XingangPan

DragGAN: Edit GAN Images by Dragging Points

DragGAN is a research tool for interactive point-based editing of images generated by StyleGAN models. It lets users move image features by dragging points while the model updates the image, making controlled edits easier than conventional prompt-based workflows.

View on GitHub Compare
Stars
35.7k
Forks
3.4k
Stars, last 30 days
-37

Our take

Health score

28/100

How it is scored
Activity
19
Community
52
Issues
21
Pull requests
20

-37 stars in the last 30 days

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

35,691

Forks

3,381

Change since Nov 16, 2026: -2 stars·0 forks

Overview

DragGAN is an interactive image-editing system built around StyleGAN generative models. Users select points on a generated image and drag them toward target positions; the system adjusts the model's latent representation to move the corresponding visual features while preserving the image's overall appearance.

It is aimed at researchers and practitioners exploring controllable image generation, as well as artists and developers who want to experiment with direct geometric edits to GAN-generated images. The repository provides the research implementation and visualizer, rather than a general-purpose editor for arbitrary photos. Editing real images requires first obtaining a suitable GAN inversion and latent code.

Key Features

  • Point-based interaction for moving features within generated images.
  • Iterative image updates as users drag selected points toward targets.
  • GUI and Gradio visualizer entry points for trying the editing workflow.
  • Support for pretrained StyleGAN models and additional weights such as StyleGAN-Human and LHQ.
  • Docker-based option for running the Gradio visualizer.
  • CPU and Apple Silicon setup guidance, in addition to the CUDA-oriented setup.
  • Watermarking functionality intended to identify results as AI generated.

Use Cases

  • GAN researchers can explore how latent representations support localized, spatially controlled edits.
  • Artists and designers can adjust positions or shapes of features in generated images without relying on text prompts.
  • Computer vision developers can prototype interactive image-manipulation workflows using StyleGAN models.
  • Users editing real photographs can use the tool after applying a separate GAN inversion method, when the image is represented by a compatible latent code.

Getting Started

conda env create -f environment.yml
conda activate stylegan3
pip install -r requirements.txt
python scripts/download_model.py
sh scripts/gui.sh

See the README for platform-specific setup, model options, and Gradio or Docker instructions. The project page and paper provide further context.

Considerations

  • The standard setup follows StyleGAN3 requirements and is oriented toward CUDA-capable GPUs. The README also describes CPU and Apple Silicon alternatives, which may have different performance.
  • The tool edits images generated by compatible GAN models. Real-image editing is not a direct workflow and requires GAN inversion tools such as PTI.
  • The repository's license field is reported as NOASSERTION, while the README describes different terms for DragGAN algorithm code and code derived from StyleGAN3. Review those terms before reuse or distribution.
  • The README says use and derivatives must preserve watermarking that shows “AI Generated.”
  • Docker is available for the visualizer, but the README warns that its image uses about 25GB of disk space.

Found this useful?

Share it with someone who would like DragGAN.