Learning a self-supervised tone mapping operator via feature contrast masking loss.