Pith. sign in

REVIEW 3 major objections 6 minor 76 references

CMamba: Learned Image Compression with State Space Models

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read CMamba fuses state space models and CNNs to beat VVC by up to 18.83% BD-Rate while cutting decoding time by 71.4%.

desk verdict A well-ablated hybrid CNN+SSM image codec whose efficiency claims over MLIC++ need measurement-protocol verification before the headline numbers can be trusted. read the letter →

arxiv 2502.04988 v1 pith:SRBGBLVK submitted 2025-02-07 eess.IV cs.CV

Zhuojie Wu , Heming Du , Shuyun Wang , Ming Lu , Haiyang Sun , Yandong Guo , Xin Yu This is my paper · ORCID
classification eess.IVcs.CV
keywords learnedimagecompressionstatespacemodelsMambaentropymodelrate-distortionoptimizationCNNandSSMhybridchannel-wiseautoregression
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

The paper introduces CMamba, a learned image compression framework that combines State Space Models (SSMs) with Convolutional Neural Networks (CNNs) to achieve both high rate-distortion performance and low computational complexity. It claims that SSMs capture global content but lose high-frequency details, while CNNs preserve local details, so the proposed Content-Adaptive SSM module dynamically fuses the two in both encoder and decoder. A second component, the Context-Aware Entropy module, uses SSMs to model spatial dependencies in latent representations and an autoregressive scan to reduce channel redundancy. The authors report that CMamba outperforms VVC by 14.95%, 18.83%, and 13.89% BD-Rate on Kodak, Tecnick, and CLIC, and reduces parameters by 51.8%, FLOPs by 28.1%, and decoding time by 71.4% compared to the previous best learned codec on Kodak.

What carries the argument

The Content-Adaptive SSM (CA-SSM) module is the core transform block: it runs a VSS block (a state space model that scans feature maps in four directions) in parallel with a ResBlock, then computes softmax weights alpha and beta from a pooled fusion to blend global and local features. The Context-Aware Entropy (CAE) module is the second piece: it splits the latent representation into channel groups, feeds the hyperprior and previously decoded groups through an SSM and feed-forward network to predict Gaussian parameters, and applies a latent residual prediction network to compensate for quantization error.

What would settle it

Re-measure encoding and decoding latency for MLIC++ and CMamba on the same RTX 3090 Ti with the same framework, batch size, and input resolution; if MLIC++ decoding time drops to near CMamba's 52 ms under matched conditions, the headline decoding-time advantage is reduced or eliminated.

Watch

Extended reading notes

Core claim

The central claim is that a hybrid CNN–SSM codec can beat both handcrafted standards and prior learned codecs on rate-distortion while being much cheaper to run. The paper argues that pure SSM backbones lose high-frequency information as depth grows, and pure CNN backbones cannot capture global content, so the architecture pairs a Visual State Space block with a ResBlock and fuses their outputs with learned per-channel weights. For entropy coding, it replaces transformer-based context models with SSMs for spatial parameterization, and processes latent channels sequentially so that each channel is conditioned on previously decoded ones, reducing inter-channel redundancy without pixel-by-pixel autoregression. Together these components yield the reported BD-Rate gains and efficiency improvements.

Load-bearing premise

The claimed efficiency gains assume that the baseline latency, FLOPs, and parameter counts in Table I were measured under the same hardware and software conditions as CMamba; if the baselines were taken from publications using different GPUs or frameworks, the 71.4% decoding-time reduction could be overstated.

Editorial extensions

If this is right

  • SSM-based entropy models can replace transformer context models in learned codecs, achieving competitive or better rate-distortion with lower latency and fewer parameters.
  • The dynamic fusion of global SSM features and local CNN features is a generally usable backbone design for compression, where preserving high-frequency detail directly reduces bitrate.
  • Channel-wise autoregressive entropy coding conditioned on SSM-extracted spatial priors gives most of the benefit of spatial autoregression at a fraction of the decoding cost.
  • The reported complexity reductions suggest that learned codecs can become practical on consumer hardware for real-time or near-real-time decoding.

