This is a Pytorch implementation of a Face Swap Autoencoder, roughly based on Shaonlu's tensorflow implementation..
- Both the autoencoder and the discriminator are using
spectral normalization
- Discriminator is being used only as a
learned preceptual loss
, not a direct adversarial loss - Conv2d has been customized to properly use spectral normalization before a pixel-shuffle
- Downsampling operations have been remove from VGG-Face to provide more detail in perceptual loss
Usage instructions found here: user manual page.