CFG scale and step count: what they actually do, and sane starting values

CFG (classifier-free guidance) scale is how hard the sampler pushes the image toward your prompt and away from the unconditioned result. Steps are how many denoising passes the sampler takes. CFG - Too low (1-3 on a normal model): soft, washed-out, vaguely related to the prompt. - Sweet spot: about 5-8 for SD 1.5 and SDXL. - Too high (12+): burnt contrast, oversaturated colour, crunchy edges, sometimes a "fried" look. The exception is LCM and other distilled models (see the LCM thread). They are trained to work at CFG about 1-2, and normal values ruin them. Steps More steps is not always more quality. With DPM++ 2M Karras, 20-30 steps is usually converged; going to 60 rarely changes anything except time. With Euler a (an ancestral sampler) the image keeps changing as you add steps, because fresh noise is injected each step, so "same seed, more steps" gives a different picture. A test grid that teaches you more than any guide Hold the prompt and seed fixed. Make a 3x3 grid: - CFG 4, 7, 10 - Steps 15, 25, 40 ComfyUI and A1111 both do this (A1111: the X/Y/Z plot script). Look at where it stops improving. That is your default for that model. What it got wrong, and the fix The most common failure posted here is "colours look burnt" — that is almost always CFG too high, or an LCM model run at normal CFG. Drop CFG before touching anything else.