If you’re anywhere around my age, I’m sure you have a bunch of photos of your parents and grandparents that have seen better days (the photos, not your ancestors). Anyhow, you can now take those old, scratched, crinkled, and torn prints and restore them without any prior Photoshop knowledge.
A neural network called GFP-GAN (Generative Facial Prior-Generative Adversarial Network) restores your old and damaged photos with impressive speed and accuracy. But just like all AI-based tools, this one has downsides as well. So, let’s check out the good and the bads and see what it has to offer.
Louis Bouchard drew our attention to the tool in a blog post he wrote about it. Of course, there’s also a research paper where you can read about this model in more detail. Put briefly, when you add your image to GFP-GAN, it creates merely a guess of the face of a person in the photo. However, in the majority of cases, they seem pretty darn close to the original image.
The AI tries to understand what’s in the photo and then add pixels or fill in the gaps. Unlike other similar models, GFP-GAN focuses on important facial features like eyes and mouth. Finally, there is a comparison of the resulting image with the original to see if there’s still the same person in both – which leads us to one of the biggest weaknesses of the technology.
Conventional image restoration methods use different technology to recreate damaged or blurry images and create new ones. However, this often results in low-quality images. GFP-GAN uses a pre-trained version of an existing model (NVIDIA’s StyleGAN-2) to inform the team’s own model at multiple stages during the image generation process. This is why the identity of people in photos remains preserved.
Still, the points out some weaknesses of the approach. The resulting images may not be very sharp sometimes, some outputs are unnatural, and there’s still a slight change of identity possible. We simply can’t be sure that the reconstructed image will be the same as the original, it is impossible. “The image will look just like our grandfather if we’re lucky enough,” Louis says, “but it may as well look like a complete stranger.” While the results are remarkable nevertheless, this is something to keep in mind when using AI tools like this. If you’d like to test out GFP-GAN, you’ll find it on GitHub.
[via Engadget]
FIND THIS INTERESTING? SHARE IT WITH YOUR FRIENDS!