You can find Artificial Intelligence (AI) software that generates new images, but how about fixing an old picture?
Recently, Louis Bouchard on YouTube featured a free tool developed at Tencent Research Institute, GFP-GAN (Generative Facial Prior Generative Adversarial Network). This new technology works by merging information from two different AI models to improve photos and fill out missing details in a matter of seconds, without compromising on the quality and accuracy.
In conventional approaches, researchers train an AI model to recognize faces and then apply the same method to any given photo. But those systems can only work if they’ve seen similar examples before.
So instead of training a system to recognize human faces, the researchers trained one to create them. They used a neural network called Nvidia’s StyleGAN 2 an existing model, to do so. Then they applied the same algorithm to every face in a picture, and the resulting images looked realistic.
You can try GFP-GAN for yourself by downloading the source code from GitHub. There are also online demos where you can try it out, like the one we used for the sample of Dr. Jose Rizal.
This project is still limited by the technology of today. While it is quite accurate, it may not produce what you expect in some cases. The researchers warn that you could get a slight variation in the output image and a lower quality than expected.