When AI images go wrong: fixes from 300 real tests
At Promptpan we test our cards by pasting each prompt exactly as written and generating the image. We now have 320 test records. In about 100 of them the first result had a problem, so we changed the prompt or ran it again.
This guide collects those records. It isn't general advice: we only list fixes we compared before and after, and we list what we could not fix at the end. Each card was tested with set example values, so your results may differ with other inputs.
1. Text you never asked for appears
This was the most common problem. A billboard ad came back with random lettering on nearby shop signs. A hanok background grew a white line that looked like writing above a door. Book spines picked up letter-like marks. One exam paper gained an "A+" nobody requested.
Image models tend to fill anything that could carry text, such as signs, books and paper. Two things help:
- State the exact words, and say that no other text may appear
- If you don't need the object that attracts text, leave that object out
The billboard card needed two rounds of edits to get there. If one change doesn't work, more specific wording often does.
2. Things you didn't choose get mixed in
On a pep-talk poster, choosing "flowers" also brought in a sun and a road. A window silhouette card added a cat, books and a lantern. When a prompt describes a mood, the model adds props that seem to fit.
What worked was naming what to leave out. These two lines removed the props from the window card:
No interior windowsill, cat, books, candles, lanterns or props.
Put a small wreath on the left window frame and keep the top center of the sky empty.
Avoid vague wording like "keep props minimal." Look at the result and write down exactly what bothers you.
3. Illustrations come out looking like photos
You want a drawing, but the skin and clothes look photographed. Our animal-ear hood character looked like a realistic doll, and a paper-cut image barely looked like paper.
"Make it a cute illustration" wasn't enough. What worked was specifying the material of the whole scene:
Draw the person, cat, outfit and entire room as flat 2D anime, with thin outlines and simple cel shading.
The whole scene is a layered matte paper-cut illustration, not a real photo.
The key word is "entire," not just the main subject. Restyling an uploaded photo is harder. The instruction to keep the face competes with the instruction to change the style, and some cards kept a photo look even after rewording.
4. Text gets pushed to the edges
Titles run past the margin or get cut off. Our bucket list wallpaper had a large title break its margin four times.
The fix was simple: stop asking for a big title and remove the decoration. Setting the title in the same small handwriting as the list items, with no sticker or underline behind it, solved it. On the pep-talk poster, a line that gathered the text in the center did the same job.
Percentages like "keep text within 12 to 88% of the width" usually widen the margin, but the result won't match those numbers exactly. Treat numbers as a direction.
5. Colors change in your photo
If a text field disagrees with your product photo, the model may change the photo. A navy product once came out light green. After adding "the attached photo's color takes priority" and rerunning the same input, the navy held.
Small changes can remain, such as different bread shine or a swapped plate. For product photos, always compare the result side by side with the original.
6. Expressions and poses come out wrong
Naming an emotion often fails. In a four-panel comic we wanted a tired character, who was already smiling in panel one. Describing the face and body worked better:
Lowered eyebrows, a small closed unsmiling mouth, slumped shoulders, holding a going-out bag.
To connect the last panel to an earlier one, describe an action, such as "puts the same bag down beside them," rather than saying "make it connect."
7. Dates and live information are wrong
"Check today's date and add it" could not be trusted. Our tests ran in November 2026, yet results kept printing 2025 dates. We also had no way to confirm any lookup really happened during image generation.
Our outfit card dropped the automatic lookup and asks you to type the date, and the date then came out correct. If a date, weather, price or opening hour matters, enter it yourself.
The same goes for sources. Asking for a source on the image produced plausible agency names, and we couldn't confirm they had said any such thing. Our seasonal food calendar fixed this by removing the source request entirely.
What we never managed to fix
Rewording didn't solve these in our tests. If you need one of them, a different method is faster than more prompt text.
| Goal | Test result | What to do instead |
|---|---|---|
| A pattern that repeats seamlessly on all four sides | Top and bottom seams showed in both tries | Use it as a single background, and test two copies side by side before printing |
| Initials on a small object such as a sword hilt | Unreadable in all four tries | Add the letters yourself afterward |
| Exact pixel dimensions | The ratio was right but the pixel count wasn't | Resize after downloading |
| Keeping every garnish in a food photo in place | Direction and placement changed | Keep the original photo and add only the text |
| Shop recommendations from live search | We got no proof a search happened | Find the places yourself and enter them |
The order we fix things in
- Run the same input once more. A second try often worked in our tests, for example when a small accessory was missing only from the first image.
- If it's still wrong, name the problem and edit the prompt. "Put the wreath on the left window frame" beats "make it nicer."
- If you hit the same wall after two or three edits, treat it as something the model can't do and switch methods, as in the table above.
For garbled lettering, see Garbled text in AI images. For faces that change, see our photo restyling guide.
FAQ
Where can I see the test results?
Cards that passed show a "Tested" label with the real input and result. Our method is explained on How we make Promptpan.
My result doesn't match your test. Why?
The same input gives a different image each time. Try the steps above, and if the card seems wrong, send its number through our contact page.