Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold

Xingang Pan · Ayush Tewari · Thomas Leimkühler · Lingjie Liu · Abhimitra Meka · Christian Theobalt

SIGGRAPH 2023 Conference Proceedings

Web Demos

Requirements

If you have CUDA graphic card, please follow the requirements of NVlabs/stylegan3.

The usual installation steps involve the following commands, they should set up the correct CUDA version and all the python packages

conda env create -f environment.yml
conda activate stylegan3

Then install the additional requirements

pip install -r requirements

Otherwise (for GPU acceleration on MacOS with Silicon Mac M1/M2, or just CPU) try the following:

cat environment.yml | \
  grep -v -E 'nvidia|cuda' > environment-no-nvidia.yml && \
    conda env create -f environment-no-nvidia.yml
conda activate stylegan3

# On MacOS
export PYTORCH_ENABLE_MPS_FALLBACK=1

Run Gradio visualizer in Docker

Provided docker image is based on NGC PyTorch repository. To quickly try out visualizer in Docker, run the following:

docker build . -t draggan:latest  
docker run -p 7860: 7860 -v "$PWD":/workspace/src -it draggan:latest bash  
cd src && python visualizer_drag_gradio.py --listen

Now you can open a shared link from Gradio (printed in the terminal console).
Beware the Docker image takes about 25GB of disk space!

Download pre-trained StyleGAN2 weights

To download pre-trained weights, simply run:

python scripts/download_model.py

If you want to try StyleGAN-Human and the Landscapes HQ (LHQ) dataset, please download weights from these links: StyleGAN-Human, LHQ, and put them under ./checkpoints.

Feel free to try other pretrained StyleGAN.

Run DragGAN GUI

To start the DragGAN GUI, simply run:

sh scripts/gui.sh

If you are using windows, you can run:

.\scripts\gui.bat

This GUI supports editing GAN-generated images. To edit a real image, you need to first perform GAN inversion using tools like PTI. Then load the new latent code and model weights to the GUI.

You can run DragGAN Gradio demo as well, this is universal for both windows and linux:

python visualizer_drag_gradio.py

Acknowledgement

This code is developed based on StyleGAN3. Part of the code is borrowed from StyleGAN-Human.

(cheers to the community as well)