ID-PreFeR: ID-Preserving Face Restoration with Mixed Data Quality

This paper introduces ID-PreFeR, a robust identity-preservingface restoration method that tackles the ill-posed face restoration prob-lem by incorporating personalized identity information. Existing approachesoften suffer from high training and storage costs and are sensitive tothe quality of reference images. To address these issues, we propose alightweight personalization injector that enables efficient personalizationwithout the need for regularization data. We also introduce an iden-tity–quality disentanglement training strategy to ensure robust identitylearning, even when some reference images are of poor quality. Further-more, we propose an identity-preserving sampling strategy to enhanceidentity fidelity during inference. Extensive experiments on both syn-thetic data and a newly collected real-world mobile-phone dataset verifythe effectiveness and practicality of the proposed method.