https://www.thispersondoesnotexist.com/
It's a computer program created by geeks at Nvidia that creates hyper-realistic photographic quality faces of nonexistent people using a technology called 'Generative Adversarial Networks'.
Click refresh/reload, and you will get another nonexistent person.
https://interestingengineering.com/this-...the-future
Technical details here:
https://arxiv.org/pdf/1812.04948.pdf
Apparently it has two AI components: a generator and a discriminator. The generator generates images (at first crude) using an initial set of preprogrammed principles and the discriminator decides if the image is real or fake, based on comparing the generator's images with real life images according to its own innate set of principles. But it doesn't stop there, since the generator and the discriminator are neural network AIs that can learn, gradually getting better and better at their tasks as they compete with each other. In effect, each network is training the other network in kind of a circular loop. The goal is to generate better and better images that become so good that the discriminator eventually can't tell the difference between a real image and a fake one.
Apparently they've succeeded very spectacularly. At least human discriminators can't tell.
Considering that humans are primed by evolution to perceive faces, making us very good discriminators of faces (remember the uncanny valley) it's going to be very hard to trick a human with a fake image, making it a good test of the technology.
The technology doesn't just work with faces, it works with any sort of image. The scientific paper includes sets of computer generated fake images of cars, cats and room decors.
They are talking about developing the Generative Adversarial Network process to produce exceedingly lifelike virtual environments for various applications like video games.
It's a computer program created by geeks at Nvidia that creates hyper-realistic photographic quality faces of nonexistent people using a technology called 'Generative Adversarial Networks'.
Click refresh/reload, and you will get another nonexistent person.
https://interestingengineering.com/this-...the-future
Technical details here:
https://arxiv.org/pdf/1812.04948.pdf
Apparently it has two AI components: a generator and a discriminator. The generator generates images (at first crude) using an initial set of preprogrammed principles and the discriminator decides if the image is real or fake, based on comparing the generator's images with real life images according to its own innate set of principles. But it doesn't stop there, since the generator and the discriminator are neural network AIs that can learn, gradually getting better and better at their tasks as they compete with each other. In effect, each network is training the other network in kind of a circular loop. The goal is to generate better and better images that become so good that the discriminator eventually can't tell the difference between a real image and a fake one.
Apparently they've succeeded very spectacularly. At least human discriminators can't tell.
Considering that humans are primed by evolution to perceive faces, making us very good discriminators of faces (remember the uncanny valley) it's going to be very hard to trick a human with a fake image, making it a good test of the technology.
The technology doesn't just work with faces, it works with any sort of image. The scientific paper includes sets of computer generated fake images of cars, cats and room decors.
They are talking about developing the Generative Adversarial Network process to produce exceedingly lifelike virtual environments for various applications like video games.