A pytorch implementation of Paper "IntroVAE: Introspective Variational Autoencoders for Photographic Image Synthesis"

• Python 2.7 or Python 3.6
• PyTorch

Use the default parameters except changing the hyper-parameter, i.e., m_plus, weight_rec, weight_kl, and weight_neg, for different image resolution settings. Noted that setting num_vae nonzero means pretraining the model in the standard VAE manner, which may helps improve the training stablitity and convergency.

The default parameters for CelebA-HQ faces at 256x256 and 1024x1024 resolutions are provided in the file 'run_256.sh' and 'run_1024.sh', respectively. Other settings are allowed as discussed in the appendix of the published paper.

If you use our codes, please cite the following paper:

@inproceedings{huang2018introvae,
  title={IntroVAE: Introspective Variational Autoencoders for Photographic Image Synthesis},
  author={Huang, Huaibo and Li, Zhihang and He, Ran and Sun, Zhenan and Tan, Tieniu},
  booktitle={Advances in Neural Information Processing Systems},
  pages={10236--10245},    
  year={2018}
}

The released codes are only allowed for non-commercial use.