tencentarc/gfpgan

GFPGAN is a Python library developed by Tencent's Applied Research Center that restores damaged, blurry, or low-quality face photos to sharp, detailed images. The problem it tackles is "blind face restoration", meaning the tool has to improve a face image without knowing exactly what kind of damage was applied (compression, blurriness, noise, low resolution, or all of the above at once). The key concept behind how it works is using a pre-trained face GAN (Generative Adversarial Network), specifically StyleGAN2, a model that has already learned what realistic human faces look like, as a guide. Rather than guessing at missing details from scratch, GFPGAN draws on StyleGAN2's built-in knowledge of face structure to fill in realistic-looking eyes, skin texture, and hair. This approach is called a "generative facial prior." The tool detects and aligns faces in an image first, applies the restoration model, and can optionally enhance the non-face background using a companion library called Real-ESRGAN. Multiple model versions exist, with newer versions producing more natural results. You would use GFPGAN when you have old scanned photos of people that are blurry or degraded, or when you have low-resolution video stills from surveillance footage or social media that you want to sharpen. It is also useful for restoring compressed or heavily edited portrait photos. The tool runs from the command line as a Python script and requires Python 3.7 or higher with PyTorch installed. A GPU with CUDA is recommended for reasonable speed, though CPU-only mode works on Windows. The library is also available as a pip package.