It feels like being ignored, and the instinct is to shout: repeat the clause, capitalise it, add "must". That rarely works, because the model is not choosing to skip your instruction. It is resolving your prompt into a single image, and some of what you wrote cannot coexist, cannot be counted, or cannot be placed. Knowing which of those three is happening tells you what to change.
Conflicts
Two clauses that describe the same property differently produce an average, and averages look like neither. "Minimalist, densely decorated" gives cluttered-but-beige. "Wide shot, extreme close-up of the eyes" gives a mid-shot. Read your prompt for pairs that argue and delete one — the deleted one is usually a word you added early and forgot about.
Counting
"Three birds" gives you two, four, or five. Exact object counts above two or three are unreliable across every current engine, including the three we publish with, because nothing in the generation process counts. It renders something that looks like a plausible arrangement of birds. If a specific number matters — a product set, a diagram — either accept regeneration until it lands, or compose the count yourself afterwards from single-object images.
Spatial relationships and attribute binding
| What you wrote | What often comes back | Why |
|---|---|---|
| a woman in a red coat and a man in a blue coat | Both coats red, or the colours swapped | Attributes leak between subjects that share a sentence |
| the cup to the left of the book | The two objects in the correct scene, arbitrary order | Left/right is weakly represented; near/far is stronger |
| a sign behind the character reading OPEN | The sign in front, or unreadable | Depth ordering and legible text are two separate hard problems stacked |
Attribute leaking has one reliable prompt-side fix: separate the subjects into their own clauses with their own full descriptions, and give them different framing roles — one in the foreground, one further back — so the model has a reason to keep them distinct. Fighting for exact left/right placement is usually not worth it; crop or flip afterwards.
The thing that is genuinely dropped
Long prompts do lose material at the end. If a clause survives when the prompt is short and vanishes when it is long, that is truncation-by-competition, not a conflict — move it earlier and cut something else rather than restating it twice.
Repeating a phrase to force it does have an effect, but the effect is usually exaggeration rather than compliance: "red, deep red, vivid red" gives a saturated red everywhere in the frame, including on things you did not want red.