NeurIPS 2020

DeepI2I: Enabling Deep Hierarchical Image-to-Image Translation by Transferring from GANs

Meta Review

Reviewers were split on this paper, with three recommended accept and one recommending rejection. The reviewers appreciated the success at getting pretraining to work for GANs, which has been an open problem. However, reviewers also had concerns about the thoroughness of the experiments, especially that structure preservation was not evaluated in the submission (R2, R3). The rebuttal did a good job addressing this concern, with a user study, although in discussion the reviewers requested a study with more participants for the final version of the paper. This and other results provided in the rebuttal convinced most reviewers and the AC of the merits of this paper. These additional results should definitely be added to the final version of the paper. Reviewers also pointed out several missing references that should at least be discussed, and ideally compared against (in particular, see R2's comments about TransGaGa). With these changes the submission could become a substantially stronger camera ready.