Reading between the lines

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

  • The Fourier-spectrum diagnostic used here could be adapted as a general tool for deciding when to hybridize a backbone: if a candidate architecture's feature spectrum decays at high frequencies, adding a local-detail branch may improve any downstream task, not just compression.
  • Because SSMs have linear complexity in sequence length, the CAE module's spatial modeling may scale to higher-resolution images or video frames where transformer context models become prohibitive.
  • A direct test of the paper's efficiency claim would be to run MLIC++ and CMamba on identical hardware and framework with the same batching; if the 71.4% decoding-time gap narrows under re-measurement, the practical advantage would be smaller than reported.
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

3 major / 6 minor

Summary. The paper proposes CMamba, a learned image compression framework that combines convolutional networks with state space models (SSMs). It introduces two components: a Content-Adaptive SSM (CA-SSM) module that dynamically fuses global features from a VSS block with local features from a ResBlock in both the analysis and synthesis transforms, and a Context-Aware Entropy (CAE) module that models spatial dependencies via SSMs and channel dependencies via channel-wise autoregression with a latent residual prediction network. The authors train on OpenImages and evaluate on Kodak, Tecnick, and CLIC, reporting BD-Rate improvements over VVC (14.95%, 18.83%, 13.89% respectively) and large efficiency gains over MLIC++: 51.8% fewer parameters, 28.1% fewer FLOPs, and 71.4% faster decoding on Kodak.

Significance. If the reported results hold, CMamba is a practically useful step in learned image compression because it combines a modest rate-distortion improvement over MLIC++ (about 1.56 percentage points BD-Rate on Kodak) with substantially lower complexity. The internal ablations are consistent: the CA-SSM module improves the VSS+ChARM baseline by 5.94 percentage points BD-Rate, and CAE adds another 2.04 points, while reducing parameters and FLOPs. The RD numbers across the three test sets are internally coherent, and the paper provides both MSE- and MS-SSIM-optimized results. However, the significance of the work hinges on the credibility of the efficiency comparison and on the availability of enough architecture detail to verify the parameter and FLOP counts. The paper currently lacks a clear measurement protocol for the baselines and omits several key architecture specifications, so the efficiency claims cannot yet be fully assessed.

major comments (3)
  1. [V-A, Table I] The paper does not state whether the latency, FLOPs, and parameter counts for the compared methods in Table I were measured on the same RTX 3090 Ti and i9-12900 platform in the authors' environment or taken from the original publications. Several entries are listed as ">1000" ms, which suggests those models were not actually run by the authors. This matters because the central claim over MLIC++ is efficiency: CMamba's BD-Rate gain on Kodak over MLIC++ is only 1.56 percentage points, while the reported decoding time reduction is 71.4%. If MLIC++'s numbers were obtained on a different GPU (e.g., A100 or V100) or with a different profiling toolchain, the comparison would be unfair and the efficiency advantage could be dominated by hardware differences. Please clarify the source and measurement protocol for every entry in Table I, and ideally re-run all baselines under the same conditions.
  2. [IV, V-B, Table I] The manuscript omits several architecture details that are necessary to reproduce CMamba and to audit the reported parameter and FLOP counts. Specifically, the channel widths of the four stages (C1-C4), the number of channel groups S in the CAE, the number of VSS blocks and ResBlocks per stage, the kernel sizes in the CNN path, and the exact configuration of the dynamic fusion MLP are not given. Without these specifications, the numbers 56.21M parameters and 355.29G FLOPs cannot be independently verified, and the design cannot be replicated. Please add a complete architecture table (channel widths, group counts, block counts, kernel sizes) or release the code with the paper.
  3. [IV-B, Eq. (9)] In the CAE module, the paper says that SSMs are used to parameterize spatial content and that channels are processed autoregressively, but it does not specify the scanning pattern or the depth of the SSM applied to the concatenated context [Phi', y_bar_<i]. It is important for an entropy model to state whether fssm is the same non-causal SS2D block used in the encoder or a causal 1D scan. If the same four-direction scan is applied to the context, the authors should explain why this does not leak information from the current group y_i, which is the group whose parameters are being estimated. Please clarify the exact definition and causality properties of fssm in Eq. (9).
minor comments (6)
  1. [V-D, Table II] The header row of Table II is garbled ("CA-SSM ✓ ✓ VVCCAE ✓ ✓"), and the column alignment is broken in the text; please reformat the table so that the ablation configuration is readable.
  2. [V-D, Table III] Table III appears to contain extra columns (e.g., values like "+1.42" and "+5.74") that are not explained in the caption or the text; please define all columns explicitly.
  3. [V-A, Table I] Please define what ">1000" means in Table I (e.g., whether it indicates a timeout or a value exceeding 1000 ms) and specify whether the VVC numbers were measured by the authors or taken from another source.
  4. [III, Eq. (5)] The discretization formula in Eq. (5) uses (Delta A)^{-1}, which is only well-defined if the matrix is invertible; since the implementation likely follows the standard Mamba/VMamba code with a different discretization, please add a note clarifying that Eq. (5) is a simplified continuous-time formulation.
  5. [Fig. 1] The y-axis of Fig. 1(b) is labeled "Relative log amplitudes," but the precise definition (difference between log amplitude at normalized frequency 0.0π and 1.0π) is only given in a footnote; please move this definition into the main figure caption for clarity.
  6. [II, reference [37]] The paper refers to MLIC++ [37] as the "previous best LIC method," but this is an ICML 2023 Workshop paper; please state the date and selection criteria for this claim, and consider comparing with more recent learned codecs that have since been published or posted.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CMamba is an empirical system paper whose performance claims come from trained models and external benchmarks, not from an equation-level derivation that reduces to its inputs.

