New breakthrough in AI can produce very realistic but fake images


Photo by Inked Pixels via Shutterstock


Photo editing in the last decade evolved to a new level, thanks to numerous software, such as Photoshop, which paved the way to image doctoring. While experts can doctor images to near perfection, doing such a task can be very challenging, especially in producing thousands of images. But with artificial intelligence, equipped by generative adversarial networks, the task becomes significantly easier and enables the production of lifelike images.

New Breakthrough in AI Technology

Generative adversarial networks or GANs are comprised of a generator network capable of producing images and a discriminator network that assesses image authenticity. The networks can allow AI systems to create very realistic but fake images by studying an enormous number of images. The generator network produces images based on the realism rating of the discriminator network.

“Neural networks are hungry for millions of example images to learn from. GANs are a (relatively) new way to automatically generate such examples,” said Oren Etzioni, the Chief Executive Office of Allen Institute for Artificial Intelligence.

The approach of GANs has been introduced by Ian Goodfellow, a research scientist now at Google Brain, in a 2014 study. Later on, several researchers have conducted experiments and tests to determine the possible uses for GANs.

GANs can also learn without human supervision and continue to upgrade its database to produce better results. However, that continuous learning can cause arrested development. When GANs fail to improve, the generator network persists in producing the same image quality.

To remedy that, NVIDIA developed a training method to allow continuous learning of GANs. The researchers at NVIDIA trained both the generator and the discriminator networks evenly, with low-resolution and progressive addition of image layers to slowly feed high-resolution details.

“We chose faces as our prime example because it is very easy for us humans to judge the success of the generative AI model -- we all have built-in neural machinery, additionally trained throughout our lives, for recognizing and interpreting faces,” said Jaakko Lehtinen, a researcher at NVIDIA involved with the project.

With GANs, filmmakers and video game developers can obtain inexpensive high-quality and realistic materials for movies and video games.