Towards Real-World Blind Face Restoration with Generative Facial Prior

Blind face restoration usually relies on facial priors, such as facialgeometry prior or reference prior, to restore realistic and faithful details.However, very low-quality inputs cannot offer accurate geometric prior whilehigh-quality references are inaccessible, limiting the applicability inreal-world scenarios. In this work, we propose GFP-GAN that leverages rich anddiverse priors encapsulated in a pretrained face GAN for blind facerestoration. This Generative Facial Prior (GFP) is incorporated into the facerestoration process via novel channel-split spatial feature transform layers,which allow our method to achieve a good balance of realness and fidelity.Thanks to the powerful generative facial prior and delicate designs, ourGFP-GAN could jointly restore facial details and enhance colors with just asingle forward pass, while GAN inversion methods require expensiveimage-specific optimization at inference. Extensive experiments show that ourmethod achieves superior performance to prior art on both synthetic andreal-world datasets.