The sizes AI image models can actually give you, and how to art direct around them

Two numbers decide whether a generated image can do the job you had in mind: how many pixels it can produce, and what shapes it will produce them in. Both are published, both are stricter than people assume, and neither is what you get by typing "4K, high resolution" into a prompt.

Here are the documented limits for the two most widely used hosted families, read on 20 September 2026, and the art direction decisions that follow from them.

Google's models: a fixed menu, with the pixels printed on it

Google's Nano Banana quickstart lists the aspect ratios its image models accept and, usefully, the pixel dimensions each one produces at the 1K setting.

Three things fall out of that table.

The default is square, unless input images say otherwise, so every image you do not deliberately shape is a 1:1. That alone explains a lot of centred, symmetrical AI output.

A 16:9 frame at 1K is 1344x768. That is smaller than a 1080p video frame. If you are making a hero banner, the 1K tier is a comp, not an asset.

Google documents four resolution tiers rather than free-form sizes: 512px on the Lite and 2 models, 1K everywhere, and 2K and 4K on Nano Banana 2 and Pro. The quickstart prints exact dimensions for the 1K tier only, so for anything higher, measure what you actually receive rather than assuming it scaled cleanly.

OpenAI's model: free-form, inside four hard rules

gpt-image-2 takes any size you pass, as long as all four of these hold:

• the long edge is under 3840px
• both edges are multiples of 16
• the ratio of long edge to short edge is no greater than 3:1
• the total pixel count is between 655,360 and 8,294,400

OpenAI adds a warning that matters more than the ceilings: above 2560x1440 the results become more variable, and it calls that territory experimental.

Work out the biggest A4 you can ask for

This is where the rules bite. A4 at 300dpi is 2480x3508, which is 8,699,840 pixels. That is over the cap, so the request is invalid no matter how you word the prompt. Hold the A4 proportion and solve for the pixel limit instead and you land near 2422x3425, which rounds down to 2416x3408 once both edges have to be multiples of 16. That is roughly 292dpi at A4 size.

So the honest answer for print is: generate at the largest legal size, then upscale outside the model, and plan the artwork so the detail that matters is not two pixels wide in the original.

The 3:1 rule is a real constraint

A web banner at 1600x400 is exactly 4:1, which gpt-image-2 will not produce. Google's models will, and go as far as 8:1. If your deliverable is a header, a billboard or a panoramic key visual, that single line in the documentation decides which model you open.

The alternative is the one photographers have always used: generate at a legal ratio with the composition planned for a crop, and leave the space you are going to cut.

Art direction follows the frame, not the other way round

The temptation is to generate square and crop later. It produces weak images, for the same reason cropping a photo to a different aspect never quite works: the model composed for the frame it was given. Subject placement, negative space, where the light falls and how much room the background gets are all decided against the edges of the canvas.

So pick the frame first, then write the prompt for that frame. A 9:16 story needs the subject low and the air above it. A 21:9 banner needs a horizon and a place for type. A 4:5 feed post needs the subject filling more of the frame than you think, because it will be viewed at thumb size.

If layout matters, say so in the prompt. OpenAI's guide recommends calling out placement in words: logo top right, subject centred with negative space to the left. That is cheaper than fixing it in the crop.

A working set of defaults

What none of this fixes

Resolution is not detail. A 4K generation of a face can still have soft eyelashes and invented stitching, because the model is inventing texture rather than resolving it. Enlarging it makes the invention bigger. When you need real fine detail, the reliable path is still a real photograph, or a generated base composited with real assets on top.

And image size has nothing to do with type quality. If a layout has small copy in it, the quality setting and the model choice matter far more than the pixel count. The prompt recipes in prompt recipes for gpt-image-2 cover how to ask for copy that survives.

Sources

• Google, Get started with Nano Banana models, read 20 September 2026
• OpenAI, GPT Image Generation Models Prompting Guide, read 20 September 2026