Rubric-CEPR: Self-Evolving Image Editing via Reward-Verified Self-Distillation
Rubric-CEPR improves image editors by having them verify their own edits using internal representations instead of human targets or external rewards. It uses a Planner to create instructions, an Editor to generate edits, and a Critic to score them based on specific checks for edit realization and content preservation. The best candidate is then distilled into the editor.