DragGAN

DragGAN is the official code release for "Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold," presented at SIGGRAPH 2023. It is built for researchers and developers who want to edit GAN-generated images by dragging points rather than writing text prompts. The tool is based on pretrained StyleGAN2 and StyleGAN3 models and includes a GUI for interactive editing.

A user downloads pre-trained StyleGAN2 weights with a script, then runs the DragGAN GUI to edit a GAN-generated image by dragging points, or uses the Gradio demo for the same functionality on Windows or Linux. To edit a real photograph instead of a generated one, a user must first perform GAN inversion with a tool such as PTI and load the resulting latent code and weights into the GUI. A Docker image is provided, built on the NGC PyTorch repository.

DragGAN lets a user drag handle points on a GAN-generated image toward target points, producing realistic edits that follow the structure of the underlying image manifold. It supports StyleGAN-Human and the Landscapes HQ dataset in addition to standard StyleGAN2 checkpoints, works with CUDA GPUs following NVlabs/stylegan3 requirements, and runs with GPU acceleration on Apple Silicon Macs or on CPU alone. All generated outputs carry a required "AI Generated" watermark.

The DragGAN algorithm's code is licensed under CC-BY-NC for non-commercial use, while most of the rest of the project, adapted from StyleGAN3, is licensed under the Nvidia Source Code License. The Docker image takes about 25GB of disk space, and users on MacOS or CPU-only systems must strip CUDA dependencies from the environment file before installing. The project credits StyleGAN3 and StyleGAN-Human as the codebases it builds on.

Use cases

• Editing a generated face or image by dragging key points
• Manipulating a real photo after GAN inversion