Stop using negation in your prompts if you want to avoid

I recently spent three weeks trying to maintain character consistency for a children's book project. The goal was simple: take one photo of a kid and place them into 30 different books, each about 11 pages long, using a free image model. The nightmare was the duplication bug. I kept getting two children or two doctors in a single scene, and my initial instinct to "fix" it actually made it worse.

The "Mention-is-Summon" Trap

When a scene in a doctor's office kept rendering two kids, I did what most people do: I got more explicit. I wrote "EXACTLY ONE child," "do not add a second doctor," and "the child appears only once." I ended up with a prompt nearly 2,000 characters long, mentioning "do not duplicate" seven different times.

The result? Still two children and two doctors.

The breakthrough happened when I stripped everything back to a single, simple sentence:

Add the child from the first image into the second image.

It worked on the first try. The realization here is that for many LLMs and image generators, mentioning a noun—even to forbid it—actually summons that noun into the latent space. When you say "no second kitten," the model focuses on "second kitten." Negation doesn't subtract; it just reinforces the presence of the object. To suppress something, stop naming it.

Describing States without Naming Objects

This "mention-is-summon" logic applies to the state of an object, too. I needed a image of a collapsed sandcastle. I tried a sandcastle collapsing into a heap and broken heap, but the model kept delivering a pristine, perfect sandcastle. The noun "sandcastle" carries such a strong canonical image that adjectives like "broken" or "collapsed" can't override it.

The fix was to describe the visual result without using the word "sandcastle" at all:

a big flattened mound of damp sand where something was built and then squashed, one broken turret fragment lying on its side

This worked 100% of the time. If you want a non-standard state (empty, abandoned, destroyed), describe the debris or the void rather than naming the object you want altered.

The Hidden Danger of Commas

This is a weird technical quirk that cost me a massive amount of time. In my AI workflow, I had an "action" field for the characters. I was using commas to list actions: kneeling by the pond, holding a jar, looking up.

The model interpreted those commas as separators for parallel subjects. Instead of one kid doing three things, I got three versions of the same kid performing those actions separately. Switching to a fluid sentence solved it:

kneeling by the pond and holding a jar while looking up

When I audited my project, I found that nearly 90% of my prompts had these commas. Most weren't "failing" obviously, but they were introducing instability into the composition.

Anchor Points and Symmetry

Finally, avoid symmetric composition if you don't want duplicates. If you describe a prop as "centered" with empty space on both sides, the model often resolves that ambiguity by mirroring the object, placing one on each side. Similarly, vague anchors like at the right of the shelf fail if the shelf spans the whole image. Be surgically precise about positioning to avoid the model "filling in the blanks" with extra characters.