GFPGAN (Generative Facial Prior GAN) is a landmark AI face restoration model created by Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan at Tencent Research (ARC Lab) and first presented at CVPR 2021.
The model solves a fundamental challenge in face restoration: how do you recover realistic, high-fidelity detail from a photo where that detail no longer exists? GFPGAN's answer is the Generative Facial Prior — a rich facial knowledge base extracted from a pre-trained StyleGAN2 that contains the natural statistics of real faces. By incorporating this prior into the restoration pipeline, GFPGAN can reconstruct eyes, skin texture, hair, and expression with extraordinary realism.
Unlike simple sharpening or upscaling tools, GFPGAN understands what faces should look like. It detects and aligns facial landmarks, encodes the specific degradation pattern of the input, applies the generative prior to fill in missing or corrupted detail, and seamlessly composites the result back onto the original image — all while working to preserve the subject's unique identity rather than inventing a plausible new face.
Since its release, GFPGAN has become one of the most widely used face restoration models in the world, cited throughout the blind-face-restoration literature and integrated into tools ranging from Stable Diffusion WebUI to ComfyUI. The weights and code are published openly by Tencent ARC on GitHub, where the current licence terms are stated.
This site is independent documentation for GFPGAN, plus a free browser-based photo enhancer. The enhancer is a classical Canvas-API filter — tone mapping, artifact smoothing and unsharp masking — that runs entirely on your device with no upload, no server and no data retention. It is not the GFPGAN neural network; our guides explain how to run the real model locally or on a hosted GPU.