (unprompted), allowing users to trade off diversity for prompt adherence.
Introduced by Ho and Salimans (2022) to improve sample quality without needing a separate classifier.
The standard method controlling the 'CFG Scale' slider found in almost all AI art tools (Stable Diffusion, Midjourney).