Image-to-image translation for cross-domain disentanglement