full rationale

CMamba does not present a first-principles derivation or a prediction; its claims are empirical rate-distortion and complexity measurements from trained models evaluated on standard benchmarks. The CA-SSM and CAE modules are defined by explicit equations whose components (VSS blocks, ResBlocks, dynamic fusion, channel-wise autoregression) are architectural choices, not quantities fitted to the reported BD-Rate numbers. The rate-distortion loss in Eq. (3) is the standard objective, and the reported gains are measured, not derived from the module definitions. One in-scope concern is that the Fourier-spectrum motivation in Fig. 1 was computed on the Kodak test set, so Kodak results are not a fully blind evaluation; this is a dataset-contamination/overfitting concern, but it does not make any claimed result equivalent to an input by construction. Similarly, Table I does not state whether baseline latency and FLOPs were re-measured on the same RTX 3090 Ti hardware, which is a comparability risk for the efficiency claims, but it is not circularity. No load-bearing self-citation chain, imported uniqueness theorem, or renamed-known-result pattern is present. The architecture ablations in Tables II-V compare against standard components, and the efficiency numbers are reported measurements rather than constructional identities. Therefore no circular step meeting the quoted-evidence standard can be identified.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim relies on standard, previously published building blocks: the LIC optimization framework, Gaussian entropy model, VMamba's SS2D scanning, and channel-wise autoregressive entropy coding. The only hand-chosen values are the Lagrange multipliers and unreported architectural dimensions (channel widths, group count S). No new physical or conceptual entities are introduced.

free parameters (4)
  • Rate-distortion Lagrange multipliers lambda (MSE) = {25, 35, 67, 130, 250, 500} x 1e-4
    Chosen by hand to produce six rate points on the RD curve; standard practice in learned image compression, not fitted to test data.
  • Rate-distortion Lagrange multipliers lambda (MS-SSIM) = {3, 5, 8, 16, 36, 64}
    Same purpose for MS-SSIM-optimized models; chosen by hand to span a bitrate range.
  • Channel widths for the four analysis and synthesis stages (C1-C4) = not reported
    Architectural hyperparameters shown in Fig. 2(a) but values are not given; the reported RD and complexity results depend on them.
  • Number of channel groups S in CAE = not reported
    The latent is split into S groups along the channel dimension; S is never specified and directly affects both BD-Rate and decode latency.
assumptions (5)
  • domain assumption Learned image compression pipeline with analysis transform, quantization, and synthesis transform (Eqn. 1).
    Adopted from Balle et al. and prior LIC works; not proven here.
  • domain assumption Latent representation follows a Gaussian distribution with parameters (mean, scale) estimated by a hyperprior and context (Eqn. 2 and 3).
    Standard in learned image compression (Minnen et al., Cheng et al.); the paper builds on this framework.
  • standard math SSM discretization via zero-order hold (Eqn. 5).
    Standard discretization of continuous-time state space models, used to integrate SSMs into deep networks.
  • domain assumption VSS/VMamba block with SS2D four-direction scanning models global content in linear time.
    Taken from VMamba [32]; the paper relies on this property for efficiency and global receptive field.
  • domain assumption Channel-wise autoregressive entropy modeling can be done efficiently enough for real-time decoding.
    Adopted from ChARM [17]; the paper assumes SSM-based spatial modeling does not reintroduce significant latency.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CMamba: Learned Image Compression with State Space Models." pith.science (2026). https://pith.science/paper/SRBGBLVK

@misc{pith2026250204988,
  author       = {Pith},
  title        = {Pith review of: CMamba: Learned Image Compression with State Space Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SRBGBLVK}},
  note         = {Machine review of arXiv:2502.04988}
}
read the original abstract

