You can boost any image assessment model's performance by adding a lightweight refinement layer that uses image geometry and internal visual features—no need to retrain the base model.
MetaRefine improves image quality assessment by refining scores from existing models using metadata (original dimensions, aspect ratio) and multi-level visual features extracted from frozen models. Lightweight residual heads predict score corrections without modifying the original assessment pipeline, achieving 7-23% improvements across 21 datasets with minimal parameter overhead.