styleganfusion

This is the official repository of "Diffusion Guided Domain Adaptation of Image Generators".

Project Page

checkpoints to start with:

Please download StyleGAN2 or EG3D checkpoints from our shared Google Drive, and put them in the download folder. They are publicly available checkpoints and are the same as the official ones.

The code is ready to run now after downloading checkpoints.

A Few Notes:

• Please use .sh files as entrances. A requirements.txt is provided to install packages.
• Results are saved in the logs folder. We also uploaded a lot of log images here.
• We shared the manually labeled "AFHQ-wild" images by their detailed types(fox/tiger/lion/wolf). See AFHQ-wild_detailed_labels. We used it to calculate FIDs.
• We use diffusers to load StableDiffusion V1.4 checkpoint from Huggingface. A TOKEN file is included to get access to the checkpoints. Please replace it with your own token file, which can be applied here.
• To calculate FID scores, please use metrics/IS_FID_prdc/src/run_FID.sh as entrance. (credit to StudioGAN)

More information: We implement our approach on two versions of StyleGAN2: the StyleGAN-NaDa (this repository) and the StyleGAN-Rosinality. To fully reproduce FIDs for animal experiments, please check our StyleGAN-Rosinality version. We share the code, training logs and images at this google drive.

If you find this work useful, please cite:

@article{song2022diffusion,
      title={Diffusion Guided Domain Adaptation of Image Generators},
      author={Song, Kunpeng and Han, Ligong and Liu, Bingchen and Metaxas, Dimitris and Elgammal, Ahmed},
      journal={arXiv preprint https://arxiv.org/abs/2212.04473},
      year={2022}
}

Cheers!

Acknowledgement: This code is built on Stylegan-NADA and DreamFusion-StableDiffusion