DraGAN

DragGAN is an innovative research project that introduces a novel interactive point-based manipulation technique for generative adversarial networks (GANs). Its core value proposition is enabling users to intuitively and precisely control the pose, shape, expression, and layout of synthesized visual content by simply dragging points on an image to target locations, with the underlying GAN model realistically deforming the object to follow these user-defined constraints. This approach provides an unprecedented level of flexible and generic controllability for image generation and editing.

Key features: The tool allows users to manipulate an image by selecting handle points and dragging them to desired positions, with the system automatically generating the intermediate content in a photorealistic manner. For example, you can change the pose of an animal by dragging its legs, adjust the smile on a generated face by moving corner points of the mouth, or modify the shape and perspective of objects like cars. It operates on the latent space of a pre-trained GAN, ensuring high-quality outputs, and includes a feature-based motion supervision to guide the points and a point tracking approach using discriminative GAN features to precisely locate the handle points throughout the manipulation.

What sets DragGAN apart from competitors is its unique methodology that does not rely on manually annotated training data or a pre-existing 3D model for control, which are common limitations in other GAN-based editing tools. This grants it superior flexibility, precision, and generality across various object categories. Technically, it is a research framework that demonstrates a powerful, less-explored direction for controlling GANs through direct user interaction in the image space, offering a more intuitive alternative to traditional slider-based or text-based editing interfaces.

Ideal for AI researchers, computer vision scientists, digital artists, and content creators who require fine-grained control over generated imagery. Specific use cases include prototyping character designs, editing product visuals, creating tailored stock photography, and conducting academic research on controllable generative models. Industries such as entertainment, advertising, and e-commerce can leverage it for rapid visual concept iteration.

As a research project from MPI-Inf, the official implementation is provided free of charge for non-commercial, academic exploration of its capabilities.