Pith. sign in

REVIEW 4 major objections 5 minor 55 references

HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper derives the first hyperprior entropy model for vector-quantized image codecs, turning continuous embedding densities into differentiable index probabilities and cutting bits by over 61% at matched perceptual quality.

desk verdict Genuinely novel way to entropy-code VQ indices, but the headline bitrate savings are not supported by the reported experiments and the routing-mask overhead is unaccounted for. read the letter →

arxiv 2512.07192 v2 pith:3QWE6XEO submitted 2025-12-08 cs.CV

Niu Yi , Xu Tianyi , Ma Mingming , Wang Xinkun This is my paper
classification cs.CV
keywords vectorquantizationhyperpriorentropymodelgenerativeimagecompressionrate-distortioncontrolVQGANcontent-adaptivecodinglearnedcross-entropyrateobjective
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

HyperVQ/HVQ-CGIC aims to give vector-quantization (VQ) based generative image codecs the same kind of content-adaptive entropy modeling that scale hyperpriors gave scalar-quantization codecs. The central move is to stop predicting probabilities for discrete codebook indices directly and instead predict a Gaussian distribution over the continuous embedding space at each location, treating the codebook vectors as fixed anchors. A Mahalanobis-distance softmax turns that continuous density into a categorical distribution over indices, so the rate term becomes a differentiable cross-entropy loss and rate-distortion control can be learned end-to-end. The paper reports large bitrate savings—about 61% fewer bits than previous state-of-the-art generative codecs at matched perceptual quality on the standard Kodak set, plus an average 18.5% saving over static-frequency coding across several VQ architectures. If the derivation holds, this would make VQ index coding content-adaptive and RD-controllable for the first time.

What carries the argument

The load-bearing mechanism is the 'anchor Gaussian' mapping from continuous hyperprior to discrete index probabilities. Codebook embeddings are fixed anchors in a learned embedding space; per-location hyperparameters (mu, sigma) describe a Gaussian 'cloud'; the squared Mahalanobis distance (simplified to scaled Euclidean distance) plus softmax converts the density into a categorical distribution over K indices. This mapping keeps probabilities differentiable in mu and sigma, lets the cross-entropy rate term drive end-to-end training, and gives an arithmetic coder content-adaptive, spatially varying probabilities. Independent hyperprior branches for coarse, medium, and fine granularities, plu

What would settle it

Take the trained model and encode the Kodak test images, measuring the total bitstream as arithmetic-coded indices plus hyper-latents plus the losslessly range-coded routing masks; compare that total bpp against the comparison methods at equal LPIPS. If the reported 61.3% saving evaporates once masks are counted, the central bitrate claim fails.

Watch

Extended reading notes

Core claim

A hyperprior network predicts the mean and variance of a multivariate Gaussian in the embedding space at each location; the K codebook vectors act as fixed anchors. The probability of choosing index k is proportional to exp(-||e_k - mu||^2/(2 sigma^2)), normalized over all anchors, so a continuous density becomes a categorical distribution over discrete indices. Because the mapping is differentiable in mu and sigma, the cross-entropy of those probabilities is a trainable rate objective. The paper presents this as the first principled hyperprior entropy model for VQ indices, one whose per-location Gaussian supplies a high-order spatial context for arithmetic coding.

Load-bearing premise

The paper's own text (Section 3.4.1) says routing masks are losslessly compressed and omitted from the rate term; the central bitrate claim assumes this overhead is negligible—if the masks cost a significant fraction of the bitstream, the headline savings are unsupported.

