If you’re anywhere my age, I’m sure you have a bunch of pictures of your parents and grandparents who had better days (photos, not your ancestors). Anyway, you can now take those old, scratched, wrinkled and torn photos and restore them without any Photoshop knowledge.
A neural network called GFP-GAN (Generating Face Priors Generative Adversarial Network) recovers your old and damaged photos with impressive speed and accuracy. But like all AI-based tools, this one has drawbacks. So, let’s look at the good and the bad and see what it has to offer.
Louis Bouchard brought this tool to our attention in a blog post he wrote. Of course, there is also a research paper where you can read about this model in more detail.In short, when you add an image to GFP-GAN, it just creates a guess Or the face of the person in the photo. However, in most cases they look very close to the original image.
AI tries to understand what’s in the photo and then adds pixels or fills in the gaps. Unlike other similar models, GFP-GAN focuses on important facial features such as eyes and mouth. Finally, the resulting image is compared to the original to see if the same person is still present in both — which leads us to one of the technology’s biggest weaknesses.
Traditional image restoration methods use different techniques to reconstruct damaged or blurred images and create new ones. However, this often results in poor image quality. GFP-GAN uses a pretrained version of an existing model (NVIDIA’s StyleGAN-2) to inform the team’s own model at multiple stages of the image generation process. This is why the identities of the people in the photos are preserved.
Nonetheless, they pointed out some weaknesses of the method. The resulting images may sometimes not be very sharp, some outputs are unnatural, and the identities may still change slightly. We simply cannot be sure that the reconstructed image is the same as the original, which is impossible. “If we’re lucky, this photo looks like our grandfather,” Lewis said, “but it could also look like a complete stranger.” While the results are striking, when using such AI tools This needs to be kept in mind. If you want to test GFP-GAN, you can find it on GitHub.