Demo AI result: the text position is a fixed estimate and the letters were drawn by your browser. With MOCK_AI set to "false", Nano Banana finds the text and redraws it in the original style.
Working…
Most edits finish in under a minute. Keep this tab open.
Each one would otherwise mean re-rolling the prompt and losing the composition you liked.
Travel posters
Harbour, lighthouse and sun-faded print texture came out perfectly; the headline came out as nonsense. One box over the title, the real word typed in — the illustration around it is untouched.
HRABOURHARBOR
Album and book covers
The colour palette and grain are exactly what the record needs, but the name has doubled letters. Re-rolling would change the whole mood; boxing the title fixes only the title.
THE MIDNIGHHT WAVVESTHE MIDNIGHT WAVES
Game key art
Generators love decorative title plates and rarely spell them right. The gold gradient and outline are redrawn around the corrected letters. For screenshots of real games see editing game fonts.
STAR FOGRESTAR FORGE
Brand mock-ups
A mock-up of your shop sign came out with the right light and texture but the wrong spelling. Box the sign, type your real name. (Fixing logos to spell someone else's trademark is not allowed.)
Farmers MarkeetFarmers Market
Image generators learn what letters look like, not how words are spelled. The prompt reaches the model as sub-word tokens and small text gets very few pixels, so you often get convincing letter shapes in the wrong order. Newer models are much better at short headlines but still slip on long words, several lines of text and non-Latin scripts.
Re-prompting redraws the whole picture, so you trade one problem for another. Fixing in place keeps the generation you liked: you mark each broken line with a box, type the correct words, and only the inside of those boxes is written back. You never have to type the gibberish itself.
About a minute from upload to download.
PNG, JPEG or WebP up to 10 MB — use the upscaled version if you have one.
Drag across the garbled text and type what it should say. Up to eight boxes per edit. Keyboard users can use “Add box” and the arrow keys.
The boxed areas are redrawn in the generation's own lettering; everything else is your original file. If one letter is still off, box it again.
Full list: Acceptable Use Policy.
Check our claims with any AI assistant
“How do I fix garbled text in an AI-generated image without re-prompting?”
Overview of how diffusion image generators work and why they struggle with text.
Documentation for the image model family EditImageText uses for in-place edits.
Background on filling and redrawing part of an image while keeping the rest.
Upload the generation, box the broken words, type what they should say. Your first fix is free.