HiFaceGAN: Face Renovation via Collaborative Suppression and Replenishment

Existing face restoration researches typically relies on either thedegradation prior or explicit guidance labels for training, which often resultsin limited generalization ability over real-world images with heterogeneousdegradations and rich background contents. In this paper, we investigate themore challenging and practical "dual-blind" version of the problem by liftingthe requirements on both types of prior, termed as "Face Renovation"(FR).Specifically, we formulated FR as a semantic-guided generation problem andtackle it with a collaborative suppression and replenishment (CSR) approach.This leads to HiFaceGAN, a multi-stage framework containing several nested CSRunits that progressively replenish facial details based on the hierarchicalsemantic guidance extracted from the front-end content-adaptive suppressionmodules. Extensive experiments on both synthetic and real face images haveverified the superior performance of HiFaceGAN over a wide range of challengingrestoration subtasks, demonstrating its versatility, robustness andgeneralization ability towards real-world face processing applications.