SD CFG Scale Advisor

Stable Diffusion CFG scale advisor

CFG scale is the dial that decides whether Stable Diffusion obeys your prompt literally or interprets it loosely. Set it too low and the model ignores you; too high and the image burns. Pick your model type and prompt style and this advisor shows a recommended range, then explains what each value on the 1–30 slider is actually doing inside the diffusion process.

How CFG scale works

CFG stands for classifier-free guidance. At each denoising step, the model makes two separate predictions: one conditioned on your text prompt, and one unconditioned (no prompt). The CFG scale controls how far the final step is pushed from the unconditioned prediction toward the conditioned one:

output = unconditioned + CFG × (conditioned − unconditioned)

At CFG 1, the unconditioned and conditioned outputs are weighted nearly equally — the model mostly follows its own learned priors. At CFG 7, the conditioned output is amplified seven times the gap. Past roughly CFG 12, the amplification overshoots in most models, and the clipping in the denoiser manifests as blown highlights, oversaturation, halos around subjects, and distorted geometry.

CFG 1–2    barely guided — creative, can ignore the prompt (turbo/LCM models)
CFG 4–5    loose adherence — good for artistic and abstract styles
CFG 6–9    balanced — standard range for SD 1.5, follows prompt with natural detail
CFG 5–8    SDXL sweet spot — slightly lower than SD 1.5 for the same quality level
CFG 10–12  strong adherence — literal, higher contrast, detail preserved but saturated
CFG 13+    overshoot zone — fried, halos, color distortion in most checkpoints

Model-specific ranges

Different model families are tuned for very different ranges. Using a range appropriate to one family on another is one of the most common sources of unexpected output:

SD 1.5 and fine-tunes: the community standard is 6–9. Fine-tuned artistic models (illustration, anime, painting styles) often work well at the lower end (6–7); photoreal checkpoints typically benefit from 7–9.

SDXL: runs best at 5–8. It was trained differently and responds to strong guidance more aggressively than SD 1.5, so the effective “balanced” range is shifted one or two points lower.

Turbo, Lightning, SDXL Turbo, LCM: these distilled models are trained to converge in very few steps at very low guidance. Their effective range is 1–2 — sometimes literally 1.0. Pushing them above 4 or 5 typically burns the image or produces flat, oversaturated output.

CFG and prompt complexity interact

Prompt length and specificity change how CFG feels. A short, vague prompt (“a forest at dusk”) gives the model little to latch onto; higher CFG (8–9) forces more adherence to the sparse terms. A long, detailed prompt with many specific descriptors already constrains the image heavily — running high CFG on top can push past the model’s capacity and produce the fried look even within the normally safe range. When a detailed prompt looks oversaturated, lower CFG by 1–2 before anything else.

Tips for tuning

• Start at the model default, then move in steps of 1. Evaluate with the same seed so you are measuring the CFG effect, not randomness.
• Fried image (halos, oversaturation, blown highlights)? Lower CFG first. If that is not enough, also check whether your prompt is fighting itself with conflicting style terms.
• Ignoring your prompt? Raise CFG, but check token ordering first — subjects buried past 75 tokens lose weight and fixing the prompt order costs nothing.
• Artistic / dreamy style? Use the lower half of the recommended range; the model’s own interpretation adds variety that high CFG erases.