Mpg

The study presents DragGAN, a novel method for interactive point-based manipulation of generative adversarial networks (GANs) to provide flexible and precise control over the pose, shape, expression, and layout of generated objects. Unlike existing approaches reliant on annotated data or prior 3D models, DragGAN enables users to "drag" points on an image to specific target positions interactively. This control is facilitated by a feature-based motion supervision guiding the handle point movement towards the target and a point tracking system based on discriminative GAN features. The method allows for manipulation of various categories like animals, cars, humans, and landscapes on the learned generative image manifold. Results indicate enhanced performance in image manipulation and point tracking compared to previous methods. Real image manipulation through GAN inversion is also demonstrated. The project is supported by various institutions and made available for non-commercial use under Creative Commons licensing. Further details can be accessed via the provided contact information.