Artificial intelligence repairing your old photos helps you recover what seemed lost forever

The job of a restorer is a fascinating art that brings an artistic manifestation, a painting, a sculpture that seems lost in the mists of time into the present. In 2022, however, even this will have fallen into the hands of artificial intelligence.

In recent months, much of the internet has revolved around artificial intelligences capable of creating paintings by dictation from a few words. Dall-E 2 is one such example, while MidJourney is another, both of which are anchored in an invitation-based mechanism or possibly a monthly subscription. But what do you do when you want to restore an old image, one that no longer looks very good, as can be seen in the image above. This is where GFP-GAN (Generative Facial Prior – Generative Adversarial Network) comes into the equation.

How to restore old AI-based images

Completely free, GFP – GAN is available here in the theoretical form developed by Tencent researchers, while a practical implementation with which you can do some experiments can be found here.The free software tool was publicized by PetaPixel and, at first glance, works great at restoring images, especially low-resolution portraits.

Basically, if you haven’t heard of GFP – GAN, the algorithm forces two artificial intelligences to compete to create missing content from a photo with realistic detail, while trying to maintain a higher level of visual accuracy and content quality.

The idea isn’t completely new, but it might work better compared to algorithms that restore images by identifying differences between the original photo and the artificial, invented one. As a difference in mechanism, the new system embraces a “pre-trained” version of a high-performance AI (NVIDIA StyleGAN-2). In just a few words, this artificial intelligence goes to great lengths to preserve the identity of the people in the photos, with a focus on facial features such as the eyes or mouth, which basically define the most memorable part of the image. Predictably, the results are far from perfect, but it works surprisingly well, even when it tweaks the facial features it’s trying to preserve a little too much.