benchrWhen it breaks
The poster looks great until you read the words on it
Also asked as text inside generated images looks wrong · the letters are nonsense · AI can't spell in images · my logo text came out mangled
Cause documented by a vendor or a published paper Not tested by benchr First logged Last checked
What is actually happening
Legible text inside an image is a specific documented capability, not a property of image generation in general. On a model that does not have it, letters are texture.
Why
- Image models that render readable text do so as a named feature; the ones that do not were never claiming to.
- Even on models that support it, the documented reliability drops with the amount of text and with unusual typography.
- Long strings, dense paragraphs and non-Latin scripts are the first things to fail.
The quick fix
Keep the text to a few words, put it in straight quotation marks in the prompt, and say where on the image it goes.
The real fix
Generate the image without the text, then set the type in an editor. This is not a workaround — it is what a designer would do anyway, and it makes the wording editable.
Step by step
- Check whether the model you are using documents readable text at all before blaming the prompt.
- Quote the exact string. Do not describe it.
- Ask for one short line rather than a headline plus a subhead plus a footer.
- Leave deliberate empty space in the composition and add the type afterwards.
- For Arabic and other connected scripts, treat generated lettering as decorative until you have checked every glyph.
Grounded in
- Generate an image with legible text inside itReadable in-image text is documented as a capability with named limits.