Nano Banana 2 vs Ideogram 4.5 for photo edits: ratios, 4K, refs

Choose Nano Banana 2 on Sume when you need a ratio like 21:9 or 8:1, up to 4K, or ten reference images. Choose Ideogram 4.5 when the edit must keep the source shape and you can live with 1K or 2K output and five images in total. Neither takes a mask_url on Sume; that parameter is advertised only on the two GPT Image 2.5 models.

Side by side

Every cell below comes from the Sume image catalog and adapter on main, checked on 2026-10-05. Prices are left out on purpose: read GET /v1/images/models/{id}/endpoints for the exact price line of the tier you will use.

What the shape rows mean

Both models turn a WxH string into a ratio from their own list, so neither promises your exact pixels. For Nano Banana, Sume documents that the exact size comes from a post-step through the job target_pixels field, which is a step after generation (Image API docs). For Ideogram 4.5, the sizes come from a fixed map, such as 1536x864 for a 16:9 edit at 1K.

Ideogram's own API describes edits that match the input dimensions exactly and copy unedited pixels from the source (read 2026-10-05), and that is the behaviour you get on Sume when you omit the shape.

Quality and cost

Ideogram's per-image prices on fal are $0.03, $0.06 and $0.22 for low, medium and high, and do not change with size (fal, read 2026-10-05). Sume bills list x 1.25, so the medium default is $0.075. If you need a 4K deliverable, that decision is already made: only Nano Banana has the tier.

When neither is right

If the edit is a small masked region on a large photo, use GPT Image 2.5 with mask_url so the rest of the picture is not repainted. If you need more than ten references, GPT Image 2.5 lists 16. And for a first pass on a tight budget, run the cheaper tier and judge a sample before you commit the rest of the batch.