Learned Image Compression (LIC) has explored various architectures, such as Convolutional Neural Networks (CNNs) and transformers, in modeling image content distributions in order to achieve compression effectiveness. However, achieving high rate-distortion performance while maintaining low computational complexity (\ie, parameters, FLOPs, and latency) remains challenging. In this paper, we propose a hybrid Convolution and State Space Models (SSMs) based image compression framework, termed \textit{CMamba}, to achieve superior rate-distortion performance with low computational complexity. Specifically, CMamba introduces two key components: a Content-Adaptive SSM (CA-SSM) module and a Context-Aware Entropy (CAE) module. First, we observed that SSMs excel in modeling overall content but tend to lose high-frequency details. In contrast, CNNs are proficient at capturing local details. Motivated by this, we propose the CA-SSM module that can dynamically fuse global content extracted by SSM blocks and local details captured by CNN blocks in both encoding and decoding stages. As a result, important image content is well preserved during compression. Second, our proposed CAE module is designed to reduce spatial and channel redundancies in latent representations after encoding. Specifically, our CAE leverages SSMs to parameterize the spatial content in latent representations. Benefiting from SSMs, CAE significantly improves spatial compression efficiency while reducing spatial content redundancies. Moreover, along the channel dimension, CAE reduces inter-channel redundancies of latent representations via an autoregressive manner, which can fully exploit prior knowledge from previous channels without sacrificing efficiency. Experimental results demonstrate that CMamba achieves superior rate-distortion performance.

Figures

Figures reproduced from arXiv: 2502.04988 by the authors.

Figure 1
Figure 1. The Fourier spectrum comparisons between SSMs and CNNs. (a) The Fourier spectrum of features obtained from the SSM-based method [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) Overview of our proposed method. (b) Detailed design of our proposed Content-Adaptive SSM (CA-SSM) module. The CA-SSM module has two [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. PSNR-Bitrate curves evaluated on Kodak, Tecnick, and CLIC datasets. The compared methods include state-of-the-art LIC models and handcrafted [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visual comparison of the decompressed kodim24.png image from the Kodak dataset using various compression methods. Opt.MSE and Opt.MS-SSIM indicate that a model is optimized with MSE and MS-SSIM, respectively. More visual comparisons are provided in the supplementary ma…
Figure 5
Figure 5. Figure 5: Rate-distortion performance evaluated on the Kodak dataset. All the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The spatial correlation map of (y − µ)/σ with models trained at λ = 0.013. The value with index (i, j) corresponds to the normalized cross￾correlation of latent representation at spatial locations (w, h) and (w +i, h+ j), averaged across all latent elements of all imag…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 58 canonical work pages

  1. [1]

    The jpeg still picture compression standard,

    G. K. Wallace, “The jpeg still picture compression standard,” Commu- nications of the ACM , vol. 34, no. 4, pp. 30–44, 1991

    work page 1991

  2. [2]

    Bpg image format,

    F. Bellard, “Bpg image format,” 2018, available at: https://bellard.org/ bpg/

    work page 2018

  3. [3]

    Versatile video coding,

    B. Benjamin, C. Jianle, L. Shan, and W. Ye-Kui, “Versatile video coding,” in JVET, 2020, p. 1

    work page 2020

  4. [4]

    End-to-end optimized image compression,

    J. Ball ´e, V . Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,” in ICLR, 2017

    work page 2017

  5. [5]

    Variable-rate deep image compression through spatially-adaptive feature transform,

    M. Song, J. Choi, and B. Han, “Variable-rate deep image compression through spatially-adaptive feature transform,” in Proc. of ICCV , 2021, pp. 2380–2389

    work page 2021

  6. [6]

    Asymmetric gained deep image compression with continuous rate adaptation,

    Z. Cui, J. Wang, S. Gao, T. Guo, Y . Feng, and B. Bai, “Asymmetric gained deep image compression with continuous rate adaptation,” in Proc. of the IEEE Conf. on CVPR , 2021, pp. 10 532–10 541

    work page 2021

  7. [7]

    End-to-end optimized versatile image compression with wavelet-like transform,

    H. Ma, D. Liu, N. Yan, H. Li, and F. Wu, “End-to-end optimized versatile image compression with wavelet-like transform,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 44, no. 3, pp. 1247–1263, 2022

    2022

  8. [8]

    Towards efficient image compression without autoregressive models,

    M. S. Ali, Y . Kim, M. Qamar, S.-C. Lim, D. Kim, C. Zhang, S.-H. Bae, and H. Y . Kim, “Towards efficient image compression without autoregressive models,” in NeurIPS, 2023

    work page 2023

Show all 76 references
  1. [9]

