Nano Banana Pro text in image: a real test, first attempt

Nano Banana Pro is the model to use when the picture has words in it. Google builds and markets it for correctly rendered, legible text directly in the image — a short tagline or a whole paragraph, across languages. We tested it with a shop sign carrying two lines of specific copy, at 1K for 8 credits, and it came back right on the first attempt: every letter, both commas, the em dash and the full stops.

The test

The prompt asked for a hand-painted enamel sign reading exactly AIKIZI — LIGHT, COLOUR, MATERIALS with a smaller line below reading OPEN FROM SIX, cream on indigo, with worn letter edges and warm side light.

What it got right: the spelling of a made-up word, letter spacing that reads as real signwriting, the em dash, both commas, the full stops, a second line at a smaller size, and enamel chipping that follows the letter edges rather than being sprinkled over them.

What drifted, honestly: we asked for one screw at each corner and it fitted six. The punctuation was tidied — we did not write the final full stops and it added them, which on a sign is the right call and is still the model deciding. Neither is a text failure; both are the model finishing a sentence you left open.

How to write a prompt with text in it

1. Write the words once, in capitals if you want capitals, and say "reading exactly" before them. Do not describe the words — quote them.
2. Say what the text is on: an enamel sign, a painted wall, a book cover, a label. A surface gives the letters a reason to sit where they sit.
3. Give the second line its own sentence and say it is smaller. Hierarchy has to be stated or you get one block.
4. Add "no other text anywhere in frame". Without it, models like to decorate the background with extra lettering.
5. Check at 1K for 8 credits before going to 2K or 4K. Text is the thing most worth verifying cheaply.

Which model, and what it costs

A sensible order: draft the composition on Lite at 2 credits with placeholder wording, then run the real copy once on Nano Banana Pro at 1K for 8. If the words are right and the picture is right, upscale for 15 rather than regenerating at 4K for 24 — a new 4K generation re-rolls the text too, and you may not win twice.

The honest limits

• Long paragraphs are harder than short lines. Every extra word is another chance to drop a letter.
• Proper nouns and invented words sometimes get "corrected" toward real ones. Quoting them exactly helps; checking is still on you.
• The model will tidy your punctuation. If the punctuation matters, say so.
• Nothing here is a typesetter. For exact kerning and a locked typeface, generate the picture and set the type over it.

FAQ

Can Nano Banana Pro put readable text in an image?

Yes — it is what Google builds and markets the model for, from a short tagline to a whole paragraph, across languages. In our test a two-line sign came back correct on the first attempt at 1K for 8 credits, including the em dash and the commas.

How do I write a prompt so the text comes out right?

Quote the words after "reading exactly" rather than describing them, say what surface the text sits on, give a second line its own sentence and say it is smaller, and add "no other text anywhere in frame".

Which model is best for text in an image?

Nano Banana Pro at 8 credits for 1K. Nano Banana 2 at 4 is usable but less reliable on long strings, and Nano Banana 2 Lite at 2 is for drafting the layout rather than the wording.

Why did the model change my punctuation?

Because it finishes sentences you leave open — ours added full stops we had not written. On a sign that was the right call, but it is still the model deciding. Say so explicitly if the punctuation matters.

Should I generate text at 4K?

Check it at 1K for 8 credits first. If the words and the picture are both right, upscale for 15 — a fresh 4K generation re-rolls the text as well as the composition, and you may not get it right twice.

Try it with your own wording in the Studio at 1K, which is the cheapest way to find out whether your copy survives. Why Pro over the cheaper model is part 2 of the generation series, and the upscale-versus-regenerate case is part 8.