StruPO: Structure-Aware Prompt Editing for Text-to-Image Models

StruPO uses calibrated scene graphs as shared representations to enable faithful text-to-image prompt editing. The method uses deletion, reordering, and insertion to modify prompts while preserving original object attributes and relations through image-grounded supervision.