According to Louis Bouchard and PetaPixel, there is a free tool that can repair time-tainted, low-resolution portraits. This tool was developed by researchers at Tencent and is called the Generative Facial Prior-Generative Adversarial Networt, or GFP-GAN for short.
Indeed, it is very annoying not to be able to restore, for example, an old family photo, but with GFP-GAN it becomes possible. To do this, the technology used is to combine information from two AI models to fill in the missing details. The most interesting thing is that everything is done in just a few seconds.
Screenshot: What is AI/YouTube
How does GFP-GAN work?
Enhanced version 1.3 of the GFP-GAN model tries to analyze the exact content of a photo in order to understand it. It will then fill in what is missing and add pixels to the missing sections as needed. PetaPixel says that “the pre-trained StyleGAN-2 model is then used to drive its own multi-scale generative model during encoding from image to latent code and thus reconstruction.”
Facial detail is greatly improved by the AI through the use of extra dimensions. AI focuses on important local features, including a person’s eyes, mouth, and nose. When it’s done, the system will start comparing the real image with the image that was just restored to make sure it’s still the same person in the generated photo. Thus, the last step is aimed at preserving the “personality” of all those who are in the photograph.
Limits of GFP-GAN
For information, anyone can try this exciting recovery technology, as the GFP-GAN demo is completely free. Moreover, it can be integrated into various human projects because the GFP-GAN code was made public by its creators.
However, you should be aware that the current AI has its limits, which means that this technology also has some limits. There may be a “little personality change” as well as a lower resolution than what you had in mind, the researchers said. Therefore, it is not recommended to rely on such a tool to print a wide format version of an old photograph of great grandparents. But it does not remain a great technological advance.