Facebook’s new AI Exemplar GANs (EXGANs) is a pathbreaker; a product of Facebook research, this AI uses exemplary information and replaces the closed eyes with open ones. The research team of Facebook calls the new approach as intelligent “in-painting”, wherein it the program fill the space as per its thinking. The output is rather mind-blowing, and the modified image exemplifies the realistic features. There is no mismatch or the artificial stitching, and the recognition part of the network very well knows how the person looks. It was almost impossible to distinguish the real photos with the AI modified images.
In testing, people are hardly able to distinguish between the real and the faked photos. Unless you are aware of the tampered one, you can hardly make a distinction between the two images. With the advancement of day-to-day technology, many tools are available that make it possible to modify the photos the way we need. But the images with closed eyes are the tricky ones, and so far no program or software has been able to offer a solution. Even though the machines have been able to replace the closed eyes with the open ones, they are far from realism. Machines do not have the capacity to consider consistency nor aware of the colour of the skin around the eye.
So, that is the reason the new novel approach of AI Exemplar GANs (Generative Adversarial Network) by Facebook’s research needs a special mention. Facebook has probably the most extensive collection of images with the eyes closed. So how is GAN able to create realistic images? The answer to this question is quite simple! One part of the GAN recognises the faces, and the other part recreates the image based on the feedback offered by the recognition part. So this results in a stunning recreation of the realistic original images.
However, the new AI still fails to address few challenges. For example, sometimes it failed to recreate the colour correctly, and in other instances, the AI gave some bizarre artefacts when a lock of hair covers the person’s face. But these are fixable problems and cannot overshadow the benefits Exemplar GANs offer.