Drag Your GAN

In the realm of synthesizing visual content to meet users' needs, achieving precise control over pose, shape, expression, and layout of generated objects is essential. Traditional approaches to controlling generative adversarial networks (GANs) have relied on manual annotations during training or prior 3D models, often lacking the flexibility, precision, and versatility required for diverse applications. In our research, we explore an innovative and relatively uncharted method for GAN control – the ability to "drag" specific image points to precisely reach user-defined target points in an interactive manner (as illustrated in Fig.1). This approach has led to the development of DragGAN, a novel framework comprising two core components: Feature-Based Motion Supervision: This component guides handle points within the image toward their intended target positions through feature-based motion supervision. Point Tracking: Leveraging discriminative GAN features, our new point tracking technique continuously localizes the position of handle points. DragGAN empowers users to deform images with remarkable precision, enabling manipulation of the pose, shape, expression, and layout across diverse categories such as animals, cars, humans, landscapes, and more. These manipulations take place within the learned generative image manifold of a GAN, resulting in realistic outputs, even in complex scenarios like generating occluded content and deforming shapes while adhering to the object's rigidity. Our comprehensive evaluations, encompassing both qualitative and quantitative comparisons, highlight DragGAN's superiority over existing methods in tasks related to image manipulation and point tracking. Additionally, we demonstrate its capabilities in manipulating real-world images through GAN inversion, showcasing its potential for various practical applications in the realm of visual content synthesis and control.