Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
# Overall Comments: This is a nice paper in the sense that it makes the AE-based models produce really high-fidelity images as good as GAN-based models. In addition, the model also inherit the nice property of AE-based models that it does not suffer from the mode collapse issue. However, it seems to me that the only difference between this paper and the VQ-VAE paper is that this work introduces the hierarchical structure to learn different levels of latent representations and priors. The novelty looks a bit low. In addition, this paper didn't provide any idea about why such a design can make the generative performance better. The loss function (2) is not a reasonable objective to optimize considering the stop gradient operator. During the optimization procedure, the loss function may increase by taking a step in the gradient directions. This makes the algorithm like a hack and not elegant at all. # Questions: - Why you call this method VQ-VAE-2 since no variational method is used in the algorithm? - Is it possible that h_bottom encodes all the information for reconstruction and the algorithm ignores h_top completely? If not, why this cannot happen. In my opinion, h_bottom has size S/4xS/4xD, where S is the original image height/width and D is the number of channels. This is already big enough to remember the training data. Of course there is still a quantization step which can forbid it from remembering the input data. However, it seems to me the capacity is already big enough for h_bottom to perfectly reconstruct the input data. - Since h_bottom may have enough capacity to remember the training data, could it be that the model is just remembering the training data rather than generating really novel samples? - Can the algorithm do interpolation like may AE-based generative models can? I cannot find a trivial way to do this. Interpolation is also an important ability of AE-based models. - The input to the pixelCNN will at least have the size S/4xS/4. This will make the generation phase relative slow. Can the authors provide the comparison of time cost between this algorithm and the corresponding GAN model?
The proposed method produces impressive samples of high quality and high diversity, with a novel approach. The text is fairly clear and makes the method easy to grasp on a high level. However, the text is light on details. A detailed architecture description of the whole two-level hiararchy is missing, where the “conditional stacks” (L135, what are those?) should be explained/visualized, and how the different inputs to the decoder are merged, how the masked attention layers are implemented exactly, etc. Also, on a more minor note, Fig. 2a) does not quite align with Algorithm 1 (encoders both have aces to x, one also to etop, decoder has two inputs, no decoder from etop to hbottom), using the same terms in the figure as in the algorithm would help. Conceptually, the method seems to have some similarities to image compression literature, where the line of work leading to [A] use a hierarchical model, including an auto regressive model for latents, for *lossy* compression, and [B] proposes a hierarchical model similar to  for *lossless* compression. Further similarities arise a) from [A] training for MSE + NLL, motivated from a rate-distortion perspective, b) from the fact that both networks could in principle be used for sampling (albeit probably with much worse results, see Fig. 3 of [B]), and c) because NLL for some levels of their hierarchies can be calculated. Given the hierarchical nature of this work, a discussion or even quantitative comparison might be interesting. The comparison to BigGAN seems thorough and is convincing. I like the proposed ‘critic’. Refs [A] Minnen, David, et al. "Joint autoregressive and hierarchical priors for learned image compression." NeurIPS 2018. https://arxiv.org/pdf/1809.02736 [B] Mentzer, Fabian, et al. "Practical full resolution learned lossless image compression." CVPR 2019. https://arxiv.org/pdf/1811.12817.pdf Minor details - Fig. 2 and Fig. 5 miss a label. - Algorithm 3 mentioned on L138 is nowhere to be found.
Model Summary: The model proposed is based on the VQ-VAE model. The VQ-VAE model is a variational auto-encoder model, with quantized latent variables, and an auto-regressive prior trained to model these latent variables. It is extended using top-down hierarchical latent variables. The second contribution of the paper is to trade-off variability against quality of images by using the confidence of a classifier to reject bad samples. The paper is well and clearly written. The state of the art and related work is well covered, and good intuitions are given about the model. Conceptually, the novelty of the proposed approach is limited. The only difference to the VQ-VAE model seems to be the addition of hierarchical top down sampling, which is a standard construction in the VAE literature. On the other hand, the results obtained are excellent in terms of image quality. So the value of the paper is mostly experimental as it scales an existing model further. However, the paper does not put much focus on ablations studies (what is important to scale things up? e.x what's the impact of batch size etc..). Results: The samples obtained are impressive especially at high resolutions on FFHQ. They demonstrate that a likelihood model with sufficient capacity can generate compelling photo-realistic images. In particular, the precision-recall curves shown in Figure 5 convincingly show that the model obtains a quality of image close to that of BigGAN, while having better diversity. This is significant: models trained by maximum-likelihood, unlike GANs, are unlikely to drop parts of the training support ('mode-dropping') which makes it harder to produce compelling samples. However, this paper is not the first to achieve that, so better ablations that clearly show why it works may be desirable. In particular, the size of the model and batch sizes used are presumably significant, and a big part of why this works. If this is what it takes to make maximum likelihood work, it is better to make it evident. Therefore, ablations on Model size could be desirable. How does this model perform compared to existing models when model sizes are comparable? In this respect, the authors have provided architectural details in the Supplementary. Similarly, I would be interested to know how this model performs without quantization. No BPD measurements (or bounds) are provided in the main paper. I find this choice a little surprising. Given that the model belongs to the family of maximum-likelihood models, it could be desirable to report these values and compare to that literature as well as BigGAN. Classifier based rejection sampling: This contribution is quite orthogonal. In practice, resampling favours model that over-generalise VS models that mode-drop: the model is no longer penalized for generating bad samples, but has better support. In particular, given enough rejections any model will eventually produce a compelling sample. But it is indeed a simple and nice way to trade-off variability for quality, and to obtain precision recall curves. In the case where the classifier is trained on an other dataset, this uses extra data. Could you elaborate on the range of thresholds used for classifier based rejection in Figure 5? Are they the same for both VQ-VAE and BigGan? Also, is classifier based rejection used in Table 1? Minor remarks: In terms of evaluating image quality, showing nearest neighbours in pixel space and vgg-feature space could be considered.