CFG scale (guidance)

At each denoising step, a diffusion model can make two predictions: one that takes your prompt into account and one that ignores it. Classifier-free guidance looks at the difference between the two and pushes the picture further in the direction of the prompt. The CFG scale is the size of that push. At zero or one the prompt has little extra weight, around five to eight is the range many models document as a sensible middle, and much higher values exaggerate everything.

The trade-off is easy to see in practice. With a low scale the result can be soft, loose and creative, but it may ignore parts of your description. With a high scale the picture follows your words closely, yet contrast and colors often become harsh, textures look burnt and the same few shapes appear again and again. If a portrait of a fictional character in a gray office looks oversaturated and brittle, lowering the scale is the first thing to try; if a requested object is missing, raise it slightly or rewrite the prompt.

Good values depend on the model, so do not carry numbers from one to another. Distilled and turbo models, trained to reach a result in very few steps, usually bake the guidance into the weights and expect a scale of one or no setting at all. Comparing is easiest with a fixed seed, changing only the scale between attempts. The negative prompt works through the same mechanism, as it is what the unguided prediction is steered away from.

A small experiment makes the effect concrete. Take a prompt such as a ceramic teapot on a windowsill, keep the seed, model and size identical, and generate the same image at several scales, from low to quite high. At the low end the teapot may be dreamy and half-formed, with a vague windowsill. In the middle, shapes are clear and the light looks natural. At the far end, edges turn crisp to the point of brittleness, highlights burn out and the glaze takes on a plastic sheen. Keeping such a contact sheet teaches you faster than any rule of thumb which range suits each model you use.

Treat the scale and the number of steps together: a higher scale on very few steps can look rough, while a moderate scale with enough steps usually gives smoother gradients and cleaner edges.

Remember also that sampler and resolution interact with guidance: a value that looks balanced on a small square may feel overcooked on a large panorama, so retest after changing either.

NeverNo does not give you a CFG scale slider. The guidance is set for each model on the server side, tuned to what the model expects, so you cannot damage a result with a value that suits another model. Your levers in the Studio are the description of up to 1,500 characters, the negative prompt of up to 500 characters on the models that support it, one of fifteen style presets and one of eight aspect ratios.

Z-Image Turbo, the default model, is a fast model: a typical 1-megapixel image takes about 3 seconds, measured on Runware, our image provider, which suits this logic of ready-made settings. If a result ignores your words, rewrite the prompt more precisely or try another model from the catalog. The images you create are stored encrypted in your browser, and the hard limits of NeverNo apply to every request.