The frame is an instruction
Models compose into the shape you give them. Hand them width and they fill it with environment: horizon, sky, foreground. Hand them height and they fill it with figure: full bodies, towers, falling light. The same prompt becomes two different pictures in two different frames.
Neither is wrong. But if you wanted one and got the other, the fix was never in the words.
Five shapes cover almost everything
- 1:1, the square. Icons, album covers, product shots, anything centered. The default on most tools.
- 3:2 and 2:3, the classic photo frame. Prints, natural crops, the shape worth reaching for when you have no reason to choose.
- 16:9 and 9:16, the screen shapes. Wallpapers and establishing shots wide; phone screens and stories tall.
- 4:5, the portrait crop social feeds use. Tighter than 2:3, made for faces and standing figures.
- 21:9, cinema. Wide vistas and film stills. The most dramatic shape and the most likely to fail; see below.
Setting it, tool by tool
Midjourney takes a flag at the end of the prompt. Any pair of whole numbers works.
Stable Diffusion has no ratio flag; you set width and height in pixels. Stay close to the model's training size and keep the total pixel count steady. For SDXL that means pairs near one megapixel:
Nano Banana, GPT-Image, and the other chat image models take the shape in words. Say wide 16:9
or portrait orientation
in the sentence. Most support a short list of shapes and snap your request to the nearest one.
Write for the shape you chose
A ratio is a promise about empty space. Keep it by telling the model what fills that space.
Wide: the sides need a world. Anchor the subject in one spot, the way the subject guide anchors things, then name what stretches away from it: a road, a shoreline, a market street.
Tall: the vertical needs a reason. Full bodies, waterfalls, stairwells, a shaft of light from the lighting wall. Say where the height goes or the model stacks filler.
Square: the frame wants a center. Symmetry, mandalas, a face, one object on a plain ground. Words from the composition wall like centered composition
and radial symmetry
earn their keep here.
Where extreme frames fail
Push past 16:9 and models start to repeat. A 21:9 frame with a vague subject returns twins or a tiled landscape; a very tall frame returns a figure with a second head where the torso ran out. The cure is the same in both directions: one anchored subject, stated once, and a named environment that owns the rest of the frame. If it keeps failing, generate at 3:2 and extend the image afterward; the chat image models do this in the edit loop.
Where this fits
The Start Here skeleton ends with --ar 3:2 because that default rarely hurts. The builder's ratio buttons write the flag for you, and the cheat sheet keeps the parameter forms for all three tool families on one printable page. Decide the shape first; the words come easier once the frame is honest.
Published here first.