Artificial intelligence is developing at an unprecedented rate. Large technology companies are increasingly using it to offer their customers solutions in the areas of security, programming, mobility, etc. GANs are a class of algorithms introduced in 2014 by Goodfellow et al. to create images with a high degree of realism with artificial intelligence.
A GAN is a generative model in which two networks compete in a game theory scenario (a mathematical field that focuses on the interactions of individual decisions, called players who are aware of the existence of these interactions). The first network is the generator, it generates a sample (e.g. an image), while its opponent, the discriminator, tries to recognize whether a sample is real or whether it is the result of the generator.
In a report released on December 12, 2008, NVIDIA describes in a supporting video how NVIDIA generates images using a style-based generator architecture for conflicting generative networks. NVIDIA’s GANs are based on a concept called style transfer. Earlier this month, Nvidia released the results of a study that showed how to combine AI-generated visuals with a traditional video game engine. The result is a hybrid graphics system that could one day be used in video games, virtual reality, and even movies.
In this report, NVIDIA researchers point out that Style Transfer creates images that can copy a painter’s brushstrokes to a photograph of a cityscape to create a new image with the same style as the artist. NVIDIA does not use the method of copying and pasting elements from different surfaces into a real character. Instead, it is based on three basic styles: The course, medium, and fine styles merge them into a new character.
In the representation of a person, high-level (coarse) styles include pose, facial shape, or hairstyle.
Medium styles include facial features such as the shape of the nose, cheeks or mouth. Finally, fine styles influence the color of facial features such as skin and hair. “We propose an alternative generator architecture for generative accumulator networks borrowed from the style of literature transfer. The new architecture leads to an unattended and automatically learned separation of high-level data (e.g. the pose and identity of a face) and the stochastic variation of the generated images (e.g. freckles, hair) allows an intuitive and controlled evaluation of the synthesis,” the report says. NVIDIA’s KI has created very impressive human facial images, animal (chat) images and room images, all of which are easily customizable.
Images with human faces and rooms are mostly successful, but images of cats have some very remarkable flaws. According to the report, NVIDIA’s modifications to its GANs have helped improve image quality, and the generator is able to distinguish between inconsistent variations and high-level variations to create new and more remarkable synthetic areas. For some Internet users, this is an incredible step forward. They hope that the tool will be improved for the general public.
Others, however, see it as an unfortunate coincidence and believe that it will only increase the number of fake characters in the world. They go on to say that if this type of technology were used, we would know more about who is real and who is not. Another also sees a serious problem in this technology and assumes that the very existence of social networks no longer makes it possible to know whether the information about a person and their images is true or not. “A niche consequence is that anyone can simulate certain profiles (Facebook, Tinder, etc.) with these random and realistic faces and searching for inverted images won’t bring you anything that makes you think the profile is real,” he said.