Researchers from several institutions around the world, including the Max Planck Institute for Informatics, MIT, University of Pennsylvania, Google AR/VR, and Saarbrücken Research Center for Visual Computing, Interaction and AI, have developed a new method for manipulating digital images in a highly flexible and precise manner. The researchers’ system, known as “DragGAN”, is now available on GitHub.
Unlike previous methods for controlling the output of generative adversarial networks (GANs), which are a type of artificial intelligence used to create images, DragGAN offers a higher level of precision, flexibility, and broad application. Traditional methods often rely on manually annotated training data or pre-existing 3D models, which can limit the system’s versatility.
What sets DragGAN apart is its ability to let users “drag” any point in an image to a specific target point interactively. This method allows anyone to alter the pose, shape, expression, or layout of various image categories such as animals, cars, humans, landscapes, etc., with exceptional precision.
DragGAN comprises two main components. The first, a feature-based motion supervision, guides the selected point to its target position. The second is a novel point tracking approach, which keeps track of the position of the selected point using GAN features.
One of the benefits of DragGAN is that it generates highly realistic outputs even in complex situations. For instance, it can create plausible representations of obscured content, or alter shapes in a way that adheres to the object’s natural rigidity.
The system has shown significant advantages over previous methods in both image manipulation and point tracking. To see example and learn more, visit the project’s website and download the sourcecode from GitHub.
