Tencent trains AI to restore old photographs

Researchers from Tencent have developed the GFP-GAN algorithm that restores damaged photos and low-resolution portraits. As reported by PetaPixel.

The technology combines information from two AI models to fill in missing details of photographs in a few seconds, preserving high accuracy and quality.

According to the scientists, conventional methods for restoring images require meticulous tuning of the existing AI model by measuring differences between generated and real images. This often leads to low-quality results, they added.

The new approach uses a pre-trained version of Nvidia’s StyleGAN-2 algorithm to inform the group’s own model at several stages of the image creation process. This helps preserve the ‘identity’ of people in the photograph, with particular emphasis on facial features such as the eyes and mouth.

The GFP-GAN demo is free. Researchers also published the source code of the model, enabling any developer to implement restoration technology in their projects.

However, the creators of GFP-GAN warned that the algorithm may distort some facial details and output the result at a lower resolution.

In June, Adobe unveiled a tool for restoring old photographs in Photoshop.

In October 2021, the Israeli startup D-ID developed a tool that makes people on portraits speak.

In April, Deep Nostalgia taught ‘live’ photos to dance and send a virtual kiss.

Subscribe to ForkLog’s news on Telegram: ForkLog AI — all the news from the world of AI!