Text Kept Sneaking Into My Images
When I ask an AI for card news backgrounds, it sometimes slips in English letters and logo-like patterns I never asked for. One day, after averaging 5 retries just to get a single usable image, I changed the prompt. I kept describing what I wanted, but added one line at the end: a 「do not include」 list.
Try It in 3 Steps
1. Gather the 10 images you rejected so far and write down the common causes. (text, fingers, watermarks, and so on)
2. Paste those causes as a single line at the very end of your prompt.
3. Generate 10 new images on the same topic and count how many are usable.
Create a background image. Pastel tones, plenty of empty space.
Do not include: letters, numbers, logos, watermarks, human hands
Result
On the same topic with 10 images, the average number of retries to get one usable image dropped from 5 to 2. The biggest change was text. We add the card news text ourselves later, so letters baked into the image were always in the way.
What I Learned
• Keep the 「do not include」 list short and use nouns only. Long sentences sometimes made the AI draw the very thing I wanted to avoid.
• Past 5 items, the effect faded. Keep only the usual offenders.
• Running it on the Windows device or the Mac mini gave similar results, since both used the same prompt file.
Today, look at just 3 rejected images, find one thing they share, and add it to the end of your prompt.