Editorial extensions

If this is right

  • Rate control becomes a learned knob: the same trained model can target different rate-distortion operating points by weighting the cross-entropy rate terms, instead of changing codebook size or downsampling.
  • Because the probability model lives in embedding space, it can be dropped onto existing VQ codecs—single-scale, large-codebook, or residual VQ—without touching their quantizers, yielding an average 18.5% bit saving over static-frequency baselines.
  • Per-location Gaussian parameters act as a high-order context source; the paper's ablation shows this outperforms explicit first- through third-order arithmetic-coding contexts while adding less compute.
  • At matched perceptual quality (LPIPS), the method uses about 61% fewer bits than the compared state-of-the-art generative codecs on the Kodak benchmark, with millisecond-level encode/decode times.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct testable extension left implicit: combining the embedding-space Gaussian prior with a lightweight autoregressive model over already-decoded indices could push the predicted distribution closer to the true posterior, since the hyperprior provides only a one-shot context.
  • If the isotropic covariance proves limiting, allowing per-dimension variances or a low-rank covariance would give the Gaussian 'cloud' a shape rather than a fixed spread; the paper gives no experiment isolating this choice, so the gain of a full covariance is an open question.
  • The routing-mask overhead should be audited: the reported 61.3% saving was computed without including losslessly compressed masks in the training rate term, so an independent bitstream-level accounting is needed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes HVQ-CGIC, a VQ-based generative image compression framework that introduces a hyperprior entropy model for discrete VQ indices. The core idea is to predict a continuous multivariate Gaussian in the embedding space, treat codebook entries as fixed anchors, and convert Mahalanobis distances into categorical index probabilities via a softmax (Eq. 3). This yields a content-adaptive, spatially varying probability model that is used for arithmetic coding and a differentiable rate term during training. The framework is evaluated on Kodak and compared with prior generative methods, with claims of large bitrate savings and improved perceptual quality.

Significance. If the quantitative claims were supported, the work would contribute a principled way to make VQ index entropy coding content-adaptive and to enable end-to-end rate–distortion optimization for VQ-based generative codecs. The conceptual idea of bridging discrete indices and continuous Gaussian priors through codebook anchors is clean and potentially applicable beyond the specific architecture. The paper also provides a sensible staged training recipe and reports faster encoding/decoding than Control-GIC. However, the experimental evidence in the main text does not substantiate the headline claims, and a key information-theoretic justification is deferred to a missing supplementary, so the significance cannot be assessed as claimed.

major comments (4)
  1. [Abstract, §1, Table 1] The headline claim 'over 61% bit reduction on Kodak with the same LPIPS as the previous SOTA' is not supported by the reported results. Table 1 compares HVQ-CGIC(RD) at 0.530 bpp / LPIPS 0.0197 with Control-GIC at 0.594 bpp / LPIPS 0.0248 — a ~10.8% bpp saving at a different and better LPIPS, not 'same LPIPS' and not 61.3%. No table or figure in the main text matches the 61.3% figure. The abstract's separate claim of an average 18.5% bitrate saving is also unexplained and appears inconsistent with the 61.3% claim, making the central quantitative contribution internally inconsistent.
  2. [§3.4.1, §3.3.3] Routing masks are losslessly compressed but 'omitted from R' (Eq. 11/15), even though the final bitstream includes them. All reported bpp numbers in Table 1 and Figures 4–6 therefore exclude a real component of the transmitted bitstream. The magnitude of this overhead is neither reported nor bounded. Without accounting for mask bits, the comparisons against Control-GIC, CDC, and HiFiC are not apples-to-apples, and the reported savings may be substantially inflated.
  3. [§3.2.3] The paper's foundational claim is a 'mathematically derive[d] hyperprior entropy model' for VQ indices, but the key step — converting the continuous Gaussian density into index probabilities via Eq. (3) — is justified only by a reference to a supplementary that is not included. Eq. (4) is a standard cross-entropy identity and does not provide that justification. Since the whole method rests on Eq. (3), the main text must at least sketch the derivation or formally state the assumptions under which this mapping is optimal.
  4. [§4.2] The abstract claims HyperVQ is a 'universal module' that works across diverse VQ architectures (single-scale, large-codebook, RVQ), but the experiments evaluate only the proposed HVQ-CGIC pipeline with its multi-granularity encoder and shared codebook. No experiments plug the entropy model into an existing VQ codec (e.g., Control-GIC or GLC) to demonstrate universality. This claim is therefore outside the evidence presented.
minor comments (5)
  1. [Abstract, §4.3] The statement 'achieving an average bitrate saving of 18.5%, which is 7.28x the saving achieved by conventional Huffman coding' is undefined. No Huffman baseline is described, and the '7.28x' factor is not derived or compared against in the experiments.
  2. [§3.3.3] The lossless compression scheme for routing masks is not specified. To assess the omitted overhead in R, the authors should describe how the masks are entropy coded and report their actual cost in the bitstream.
  3. [Table 1] The 'No Compression' row reports 0.625 bpp, which is not a no-compression baseline for typical 8-bit images. This should be clarified (presumably a downsampled/uncompressed representation). Using this as the reference for 'relative improvements' is misleading without explanation.
  4. [§3.2.4] The phrase 'ultra-high-dimensional context' is informal. The paper does not compare against a conventional high-order context model beyond the 1st–3rd order ablation in Table 1, and the claim that the hyperprior is 'superior to any hand-crafted context' is not supported by theory or exhaustive experiment.
  5. [References] References are inconsistently formatted: some are arXiv preprints without venue or version (e.g., [5], [30]), while the text refers to [30] as ICLR'25. Please standardize citations and include all necessary bibliographic details.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the VQ hyperprior is an explicit modeling ansatz trained by cross-entropy, not a prediction forced by its own inputs.

