Nano Banana 2.1 small text looks blurry? A three-row test on Sume

Yes, Google's own model card for Nano Banana 2.1 says small text often comes out blurry, so a poster with fine print is the first thing to test before you commit a job to it. The same card lists text rendering for posters and diagrams as a strength, which is not a contradiction: large headlines and small body copy are different tests.

This post turns that one limitation line into a cheap experiment on Sume: one prompt set, three image rows, a priced plan, and a rule for choosing. The Google statements come from the Nano Banana 2.1 model card and the Gemini API image generation guide, both read on 2026-10-10. The Sume rows and prices come from the repository's image catalog and the Image API docs.

What the model card actually says

Reading the card, the limitations list is short and specific. Small text often appears blurry. Character consistency between the input images and the generated image is not always perfect. 3D reasoning and world knowledge are limited, spatial localization can be confused, and the knowledge cutoff is March 2026. On the strengths side the card names text rendering for posters and diagrams, multi-language text localization, and mask or ink-style editing.

The Gemini API guide adds that Nano Banana 2.1 takes up to 10 object images, 4 character images and 3 style references, and supports 1K, 2K and 4K output. None of that tells you whether your 9-point footer will be readable, because neither page gives a size threshold. That is why the test below uses your own layout.

Pick three Sume rows to compare

Sume lists nano-banana-2.1, ideogram-v4.5 and flux-2-pro among others in its image catalog. The Image API documents Ideogram 4.5 as taking text-only calls or up to five images for an edit, with quality set to low, medium or high (default medium). Those three give you a Google row, a row with a selectable quality ladder, and a third architecture. The prices below are the catalog list times 1.25, before cent rounding.

Price the test before you run it

Write 12 prompts that put the same text at the sizes your design needs: a headline, a subhead, a price line, and 8 to 10 words of body copy or a legal line. For one pass of those 12 prompts at Nano Banana 2.1 1K, Ideogram 4.5 medium and Ideogram 4.5 high, the cost is 12 x ($0.10 + $0.075 + $0.275) = 12 x $0.45 = $5.40. Per row that is $1.20, $0.90 and $3.30. Adding Flux 2 Pro would add 12 x $0.0375 = $0.45.

Send every request through POST /v1/images. It blocks up to 30 seconds and returns 200 with image URLs; if the configuration is slow it returns 202 and a job to poll, so check the status code, not the body shape. Keep aspect_ratio identical across rows, and read each row's supported_parameters first, because a parameter a model does not list returns 400 unsupported_parameter.

Judge it the way a customer would see it

That last option is often cheaper than paying a higher tier for every image.

• View each result at the size it will ship: a phone feed is not a 4K monitor, so crop to 100 percent only for the footer line.
• Mark a cell failed if any character in a required string is wrong, not just blurry. Prices, dates and legal lines tolerate no typos.
• Count only the prompts that pass first try; cost per usable image is total spend divided by passes.
• If Nano Banana 2.1 passes the headline rows but fails the small-text rows, generate the art there and place the fine print in a layout tool or the editor afterward.

What this test does not show

It does not rank the models in general; it ranks them on your copy. The Google card is also a statement of limits at launch, and Google can revise it, so re-read the card when you reuse the plan later. For region edits, remember that Sume's Nano Banana 2.1 row takes a prompt and references, not a mask field; the mask_url field belongs to the GPT Image 2.5 rows.