GFPGAN

GFPGAN tackles blind face restoration: recovering realistic facial detail from old, damaged, or low-resolution photos. Developed by Tencent ARC, it uses a generative facial prior from a pretrained StyleGAN2 to reconstruct faces in a single pass, and it can pair with Real-ESRGAN to enhance the non-face background at the same time. The repository ships a Python package, inference scripts, training code, and several pretrained weights: V1.3 gives the most natural results on very low or high quality inputs, V1.2 produces sharper output with a beautification effect, and the original V1 model from the paper includes colorization. It runs on Python 3.7 and newer with PyTorch, works on CPU or CUDA GPUs, and a clean version avoids custom CUDA extensions entirely. Released under the Apache 2.0 license, GFPGAN is a standard component in photo restoration and image enhancement pipelines, with hosted demos on Hugging Face Spaces and Replicate.