Fig. 3

Overview of the MichiGAN architecture. We first train a model, such as β-TCVAE, to learn a disentangled representation of the real data. We then use the resulting latent codes to train a conditional GAN with projection discriminator, so that the GAN generator becomes like a more accurate decoder. Because the VAE and GAN are trained separately, training is just as stable as training each one individually, but the combined approach inherits the strengths of each individual technique. After training, we can generate high-quality samples from the disentangled representation using the GAN generator