Bluffing! Here is DragGAN, an AI capable of inventing what does not exist in a photo with a click and a movement of the mouse.
In the world of artificial intelligence specialized in image generation, here is a newcomer. It’s called DragGAN and its objective is not to create an image from words, but rather to reorient a photo and modify certain elements of it by artificially regenerating them with a few mouse movements. With it, it is possible to make you smile again in a photo, for example, or to make a cat wink and a lion turn its head. In principle, the tool is reminiscent of Photoshop Warp from Adobe, but with a big plus. Instead of stretching pixels with the mouse, the onboard AI will totally generate non-existent areas of the image. To design this impressive tool, it took researchers from major institutions and companies, such as Google, MIT and the University of Pennsylvania working together. They have published a page where you can see how to manipulate the image and let the AI do the editing. DragGAN’s GAN is an acronym for “Generative Adversarial Networks”. With them, a first network generates an image and its adversary hunts down the details that are wrong.
To fool the discriminator, the generative network must ensure that the result is high quality and ultra-realistic. In the case of this research work with DragGAN, researchers have come up with a new way to manipulate these GANs. The user will click on the part to be modified, to create an anchor, then perform a movement which will allow this anchor to be moved. The GANs will then observe these movements. But that’s not all. The discriminating network will follow and carry out the critical analysis of the points of this anchor during their displacement. With this process, it is possible to distort an image with precise and realistic control of the new location of the pixels. A dog photographed from the front can thus be repositioned three-quarters in a photorealistic way. Still with this principle, the researchers explain that AI is capable of inventing what does not exist in the image. This would be the case of the teeth of a lion whose mouth is opened, for example.