When you describe a scene to an AI image generator, small changes to your description can accidentally alter important details like the color of an object or which objects are present. Automatic prompt rewriting often has this problem because it focuses on overall similarity instead of preserving the specific scene.
StruPO, a new prompt editor, first extracts a scene graph, a map of objects, their attributes, and relationships, from the original prompt. It then learns to make edits that modify only non-essential parts while keeping the core scene intact. The system uses feedback from generated images to improve its editing policy.
In tests across different generators and thousands of prompts, StruPO consistently preserved object relationships better than other methods, according to both automated checks and human judges. This means users get higher-quality images without losing the scene they intended.