How It Started
Whenever I generated background images for our card news, I tacked these lines onto the end of the prompt:
No text. No human hands. No dark colors.
Yet when I opened the results, there were text-like patterns in the background, and a few fingers poked in from somewhere. I had clearly said not to.
A Small Experiment
Over one week I generated the same 10 scenes in two different ways and compared them.
• **Version A (negative):** «Do not include text, hands, or dark colors»
• **Version B (positive):** «Cream paper texture, generous empty space in the center, a still life of objects only, bright pastel tones»
There was just one measure: how many times I had to regenerate before getting an image I could use right away.
Version A average regenerations: 3.4
Version B average regenerations: 1.8
Why Might This Happen
This is only my guess, but the moment you hand the model the word «hand», that concept seems to step into the scene. Describing the scene you actually want worked far better than naming what you want removed.
A New Homework Item
Version B results all looked a bit alike. So I decided to change just one prop per scene: a pencil on the first card, a teacup on the second, a single dried leaf on the third.
Today's Notes
• Do not name what you want gone. Fill the frame with what you want instead.
• Change only one prop at a time.
• Ten images is a tiny sample, so treat the numbers as a rough guide.