Guidance scale: how strictly does AI follow your prompt?

The guidance scale, also called CFG or classifier-free guidance, sets how strictly an AI image model follows your prompt: a low value gives the model lots of freedom, a high value forces it to follow your words almost literally. It's the knob behind two common complaints: "the AI ignores my prompt" and "my image looks harsh and burnt". This guide explains what the value does, why too high is just as bad as too low, and how to steer even when you can't see the slider.

What exactly is the guidance scale?

The guidance scale is a number that controls how hard the model pulls toward your prompt while generating. Under the hood the model actually makes two estimates: one with your prompt and one without. The guidance scale decides how far the result is pulled toward the prompt version. Higher means following your words more literally, lower means more room for the model's own interpretation.

The technique comes from the paper "Classifier-Free Diffusion Guidance" by Jonathan Ho and Tim Salimans (26 July 2022). In Stable Diffusion the value defaults to 7.5, and in practice a range of roughly 7 to 12 often works well. That's not a rule: the best value depends on the model, your prompt, and your subject.

What happens at a value that's too low?

At a low guidance scale the model largely ignores your prompt and mostly does its own thing. Around a value of 1 your text barely counts: the composition turns soft and vague, and details you explicitly asked for disappear. You get a nice but random image that has little to do with your description.

Sometimes that's exactly what you want. If you're after variation or a surprise and don't have a fixed end result in mind, a lower value gives the model breathing room. If you're chasing a specific result, too low is usually frustrating: you ask for a red jacket and get some random color.

Why a value that's too high ruins your image

At a guidance scale that's too high the model overshoots: colors become oversaturated, contrast turns harsh, and you get artifacts and "color burn" where highlights blow out. Research and practice show the same pattern: as you push the value up, images look more saturated and variety drops. Above roughly 12, that effect becomes visible on many models.

It's a misconception that "following the prompt more strictly" is always better. Past a certain point, extra guidance buys you no extra faithfulness, only an unnatural, hard-edged image. If your portrait suddenly looks plastic and too intense, a guidance scale that's too high is one of the first suspects.

Which value fits which model?

There's no universal "best" value: it varies by model family. Reported ballpark values look roughly like this:

• Stable Diffusion 1.5: around 7 to 11.
• SDXL: often lower, around 5 to 8.
• Stable Diffusion 3.5: lower still, around 3.5 to 4.5.
• Fast "turbo" and LCM models: around 1 to 2, because they run in very few steps.
• Flux: uses a distilled form with its own scale, with a default around 3.5.

The lesson isn't to memorize these numbers, but to realize that a value that's perfect on one model can be far too high on another. Always take a new model on its own recommended range.

What if you don't see the slider?

Many modern and commercial generators don't show a CFG slider and set the value internally at a sensible level. Then you steer not with a number, but with your prompt and your approach. Three practical moves:

1. Rewrite instead of pushing harder. If you can't get a detail to appear, move it to the front of your prompt or describe it more concretely, rather than reaching for stronger guidance.
2. Test small before you scale up. Run a few variants on an affordable model, paying per render, see which direction is right, and only then render your premium version.
3. Let your prompt be structured for you. A structured prompt (subject, light, composition) is easier for the model to follow than a wall of text. A prompt generator helps with that.

If you want to feel the balance between guidance and saturation yourself, the fastest way is to generate the same prompt side by side in the photo generator.

Frequently asked questions

What's a good guidance scale to start with?

For most Stable Diffusion models, 7 is a safe starting point, close to the default of 7.5. Then generate one variant slightly higher and one slightly lower and compare; that quickly reveals the value that fits your prompt and model.

Why does my AI image look burnt or oversaturated?

Usually the guidance scale is too high. Above roughly 12, colors on many models oversaturate and highlights blow out. Lower the value, or if you have no slider, make your prompt calmer and less stacked.

Do I always have to set the guidance scale myself?

No. Many tools set the value well internally and don't show a slider. In that case the concept is mainly useful for understanding why your prompt is sometimes followed too loosely or too literally, and for rewriting with intent.

The guidance scale isn't a magic quality dial, but a balance between faithfulness to your prompt and a natural image. Want to see the difference yourself? Create an account and compare a few renders side by side.