    Lossy image compression with compressive autoencoders,

    L. Theis, W. Shi, A. Cunningham, and F. Husz ´ar, “Lossy image compression with compressive autoencoders,” in ICLR, 2017

    2017

  2. [10]

    Conditional probability models for deep image compression,

    F. Mentzer, E. Agustsson, M. Tschannen, R. Timofte, and L. Van Gool, “Conditional probability models for deep image compression,” in Proc. of the IEEE Conf. on CVPR , 2018, pp. 4394–4402

    2018

  3. [11]

    Efficient and effective context-based convolutional entropy modeling for image compression,

    M. Li, K. Ma, J. You, D. Zhang, and W. Zuo, “Efficient and effective context-based convolutional entropy modeling for image compression,” IEEE Trans. Image Process. , vol. 29, pp. 5900–5911, 2020

    2020

  4. [12]

    Enhanced standard compatible image compression framework based on auxiliary codec networks,

    H. Son, T. Kim, H. Lee, and S. Lee, “Enhanced standard compatible image compression framework based on auxiliary codec networks,” IEEE Trans. Image Process. , vol. 31, pp. 664–677, 2021

    2021

  5. [13]

    Dynamic neural network for lossy-to-lossless image coding,

    T. Dardouri, M. Kaaniche, A. Benazza-Benyahia, and J.-C. Pesquet, “Dynamic neural network for lossy-to-lossless image coding,” IEEE Trans. Image Process. , vol. 31, pp. 569–584, 2021

    2021

  6. [14]

    End-to-end optimized image compression with attention mechanism

    L. Zhou, Z. Sun, X. Wu, and J. Wu, “End-to-end optimized image compression with attention mechanism.” in CVPR workshops, 2019, p. 0

    2019

  7. [15]

    Learned image compression with mixed transformer-cnn architectures,

    J. Liu, H. Sun, and J. Katto, “Learned image compression with mixed transformer-cnn architectures,” in Proc. of the IEEE Conf. on CVPR , 2023, pp. 14 388–14 397

    2023

  8. [16]

    Checkerboard context model for efficient learned image compression,

    D. He, Y . Zheng, B. Sun, Y . Wang, and H. Qin, “Checkerboard context model for efficient learned image compression,” in Proc. of the IEEE Conf. on CVPR , 2021, pp. 14 771–14 780

    2021

  9. [17]

    Channel-wise autoregressive entropy models for learned image compression,

    D. Minnen and S. Singh, “Channel-wise autoregressive entropy models for learned image compression,” in IEEE International Conf. on Image Processing. IEEE, 2020, pp. 3339–3343

    2020

  10. [18]

    Learned image com- pression with discretized gaussian mixture likelihoods and attention modules,

    Z. Cheng, H. Sun, M. Takeuchi, and J. Katto, “Learned image com- pression with discretized gaussian mixture likelihoods and attention modules,” in Proc. of the IEEE Conf. on CVPR , 2020, pp. 7939–7948

    2020

  11. [19]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. D. M.-W. C. Kenton and L. K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Proc. of NAACL-HLT, 2019, pp. 4171–4186

    2019

  12. [20]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly et al. , “An image is worth 16x16 words: Transformers for image recognition at scale,” in ICLR, 2020

    2020

  13. [21]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proc. of ICCV , 2021, pp. 10 012–10 022

    2021

  14. [22]

    Transformer-based transform coding,

    Y . Zhu, Y . Yang, and T. Cohen, “Transformer-based transform coding,” in ICLR, 2022

    2022

  15. [23]

    The devil is in the details: Window- based attention for image compression,

    R. Zou, C. Song, and Z. Zhang, “The devil is in the details: Window- based attention for image compression,” in Proc. of the IEEE Conf. on CVPR, 2022, pp. 17 492–17 501. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 10

    2022

  16. [24]

    Frequency-aware transformer for learned image compression,

    H. Li, S. Li, W. Dai, C. Li, J. Zou, and H. Xiong, “Frequency-aware transformer for learned image compression,” in ICLR, 2024

    2024

  17. [25]

    End-to- end learnt image compression via non-local attention optimization and improved context modeling,

    T. Chen, H. Liu, Z. Ma, Q. Shen, X. Cao, and Y . Wang, “End-to- end learnt image compression via non-local attention optimization and improved context modeling,” IEEE Trans. Image Process. , vol. 30, pp. 3179–3191, 2021

