Developments in artificial intelligence they move at an impressive rate, so much so that it is often difficult to watch. Cameras have now been able to produce incredibly realistic photographs of humans, but they have never been on Earth.
These realistic but non-existent faces in the photo above are the work of researchers from Nvidia. In their work on artificial intelligence. which they presented publicly last week, describe her amendment of the GAN architecture (Generative Adversarial Network) to create these fictitious images. Take a better look. If you didn't know they were fake, could you tell?
In the image above you can see what progress is within four years of entering the artificial intelligence into the image. The rough black and white faces on the left are from 2014 and were published in one paper, which then launched the AI tool called Generic Adversarial Network (GAN). On the right you see a similar job that was published this month, which used the same method (GAN), but the image quality difference is clearly huge.
What is particularly interesting is that these fake faces can also be easily adjusted. Nvidia engineers have incorporated in their work a method known as style transfer, in which the features of one image are blended with another. You may recognize the term from various image filters that have been popular in recent years in applications such as Prisma and Facebook, which can make yourself look like an Impressionist painting or a cubist work of art.
Applying style transfer to people, allowed Nvidia researchers to adapt them remarkably. In the following grid, you can see the evolution. By taking a picture as a source of a real person (the top row) and mixing them with others of a different real person (right column) see the result produced. Features such as the color of the skin and the hair etc are mixed together, creating a whole new face.
Of course, the ability to create realistic faces through artificial intelligence raises troubling questions. These tools could be used for misinformation and propaganda and could undermine public confidence in visual elements. This could damage the judiciary as well as politics. Especially in today's era that fake news they have managed to fool even news agencies.
As for construction difficulty and fidelity: Nvidia researchers took a week to train their model on a machine with eight Tesla GPUs to create these faces. There are also indications which we can search to spot the fake photos. Hair, for example, is very difficult to fake. They often look like they are painted with a brush or very blurry. But Nvidia's work shows how fast artificial intelligence is evolving in this area, and at this rate it won't sloware to create algorithms that will be able to prevent such mistakes.
It is obvious that there will be a battle in the next few decades about the identity of an image, between fake creatures of artificial intelligence and security researchers. And at this time, artificial intelligence has the lead.
