Decide the shape and the size before you write a word of the prompt

Proportion changes the picture, not just the crop
A square and a wide banner are not two crops of the same image. The proportion is supplied before generation and the composition is built to fill it, so a tall frame produces a vertically arranged subject and a wide one spreads the scene out. Ask for the wrong shape and you get a well-composed picture that cannot be cropped into the space you have.
Cropping afterwards is where the damage shows. The subject was centred for the original frame, so a wide crop of a square image loses the top and bottom of whatever mattered, and enlarging to compensate softens the whole thing. Two minutes of layout thinking prevents this entirely.
The practical habit is to establish the frame first, from the place the image will actually appear: a page header, a card in a grid, a slide, a social preview. Each of those has a fixed proportion, and generating directly into it removes an entire category of rework.
Leave the space the layout needs
Most published images have something on top of them — a headline, a logo, a button, a caption bar. If the composition fills the frame edge to edge with detail, that overlay will be unreadable, and no amount of adjusting the text colour will fix a busy background.
The instruction that solves it is compositional rather than stylistic: place the subject to one side, keep the upper third plain, leave the background uncluttered across the top. Asking for negative space in a named region is a spatial fact the process can act on, unlike asking for a design-friendly image.
It also pays to check the automatic crops your image will receive. Preview cards, thumbnails and small grid tiles usually crop towards the centre, so a subject placed near an edge for the sake of overlay space can disappear from the version most people see. Composing for both cases means keeping the essential content away from the extremes.
Resolution is decided by the destination
For screen use, an image needs to be comfortably larger than the space it occupies and no larger, because oversized files slow pages down for no visible benefit. For print, the requirement is different in kind, and an image that looks excellent on a screen can be plainly insufficient at page size.
Enlargement after the fact has limits that are frequently oversold. The tools that increase resolution invent plausible detail, which is fine on texture and unreliable on anything a viewer will inspect closely — faces, small text, fine mechanical parts. Generating at the size you need, where the tool allows it, avoids the question.
File format is part of the same decision. Flat illustration with hard edges and photographic material behave differently under compression, and a poor format choice can undo the quality you generated at. This is ordinary production knowledge rather than anything specific to generated images, and it is the step most often skipped.
Very wide and very tall frames misbehave
Extreme proportions are the reliable trouble spot. Long banners tend to produce repeated elements or a subject stretched to fill the width, and very tall frames produce awkward vertical stacking. The process is generally most stable near the common proportions and degrades as you move away from them.
The workaround is to generate near a standard shape and extend the frame afterwards, either by outward extension where the tool supports it or by composing the wide version in an editor over a plain ground. Both approaches give better results than requesting the extreme proportion outright.
For a very wide header the simplest answer is often not a wide picture at all. A subject on a solid background, placed to one side with the rest of the strip left plain, reads better at banner proportions than any full-bleed scene, and it survives being cropped on a narrow screen.
Plan the set, not the picture
Most images are needed in several sizes at once: a header, a card, a thumbnail, a preview. Deciding that at the start means generating one composition with room to be cropped in known ways, rather than discovering a week later that the thumbnail cuts the subject in half.
A short note recording the proportions each destination requires, and the safe areas within them, is worth more than any prompt refinement. It is the kind of specification that has existed in design work for decades, and generated images have not changed the requirement.
The general point is that the prompt is the last decision rather than the first. Frame, safe area, resolution and format are determined by where the picture is going, and getting them right in advance removes most of the work that otherwise happens after an image is already made.
Common questions
Can I not just crop a square image to whatever shape I need?
Sometimes, but the composition was built for the original proportion, so a crop usually removes part of the subject or forces an enlargement that softens the result. Generating into the proportion you actually need avoids both problems and takes no extra time.
How do I get room for a headline over an image?
Ask for it spatially: subject to one side, upper third plain, background uncluttered where the text will sit. Describing the region to keep empty works, whereas asking for an image suitable for text overlay does not constrain the composition at all.
Is upscaling a reliable way to reach print resolution?
It is reliable for texture and unreliable for anything examined closely, because the added detail is invented rather than recovered. Where the tool allows generating at a higher resolution, do that instead, and treat enlargement as a last resort for material that will not be inspected.
Senior writer, Prompt After Prompt
Maya writes about prompt craft, writing with ai, images & audio, mostly the parts other people skip and thinks most subjects are more interesting once you know how they work.