full rationale

The derivation chain in Section 3.2 is an explicit modeling choice, not a circular reduction. The paper states "We hypothesize that the probability of selecting index k at location (i,j) is related to the 'distance' of its anchor e_k from the predicted Gaussian" and then defines the categorical probability via a softmax over Mahalanobis distances (Eqs. 1-3). This is a transparent ansatz, not a quantity defined in terms of the target bitrate. The rate term in Eq. (4) is the cross-entropy between the true index distribution and the predicted distribution, and training minimizes this cross-entropy; using the same learned probability model for arithmetic coding is standard practice, not circularity. There are no load-bearing self-citations: the cited baselines and prior methods are external works, and no uniqueness theorem from the authors' own prior work is invoked. The concerns about the reported 61.3% bit-reduction figure and the omission of routing-mask overhead from R are quantitative-support or evaluation-consistency issues, not cases where a prediction reduces by construction to its input. Therefore, no significant circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework introduces no new physical or mathematical entities. It adds a modeling choice (Gaussian in embedding space) and a practical simplification (isotropic covariance), both of which are architectural assumptions rather than invented objects.

free parameters (3)
  • Per-granularity rate weights λ_y,g and λ_z,g = 1.5e-3 to 1.5e-2; λ_GAN=0.1, λ_VQ=1 in Sec 4.3
    Hand-chosen loss weights control the RD trade-off; the reported savings depend on these operating points.
  • Granularity training ratios = Stage A: (0.1,0.3,0.6); Stages B/C: (0.3,0.3,0.4)
    Manually chosen mix of coarse/medium/fine tokens; affects rate-distortion behavior and is not derived.
  • VQ commitment coefficient β = 0.25
    Standard VQ commitment weight, chosen by hand.
assumptions (4)
  • ad hoc to paper Index probabilities are obtained by evaluating a Gaussian density at discrete codebook anchors and normalizing (Eq. 3)
    The mapping from continuous Gaussian to categorical probabilities is postulated, not derived in the main text; the promised information-theoretic justification is deferred to a missing supplementary.
  • domain assumption VQ indices are spatially conditionally independent given the hyperprior z (rate is a sum over locations in Eq. 4)
    The rate model factorizes over spatial locations; any residual spatial correlation not captured by z is ignored.
  • ad hoc to paper Routing-mask overhead can be omitted from the rate during training
    The loss omits losslessly compressed masks, assuming their bit cost is negligible relative to index and hyper-latent bits.
  • standard math Arithmetic coding with the learned P(y|z) achieves code length close to Shannon entropy
    Standard information-theoretic result, used to justify R_y as the expected bitrate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression." pith.science (2026). https://pith.science/paper/3QWE6XEO

@misc{pith2026251207192,
  author       = {Pith},
  title        = {Pith review of: HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3QWE6XEO}},
  note         = {Machine review of arXiv:2512.07192}
}
read the original abstract

Vector Quantization (VQ) based generative image compression has achieved remarkable perceptual quality. However, existing VQ codecs suffer from two fundamental limitations. First, they lack efficient content-adaptive entropy modeling and rely on static frequencies, leading to low coding efficiency. Second, the inherent conflict between discrete indices and continuous priors prevents true end-to-end joint Rate-Distortion (RD) optimization. To resolve these issues, we propose HyperVQ, a principled framework that establishes a high-performance hyperprior entropy foundation for VQ-based codecs. The core insight of HyperVQ is to shift probability modeling entirely into the continuous embedding space. Instead of directly predicting probabilities for discrete symbols, HyperVQ predicts a high-dimensional continuous multivariate Gaussian distribution for the continuous latents. By treating the discrete codebook entries as fixed "anchors" in this space, we convert the continuous Gaussian density into categorical index probabilities based on relative distances. This elegant formulation provides a powerful, spatially-adaptive entropy engine and renders the cross-entropy rate objective fully differentiable, empowering the network to actively and dynamically optimize the RD trade-off during training. To ensure practicality, we design the lightweight H Block and the Probability Estimation Engine (PEE) to facilitate highly parallel, millisecond-level inference. Experiments demonstrate that HyperVQ acts as a universal module across diverse VQ architectures (single-scale, large-codebook, RVQ), achieving an average bitrate saving of 18.5%, which is 7.28x the saving achieved by conventional Huffman coding. This establishes a robust, RD-controllable foundation for next-generation generative image compression.

Figures

Figures reproduced from arXiv: 2512.07192 by the authors.

Figure 1
Figure 1. Our HVQ-CGIC breaks through the limitation that VQ cannot design hyperprior entropy models, and achieves significantly lower bitrates than [30] on Kodak under the same downsampling factor. term is unavailable or constant during training. As a result, the bitrate is steered by manual knobs (code￾book size, downsampling) rather than a differentiable rate term. In contrast, our hyperprior provides a learned P(y | z) ov… view at source ↗
Figure 2
Figure 2. HVQ-CGIC flowchart, details see 3.3. Θ: represents element-wise multiplication here. 2 Related Works and Discus￾sions 2.1 Learned Image Compression The foundations of lossy image compression are grounded in Shannon’s rate-distortion theory [42]. The seminal work of Ballé et al. [4] established the now-dominant paradigm of end-to-end optimization based on Scalar Quantization (SQ). Subsequent re￾search has rapidly adv… view at source ↗
Figure 3
Figure 3. Scalar quantization hyperprior model (a) and our proposed vector quantization hyperprior models (b)(c). Detailed introduction is in §3.2.3. RD-balance regulation. Jia et al. [22] (GLC) uses a VQ-VAE as a ”frontend” to map an image to a continuous latent space, which is then compressed by a standard SQ-Hyperprior pipeline. This hybrid approach is essentially still reliant on scalar quanti￾zation for entropy coding, n… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of Methods on Kodak Dataset [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Conduct quantitative comparisons with the listed Baseline Methods. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: In addition to the context designs shown in the above figure, these three entropy coding strate￾gies also perform adaptive arithmetic coding by up￾dating encoding frequencies in real-time based on the data used for testing, building upon statistical en￾tropy coding. ch…
Figure 6
Figure 6. Figure 6: Comparison of Methods on Kodim21. son results for high-resolution images on DIV2K and CLIC2020 in the Supplementary Material. 4.3 Ultimate Entropy Coding Perfor￾mance To demonstrate where the limits of hyperpriors lie, we increased the weights of λy,g and λz,g from 1.5…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 8 linked inside Pith

  1. [1]

    Soft-to-hard vector quantization for end-to-end learning compressible representations

    Eirikur Agustsson, Fabian Mentzer, Michael Tschan- nen, Lukas Cavigelli, Radu Timofte, Luca Benini, and Luc V Gool. Soft-to-hard vector quantization for end-to-end learning compressible representations. Advances in neural information processing systems , 30, 2017

    2017

  2. [2]

    Ntire 2017 challenge on single image super-resolution: Dataset and study

    Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages 126–135, 2017

    2017

  3. [3]

    Gen- erative adversarial networks for extreme learned im- age compression

    Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Gen- erative adversarial networks for extreme learned im- age compression. In Proceedings of the IEEE/CVF international conference on computer vision , pages 221–231, 2019

    2019

  4. [4]

    End-to-end optimized image compression

    Johannes Ballé, Valero Laparra, and Eero P Si- moncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704 , 2016

    arXiv 2016

  5. [5]

    Variational image compression with a scale hyperprior

    Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 , 2018

    arXiv 2018

  6. [6]

    A note on the inception score

    Shane Barratt and Rishi Sharma. A note on the inception score. arXiv preprint arXiv:1801.01973 , 2018

    arXiv 2018

  7. [7]

    BPG image format

    Fabrice Bellard. BPG image format. https:// bellard.org/bpg/, 2018

    2018

  8. [8]

    Demystifying mmd gans

    Mikołaj Bińkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401 , 2018

    arXiv 2018

Show all 55 references
  1. [9]

    Overview of the versatile video coding (vvc) standard and its applications

    Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology , 31(10):3736– 3764, 2021

    2021

  2. [10]

    Towards image compres- sion with perfect realism at ultra-low bitrates

    Marlene Careil, Matthew J Muckley, Jakob Verbeek, and Stéphane Lathuilière. Towards image compres- sion with perfect realism at ultra-low bitrates. In The Twelfth International Conference on Learning Representations, 2023

    2023

  3. [11]

    Learned image compression with discretized gaussian mixture likelihoods and atten- tion modules

    Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and atten- tion modules. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 7939–7948, 2020

    2020

  4. [12]

    Variable rate deep image compression with a condi- tional autoencoder

    Yoojin Choi, Mostafa El-Khamy, and Jungwon Lee. Variable rate deep image compression with a condi- tional autoencoder. In Proceedings of the IEEE/CVF international conference on computer vision , pages 3146–3154, 2019

    2019

  5. [13]

    Kodak lossless true color image suite

    Eastman Kodak Company. Kodak lossless true color image suite. http://r0k.us/graphics/kodak/,

  6. [14]

    Image quality assessment: Unifying structure and texture similarity

    Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transac- tions on pattern analysis and machine intelligence , 44(5):2567–2581, 2020

    2020

  7. [15]

    Taming transformers for high-resolution image syn- thesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image syn- thesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 12873–12883, 2021

    2021

  8. [16]

    Nvtc: Nonlinear vector transform coding

    Runsen Feng, Zongyu Guo, Weiping Li, and Zhibo Chen. Nvtc: Nonlinear vector transform coding. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 6101– 6110, 2023

    2023

  9. [17]

    Neu- ral image compression via attentional multi-scale back projection and frequency decomposition

    Ge Gao, Pei You, Rong Pan, Shunyuan Han, Yuanyuan Zhang, Yuchao Dai, and Hojae Lee. Neu- ral image compression via attentional multi-scale back projection and frequency decomposition. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision , pages 14677–14...

    2021

  10. [18]

    Elic: Efficient learned image compression with unevenly grouped space- channel contextual adaptive coding

    Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space- channel contextual adaptive coding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5718–5727, 2022

    2022

  11. [19]

    Checkerboard context model for efficient learned image compression

    Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 14771– 14780, 2021. 10

    2021

  12. [20]

    Gans trained by a two time-scale update rule con- verge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Un- terthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule con- verge to a local nash equilibrium. Advances in neural information processing systems , 30, 2017

    2017

  13. [21]

    Context-based trit-plane coding for progressive image compression

    Seungmin Jeon, Kwang Pyo Choi, Youngo Park, and Chang-Su Kim. Context-based trit-plane coding for progressive image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14348–14357, 2023

    2023

  14. [22]

    Generative latent coding for ultra-low bitrate image compression

    Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Generative latent coding for ultra-low bitrate image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26088–26098, 2024

    2024

  15. [23]

    Mlic: Multi- reference entropy model for learned image compres- sion

    Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi- reference entropy model for learned image compres- sion. In Proceedings of the 31st ACM International Conference on Multimedia , pages 7618–7627, 2023

    2023

  16. [24]

    Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks

    Nick Johnston, Damien Vincent, David Minnen, Michele Covell, Saurabh Singh, Troy Chinen, Sung Jin Hwang, Joel Shor, and George Toderici. Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In Proceedings of the IEEE conference...

    2018

  17. [25]

    Joint global and local hierarchical priors for learned image compression

    Jun-Hyuk Kim, Byeongho Heo, and Jong-Seok Lee. Joint global and local hierarchical priors for learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pages 5992–6001, 2022

    2022

  18. [26]

    Openimages: A public dataset for large-scale multi-label and multi-class image clas- sification

    Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Fer- rari, Sami Abu-El-Haija, Alina Kuznetsova, Has- san Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, et al. Openimages: A public dataset for large-scale multi-label and multi-class image clas- sification. Dataset available ...

    2017

  19. [27]

    Context-adaptive entropy model for end-to- end optimized image compression

    Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beack. Context-adaptive entropy model for end-to- end optimized image compression. arXiv preprint arXiv:1809.10452, 2018

    2018 arXiv

  20. [28]

    Selective compression learning of latent represen- tations for variable-rate image compression

    Jooyoung Lee, Seyoon Jeong, and Munchurl Kim. Selective compression learning of latent represen- tations for variable-rate image compression. Ad- vances in Neural Information Processing Systems , 35:13146–13157, 2022

    2022

  21. [29]

    Approaching rate-distortion limits in neu- ral compression with lattice transform coding

    Eric Lei, Hamed Hassani, and Shirin Saeedi Bidokhti. Approaching rate-distortion limits in neu- ral compression with lattice transform coding. arXiv preprint arXiv:2403.07320 , 2024

    2024 arXiv

  22. [30]

    Once-for-all: Controllable gener- ative image compression with dynamic granularity adaption

    Anqi Li, Feng Li, Yuxi Liu, Runmin Cong, Yao Zhao, and Huihui Bai. Once-for-all: Controllable gener- ative image compression with dynamic granularity adaption. arXiv preprint arXiv:2406.00758 , 2024

    2024 arXiv

  23. [31]

    Content-oriented learned image com- pression

    Meng Li, Shangyin Gao, Yihui Feng, Yibo Shi, and Jing Wang. Content-oriented learned image com- pression. In European Conference on Computer Vi- sion, pages 632–647. Springer, 2022

    2022

  24. [32]

    Learning convolutional networks for content-weighted image compression

    Mu Li, Wangmeng Zuo, Shuhang Gu, Debin Zhao, and David Zhang. Learning convolutional networks for content-weighted image compression. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 3214–3223, 2018

    2018

  25. [33]

    A spatial rnn codec for end-to-end im- age compression

    Chaoyi Lin, Jiabao Yao, Fangdong Chen, and Li Wang. A spatial rnn codec for end-to-end im- age compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pages 13269–13277, 2020

    2020

  26. [34]

    Learned image compression with mixed transformer-cnn ar- chitectures

    Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn ar- chitectures. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 14388–14397, 2023

    2023

  27. [35]

    Conditional probability models for deep image compression

    Fabian Mentzer, Eirikur Agustsson, Michael Tschan- nen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In Proceedings of the IEEE conference on computer vi- sion and pattern recognition , pages 4394–4402, 2018

    2018

  28. [36]

    High-fidelity generative image compression

    Fabian Mentzer, George D Toderici, Michael Tschan- nen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in neural information processing systems, 33:11913–11924, 2020

    2020

  29. [37]

    Joint autoregressive and hierarchical priors for learned image compression

    David Minnen, Johannes Ballé, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems , 31, 2018

    2018

  30. [38]

    Channel-wise au- toregressive entropy models for learned image com- pression

    David Minnen and Saurabh Singh. Channel-wise au- toregressive entropy models for learned image com- pression. In 2020 IEEE International Conference on Image Processing (ICIP) , pages 3339–3343. IEEE, 2020

    2020

  31. [39]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters , 20(3):209– 212, 2012

    2012

  32. [40]

    Content adaptive latents and decoder for neural im- age compression

    Guanbo Pan, Guo Lu, Zhihao Hu, and Dong Xu. Content adaptive latents and decoder for neural im- age compression. In European Conference on Com- puter Vision , pages 556–573. Springer, 2022

    2022

  33. [41]

    Real-time adap- tive image compression

    Oren Rippel and Lubomir Bourdev. Real-time adap- tive image compression. In International conference on machine learning , pages 2922–2930. PMLR, 2017

    2017

  34. [42]

    Coding theorems for a dis- crete source with a fidelity criterion

    Claude E Shannon et al. Coding theorems for a dis- crete source with a fidelity criterion. IRE Nat. Conv. Rec, 4(142-163):1, 1959

    1959

  35. [43]

    Expanded adaptive scaling normalization for end to end image compression

    Chajin Shin, Hyeongmin Lee, Hanbin Son, Sangjin Lee, Dogyoon Lee, and Sangyoun Lee. Expanded adaptive scaling normalization for end to end image compression. In European Conference on Computer Vision, pages 390–405. Springer, 2022. 11

    2022

  36. [44]

    Clic 2020: Chal- lenge on learned image compression

    George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Ballé, Wenzhe Shi, and Radu Timofte. Clic 2020: Chal- lenge on learned image compression. Retrieved March, 29:2021, 2020

    2020

  37. [45]

    Evc: Towards real-time neural image compression with mask decay

    Guo-Hua Wang, Jiahao Li, Bin Li, and Yan Lu. Evc: Towards real-time neural image compression with mask decay. arXiv preprint arXiv:2302.05071 , 2023

    2023 arXiv

  38. [46]

    Slimmable compressive autoen- coders for practical neural image compression

    Fei Yang, Luis Herranz, Yongmei Cheng, and Mikhail G Mozerov. Slimmable compressive autoen- coders for practical neural image compression. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 4998– 5007, 2021

    2021

  39. [47]

    Lossy image compression with conditional diffusion models

    Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models. Ad- vances in Neural Information Processing Systems , 36:64971–64995, 2023

    2023

  40. [48]

    The unreasonable ef- fectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable ef- fectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 586–595, 2018

    2018

  41. [49]

    Attention-guided im- age compression by deep reconstruction of compres- sive sensed saliency skeleton

    Xi Zhang and Xiaolin Wu. Attention-guided im- age compression by deep reconstruction of compres- sive sensed saliency skeleton. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13354–13364, 2021

    2021

  42. [50]

    Multi-modality deep restoration of extremely compressed face videos

    Xi Zhang and Xiaolin Wu. Multi-modality deep restoration of extremely compressed face videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2024–2037, 2022

    2024

  43. [51]

    Lvqac: Lattice vector quantization coupled with spatially adaptive com- panding for efficient learned image compression

    Xi Zhang and Xiaolin Wu. Lvqac: Lattice vector quantization coupled with spatially adaptive com- panding for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 10239– 10248, 2023

    2023

  44. [52]

    Learning optimal lattice vector quantizers for end-to-end neural image com- pression

    Xi Zhang and Xiaolin Wu. Learning optimal lattice vector quantizers for end-to-end neural image com- pression. Advances in Neural Information Processing Systems, 37:106497–106518, 2024

    2024

  45. [53]

    Unified multivariate gaussian mixture for efficient neural image compression

    Xiaosu Zhu, Jingkuan Song, Lianli Gao, Feng Zheng, and Heng Tao Shen. Unified multivariate gaussian mixture for efficient neural image compression. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 17612– 17621, 2022

    2022

  46. [54]

    Transformer-based transform coding

    Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In Interna- tional conference on learning representations , 2022. 12

    2022

  47. [1993]

    Original images, commonly used as a test set for image compression

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.