Aim: Produce images of Pegasi from the CIFAR10 dataset.
Chosen Network: VAE-GAN [https://arxiv.org/pdf/1512.09300.pdf]
Conclusion: Due to the large intra-class variability, images were blurrier than expected. Having to interpolate between encodings/latent spaces meant the target feature (wings) were not as common. Nevertheless, the quality of results are better than the standard VAE (and was easier to train than a GAN).
Encoder Network
Decoder Network
Discriminator Network
Complete Network
Results
