CFG Scale controls how strictly an AI image model follows your prompt versus how much creative freedom it takes. Low values (1-4) produce loose, dreamlike results that may ignore parts of the prompt. Mid values (5-9) are the sweet spot for most subjects. High values (10-20) force the model to obey every word but often produce oversaturated, plastic-looking output. Stable Diffusion exposes this as a slider; Midjourney does similar work via the --stylize parameter; Flux 2 and Imagen 4 tune it internally and rarely expose it.