Figure 1

Proposed AC-GAN architecture. The noise z and the class one-hot encoding c are concatenated and sent to the generator (G). The generator creates class-conditioned fake images. The classes c and the source information s are jointly predicted by the discriminator (D) given a fake (left) or a real (right) input. SNConv implies convolutional layers using spectral normalization51, which introduces a minor modification to the original AC-GAN architecture41.