CFG scale
CFG scale (classifier-free guidance scale) controls how strongly the image follows your prompt. Low values are loose and creative; high values stick closely to the prompt but can look over-processed.
CFG stands for classifier-free guidance. The CFG scale (also called guidance scale) is a single number that sets how hard the model tries to match your prompt versus how much freedom it has to do its own thing. It is one of the most important quality dials in any generator.
How it works
At each denoising step the model makes two predictions: one guided by your prompt and one unconditioned (ignoring the prompt). CFG scale amplifies the difference between them. A higher value exaggerates the prompt-driven direction, pulling the image harder toward your words - and toward your negative prompt away from what you do not want.
Choosing a value
- Low (1-4): loose, sometimes more photographic and natural, but may drift from your prompt.
- Mid (5-9): the usual sweet spot - good prompt adherence with natural-looking results.
- High (10+): very literal to the prompt, but often over-saturated, harsh or "fried" looking.
The ideal range depends on the model and sampler; some newer models are tuned for low CFG. If results look burnt or contrasty, lower the CFG before changing anything else.
Try it in the generator
Put cfg scale to work right now - free daily generations, commercial license included.
Frequently asked questions
What CFG scale should I use?
Start around 7 for most models and adjust. Lower it toward 4-5 if images look over-saturated or harsh; raise it toward 9-10 if the model is ignoring parts of your prompt.
Related terms
- Negative promptA negative prompt is a list of words describing what you do NOT want in the image. The model steers away from those concepts while it generates, which removes common flaws and unwanted elements.
- StepsSteps (sampling steps) are the number of denoising passes the model runs to turn noise into an image. More steps can mean more detail, but past a point they only add time, not quality.
- SamplerA sampler is the algorithm that decides how noise is removed at each step of generation. Different samplers reach the final image by different paths, trading off speed, detail and consistency.
- SeedA seed is the number used to initialize the random noise an image is built from. The same seed plus the same prompt and settings produces the same image every time, which makes results reproducible.