Start at the placement, not the picture

“Something for the launch”

That is the brief you will actually receive. It contains no dimensions, no placement, no ratio and no idea of whether the thing will be seen at 400 pixels wide on a phone or at two metres on a wall. Generating against it wastes an afternoon, because the first question at review will be “can we get it in a square for Instagram too”, and the answer will be no.

The fix is one question asked first: where does this appear? Everything else follows from the answer.

The numbers you will actually meet

Keep these somewhere. They change slowly and they decide your generation size.

• Instagram feed portrait: 1080 × 1350 (4:5). The single most common social deliverable.
• Instagram/TikTok story or reel cover: 1080 × 1920 (9:16).
• Open Graph card for a link preview on most platforms: 1200 × 630 (roughly 1.91:1). Cropped hard on both sides on some surfaces.
• YouTube thumbnail: 1280 × 720 (16:9), and it will be displayed as small as 246 × 138.
• Website hero: commonly 2560 wide for a desktop banner, with the height decided by the section, and a separate mobile crop.
• A4 page at 300 dpi: 2480 × 3508. A5 is 1748 × 2480.
• A3 poster at 300 dpi: 3508 × 4961.
• Roll-up banner, 850 × 2000 mm: at 150 dpi that is 5020 × 11811, which no generator produces directly and which you reach by upscaling a correctly proportioned render.

Two of these deserve a note. The YouTube thumbnail is a 1280-pixel image that most people see at a fifth of that size, so detail below a certain scale is wasted effort and contrast is everything. And large-format print does not need 300 dpi: a roll-up banner viewed from two metres is perfectly sharp at 100 to 150 dpi, and a billboard at 30 metres is fine at 15. People generate uselessly enormous files because they learned one number and applied it everywhere.

What the placement tells you beyond size

How long it is looked at. A story frame gets under two seconds. A poster on a wall gets thirty. That difference decides how many elements the composition can carry, and it is a stronger constraint than any style choice.

Where the type goes. If a headline will sit across the top third, that third needs to be quiet (a plain sky, a wall, a gradient) and you should generate it that way rather than blurring it afterwards. Asking for “negative space in the upper third for a headline” in the prompt works surprisingly well and costs nothing.

What the format has to be. A photographic hero delivers as JPEG or WebP. A product cut-out for a shop listing needs a transparent PNG, which no diffusion model outputs directly: it comes from generating on a plain background and cutting it out afterwards. Discovering this at delivery is an unpleasant hour.

Whether it must tile or repeat. A background pattern has a constraint no amount of prompting fixes casually, and it needs a seamless-tiling workflow chosen at the start.

Ask these four questions

Before generating anything, you want four answers in writing. They take one message to obtain:

1. What are the final pixel dimensions, and are there other sizes it must also work in?
2. Will type sit over it, and where?
3. Print or screen? If print, what process and what size?
4. Does anything real have to be in it: a specific product, a named person, a building?

That fourth question is the one that saves the most time, and the next lessons in this block are largely about what to do when the answer is yes.

The limit worth stating plainly

A single generated image cannot serve every placement. A 1080 × 1350 portrait crops badly to 1200 × 630, and stretching it is visible. The professional answer is to generate wider than you need and crop down, which the next lesson covers, or to generate each placement separately from a shared reference.

What you should not do is promise one image will cover six placements before checking the ratios. It is the commonest small failure in this work, it is discovered at the worst possible moment, and it is avoided by thirty seconds of arithmetic at the start.