Why Does the AI Ignore My Negative Prompt?

Updated 2026-10-27·3 min read·Rendering

Why it happens

Image models learned from pictures paired with captions. Captions describe what is in a picture. Almost nobody captions a photo with what isn't there, so a model has a strong idea of "a man holding a handgun" and a very weak idea of "not." Write "no handgun" and the strongest word in that clause is still handgun.

Many still models have no separate negative field at all, and the ones that do weight it lightly. We ran into this on a shot where a character was meant to hold a plain, unbranded pistol. Telling the model what the gun should not be kept returning a recognizable branded handgun. It went away only when the description of the gun moved into the positive prompt.

Two other things make negatives fail:

• References win. If a character sheet, location still or prop image sent with the shot shows the thing you're removing, the picture beats your words. A pink phone stays pink if its reference image is pink.
• A loose list binds to nothing. "No extra people, no text, no blur" tacked onto the end floats over the whole frame. And in a long list, the category words that matter ("humanoid," "dinosaur") get buried under synonyms written first.

How to fix it

1. Say what IS there. Replace every "no X" with what fills that space. A model draws what you name, so "an empty dark office behind him, rows of unlit desks" works where "no people in the background" doesn't.
2. Check the references before you rewrite. List every image going up with the shot. If one of them shows the unwanted thing, swap or drop that reference first.
3. Bind a constraint to its subject. Put it next to the name it's about ("the dragon, a four-legged winged reptile") instead of in a list at the end.
4. Keep the negatives you do use short and categorical. Short category words do more than five ways of saying the same detail. Drop boilerplate like "blurry, watermark, low quality"; it constrains nothing about your subject.
5. Don't re-describe a referenced character in text. If their sheet is already attached, describing their look again is a second identity signal, and the model can render them twice.
6. If it still fails, make the prompt shorter. Adding one more clause on each attempt just piles up noise.

Example

Before:

A detective at his desk at night. No gun, no badge, no people in the background.

After:

A detective alone at his desk at night, both hands wrapped around a coffee mug, his jacket hung on the chair. Behind him, a dark open-plan office with rows of empty, unlit desks.

The second prompt never mentions a gun or a crowd, so there is nothing for the model to latch onto.

How ShowMaker handles it

In ShowMaker, cast, locations and props are @references with their own stills, so you can see and change the images each shot sends instead of guessing at them. If a frame comes back with something it shouldn't have, Shoot with AI lets you say it in plain words ("take the gun out of his hand, he's holding a mug") and it redraws the frame, showing the credit cost before you spend. The checks on every shot also flag a frame that shows something the text never mentioned. More on building shots in the storyboard docs, or try the AI storyboard generator.