    2021

  18. [26]

    Joint autoregressive and hierarchical priors for learned image compression,

    D. Minnen, J. Ball ´e, and G. D. Toderici, “Joint autoregressive and hierarchical priors for learned image compression,” NeurIPS, vol. 31, 2018

    2018

  19. [27]

    Entroformer: A transformer-based entropy model for learned image compression,

    Y . Qian, X. Sun, M. Lin, Z. Tan, and R. Jin, “Entroformer: A transformer-based entropy model for learned image compression,” in ICLR, 2022

    2022

  20. [28]

    Mlic: Multi- reference entropy model for learned image compression,

    W. Jiang, J. Yang, Y . Zhai, P. Ning, F. Gao, and R. Wang, “Mlic: Multi- reference entropy model for learned image compression,” in Proc. of ACM MM, 2023, pp. 7618–7627

    2023

  21. [29]

    Contextformer: A transformer with spatio-channel attention for context modeling in learned image compression,

    A. B. Koyuncu, H. Gao, A. Boev, G. Gaikov, E. Alshina, and E. Stein- bach, “Contextformer: A transformer with spatio-channel attention for context modeling in learned image compression,” in ECCV. Springer, 2022, pp. 447–463

    2022

  22. [30]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv:2312.00752, 2023

    2023 arXiv

  23. [31]

    Vision mamba: Efficient visual representation learning with bidirectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,” in ICML, 2024

    2024

  24. [32]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, and Y . Liu, “Vmamba: Visual state space model,” NeurIPS, 2025

    2025

  25. [33]

    How do vision transformers work?

    N. Park and S. Kim, “How do vision transformers work?” in ICLR, 2021

    2021

  26. [34]

    Kodak lossless true color image suite,

    R. Franzen, “Kodak lossless true color image suite,” 1999

    1999

  27. [35]

    Testimages: a large-scale archive for testing visual devices and basic image processing algorithms

    N. Asuni and A. Giachetti, “Testimages: a large-scale archive for testing visual devices and basic image processing algorithms.” in STAG, 2014, pp. 63–70

    2014

  28. [36]

    Clic, workshop and challenge on learned image compression,

    L. Theis and G. Toderici, “Clic, workshop and challenge on learned image compression,” in Proc. of the IEEE Conf. on CVPR , 2021

    2021

  29. [37]

    Mlic++: Linear complexity multi-reference entropy modeling for learned image compression,

    W. Jiang and R. Wang, “Mlic++: Linear complexity multi-reference entropy modeling for learned image compression,” in ICML 2023 Workshop Neural Compression , 2023

    2023

  30. [38]

    Lc-fdnet: Learned lossless image compression with frequency decomposition network,

    H. Rhee, Y . I. Jang, S. Kim, and N. I. Cho, “Lc-fdnet: Learned lossless image compression with frequency decomposition network,” in Proc. of the IEEE Conf. on CVPR , 2022, pp. 6033–6042

    2022

  31. [39]

    Dpict: Deep progressive image compression using trit-planes,

    J.-H. Lee, S. Jeon, K. P. Choi, Y . Park, and C.-S. Kim, “Dpict: Deep progressive image compression using trit-planes,” in Proc. of the IEEE Conf. on CVPR , 2022, pp. 16 113–16 122

    2022

  32. [40]

    Learned image compression with gaussian-laplacian- logistic mixture model and concatenated residual modules,

    H. Fu, F. Liang, J. Lin, B. Li, M. Akbari, J. Liang, G. Zhang, D. Liu, C. Tu, and J. Han, “Learned image compression with gaussian-laplacian- logistic mixture model and concatenated residual modules,” IEEE Trans. Image Process., vol. 32, pp. 2063–2076, 2023

    2023

  33. [41]

    Enhanced invertible encoding for learned image compression,

    Y . Xie, K. L. Cheng, and Q. Chen, “Enhanced invertible encoding for learned image compression,” in Proc. of ACM MM , 2021, pp. 162–170

    2021

  34. [42]

    Lossy image compression with conditional diffusion models,

    R. Yang and S. Mandt, “Lossy image compression with conditional diffusion models,” NeurIPS, vol. 36, 2024

    2024

  35. [43]

    Mambavc: Learned visual compression with selective state spaces,

    S. Qin, J. Wang, Y . Zhou, B. Chen, T. Luo, B. An, T. Dai, S. Xia, and Y . Wang, “Mambavc: Learned visual compression with selective state spaces,” arXiv:2405.15413, 2024

