DragGAN

DragGAN is the official code release for Drag Your GAN: Interactive Point based Manipulation on the Generative Image Manifold, a research paper by Xingang Pan and colleagues published in the SIGGRAPH 2023 Conference Proceedings. It lets a user edit an image by placing handle points and target points, then dragging: the model moves the pose, shape, expression or layout of the subject toward the targets while keeping the result photorealistic, because every edit stays on the manifold learned by a generative adversarial network. The repository suits researchers, graphics engineers and technically confident hobbyists who want to experiment with point based image editing. It ships a desktop GUI and a Gradio visualizer, a Dockerfile for running the Gradio demo in a container, and a script that downloads pretrained StyleGAN2 weights, with optional StyleGAN Human and Landscapes HQ checkpoints. It runs best on an NVIDIA CUDA graphics card, and the README documents a fallback for Apple Silicon GPU acceleration or plain CPU. The GUI edits images the GAN itself generated; editing a real photo first requires GAN inversion with a separate tool such as PTI. There is no price. The code is free to download from GitHub. The DragGAN algorithm code is licensed under the Creative Commons Attribution NonCommercial license, and the code derived from StyleGAN3 falls under the Nvidia Source Code License, so commercial use is not permitted under the published terms. One condition is worth knowing before building on it: the license requires any use or derivative of the code to preserve the watermarking feature that labels output as AI Generated. The Docker image also needs roughly 25GB of disk space.