DragGAN

The tool introduces a new paradigm for image editing where users can drag any point on a generated image to a target position, causing the image to deform naturally and realistically along the generative model's learned manifold.

Unlike pixel-level editing tools, DragGAN manipulates images at the semantic level moving a person's leg causes the entire pose to adjust believably, rather than producing distorted pixels.

The technical approach combines a motion supervision loss that guides feature points toward their targets with a point tracking mechanism that maintains correspondence during the deformation process.

This operates entirely within the latent space of a pre-trained GAN (specifically StyleGAN2), meaning the edits always remain photorealistic and consistent with the training distribution.

The result is fluid, interactive manipulation of faces, animals, cars, and landscapes that preserves structural coherence the editing tools cannot achieve.

DragGAN requires a CUDA-capable GPU for real-time performance and runs as a Gradio web interface that launches locally. The project is intended for researchers and digital artists exploring fine-grained generative image control, and it has influenced subsequent work on diffusion model editing methods.

As a SIGGRAPH publication, the codebase is a reference implementation prioritizing research correctness; production use cases may require additional optimization for deployment at scale.