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

Chengxuan Zhu, Yuchen Hong, Qingnan Fan, Qi Zhang, Bingtao Fu, Jinxiu Liang, Jinwei Chen, Huaqi Zhang, Chao Xu, Boxin Shi

Keywords: face restoration, diffusion, image enhancement

TL;DR: A face restoration method based on a personalized diffusion model, and is able to handle low-quality reference images.

Abstract: This paper introduces ID-PreFeR, a robust ID-preserving face restoration method that addresses the ill-posed face restoration problem by introducing personalized information. Existing methods often suffer from computationally expensive training and storage requirements while being sensitive to the quality of reference images. We present a lightweight personalized injector to enable efficient personalization without the burden of regularization data. Besides, we propose an ID-quality disentanglement training strategy to ensure robust identity learning, even when some of the reference images are of low-quality. An ID-preserving sampling strategy is further proposed to enhance the identity fidelity during inference. Experiments on both synthetic and a newly collected real-world mobile phone dataset validate the effectiveness and practicality of the proposed method.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 7034