Aspect ratio looks like a technical setting you deal with at the end, after the picture is right. In practice it is one of the first decisions, because the shape of the frame changes what the model puts inside it. A square and a wide rectangle given the same prompt do not produce the same image cropped differently. They produce different compositions.

The reason is simple enough. Training data has habits. Tall frames are full of portraits, standing figures and phone photos. Wide frames are full of landscapes, film stills and group shots. When you choose a shape, you are quietly nudging the model toward the kind of picture that usually arrives in that shape.

Match the ratio to the subject

A single person, a bottle, a tree, a tall building — anything whose interesting axis runs up and down wants a vertical frame. Give it a wide one and the model has to fill the sides with something, and that something is often invented clutter or an oddly empty background.

Horizons, interiors, groups and anything about context want a wide frame. Squeeze a landscape into a tall crop and you get a strip of sky over a strip of ground, with the sense of space that made it worth drawing removed.

Square sits between the two and is genuinely good at one thing: a centred subject with a bit of breathing room. It is also the safest choice when you have no idea yet what the composition should be, because it commits to neither axis.

Let the destination decide

The other half of the question is where the image ends up. A site hero is wide. A phone-first social post is tall. A product card is often square. Generating in roughly the right shape from the start saves you from cropping later, and cropping is where good compositions go to die — the model centred the subject for the frame it was given, and your crop moves it somewhere it was never meant to be.

If you already know the image will be cropped in more than one shape, generate the widest version and leave deliberate empty space where the crop will bite. Asking for a subject placed to one side with clean negative space on the other gives you room to work with. It is a small extra sentence in the prompt and it saves a regeneration.

Say it plainly, and check what you got

Whether the ratio comes from a dropdown or from words in the prompt, be direct: vertical portrait, wide landscape, square. Vague phrases like cinematic can imply a wide frame but mostly imply colour and mood, so do not rely on them to set shape.

And be aware that a ratio you can request is not always a ratio you get exactly. Tools round to the sizes they can actually render. If precise dimensions matter — a fixed slot in a layout, say — generate a little larger than you need in the closest available shape and trim, rather than fighting the generator for the last few pixels.