All articles

Explainers

Denoising Steps Explained: What Effort Changes

Learn what each denoising step changes, why turbo models need fewer steps, and how to choose a practical step count on an iPhone.

By · · Updated · 6 min read

Each denoising step is one more pass the same network makes over a picture that is still forming, so more steps polish the image rather than make the model think harder. Past a certain point, extra passes stop improving it. Almost every image generator has a steps slider, and almost everyone misreads it: people carry over the chatbot instinct that more time means harder thinking, and that instinct does not apply here.

This article pins down what a single pass computes. It shows why the step count follows a curve of shrinking returns. It also covers how turbo models squeezed a 50-step process down to one to four. At the end there is practical guidance for setting the slider.

What one step computes

An image model starts from random static and clears it away a little at a time. Our plain-language guide to diffusion models walks through that process. One step is one turn of the loop, and each turn has exactly two parts. (The technical name for one turn is a denoising step.)

First, the network makes a prediction. It looks at the messy canvas and at your prompt, then estimates which parts of what it sees are noise. That estimate runs the picture through the whole network once, which takes billions of multiplications. Nearly all of the time and battery goes here.

Second, a small piece of bookkeeping applies the estimate. It subtracts a calculated fraction of the noise and hands the cleaner canvas to the next step. (The bookkeeper is called the scheduler.) The schedule decides how big each subtraction is. Early steps settle the composition, meaning where the big masses go. Later steps resolve edges and texture. The network itself is identical every time. Step 27 differs from step 3 only in receiving a cleaner input.

More steps is not more thinking

When a chatbot spends longer on a problem, it writes out intermediate reasoning. It explores alternatives and can catch its own mistakes. Extra compute buys deliberation. A diffusion model carries no memory between steps, weighs no alternatives, and never reconsiders the prompt. Every step is the same reflexive judgment, “this looks like noise,” applied to a slightly cleaner input.

Onhand presents model-specific quality choices. They change the sampling work rather than adding a reasoning process. More steps may affect detail, but they do not guarantee correct counting, spelling, or anatomy.

The diminishing returns curve

Plot image quality against step count for a classic diffusion model and the shape comes out the same every time. There is a steep climb, then a bend, then a long flat shelf. (That bend is often called the knee.) The first steps do enormous work, dragging pure static into a composed scene. The middle steps sharpen it. Past the bend, typically somewhere in the 20s for older models, each added step changes fewer and fewer pixels. Before long, one output is nearly indistinguishable from the next. Doubling from 25 to 50 steps doubles your wait for a difference you would need a side-by-side comparison to spot.

The shelf is not perfectly flat, and very high step counts can subtly shift details. But it is flat enough that “maximum steps for maximum quality” has always been a superstition. On a server that superstition wastes money. On a phone, where each step costs full-power GPU time, it wastes battery.

How turbo models collapsed the schedule

Classic models need many steps because of how they were trained. Each prediction was only ever graded on removing a thin slice of noise, so each step makes one cautious correction. It takes dozens of cautious corrections to cross from static to an image.

Turbo models were retrained to make the whole crossing at once. (The technical name for that retraining is distillation.) A student network learns to jump straight to the answer a slow teacher reaches in dozens of steps. A second network acts as a critic during training, forcing the jump to land on something crisp instead of a safe blur. (That critic is an adversarial discriminator, and the full story is in our SDXL Turbo explainer.) The result: step one produces a complete image, and steps two through four act as short refinement passes. The model lightly re-noises its own output and cleans it up again, firming hands, fabric, and fine structure. Beyond four there is little left to firm. The bend in the curve has moved from step 25 to roughly step 2.

Choosing a step count in practice

For a turbo model there is little to decide.

  • 1 step while you experiment. Composition, palette, and subject are all fully visible, and you can burn through prompt variations as fast as the hardware allows.
  • 2–3 steps as the everyday default. Fine detail holds together noticeably better for a modest cost.
  • 4 steps for a keeper. This is the practical ceiling. The model was retrained for exactly this range, and pushing further buys nothing visible.

Explore at a lower quality setting, then compare another setting if useful. Changing the step schedule can change composition even with a fixed seed. If you want to preserve the actual image as input, use a supported img2img or Refine workflow and inspect the result.

Effort in Onhand

Onhand presents quality choices such as Fast, Balanced, and Detailed. Their step counts depend on the selected model: Turbo and Detailed Images use one, two, or four steps; other models use different schedules. More steps increase work and may change composition as well as detail.

There is no consumer numeric step slider or seed-lock control. Start with Fast to explore, compare another quality choice when useful, and use Refine if you want to work from a finished image. Generation time varies by model, device, loading, and temperature. Guidance-free Turbo behavior does not describe every model in the library.

If you want to set the step count yourself, Onhand vs Conjure compares Onhand with an app whose App Store listing exposes steps, guidance and schedulers. The offline AI image generator for iPhone page covers what to expect from Onhand in speed and quality.

Sources

Primary sources for the claims above.