What Are Steps in ComfyUI?

For a standard (non-distilled) SDXL or Flux-class checkpoint, 20–30 steps is a reasonable starting point. Distilled or "turbo" models are built for far fewer — often 4–8 — and using 30 steps on one of those wastes time without improving the result.

Distilled models are specifically trained to produce a clean result in very few denoising passes, trading a small amount of flexibility for major speed gains. Running them at high step counts doesn’t use that training advantage and just adds wasted computation.

No. Quality improves quickly at first, then plateaus — past a model’s sweet spot, extra steps add generation time without a visible difference, and in rare cases can introduce artifacts instead of removing them.