More steps will not fix a soft image

The first slider people reach for on a mushy image is sampling steps. It feels like quality. It maps onto time. It is almost never the cause.

Generation is iterative: each step removes a little noise and re-estimates what is underneath. The early steps decide where things are and how big they are. Later steps polish texture. Past the model's comfortable range the picture stops improving and can start drifting. A very high number is usually wasted time rather than insurance. If the face is in the wrong place, more polish will not move it. If the file is soft because it was generated below the detail the brief needs, more polish on a small latent will not invent the missing pixels.

Ask three questions, in order: is this a resolution ceiling, a sampler mismatch, or a guidance problem? Only if all three answers are "no" do you add steps, and even then you add them with the seed locked so you can see whether they did anything.

What extra steps actually buy

Structure is decided early. Detail is decided late. That split is the whole diagnostic.

Composition wrong (crop, pose, number of people, the product on the wrong side): more steps will not fix it. Change the prompt, the seed, or the aspect ratio.
Texture under-resolved (the crop is right, but fabric and skin look like they stopped halfway): a few more steps on a full, non-distilled model can help, up to the plateau.
Already past the plateau: adjacent step counts look the same, or the later one has started to wander. You are paying for minutes, not pixels.
Distilled, turbo, or few-step models: they are trained to converge in a handful of passes. Raising the count can make them worse. That is the distillation bargain.

Time scales roughly with the step count, so it is the right thing to cut when drafting and the wrong thing to max "just in case". Find the plateau once per model, write it down, stop treating the slider as a quality knob.

A lock-the-seed sweep is the only honest measurement. Same prompt, same seed, 8 / 16 / 24 / 32 (or whatever band the model documents). When two adjacent outputs agree, you are done. If the 8-step image is a different picture from the 32-step image, you are looking at under-sampling. If they are the same picture and both are soft, steps are not the lever.

Question 1: is the softness a resolution ceiling?

Softness that survives a sensible step count is often a size problem wearing a sampling costume.

The model has a native size it was trained around. Pushing far past it produces a large image with soft or repeating detail rather than a genuinely more detailed one. The glossary on resolution is blunt about this: the reliable route to a big, sharp file is often to generate at the comfortable size and upscale. Generating "at 4K" on a model whose latent is happiest much smaller does not give you 4K of information. It gives you a stretched guess.

Latent compression discards pore-scale detail by design. Generation happens in a compact representation; the picture you see is decoded at the end. Very fine texture, small text, and single-pixel detail live below what that representation can hold. Blaming the step count for missing micro-detail is blaming the wrong stage. Some of that loss happened in the encode/decode round trip.

A 2x upscale of a sharp small file beats a native large file that the model could not fill. That is the same rule as upscale the render or generate at full resolution, applied to stills. If the brief is print or a tight crop on a face, generate at the model's native size with a finished composition, then run a dedicated upscaler such as Topaz Upscale Image. Two 2x passes usually beat one 4x pass.

If the source is already soft, upscaling will invent texture on top of mush. That looks sharp at a glance and wrong at 100%. Steps cannot recover what the latent never held; an upscaler cannot recover what the source never held. Fix size first, then sample, then enlarge.

Question 2: is the sampler mismatched?

The step count only means something relative to the sampling method the model was tuned for. "30 steps" is not a universal quality setting. It is 30 iterations of a particular solver.

Three mismatches produce softness, wander, or both:

Distilled and turbo samplers live in a narrow step band. They were taught to imitate a larger model in few passes. Raise the count toward the parent's favourite 30 and you pull them off the trajectory they were trained on. The result is often over-cooked, smeared, or strangely soft in the midtones. Use the low count the variant expects. Draft on the fast path, finish on the full model if the tail of the brief needs it.

Stochastic (ancestral) samplers keep injecting noise. Extra steps do not converge on a sharper image. They take another walk. The picture changes; it does not lock. If you raise steps on an ancestral sampler expecting crispness, you get a different, equally-soft relative. Switch to a deterministic sampler when the goal is to finish the current picture rather than to resample it.

Deterministic samplers can oversmooth past the plateau. Once the noise is gone, further steps polish toward an average. Skin and fabric lose the last of their grit. That is softness from too many steps, and it is why the seed-locked sweep matters: you will see the image peak and then go bland.

If the endpoint does not let you pick a sampler or a step count (most hosted APIs, including a lot of what sits behind the image generator), you still have the distilled-versus-full choice. A Flux Schnell-class path is a few-step model. A full Flux or SDXL-class path is not. Do not copy a step count across that boundary.

Question 3: is guidance the culprit?

People call two different pictures "soft". One is mushy and under-defined. One is waxy, over-smooth, and a bit glowing. They are opposite CFG failures.

Too low. The unconditioned prior wins. Edges do not commit. The image looks pretty and slightly drunk. Raising CFG a little, still inside the model's middle band, does more for this than adding steps.

Too high. Oversaturation, hard contrast, a burnt look, plastic skin. That is not softness, but it is often what someone is trying to "sharpen away" with more steps. Extra steps on a CFG-12 portrait polish the wax. Drop guidance into the working band (3–7 on many full checkpoints; much lower on distilled) and leave the step count alone.

Prompted softness. "Soft focus, dreamy, creamy bokeh, atmospheric haze" is an instruction to smear. No sampler setting will disobey it. If you wanted a sharp subject and a smooth background, say that: "sharp on the eyes, 85mm, background a smooth falloff."

Guidance and steps interact. Very few steps at high denoise, especially on image-to-image, is noise the model never finished resolving. That pair looks soft because it is unfinished, and the fix is either more steps or less denoise, not both maxed.

A sequence that does not waste the run

Confirm the canvas. Native size, supported dimensions, the right aspect ratio. If the file has to be large, plan an upscale after, not a native size the model cannot fill.
Confirm the sampler class. Distilled/turbo: low steps, low CFG. Full model: the model's documented band. Ancestral: do not expect extra steps to sharpen.
Confirm guidance. If the picture is mushy and undefined, CFG may be low. If it is waxy or burnt, CFG is high. If the prompt asked for dreamy, the prompt is the smear.
Only then sweep steps with the seed locked. Stop at the plateau.
Enlarge last, on a file that is already sharp at its native size. Dedicated upscalers are for this stage, not for rescuing an unfinished latent.

If you change two of these in one reroll, you will not know which one worked. The whole point of the three questions is to stop spending credits on a slider that was never the diagnosis.

FAQ

How many steps is enough?

Enough is the first count at which a seed-locked neighbour looks the same. That number is per model and per sampler, which is why a screenshot of someone else's 40 is not a setting. Distilled variants often want a handful. Full models plateau well before their advertised maximum. Past the plateau you are buying time.

Why did raising steps make a turbo model worse?

Because the turbo model was trained to finish in few steps. Extra iterations pull it off that trajectory. Use the low count it expects, or switch to the full model for the final take. Comparing a fast variant to its parent on the parent's settings is how you conclude the fast variant is "worse" when it is only mis-set.

Is a soft image always under-sampled?

No. Under-sampling is one of four common causes, and it is the one people jump to. The other three are a resolution ceiling (including latent compression), a sampler or distilled-band mismatch, and guidance or prompt language that asked for smear. A seed-locked 8 / 16 / 32 strip tells you whether steps are even in play. If 16 and 32 agree and both are soft, look at size, sampler, and CFG.

Should I upscale instead of adding steps?

If the composition is right and the native file is sharp at 100%, yes: enlarge with an upscaler rather than sampling a bigger latent the model cannot fill. If the native file is mushy at 100%, upscaling will invent detail on top of mush. Finish the small file first.