Edit-Compass & EditReward-Compass: A Unified Benchmark for Image Editing and Reward Modeling
Researchers have introduced Edit-Compass and EditReward-Compass, a unified benchmark designed to more accurately evaluate image editing models and their associated reward models. The new suite addresses limitations in existing benchmarks, which often fail to reflect human judgment due to insufficient task difficulty and coarse evaluation methods. Edit-Compass features 2,388 annotated instances across six difficulty levels, assessing capabilities like reasoning and multi-image editing with a fine-grained multidimensional framework. EditReward-Compass includes 2,251 preference pairs to simulate realistic reward modeling scenarios for reinforcement learning optimization. AI
IMPACT Provides a more robust evaluation framework for image editing and reward models, potentially leading to more accurate assessments and improved model development.