Guidance Scale

Also known as: guidance_scale, guidance, cfg_scale, cfg

Guidance scale controls how strongly the model follows your text prompt during image generation. Low values produce creative, sometimes unexpected results. High values force strict prompt adherence but can look oversaturated or artificial.

What It Does

Guidance scale is the core control behind Classifier-Free Guidance (CFG), the technique that makes text-to-image generation actually follow your prompts. At each denoising step, the model produces two separate predictions: one conditioned on your text prompt and one completely unconditioned (as if you gave it no prompt at all). The guidance scale determines how much the conditioned prediction is amplified relative to the unconditioned one.

At a guidance scale of 1.0, the model effectively ignores your prompt and generates whatever it wants. As you increase the value, the model pays more and more attention to what you asked for. The sweet spot for most models falls between 3.0 and 7.0, where you get good prompt adherence while still allowing the model to make natural, visually pleasing choices on its own.

Push the value above 10 to 15 and things start to go wrong. Colors become oversaturated, contrast gets extreme, and fine details can warp into harsh artifacts. The model is essentially overcommitting to the prompt signal at the expense of image quality. Think of it like turning up the volume on a speaker: a little boost makes the music clearer, but crank it too far and everything distorts.

Value Ranges

Low (1.0 – 2.0)

Dreamy, loose, creative. The model wanders freely from your prompt, often producing unexpected compositions and color choices. Useful for abstract or experimental work.

Mid (3.0 – 7.0)

Balanced. Good prompt adherence with natural-looking results. This is the default range for most models and where you should start before making adjustments.

High (8.0 – 20.0+)

Very strict prompt following. Colors may oversaturate and details can look exaggerated or artificial. Useful when the model isn't capturing a specific element you need, but use sparingly.

Visual Comparison

Image pending

guidance_scale = 1.5

Image pending

guidance_scale = 3.5

Image pending

guidance_scale = 10.0

Tips

Start with the model's default (usually 3.5 to 7.5) and adjust in small increments.
If your image looks nothing like your prompt, increase guidance scale.
If the image looks harsh, oversaturated, or has artifacts, decrease it.
Flux models tend to work well at lower guidance (2 to 5) compared to SDXL models (5 to 12).