Fix text in AI generated images

Upload the image here, press Read text so the model lists each line as it reads, type the correct wording next to any line that is wrong, and press Fix the text. The image model removes the wrong letters and draws the right words in the same place, in the same lettering, colour and effects. You get the same picture back with the text corrected; nothing else is regenerated.

Image generators draw letters as shapes rather than spelling words, so a plausible-looking shape in the right spot satisfies them. Short, common words usually come out right; long words, invented names, numbers and small print are where the mistakes land. Newer models make fewer of them, but a wrong date or a doubled letter still slips through.

Use a text edit instead of a new generation. On this page you type the misspelled line and the correct line, and only those letters are redrawn; the composition, colours and everything else stay as they were. Regenerating, even with the text in quotes, gives you a different picture each time.

Type it roughly as it reads (Read text will give you the model's best guess) and put the words you want in the second box. The edit replaces whatever is in that spot with the right words in the same style. If the gibberish is long and the replacement is short, the new line is set at the same size and centred where the old one was.

The new letters are drawn to match the existing lettering by eye: style, weight, colour, size and position, plus glow, outline, gradient and shadow. In our 26-design text-edit test the lettering carried over in 25 of 26 before/after crops on Nano Banana 2. You do not need to know or own the font.

It is asked to change nothing else, and in our tests the rest stays put; the median share of the picture that changed outside the edited lines was under one percent. Because the whole image is regenerated by the model, tiny details can shift slightly, so compare before and after and keep the original.

Yes. The fix is done on the exported picture, so the generator does not matter. Upload the JPG, PNG or WebP you downloaded from Midjourney, ChatGPT, Gemini, Nano Banana, GPT Image, Seedream, Flux, Ideogram, Stable Diffusion, Canva, Firefly or Leonardo.

Yes. The page works in the phone browser: upload from your camera roll, type the correction and download the PNG. No app is needed.

Start with Seedream 5.0 Flash (1 credit). On the two posters above it corrected every misspelled line in one pass. Nano Banana 2 draws the closest lettering on ordinary edits, but it reads a near-miss such as “OPENNING” or “BLUBERRY” as if it were already spelled right and leaves it alone: three runs on the bakery poster and two on the chalkboard each fixed one line and skipped the other. If Flash changes the lettering more than you like, try Nano Banana 2 with the wrong word on its own, or GPT Image 2.

Sign in with Google (a verified Gmail account) and your first text edit, or one layer split, is free, with no card. After that each edit uses credits: 1 to 3.1 per image depending on the model, from a one-time pack starting at $9.90 for 100 credits.