    2024 arXiv

  36. [44]

    Variational image compression with a scale hyperprior,

    J. Ball ´e, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,” arXiv:1802.01436, 2018

    2018 arXiv

  37. [45]

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

    D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y . Wang, “Elic: Efficient learned image compression with unevenly grouped space- channel contextual adaptive coding,” in Proc. of the IEEE Conf. on CVPR, 2022, pp. 5718–5727

    2022

  38. [46]

    Efficient contextformer: Spatio-channel window attention for fast context mod- eling in learned image compression,

    A. B. Koyuncu, P. Jia, A. Boev, E. Alshina, and E. Steinbach, “Efficient contextformer: Spatio-channel window attention for fast context mod- eling in learned image compression,” IEEE Trans. Circuits Syst. Video Technol., 2024

    2024

  39. [47]

    Hippo: Recurrent memory with optimal polynomial projections,

    A. Gu, T. Dao, S. Ermon, A. Rudra, and C. R ´e, “Hippo: Recurrent memory with optimal polynomial projections,” NeurIPS, vol. 33, pp. 1474–1487, 2020

    2020

  40. [48]

    Combining recurrent, convolutional, and continuous-time models with linear state space layers,

    A. Gu, I. Johnson, K. Goel, K. Saab, T. Dao, A. Rudra, and C. R ´e, “Combining recurrent, convolutional, and continuous-time models with linear state space layers,” NeurIPS, vol. 34, pp. 572–585, 2021

    2021

  41. [49]

    It’s raw! audio generation with state-space models,

    K. Goel, A. Gu, C. Donahue, and C. R ´e, “It’s raw! audio generation with state-space models,” in ICML. PMLR, 2022, pp. 7616–7633

    2022

  42. [50]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. Re, “Efficiently modeling long sequences with structured state spaces,” in ICLR, 2021

    2021

  43. [51]

    On the parameterization and initialization of diagonal state space models,

    A. Gu, K. Goel, A. Gupta, and C. R ´e, “On the parameterization and initialization of diagonal state space models,” NeurIPS, vol. 35, pp. 35 971–35 983, 2022

    2022

  44. [52]

    Diagonal state spaces are as effective as structured state spaces,

    A. Gupta, A. Gu, and J. Berant, “Diagonal state spaces are as effective as structured state spaces,” NeurIPS, vol. 35, pp. 22 982–22 994, 2022

    2022

  45. [53]

    Simplified state space layers for sequence modeling,

    J. T. Smith, A. Warrington, and S. Linderman, “Simplified state space layers for sequence modeling,” in ICLR, 2022

    2022

  46. [54]

    Liquid structural state-space models,

    R. Hasani, M. Lechner, T.-H. Wang, M. Chahine, A. Amini, and D. Rus, “Liquid structural state-space models,” in ICLR, 2022

    2022

  47. [55]

    Long range language modeling via gated state spaces,

    H. Mehta, A. Gupta, A. Cutkosky, and B. Neyshabur, “Long range language modeling via gated state spaces,” in ICLR, 2023

    2023

  48. [56]

    Lo- calmamba: Visual state space model with windowed selective scan,

    T. Huang, X. Pei, S. You, F. Wang, C. Qian, and C. Xu, “Lo- calmamba: Visual state space model with windowed selective scan,” arXiv:2403.09338, 2024

    2024 arXiv

  49. [57]

    Mambair: A simple baseline for image restoration with state-space model,

    H. Guo, J. Li, T. Dai, Z. Ouyang, X. Ren, and S.-T. Xia, “Mambair: A simple baseline for image restoration with state-space model,” in ECCV. Springer, 2025, pp. 222–241

    2025

  50. [58]

    Activating wider areas in image super- resolution,

    C. Cheng, H. Wang, and H. Sun, “Activating wider areas in image super- resolution,” arXiv:2403.08330, 2024

    2024 arXiv

  51. [59]

    Cu-mamba: Selective state space models with channel learning for image restoration,

    R. Deng and T. Gu, “Cu-mamba: Selective state space models with channel learning for image restoration,” arXiv:2404.11778, 2024

    2024 arXiv

  52. [60]

    Vmambair: Visual state space model for image restoration,

    Y . Shi, B. Xia, X. Jin, X. Wang, T. Zhao, X. Xia, X. Xiao, and W. Yang, “Vmambair: Visual state space model for image restoration,” arXiv:2403.11423, 2024

    2024 arXiv

  53. [61]

    3dmambacomplete: Exploring structured state space model for point cloud completion,

