Drag Your GAN

Introducing DragGAN: Enhancing GAN Control Through Image Point ManipulationIn the world of visual content synthesis, having precise control over the shape, pose, expression, and layout of generated objects is crucial. However, traditional approaches to controlling generative adversarial networks (GANs) often lack flexibility, precision, and versatility. In our research, we introduce a novel method for GAN control called DragGAN, which allows users to “drag” specific image points to reach user-defined target points interactively. DragGAN is composed of two core components: Feature-Based Motion Supervision and Point Tracking. The former guides handle points toward their intended target positions through feature-based motion supervision, while the latter continuously localizes the position of handle points using discriminative GAN features. This framework enables users to manipulate images with remarkable precision, allowing for the deformation of pose, shape, expression, and layout across diverse categories such as animals, cars, humans, landscapes, and more. By operating within the learned generative image manifold of a GAN, DragGAN produces realistic outputs, even in complex scenarios such as generating occluded content and deforming shapes while adhering to the object’s rigidity. Our comprehensive evaluations, including both qualitative and quantitative comparisons, demonstrate DragGAN’s superiority over existing methods in tasks related to image manipulation and point tracking. Furthermore, we showcase DragGAN’s potential for practical applications in the realm of visual content synthesis and control by demonstrating its capabilities in manipulating real-world images through GAN inversion. With DragGAN, users can achieve greater control and flexibility in synthesizing visual content, making it a valuable tool for a wide range of applications in various fields.