What do seed, aspect ratio, guidance and steps actually do?

Most confusion about generation settings comes from mixing two different eras of tooling. The parameter vocabulary people search for — seed, CFG scale, sampler, steps — comes from local Stable Diffusion interfaces, where the whole diffusion loop is exposed to you. The models most people actually use in 2026 are hosted, and expose a much smaller surface. Both are worth understanding, because the underlying mechanics still explain what you're seeing even when the slider isn't there.

Seed: the reproducibility control

An image model starts from random noise. The seed is the number that generates that noise. Same seed + same prompt + same model + same settings = the same image, every time. Change one word of the prompt while holding the seed, and you get a controlled comparison: the composition stays broadly recognisable and you see what your edit actually did. Change the seed and you get an unrelated roll of the dice.

That is the entire practical value of a seed, and it is worth more than it sounds. Without it, prompt iteration is guesswork — you change three words, the image changes completely, and you can't tell which word did it. With a fixed seed you can isolate variables one at a time, the same way you'd debug anything else.

Fixed seeds are not a consistency mechanism for characters. Same seed with a substantially different prompt gives you a different person in a similar layout, not the same person in a new scene. That problem needs a different approach.

Guidance (CFG): how literally the model obeys you

Classifier-free guidance is the strength of the pull toward your prompt and away from an unconditioned image. Low guidance (roughly 2–5 on the usual scale) gives the model latitude: more natural lighting, softer detail, more of its own aesthetic priors, and a real chance it quietly ignores part of what you asked for. High guidance (12+) forces literal compliance and tends to produce oversaturated colour, crunchy contrast, and a stiff, posed look — the model is being pushed into a corner of its distribution where good images are rare.

The failure mode people misdiagnose: cranking guidance because a detail is being ignored. It rarely fixes an omission and usually degrades everything else. An ignored detail is nearly always a prompt-structure problem — the detail is buried at the end of a long sentence, or it competes with another instruction — not an obedience problem.

Steps: refinement passes, with sharply diminishing returns

Steps are how many denoising passes the model makes on the way from noise to image. Too few and you get mush or an unresolved, painterly blur. Past a certain point — typically somewhere in the 25–40 range for modern samplers — extra steps buy you almost nothing and cost linear time and money. If an image looks wrong at 30 steps, it will look wrong in the same way at 100. Steps fix undercooked, not misconceived.

Aspect ratio: a composition control, not a crop

Changing the aspect ratio does not crop a fixed scene — it changes what the model composes in the first place. Ask for a portrait frame and you get a subject-dominant composition with tight headroom. Ask for a wide frame and you get environment, negative space, and a subject placed within a setting. The same prompt at 9:16 and at 16:9 will differ in subject scale, in how much background exists, and often in what the background is at all.

What our own stack actually exposes

Every prompt in this gallery is generated through hosted inference against three models, and it's worth being precise about what's under our control there rather than describing a generic slider panel:

That gap is not an oversight on our side. OpenAI's gpt-image-2 and Google's Nano Banana 2 don't expose a diffusion loop to their API callers at all; you send text, a size, and optionally a reference image. FLUX 2 Pro exposes more in some deployments. This is the general direction of the field: fewer knobs, better defaults, and the prompt itself carrying more of the load. It's also why the prompt text on every page here is written to be self-sufficient — it has to be, because there's no CFG dial to compensate with.

So where does the control actually live?

In the prompt's structure. Every prompt on this site is broken into subject, setting, lighting, camera and style modifiers precisely because those are the levers that still work across all three models. Lighting language changes mood far more reliably than guidance ever did. Camera language ('35mm', 'shot from below', 'shallow depth of field') controls framing and depth more predictably than a step count. If you're used to tuning parameters, this is where that effort now goes.