    Y . Li, W. Yang, and B. Fei, “3dmambacomplete: Exploring structured state space model for point cloud completion,” arXiv:2404.07106, 2024

    2024 arXiv

  54. [62]

    Pointmamba: A simple state space model for point cloud analysis,

    D. Liang, X. Zhou, W. Xu, X. Zhu, Z. Zou, X. Ye, X. Tan, and X. Bai, “Pointmamba: A simple state space model for point cloud analysis,” in NeurIPS, 2024

    2024

  55. [63]

    Point mamba: A novel point cloud backbone based on state space model with octree-based ordering strategy,

    J. Liu, R. Yu, Y . Wang, Y . Zheng, T. Deng, W. Ye, and H. Wang, “Point mamba: A novel point cloud backbone based on state space model with octree-based ordering strategy,” arXiv:2403.06467, 2024

    2024 arXiv

  56. [64]

    Point could mamba: Point cloud learning via state space model,

    T. Zhang, X. Li, H. Yuan, S. Ji, and S. Yan, “Point could mamba: Point cloud learning via state space model,” arXiv:2403.00762, 2024

    2024 arXiv

  57. [65]

    Video mamba suite: State space model as a versatile alternative for video understanding,

    G. Chen, Y . Huang, J. Xu, B. Pei, Z. Chen, Z. Li, J. Wang, K. Li, T. Lu, and L. Wang, “Video mamba suite: State space model as a versatile alternative for video understanding,” arXiv:2403.09626, 2024

    2024 arXiv

  58. [66]

    Videomamba: State space model for efficient video understanding,

    K. Li, X. Li, Y . Wang, Y . He, Y . Wang, L. Wang, and Y . Qiao, “Videomamba: State space model for efficient video understanding,” in ECCV. Springer, 2025, pp. 237–255

    2025

  59. [67]

    Rhythmmamba: Fast remote phys- iological measurement with arbitrary length videos,

    B. Zou, Z. Guo, X. Hu, and H. Ma, “Rhythmmamba: Fast remote phys- iological measurement with arbitrary length videos,” arXiv:2404.06483, 2024

    2024 arXiv

  60. [68]

    U-mamba: Enhancing long-range depen- dency for biomedical image segmentation,

    J. Ma, F. Li, and B. Wang, “U-mamba: Enhancing long-range depen- dency for biomedical image segmentation,” arXiv:2401.04722, 2024

    2024 arXiv

  61. [69]

    Medmamba: Vision mamba for medical image classification,

    Y . Yue and Z. Li, “Medmamba: Vision mamba for medical image classification,” arXiv:2403.03849, 2024

    2024 arXiv

  62. [70]

    Semi-mamba-unet: Pixel-level contrastive and pixel-level cross-supervised visual mamba-based unet for semi- supervised medical image segmentation,

    C. Ma and Z. Wang, “Semi-mamba-unet: Pixel-level contrastive and pixel-level cross-supervised visual mamba-based unet for semi- supervised medical image segmentation,” arXiv prints , pp. arXiv–2402, 2024

    2024

  63. [71]

    J. P. Hespanha, Linear systems theory. Princeton university press, 2018

    2018

  64. [72]

    Searching for activation functions,

    P. Ramachandran, B. Zoph, and Q. V . Le, “Searching for activation functions,” arXiv:1710.05941, 2017

    2017 arXiv

  65. [73]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. of the IEEE Conf. on CVPR , 2016, pp. 770–778

    2016

  66. [74]

    Openim- ages: A public dataset for large-scale multi-label and multi-class image classification,

    I. Krasin, T. Duerig, N. Alldrin, V . Ferrari, S. Abu-El-Haija, A. Kuznetsova, H. Rom, J. Uijlings, S. Popov, A. Veit et al. , “Openim- ages: A public dataset for large-scale multi-label and multi-class image classification,” Dataset available from https://github. com/openimag...

    2017

  67. [75]

    Adam: a method for stochastic optimization,

    D. Kingma, “Adam: a method for stochastic optimization,” in ICLR, 2015

    2015

  68. [76]

    Video quality evaluation methodology and verification testing of hevc compression performance,

    T. K. Tan, R. Weerakkody, M. Mrak, N. Ramzan, V . Baroncini, J.- R. Ohm, and G. J. Sullivan, “Video quality evaluation methodology and verification testing of hevc compression performance,” IEEE Trans. Circuits Syst. Video Technol. , vol. 26, no. 1, pp. 76–90, 2015

    2015

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.