Two settings sit above the prompt box, and between them they decide more than most of the words you type. One changes what the picture is. The other changes what it costs. People tend to leave both alone and then wonder why the result needs cropping.
Aspect ratio is composition, not cropping
A model does not render a square and trim it. It composes for the frame you asked for. A 16:9 prompt gets a horizon and negative space at the sides; the same prompt at 9:16 gets a subject stacked vertically with air above and below. Different pictures, same words.
This is why cropping a square down to a banner so rarely works. You are not removing empty space, you are removing composition the model deliberately put there.
Pick the ratio from where the image will live
| Destination | Ratio | Note |
|---|---|---|
| Blog header, link preview | 16:9 | Crops safely to 1200x630 for sharing cards |
| Instagram feed | 1:1 or 4:5 | 4:5 takes more screen height |
| Stories, Reels, TikTok | 9:16 | Keep the subject clear of the top and bottom |
| Marketplace listing | 1:1 | Most platforms crop everything to square anyway |
| Site hero, wide banner | 21:9 | Only some models offer it natively |
| Print, poster | 2:3 or 3:4 | Matches standard paper sizes |
If one image has to serve several of those, generate it wide and generate it again tall. Two renders on a cheap model cost less than one render plus an afternoon of cropping that does not work.
Not every model offers every frame
The range varies more than you would expect. Nano Banana 2 and the Wan models expose fifteen ratios, including extremes like 1:4 and 8:1 for banners and vertical strips. GPT Image 2.5 offers twelve, including 21:9 and the unusual 9:8. Nano Banana Pro and Nano Banana offer eleven.
At the other end, GPT Image 1.5 offers exactly three — square, 2:3 and 3:2 — because that is what the underlying model supports. It is a good model; it is not the one to pick for a 21:9 hero.
The catalogue lists what each engine accepts. It is worth checking before you write a long prompt, not after.
The "auto" option
Several models offer auto, which lets the engine choose based on the prompt. It is a reasonable default for exploration and a poor one for production, because two runs of the same prompt can come back in different frames. If the image has a slot to fill, name the ratio.
Resolution is a budget decision
Resolution is where the credits go. The same model at 1K, 2K and 4K costs 8, 12 and 20 credits on Nano Banana Pro; 6, 10 and 16 on GPT Image 2.5; 4, 8 and 12 on GPT Image 2. You are not buying a better model at the top tier. You are buying more pixels from the same one.
| Tier | Roughly | Use it when |
|---|---|---|
| 1K | ~1024px on the long edge | Drafts, social posts, thumbnails, most web images |
| 2K | ~2048px | Hero images, anything that will be cropped later |
| 4K | ~4096px | Print, zoomable product shots, large-format display |
A browser showing a 1,200-pixel-wide card gains nothing from a 4K source. The file is heavier, the page is slower, and the extra detail is discarded on the way to the screen. If you need a smaller file from a large render, resize and compress handle it locally without spending credits.
What resolution does not fix
Soft or muddy output at 1K usually means the prompt was vague, not that the model needed more pixels. Rendering the same vague prompt at 4K produces a larger version of the same problem at two and a half times the price.
Detail the model never had is also not recoverable. If a logo is unreadable at 1K because the model was inventing it, 4K will render an unreadable logo more crisply. Supply it as a reference through image to image instead.
A sensible default workflow
Draft at 1K, always, on a cheap model. Two credits per attempt means you can afford to be wrong six times while you find the framing. Commit the winning frame at the resolution the destination actually needs. Then, if the same image has to work in another slot, re-run it in the second ratio rather than cropping.
For a product listing that means: 1K drafts to choose the angle, 2K for the hero because customers zoom, 1K for the supporting shots because they are displayed small. That mix usually costs less than half of what a flat "everything at 4K" policy costs, and the listing looks identical.
Frames, files and formats
Once the image exists, the remaining decisions are about files rather than pixels. Convert handles format changes when a platform demands JPEG or PNG. Rotate fixes orientation. All of these run in the browser and cost nothing, which is worth remembering before you regenerate an image to fix something that is not a generation problem.
For sharing cards specifically, 1200x630 is the size that every major platform crops toward. Generating at 16:9 and exporting to that box is the least lossy route.
The short checklist
Name the ratio, do not leave it on auto. Pick the ratio from the destination, not from habit. Check that your model offers that ratio before writing the prompt. Draft at 1K. Finish at the size the image will really be seen at. Use the file tools for file problems.
Doing that consistently is worth more than any single model choice. The workspace keeps both settings next to the prompt box, and the plans page shows what each tier of credits buys once you know which resolution you actually need. Our credit breakdown works through the arithmetic model by model.
One subject, two frames
Campaign work almost always needs the same image wide and tall. The reliable way to get both is to render twice rather than crop once: keep the prompt, the model and the references identical, change only the ratio, and run it again. The model recomposes for each frame and both versions look intentional.
Cropping a 16:9 render to 9:16 does the opposite. You lose the sides the model composed with and gain nothing at the top and bottom, which is why cropped vertical ads so often have the subject jammed against an edge.
Sharing cards and the 1200x630 box
When a link is pasted into a chat app or a social feed, the preview is generated from an image that platforms crop toward roughly 1200 by 630 pixels. Anything far from that ratio gets cut, usually across the middle of your subject.
Rendering at 16:9 and exporting to that box is the least destructive path. Keep important detail away from the outer edges, because different platforms trim by different amounts.
Print needs real dimensions, not "4K"
Print is measured in inches and dots per inch, not tiers. At 300 dpi, an A4 page needs roughly 2,480 by 3,508 pixels — which a 4K render covers comfortably and a 1K render does not come close to. A 5x7 card needs about 1,500 by 2,100, which 2K handles.
The practical rule: multiply the printed size in inches by 300, and pick the tier that exceeds it. Anything beyond that is storage, not sharpness.
| Output | Pixels needed at 300 dpi | Tier |
|---|---|---|
| Business card | ~1,050 x 600 | 1K |
| 5x7 card | ~1,500 x 2,100 | 2K |
| A4 page | ~2,480 x 3,508 | 4K |
| A3 poster | ~3,508 x 4,961 | 4K, and choose a model that holds detail |