Rabitə və İnformasiya Texnologiyaları Nazirliyinin elektron xəbər xidməti

This website uses AI to generate portraits of people who don’t actually exist


A new website called This Person Does Not Exist went viral this week, and it has one simple function: displaying a portrait of a random person each time the page is refreshed. The website is pointless at first glance, but there's a secret behind its seemingly endless stream of images. According to a Facebook post detailing the website, the images are generated using a generative adversarial networks (GANs) algorithm.
 
In December, NVIDIA published research detailing the use of style-based GANs (StyleGAN) to generate very realistic portraits of people who don't exist. The same technology is powering This Person Does Not Exist, which was created by Uber software engineer Phillip Wang to 'raise some public awareness for this technology.'
 
 
In his Facebook post, Wang said: “Faces are most salient to our cognition, so I've decided to put that specific pretrained model up. Their research group have also included pretrained models for cats, cars, and bedrooms in their repository that you can immediately use.”
 
Each time you refresh the site, the network will generate a new facial image from scratch from a 512 dimensional vector.
 
Generative adversarial networks were first introduced in 2014 as a way to generate images from datasets, but the resulting content was less than realistic. The technology has improved drastically in only a few years, with major breakthroughs in 2017 and again last year with NVIDIA's introduction of StyleGAN.
 
This Person Does Not Exist underscores the technology's growing ability to produce life-like images that, in many cases, are indistinguishable from portraits of real people.
 
As described by NVIDIA last year, StyleGAN can be used to generate more than just portraits. In the video above, the researchers demonstrate the technology being used to generate images of rooms and vehicles, and to modify 'fine styles' in images, such as the color of objects. Results were, in most cases, indistinguishable from images of real settings.






18/02/19    Çap et