Rendering objects and editing digital images can sometimes take hours, but a team of researchers at MIT and IBM are looking to change that. They trained the AI to generate photographic images from scratch and intelligently edit individual objects within those images.
First of all, this invention can be useful for artists and designers, but it also provides insight into the algorithm for studying the context of neural networks, and the team of authors hopes to use this tool to identify fake or altered images.
The tool, dubbed GANpaint Studio, is currently available as a free demo. Instead of manually adding the missing tree to the image, you can tell the AI where you think the object should be, and it will add a tree that matches the big picture. You can also wash certain objects, for example, stools from a picture of a kitchen interior.
While the process of developing the tool continues, but scientists hope that one day GANpaint Studio will be able to edit even video clips. If, for example, some important props are missing in a particular scene, editors can add them later using artificial intelligence.
In the process of creating GANpaint Studio, the researchers were surprised to find that the system had learned simple rules regarding the relationship between objects. For example, she somehow knows that the door is not connected to the sky. Because GANpaint Studio uses GAN, a collection of neural networks designed to compete with each other, the AI has to use its internal "common sense" to make decisions. For example, it depends on it whether a cloud appears in the grass. This information can help researchers better understand how neural networks learn context and what we perceive as common sense.
While the artificial intelligence of GANpaint Studio makes it easy to create fake images, it can also help computer scientists and programmers learn to detect them. “You have to study your opponent well before you can defend against him,” said Jun Yang Zhu, one of the co-authors of the GANpaint Studio article. Researchers will present their work at a conference next month. In the meantime, you can test their invention in practice.
Igor Abramov
