DragGAN
by Max Planck Institute
Interactive point-based GAN image editing for intuitive drag-to-modify control
Visit Product
16 upvotes
1,869 views
About
DragGAN is a groundbreaking interactive image editing method that allows users to manipulate images by dragging points to desired positions — intuitively controlling the pose, shape, expression, and layout of objects in GAN-generated images. Developed by researchers at the Max Planck Institute for Informatics, it represents a major advance in how humans can interact with and edit AI-generated images.
The method works within the latent space of a pretrained GAN: users place 'handle points' on parts of an image they want to move and 'target points' where they want them to go. The system then optimizes the GAN's latent code to move the selected image content to the target position while maintaining realistic image quality and coherence — essentially sculpting the image through intuitive direct manipulation.
DragGAN was released as a research demonstration in 2023 and quickly captured global attention for making complex image manipulation as intuitive as touching a screen and dragging. It demonstrated the potential for fundamentally new interfaces for AI image editing that go beyond text prompts to direct physical manipulation of image content.
The method works within the latent space of a pretrained GAN: users place 'handle points' on parts of an image they want to move and 'target points' where they want them to go. The system then optimizes the GAN's latent code to move the selected image content to the target position while maintaining realistic image quality and coherence — essentially sculpting the image through intuitive direct manipulation.
DragGAN was released as a research demonstration in 2023 and quickly captured global attention for making complex image manipulation as intuitive as touching a screen and dragging. It demonstrated the potential for fundamentally new interfaces for AI image editing that go beyond text prompts to direct physical manipulation of image content.
Product Features
- Drag-to-edit interaction for intuitive image manipulation
- Works within GAN latent space for realistic results
- Point-based handle and target position control
- Real-time preview of manipulation results
- Preserves image realism during editing
- Suitable for face, body, landscape, and object manipulation
- Open-source code available for research
- Compatible with StyleGAN and other GAN architectures
- Interactive demonstration available online
- Research paper with full methodology
- Works within GAN latent space for realistic results
- Point-based handle and target position control
- Real-time preview of manipulation results
- Preserves image realism during editing
- Suitable for face, body, landscape, and object manipulation
- Open-source code available for research
- Compatible with StyleGAN and other GAN architectures
- Interactive demonstration available online
- Research paper with full methodology
About the Publisher
DragGAN was developed by Xingang Pan, Ayush Tewari, Thomas Leimkühler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt at the Max Planck Institute for Informatics in Germany. The research was presented at SIGGRAPH 2023 and attracted enormous public attention for its intuitive image manipulation interface. The Max Planck Society is Germany's premier fundamental research organization, operating over 80 research institutes across sciences.