Prompt Debt
Image generation gets harder when the conversation gets longer.
With image generation, iteration is not always improvement.
You start with a clear idea. Then: keep this, change that, make it darker, move the subject, restore the previous composition, add one element, remove another. After enough turns, the prompt is no longer a specification. It becomes a history.
That history can create drift. Earlier constraints remain in play, new ones compete with them, and the model has to infer which parts still matter. The result may become less coherent even as the instructions become more detailed.
The best defense is simple: get the image specification right as early as possible.
Define the composition, subject, proportions, camera angle, palette, lighting and exclusions in the first prompt. During revisions, distinguish clearly between what must remain unchanged and what should change.
And when the image starts drifting, stop patching it.
Take what worked, rewrite the complete prompt from scratch, and generate again.
In visual AI, a clean specification is often more valuable than another correction.
