What the number actually controls
At every denoising step a diffusion model quietly makes two predictions: one conditioned on your prompt, one conditioned on nothing at all. Classifier free guidance takes the difference between them and amplifies it. The scale is the amplifier. At 1 you get the conditioned prediction untouched. At 7 you get the prompt direction exaggerated sevenfold, so the output leans harder into whatever makes it match your words.
That framing explains everything else on this page. The setting does not add information. It exaggerates a direction the model already computed, and exaggeration has a ceiling before it turns into distortion.
Ranges that work
| Model family | Typical | Notes |
|---|---|---|
| SD 1.5 | 6 to 9 | Tolerates high values, needs a negative prompt |
| SDXL | 5 to 8 | 7 is a safe default |
| Flux dev style | 2.5 to 4 | Distilled guidance, different meaning |
| Turbo / lightning | 1 to 2 | Effectively fixed, do not raise |
| Video models | 4 to 7 | High values amplify motion artefacts |
Treat these as starting points, not settings. The only reliable method is to lock the seed, change guidance alone, and compare a small ladder of values such as 3, 5, 7, and 9 on the same prompt.
Reading a bad result
Too low looks under committed: washed out colour, soft edges, a composition that only half matches the prompt, backgrounds that drift into abstraction. People usually diagnose this as a weak prompt and start adding words, which makes it worse.
Too high looks fried: clipped saturation, blown highlights, dark outlines around subjects, plastic skin, and mangled hands. In video it shows up as jitter and popping, because guidance is being exaggerated independently on frames that need to agree with each other.
Both extremes are easier to spot side by side than alone, which is the argument for running the ladder once per model and remembering the answer.
When it cannot help
Three situations where turning the dial is wasted effort. First, missing knowledge: if the checkpoint has never learned a concept, no amount of amplification will produce it. Second, dropped clauses: a prompt with six competing instructions loses some of them to attention dilution, and guidance amplifies the surviving direction rather than restoring the lost ones. Third, distilled models, where the value is fixed by training and the field is decoration.
The practical habit worth building: treat CFG scale as a per model constant you find once, then leave alone. Prompt content, seed, and steps are the things you actually vary run to run. Almost nobody needs to touch guidance twice on the same model, and treating it as a creative control is how sessions turn into forty generations of the same burned image.