SADIM

An unofficial implementation and experimentation on the model proposed in "Swapping Autoencoder for Deep Image Manipulation" by Park et al. (2020)

A considerable portion of this project is performed thanks to rosinality. The official implementation is also available. Details were provided by Park et al. in the official GitHub repository.

For further details of the architecture and training, please refer to the paper.

@article{DBLP:journals/corr/abs-2007-00653,
  author    = {Taesung Park and
               Jun{-}Yan Zhu and
               Oliver Wang and
               Jingwan Lu and
               Eli Shechtman and
               Alexei A. Efros and
               Richard Zhang},
  title     = {Swapping Autoencoder for Deep Image Manipulation},
  journal   = {CoRR},
  volume    = {abs/2007.00653},
  year      = {2020},
  url       = {https://arxiv.org/abs/2007.00653},
  eprinttype = {arXiv},
  eprint    = {2007.00653},
  timestamp = {Mon, 06 Jul 2020 15:26:01 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2007-00653.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Qualitative Results with Changing Patch Sizes for the Cooccurrence Discriminaator:

For further ablation studies, I tested the qualitative effects of patch size that is utilized by the Patch Cooccurrence Discriminator. The qualitative results can be found in the results:

Exact reconstruction with patch size 1/4:

Exact reconstruction with patch size 3/4:

Latent traverse with patch size 1/4:

Latent traverse with patch size 3/4: