NeurIPS 2020

High-Fidelity Generative Image Compression


Review 1

Summary and Contributions: The authors study how different elements of an optimization algorithm effect optimized generative models for lossy image compression (including architectural decisions, training strategies, and perceptual loss functions). They claim their results achieve a state of the art (evaluated with both perceptual metrics and a user study) and also say that their methodology works on high resolution images.

Strengths: --- Updated Review --- I want to thank the authors for addressing some of my concerns. I believe as I did before that this is quite a strong submission and my review of it is unchanged. A clear accept. This is a very strong submission with both extremely compelling practical results, and interesting theoretical progress and observations that advance our understanding of why previous methods have had the types of failures that they have had. I believe this is of high significance, novelty and relevance to the entire NeurIPS community, especially those working in image compression and perceptual metrics and loss functions.

Weaknesses: I'm a bit concerned by the perceptual study methodology. The authors ask users to search for "interesting" areas of the image that will be shown Is it assumed that there are no meaningful compression errors in flat/featureless areas of the image? Artifacts in these areas could be the most noticeable artifacts from a pure perception standpoint (as they are not masked by local contrast though may be masked by luminance). Additionally, are any considerations taken into account to ensure consistency of display and viewing conditions (ie resolution, color space, size of display, viewing distance, lighting conditions? ) It's also a bit strange to compare full image ratings from the metrics to crops rated by the individuals. Can the authors address this?

Correctness: I'd like the authors to address the concerns listed above with the user study, but I believe the visible results to be quite impressive regardless of the quantification.

Clarity: Please fix the labeling on figure 3 (cannot see some of the text below the x axis). The axes in figure 5 are not labeled and thus the figures are not entirely clear (though one can guess what they are). Please fix and make more clear.

Relation to Prior Work: Prior work is clearly discussed, critiqued, built upon and it is shown how this work moves to a novel contribution.

Reproducibility: Yes

Additional Feedback:


Review 2

Summary and Contributions: The authors present a method for lossy neural image compression that combines conditional generative adversarial networks with learned compression. The method is able to achieve competitive rate distortion performance perceptually as evidenced by an extensive evaluation. It also operates at fairly high resolutions (e.g. 2k x 2k) and the authors additionally show an analysis of the different architectural components.

Strengths: The topic of neural image compression is currently very significant and gains in rate distortion performance through new methods would have a positive impact e.g. on bandwidth used regarding transmission of visual content. In my opinion, the strength of the paper is to propose a system that achieves good practical performance when measuring rate distortion performance perceptually. The evaluation seems thorough and the results look very convincing outperforming previous compression methods that employ generative adversarial networks. The analysis of the individual components used is sufficiently detailed.

Weaknesses: While the approach seems sound and practically achieves good performance, the idea of combining generative adversarial networks and learned compression is (also as the authors show in related work) not new. The novelty here lies in the details of how the system is designed exactly and arguably in the conditioning of the GAN. Despite a potential lack in novelty, I still tend towards acceptance since the system is setting a new state of the art and reading the paper is insightful.

Correctness: The claims and method look correct to me.

Clarity: The paper is well written.

Relation to Prior Work: Related work is clearly discussed.

Reproducibility: Yes

Additional Feedback: With VVC being finalised, it would be interesting to know how this neural image compression method compares to IFrame coding in VVC. If this is still too early to do such comparisons, I do not insist to include this. To get a better feeling of the reconstruction the GAN encourages, I would be interested to see what happens temporally when compressing an image sequence. Do the authors expect temporal artefacts or do smooth changes of the input carry over to temporally coherent output? In case there are temporal instabilities, are they only observed for a certain low bit rate?


Review 3

Summary and Contributions: The submission presents an end-to-end learned lossy image compression model that achieves state of the art bit-rates while maintaining high perceptual quality. Update after rebuttal: I am happy with the promised inclusion and discussion of failure modes in the main text and maintain my positive assessment.

Strengths: - Impressive visual quality of compressed images - Extensive and sound evaluation showing clear average performance improvements over classical image compression methods as well end-to-end learned methods.

Weaknesses: - Little conceptual novelty, mainly skillful engineering using existing model components. - No examples and discussion of model limitations in the main text. As typical for GAN-based image synthesis, the model has issues generating fine-detail that is not texture. This becomes particularly visible for human bodies/faces (Appendix page 23, Kodak/14) or text (Appendix page 31, CLIC 25bf4). In both cases I would argue that the comparable BPG compression is clearly superior. I believe that these quality shortcomings for important sub-categories of images are a major obstacle to deploying such models as a generic image compression algorithm in real world settings and thus should be adequately discussed in the main text.

Correctness: Yes

Clarity: Yes

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback: n/a


Review 4

Summary and Contributions: This paper proposes a generative compression method to achieve high quality reconstructions, In a user study, the paper shows that the proposed approach is visually preferred. Several distortion metrics are used to evaluate the method quantitatively. and it extensively studies the proposed architecture, training strategies, as well as the loss, in terms of perceptual metrics and stability.

Strengths: This paper improves the GAN based image compression method using the conditional GAN, appling a perceptual distortion and modifies several layers. User study experiments and evaluation using different metrics are clarified specifically.

Weaknesses: Two comments are listed as below: 1) As the title named High-Fidelity Generative Image Compression, I can not understand why it is called high-fidelity. With the GAN loss and perceptual loss added, the compressed image can be more perceptually preferred as proved before. However, texture generated by GAN is usually fake and it may not be called high-fidelity. I think an additional section is needed to discuss texture generated by the proposed method is much closer to input than others apart from perceptual evaluation. 2) Comparison with the prior work which utilized GAN is necessary to include in section 4 to illustrate the advantage of the proposed method.

Correctness: Yes.

Clarity: Yes.

Relation to Prior Work: The proposed one utilizes Conditional GAN instead of GAN, single scale instead of multi-scale, channel-averaging LayerNorm instead of InstanceNorm and additionally a perceptual distortion. However, it may lack of proof of the above changes. Maybe ablation study is necessary.

Reproducibility: Yes

Additional Feedback: