GFPGAN

GFPGAN (Generative Facial Prior Generative Adversarial Network) is a blind face restoration model developed by Tencent's Applied Research Center (ARC). It is designed to restore low-quality, blurry, or damaged facial images to high-resolution outputs with a single forward pass. The model is highly useful for enhancing historical family photos, repairing heavily compressed web images, and post-processing facial artifacts generated by text-to-image or generative adversarial networks.

Architecture and Core Innovations

The core innovation of GFPGAN lies in leveraging Generative Facial Priors (GFP) encapsulated within a pre-trained face generator (such as StyleGAN2) rather than relying solely on geometric or reference priors. GFPGAN employs a U-Net degradation removal module to eliminate artifacts and extract two distinct sets of features: latent features that map the input to its closest latent code in StyleGAN2, and multi-resolution spatial features. These features are integrated into the restoration process via Channel-Split Spatial Feature Transform (CS-SFT) layers. By applying SFT modulation to only a subset of features while letting the remaining features pass through unaltered, the network achieves a balanced trade-off between visual realism and reconstruction fidelity.

Objectives and Capabilities

During training, GFPGAN incorporates multiple optimization objectives to guide the restoration. The training framework utilizes intermediate L1 reconstruction losses applied at multiple scales within the U-Net module to wipe away complex degradation. It also features facial component losses using localized discriminators targeting perceptually significant areas like the eyes, nose, and mouth to enforce crisp details. Additionally, an identity-preserving loss leveraging feature embeddings from a pre-trained ArcFace face recognition network ensures the restored identity matches the original.

GFPGAN was trained on the FFHQ dataset using synthetic degradation profiles. The resulting system is capable of restoring facial textures, hair, and eyes while simultaneously correcting color shifts in a single forward pass, bypassing the computationally expensive, image-specific optimization steps required by traditional GAN inversion techniques.