What is CFG scale?
Also known as: Classifier-free guidance, Guidance scale
Classifier-free guidance scale — a parameter controlling how strictly a diffusion model follows the text prompt. Higher values produce more literal interpretations; lower values allow more creative variation.
In detail
CFG scale is the control parameter in classifier-free guidance, a technique that improves prompt adherence in diffusion models. The model generates two predictions at each denoising step — one conditioned on the prompt, one unconditioned — and CFG scale determines how strongly the conditioned prediction overrides the unconditioned. CFG 1.0 ignores the prompt entirely (produces unconditioned random outputs); CFG 7-8 is the typical default for balanced prompt-fidelity and creative variation; CFG 12+ produces strong prompt adherence but can over-saturate colors and reduce quality; CFG 3-5 allows creative drift away from the prompt. For textile work, CFG 5-8 is the sweet spot — prompts produce recognizable results but the AI has enough freedom to produce visually appealing compositions. Pushing CFG too high produces 'fried' outputs with unnatural saturation and contrast.
Example
Prompt: 'soft watercolor peony floral, dusty pink and sage.' At CFG 7, the model produces a balanced soft watercolor look. At CFG 12, it produces an over-saturated, hyper-pink result with cartoonish edges. At CFG 4, it drifts toward generic florals that ignore the watercolor and color cues.