Skip to main content
Text placed inside a generated image can come out broken, or different from the wording you asked for.

​
1. Difficulty depends on the script and the length

Numbers and Latin letters usually come out accurately, while scripts with more strokes, such as Korean, are harder to draw inside an image. The more characters and the longer the phrase, the more often it fails. Results vary a lot by model, and the high performance image models available on paid plans render in image text far more reliably.

When Korean text breaks

A closer look at why Korean text breaks and what to do on each plan.

​
2. How to improve results

Paid plans can specify a high performance model that renders in image text well.
Example prompt: “(Enter desired content) Use Nanobanana Pro”
The free plan can only use the Carat default image model. If you specify another model, you will see a notice that it is available on paid plans. These approaches raise your success rate on the default model:
  • Ask for shorter text in the image.
  • Upload an image that already has the text and ask for an edit. The original text is preserved.
When exact wording matters most, generate the image without text and add the wording yourself in an editing tool. If you need in image text regularly, subscribing to a paid plan gives you access to the high performance image models.