Generative Facial Prior for Large-Factor Blind Face Super-Resolution

Most existing face super-resolution (SR) methods are developed based on an assumption that the degradation is fixed and known (e.g., bicubic down sampling). However, these methods suffer a severe performance drop in various unknown degradations in real-world applications. Previous methods usually rely on facial priors, such as facial geometry prior or reference prior, to restore realistic face details. Nevertheless, low-quality inputs cannot provide accurate geometric priors while high-quality references are often unavailable, which limits the use of face super-resolution in real-world scenes. In this work, we propose GPLSR which used the rich priors encapsulated in the pre-trained face GAN network to perform blind face super-resolution. This generative facial priori is introduced into the face super-resolution process through channel squeeze-and-excitation spatial feature transformation layer (SE-SFT), which makes our method achieve a good balance between realness and fidelity. Moreover, GPLSR can restores facial details with single forward pass because of powerful generative facial prior information. Extensive experiment shows that when the magnification factor is 16, this method achieves better performance than existing techniques in both synthetic and real datasets.