Edit-Compass & EditReward-Compass: A Unified Benchmark for Image Editing and Reward Modeling
The research created a unified benchmark suite, Edit-Compass and EditReward-Compass, to rigorously test image editing systems and their reward models. Edit-Compass evaluates editing capabilities across 2,388 tasks using five metrics like instruction following. EditReward-Compass simulates realistic RL reward scenarios using this data to assess how well different models learn to generate preferred edits.