Why AI images come out with broken text — and how to ask for the right language

Ask for a poster and you get a poster: a title, a credit block, sometimes a release line. The model invents all of it, because a prompt that describes a layout rarely spells out the words. That is the first surprise. The second is that the invented words are usually English — and that if you want them in Japanese, Korean or Chinese, you have to say so.

The sentence that decides it

One sentence, appended to whatever the prompt already asks for: “Any visible text in the image must be written in Japanese (日本語). Keep the typography, placement and layout of the design, and do not add text that is not part of the design.” Naming the language in English, and giving it in its own script in brackets, is deliberate — the prompt is English, and the bracket removes any ambiguity about which script is meant.

What it does, tested rather than asserted

We ran the same template prompt twice on the same model, changing nothing but that sentence. Without it, the one-sheet came back reading “SILENT ECHO” over “THIS AUTUMN”, in a heavy condensed face — an English-language poster, inventing English words. With the Japanese directive, the same prompt returned a poster with a Japanese title set in Japanese typography, and the credit line beneath it in Japanese as well.

Two honest caveats, both visible in the pair of images. The model still invents the words: the English one happened to be plausible, the Japanese one was nonsense — kanji and kana, correctly formed, meaning nothing. And asking for Japanese also changed the subject and the styling, because the model read the directive as context about the whole picture. Neither is a defect you can prompt away; both are worth knowing before you spend credits on a language you cannot check.

What to do with that

• Keep in-image text short. A two-word title renders far more reliably than a paragraph, in every script.
• Check the words before you judge the picture, and regenerate rather than reuse a frame with gibberish — a poster that reads as nonsense undermines the design around it.
• For Latin scripts, expect the model to be closer to correct; for Japanese, Korean, Chinese, Arabic or Devanagari, check character by character.
• If the text does not matter, say so: “no text anywhere in the image” is a legitimate instruction, and it removes the failure mode entirely.
• If you will add your own lettering in an editor afterwards, ask for the layout without words — an empty space reserved for a title is easier to fill than a wrong title is to remove.

In the Image Prompts Studio, “text inside the image” is a control next to the prompt: keep the template’s design, remove all text, or pick a language. Write your own line in a non-Latin script and the control follows it automatically — and stops following the moment you choose something yourself. The studio is at /image-prompts-studio.