Garbled text in AI images — how to get exact words on a poster
AI image tools are great at scenes and terrible at spelling. A poster comes back looking lovely, but "Croissant" turns into "Croisant", a price of 13.00 becomes 18.00, or one letter appears twice. On a sign that customers read, that is not a small problem.
It happens because an image model does not type letters. It paints shapes that look like letters. The longer and smaller the text, the more chances it has to paint the wrong shape. This guide collects the habits we use when we build and check the cards on Promptpan.
1. Keep the text short — two lines at most
Every extra word is another place to go wrong. On a poster, keep only what a passer-by must read.
| What you want to say | What goes inside the image |
|---|---|
| We are excited to introduce our new winter citrus tea | Citrus tea is back |
| Please note we are closed every Monday | Closed Mondays |
| Buy four cans for a special price of 13.00 | 4 cans 13.00 |
Put the long explanation in the caption or on a separate sign. Promptpan poster cards split the text into separate fields — a big title, a price, one short line — for exactly this reason, as in the fresh salt bread poster or the yuja and ginger tea launch poster.
2. Put exact words inside quotation marks
If you write "make the title feel wintery", the model will invent its own words. When the wording is fixed, quote it and forbid anything else.
Put only two lines of text in the image: "Citrus tea is back" and "5.00". Do not add any other text.
Promptpan cards already wrap whatever you type in a field in quotation marks and tell the model not to add other text, so you only need to fill in the fields.
3. Make the text big and the background quiet
Small text breaks first. Ask for a title wider than half the image, and keep the area behind the letters plain. Busy patterns behind text make strokes merge into the background.
Make the title wider than half the image width, on a plain background.
4. Check every character before you print
Zoom in and compare the result with what you typed, character by character. Pay special attention to:
- Digits: 13.00 vs 18.00, 1 vs 7, missing decimal points
- Dates and days: the date and the weekday must match the calendar
- Repeated or missing letters: "Coffe", "Bakeery"
- Punctuation: commas in prices, periods in dates
If the price or date is wrong, do not use the image, however good it looks.
5. Ask to fix only what is wrong
Regenerating from scratch also changes the parts you liked. In the same chat, point to the exact mistake.
Change "Croisant" to "Croissant". Keep the picture and layout the same.
Rewrite the price exactly as "13.00". Do not change anything else.
Fixing one or two things at a time is more reliable than listing everything at once. Read the whole image again after each fix.
6. If it keeps failing, add the text yourself
When the same word stays wrong after a few tries, change the method. Ask for the image without text, with an empty space where the title will go, and add the words in any photo editor or design tool. The text will be exactly right, and next month you can update the dates without regenerating the artwork.
Do not include any text. Keep the top quarter as a plain area for a title.
7. When text is the star, draw it with code
For results where the words are the main subject, such as animated text videos, code is more accurate than an image model. Promptpan motion cards such as One-sentence ad ask an AI to set real fonts and animate them with code, so letters never warp. Whether a service can run code and save a video file depends on the service and plan, so check its help pages first.
FAQ
Does a paid plan fix spelling?
No plan changes how image models draw text, so mistakes still happen. Always proofread the final image.
What about menus with lots of text?
For text-heavy items like menus and price lists, make only the background with AI and add the text yourself (step 6). A wrong price can cause real trouble at the counter.