Under the hood the model produces two predictions on every step — one informed by your prompt, one ignoring it — and CFG is the weight on the difference between them. Low values let the unconditioned prediction dominate, giving loose, sometimes more natural results that only partly match what you asked for. High values exaggerate the gap, which is why the output starts to look forced: over-saturated colour, hard contrast, a burnt quality that is the visual signature of too much guidance.
There is a usable middle band on every model and it differs by model, so a number copied from someone else's workflow is a starting guess. The sane way to find yours is to fix the seed and the prompt and step through values, because CFG is one of the few parameters whose effect is legible in a strip of comparisons.
Distilled and turbo variants are the exception. They are trained to work at very low guidance, and feeding them the value that suits a full model produces the burnt look immediately.
In practice
- Too low: pretty but off-brief. Too high: on-brief but harsh and over-contrasted.
- Change it in steps against a locked seed — sweeping it with a floating seed tells you nothing.
- Fast and distilled variants want much lower values than their full-size siblings.
The mistake to avoid
Raising CFG to fix a prompt the model is ignoring. Guidance amplifies what the model already understood; if a term meant nothing to it, more guidance amplifies nothing.
Related terms
Prompt
A prompt is the written instruction a generative model reads to decide what to make — the one input almost every model requires.
Sampling steps
Sampling steps is how many passes a model takes to turn its starting noise into a finished output — more passes, more refinement, more time.
Denoising strength
Denoising strength decides how much of your input picture gets thrown away before regeneration — low keeps it nearly intact, high keeps only the general shape.
Prompt adherence
Prompt adherence is how faithfully a model does what the prompt actually said, as opposed to producing something attractive in the same neighbourhood.
Seed
A seed is the number that decides the random starting noise for a generation, so the same seed with the same settings reproduces the same output.
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.