The paper proposes a dual variational autoencoder, used to generate new synthetic training data for heterogeneous face recognition, by preserving generation both in the image and embedding spaces and providing variation for the training data of the downstream recognition task. The authors claim a big improvement in performance. Reviewers initially were convinced on the goodness of the paper and then after the rebuttal one of the reviewer increased its rate. Thus the consensuswas reached and also the area chair agreed with the acceptance rate.