You generated a logo for a coffee brand. Midjourney returned a beautiful image with the word "CFOFEE" stamped across the cup.
You tried again. "COFFEEEE." Then "COFEFE." Then a perfectly drawn but completely fake logo with letters that are not in any alphabet.
This is one of the most well-known image-model failure modes. SDXL, DALL-E 3, and Flux all do versions of it. None of them generate legible text reliably, and all of them will add text to your image whether you asked for it or not.
Here is what causes it and how to suppress it.
Why image models invent text
Text is just another visual pattern. The model has seen millions of images with text on them. Packaging. Signs. T-shirts. Posters. Book covers. So it learned that coffee cup images often have words on them.
When you ask for a coffee cup, the model fills that typical visual region with what looks like text. It has no understanding of letters, spelling, or words. It produces shapes that resemble characters and arranges them where text usually appears.
Midjourney v6 and v7 got better at intentional text. You can now ask for "hello world" written on a sign and get it mostly right. Unintentional text is still a daily occurrence.
The fix
Add these two things to every prompt where you do not want text:
--no text, no letters, no words, no logos, no signs, no labels, no typography
And in the prompt body:
clean composition with no text or written elements
That covers Midjourney (the --no flag) and most other image models (the prose constraint). Together they suppress text generation in around 90% of generations.
When text still appears
Some cases leak through even with both fixes. Each has a specific solution.
Fake logo on the bottle. The model saw "bottle" and pulled a packaged-product pattern. Tell it what the bottle should look like instead:
a plain, unlabeled brown glass coffee bottle on a counter, no markings, no logos, no labels, blank surface
Writing on the building or wall behind the subject. The model added signage to a streetscape because city backgrounds usually have storefronts. Name a different background:
a person walking on a quiet residential side street, brick wall behind, no shops, no signs, no storefronts
Words floating in mid-air. Some prompts trigger text more than others. "Abstract" and "concept art" are common culprits. Replace with visual metaphors:
concept art of innovation, visual metaphors only (gears, light, growth shapes), no text, no symbols, no typography
Fake brand text on clothing. Clothing prompts pull from product photography. Specify plain garments:
a man in a plain solid-color heather-gray hoodie, no logos, no graphics, no brand markings
When you actually want text
If you want intentional text, like the word "OPEN" on a sign:
- Put the exact text in quotes:
a wooden sign that reads "OPEN" in white painted letters. - Keep it short. One to three words max. Longer text gets garbled.
- Use Midjourney v6.1 or v7. Earlier versions cannot do legible text reliably.
- Generate four variations and pick the legible one. Even v7 misses around 30% of the time on text.
Anything longer than three words is a bad idea. Generate the image without text, then add text in Figma or Photoshop in post.
The aspect ratio trick
This one surprises people. Certain aspect ratios trigger more text generation than others.
--ar 16:9 (landscape) often adds banner or sign text. --ar 1:1 (square) tends to add logo-like marks. --ar 2:3 (portrait) and --ar 4:5 (social) are usually clean.
If you keep getting text leakage and the composition allows it, switch to a portrait ratio.
A faster way to check
Paste an image prompt into FixMyPrompt with image mode on. The rubric flags missing --no text negative prompts, missing "no signs, no logos, no labels" exclusions, subject types known to trigger text leak (bottles, books, signs, clothing, devices), and aspect ratios that raise the risk.
Three free image-prompt reports per day. No signup.