RETHINKING STYLE AND CONTENT DISENTANGLEMENT IN VARIATIONAL AUTOENCODERS

disentanglement creteria/definition

• label & style preserving, i.e. for fixed style & label, model outputs images of same style and label

label preservation

• rewrite objective -> posterior regularization

• p(y

  x) -> p_D(y

  x)

  • hence the style can only be encoded in z
  • but style preservation may not hold

style preservation

• as domain adaptation: unsupervised
• VAE does not enforce style preservation

…so why VAE consistently learn disentangled representations?

• a shared encoder or decoder is sufficient for disentanglement