A persistent distribution shift between standard AIOIR (and broader low-level vision) training sets and real-world degradations constrains generalization; we plan to mitigate this gap in subsequent ...