This Free AI Tool Restores Family Photos With a Bit of Guesswork

An example of how other restoration tools compare to GFP-GAN.
An example of how other restoration tools compare to GFP-GAN. Wang, Li, Zhang, Shan

We’ve reached a point where AI photo restoration tools are relatively common. But if you want to fix a family photo for free, you’re almost guaranteed to end up with terrible results. That’s why Tencent’s AI researchers launched GFP-GAN, a free and open-source restoration tool that takes just seconds to fix an old photo.

The GFP-GAN (Generative Facial Prior-Generative Adversarial Network) is quite unique. Unlike most AI restoration tools, it aims to fix old photos without obscuring the identity of their subject. That means clearing noise, patching creases, bringing out detail, and enhancing color with just a bit of guesswork.

As you can see in the above images, GFP-GAN doesn’t add much to photos. It simply cleans things up and, when necessary, fills in some gaps. Subjects aren’t losing their facial hair or mutating into a stranger. This is partially thanks to NVIDIA’s StyleGAN-2, which Tencent researchers use alongside their own GFP-GAN model. (NVIDIA’s model made waves in 2020 when it recreated Pac-Man, the arcade game, from scratch. It’s good at making safe, educated guesses.)

Now, AI restoration tools are never perfect, and GFP-GAN does its fair share of guesswork. Tencent researchers warn that restored photos may not be a satisfactory resolution, and that in some cases, family members may look a bit off. If you’re unlucky, they could even transform into another person.

You can try GFP-GAN now in your browser or download the source code at Github. In theory, anyone can take GFP-GAN and tweak it to fulfill different tasks or integrate it with new software.

Source: Tencent via Engadget