NeurIPS 2020

Swapping Autoencoder for Deep Image Manipulation


Meta Review

*The paper uses the broader statement section to report additional results. This is not acceptable and some papers have been desk-rejected for doing the same thing. These additional results have to be moved elsewhere.* The reviewers' opinions on this paper diverge even after considering the rebuttal and discussing. The meta-review is thus unusually detailed. The paper proposes an approach to image editing by disentangling the structure and texture using an autoencoder with the latent space decomposed into two parts - corresponding to texture and structure. When both are taken from the same image, the model is required to reconstruct the image. However, when the texture representation is taken from a different image, the model is required to generate a realistic image with local statistics similar to the texture image (enforced via a patch-based discriminator). The method demonstrates good qualitative and quantitative results on several datasets. Pros: 1) High quality qualitative results (R3, R4). Many of these are presented in the supplementary material. For a paper on image editing this is crucial. 2) Good quantitative results, largely based on user studies (R4) 3) Thorough and convincing experiments (R3, R4) 4) Clear presentation (R3, R4) 5) Fast runtime compared to some of the competing methods Cons: 1) Limited novelty relative to prior methods, e.g. FUNIT (R1, R3) 2) Encoding texture as a fixed-length vector may not be sufficient to represent complex textures (R1) the choice of the representation of the texture information (R1) 3) Limited comparison to (photorealistic) style transfer methods (R4) 4) Missing references (R1) Limited novelty is the biggest concern, but I believe even though the components are quite known, this combination of them is new, intuitive, clean, and works well. Moreover, two of the reviewers have been quite satisfied with the authors' rebuttal and the third, most negative, reviewer (R3), has not responded to the rebuttal and has not participated in the discussion. After carefully considering the reviews, the rebuttal, and the paper itself, I recommend acceptance. Authors are, however, required to remove additional results from the broader statement section - that's not what the section is intended for. Moreover, the authors are encouraged to address the reviewers' concerns in the final version of the paper.