GFPGAN

GFPGAN is an AI face restoration tool for repairing low-quality photos and adding restoration to image workflows. It reconstructs facial detail using a pretrained generative model, so results can alter a person's likeness.

• Face restoration: Restore degraded faces using models that make different tradeoffs between sharpness, natural appearance and identity preservation.
• Background enhancement: Pair face restoration with Real-ESRGAN to improve the surrounding image and other non-face regions.
• Colorization choice: Choose models with or without face colorization; some sharper outputs can look unnatural, while more natural outputs may change identity slightly.
• Model development: Use the included training code to adapt GFPGAN, or use the supplied RestoreFormer inference support.

GFPGAN runs locally with Python and PyTorch, supports Linux and can optionally use an NVIDIA GPU through CUDA. Its clean implementation doesn't require custom CUDA extensions. Hosted demonstrations are available through Hugging Face Spaces with Gradio and Google Colab.