Avoiding negation in prompts prevents unwanted duplication in generated images

The challenge of maintaining character consistency in a children’s book project revealed a critical flaw in how image models process prompts. Over three weeks, inserting a single child’s photo into thirty books—each with eleven pages—proved difficult due to persistent duplication errors. Scenes repeatedly generated two children or two doctors, even after adding explicit instructions like "EXACTLY ONE child" or "the child appears only once."

The root issue stems from a "mention-is-summon" effect: naming a noun—even to exclude it—triggers the model to focus on that noun. A prompt like "do not add a second doctor" failed because the model treated "second" and "doctor" as key terms, reinforcing their presence rather than suppressing them. The fix required stripping all negation and reducing the prompt to a simple, direct instruction:

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

This minimal approach resolved the duplication on the first attempt. The lesson applies beyond exclusions: describing an object’s altered state (e.g., "collapsed sandcastle") often fails because the noun itself dominates the output. Instead, focus on the visual aftermath—such as "a big flattened mound of damp sand where something was built and then squashed"—to achieve non-standard results.

Even minor syntax, like commas in action lists, can misdirect the model. A prompt like kneeling by the pond, holding a jar, looking up produced three separate versions of the same action rather than a single cohesive scene. Rewriting it as a continuous phrase—kneeling by the pond and holding a jar while looking up—eliminated the fragmentation.

Precision in positioning also prevents unintended duplicates. Symmetric descriptions (e.g., "centered") or vague anchors (e.g., "at the right of the shelf") may trigger the model to mirror or distribute elements symmetrically. Specifying exact coordinates or avoiding symmetry entirely ensures consistency in output.