Plain-English explanation
Guidance scale is the dial that decides how literally the model takes your prompt. At a low value, the model treats the prompt as more of a loose suggestion and generation has more freedom to wander, sometimes producing more natural-looking but less prompt-accurate images. At a high value, the model is pushed hard to match the prompt as closely as possible, which can improve accuracy to the wording but, past a certain point, tends to produce oversaturated colors, exaggerated contrast, or artificial-looking results.
How it works
Guidance scale is implemented through classifier-free guidance. At each denoising step, the model computes two predictions: one conditioned on the prompt (and negative prompt, if set), and one unconditional or negative-conditioned prediction. It then extrapolates away from the unconditional or negative direction toward the conditional one, scaled by the guidance value. Hugging Face's Diffusers documentation shows this directly in a basic Stable Diffusion pipeline: the final noise prediction is computed as the unconditional prediction plus the guidance scale multiplied by the difference between the prompt-conditioned and unconditional predictions. A guidance scale of 1 effectively disables this extrapolation.
Adult-AI use
Guidance scale (often just labeled "CFG" or "CFG scale") is a literal slider in nearly every adult-AI image generator, sitting right alongside the prompt box. It's one of the first settings people tune when a generation isn't matching their prompt closely enough, or conversely when results look artificial and over-processed at high settings. For character work specifically, guidance scale interacts with how strictly the model honors specific described features (outfit, pose, expression) versus letting the model's own priors take over: a common source of confusion when a detailed prompt still doesn't come through clearly.
Common uses
- Raising guidance scale to enforce a detailed or specific prompt more literally.
- Lowering guidance scale for a softer, more natural, less "AI-typical" look.
- Combining guidance scale with image-to-image strength: high strength and high guidance for maximum creative reinterpretation, or low values on both to stay close to a source image without being tightly bound to the prompt.
Limitations and misconceptions
- Higher is not simply "better": very high guidance scale values commonly produce oversaturated, high-contrast, or unnaturally sharpened images rather than more accurate ones.
- The ideal guidance scale range differs across model architectures; a value that looks balanced on one model can look badly over- or under-guided on another. Some newer model families are tuned to expect much lower guidance values than older Stable Diffusion checkpoints.
- Guidance scale and negative prompt work together, not independently, changing one changes how the other behaves, since guidance scale is what determines how strongly the negative prompt's "push away" is applied.