Pith. sign in

REVIEW 4 major objections 7 minor 51 references

Linear Attention Modeling for Learned Image Compression

T0 review · 4 major / 7 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 A linear-attention image codec beats VTM-9.1 by 15% to 18% in BD-rate on three datasets.

desk verdict First RWKV-based image compression codec with plausible RD gains; the efficiency claim rests on shaky FLOPs measurement. read the letter →

arxiv 2502.05741 v2 pith:ZXPAXYND submitted 2025-02-09 cs.CV

Donghui Feng , Zhengxue Cheng , Shen Wang , Ronghua Wu , Hongwei Hu , Guo Lu , Li Song This is my paper · ORCID
classification cs.CV
keywords learnedimagecompressionlinearattentionRWKVBiWKVentropymodelrate-distortionoptimizationlow-complexitycodingspatial-channelcontext
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

This paper argues that linear-attention architectures, specifically RWKV-style bidirectional attention, can replace quadratic self-attention in learned image compression without a rate-distortion penalty. The proposed LALIC codec uses Bi-RWKV blocks as the nonlinear transform backbone and an RWKV-based spatial-channel context entropy model, obtaining a global receptive field at computational cost that grows linearly with pixel count. On Kodak, CLIC, and Tecnick, it reportedly outperforms the VVC reference encoder VTM-9.1 by 15.26%, 15.41%, and 17.63% in BD-rate, with a decoding time of 150 ms on an RTX 4090. If these results hold, the work matters because it provides a practical route to high-resolution learned codecs that keep long-range modeling without the quadratic cost of transformers.

What carries the argument

The central object is the BiWKV attention mechanism, a bidirectional form of RWKV's WKV attention: it computes a weighted average of value vectors, with the current token amplified by a learnable vector and other tokens decayed exponentially according to their distance, all in time linear in sequence length. The Bi-RWKV block wraps this mechanism in a Spatial-Mix branch (with an Omni-Shift reparameterized depthwise convolution to inject local 2D context) and a Channel-Mix branch (a gated MLP with squared ReLU), and this block is reused both inside the analysis and synthesis transforms and inside the RWKV-SCCTX entropy model. The BiWKV attention is what carries the argument: it gives a global effective receptive field while keeping the operation count proportional to $L \cdot D$ rather than $L^2$, which is the efficiency claim that separates this codec from window-based transformers.

What would settle it

Run an independent profiler that counts every tensor operation in the released model across increasing resolutions, or time a 4K decode: if FLOPs or latency grow super-linearly with pixel count, the linear-complexity claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that replacing quadratic attention with linear bidirectional WKV attention in both the transform network and the entropy model yields state-of-the-art rate-distortion performance at moderate complexity. The paper introduces Bi-RWKV blocks that combine a Spatial-Mix branch, which uses a reparameterized 5x5 depthwise convolution (Omni-Shift) for local 2D context and BiWKV attention for global dependency, with a Channel-Mix branch for cross-channel fusion. It also proposes RWKV-SCCTX, an entropy model that predicts Gaussian parameters using checkerboard spatial context and chunked channel-wise context processed by the same Bi-RWKV blocks. The reported experiments show LALIC surpassing VTM-9.1 by about 15 to 18 percent in BD-rate on Kodak, CLIC, and Tecnick, and matching or beating recent learned codecs including ELIC, TCM-large, FAT, MLIC++, and MambaVC while keeping decoding time at 150 ms and forward FLOPs at 286 G.

Load-bearing premise

The complexity advantage is measured with a FLOPs counter that cannot see attention operations, matrix multiplications, or CUDA kernels; if the true FLOPs of BiWKV are much higher than the theoretical linear count, the paper's efficiency claim weakens.

Editorial extensions

If this is right

  • Learned image codecs can obtain a near-global receptive field with linear complexity in pixel count, removing the need for window partitioning or selective-scan overhead.
  • The RWKV-based entropy model improves rate-distortion performance over convolutional spatial-channel context models of similar parameter count, per the ablation study.
  • The reported BD-rate gains over VTM-9.1 suggest learned linear-attention codecs are now a practical alternative to traditional codecs on high-resolution content.
  • The architecture's moderate decoding latency (150 ms) makes near-real-time decoding of high-resolution images plausible on a consumer GPU.
  • The chunked channel-wise context design, allocating fewer channels to the most-referenced early slices, offers a general recipe for efficient autoregressive entropy models.

Reading between the lines

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

  • If the paper's FLOPs measurements undercount attention operations, as it itself notes, an independent profiler that counts every matrix multiplication could rank BiWKV's true cost higher than reported relative to its baselines.
  • The Bi-RWKV block is a drop-in component that could be transferred to other pixel-wise tasks such as video compression, super-resolution, or image restoration, where long-range dependencies and low decoding latency both matter.
  • RWKV-SCCTX only changes how context is mixed, not how the prior is parameterized, so it can likely be combined with latent residual prediction or conditional priors for further gains.
  • The paper's scaling plots imply a testable prediction: decode latency and FLOPs should grow roughly linearly with pixel count; measuring them at 4K resolution against the reported 1024x1024 numbers would confirm or refute that scaling.
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 / 7 minor

Summary. The paper proposes LALIC, a learned image compression model that uses Bi-RWKV linear-attention blocks in both the nonlinear transforms and the entropy model, together with a RWKV-based spatial-channel context model. The authors report BD-rate gains over VTM-9.1 of -15.26%, -15.41%, and -17.63% on Kodak, CLIC, and Tecnick, respectively, and claim state-of-the-art performance while maintaining moderate decoding latency (150 ms) on an RTX 4090. The paper includes extensive rate-distortion experiments, ablations, ERF visualizations, and a public code repository.

Significance. If the reported rate-distortion and complexity results hold, this is a meaningful contribution: it is, to my knowledge, the first application of RWKV-style linear attention to learned image compression, and it shows that a global-receptive-field linear-attention backbone can match or beat strong transformer- and Mamba-based codecs on standard benchmarks. The public code, multi-dataset evaluation, and ablation studies are valuable. However, the complexity advantage is currently supported by FLOPs counts that exclude the attention mechanism itself, and there are inconsistencies in the headline BD-rate numbers. The efficiency contribution is therefore not yet firmly established, even though the RD evidence is substantial.

major comments (4)
  1. [Section 4.3 and Supplementary A, Table 1] The FLOPs comparison is not a valid quantitative measure of the paper's central complexity claim. Supp. A states that the thop package cannot count FLOPs from matrix multiplications, mathematical functions, or CUDA-specific implementations, yet the BiWKV attention in Eq. (6) consists almost entirely of such operations (exponentials, elementwise products, a weighted sum over all tokens, and a division). The reported 286.16 GFLOPs for LALIC therefore excludes the very attention mechanism that is the paper's core contribution, and the comparison against MLIC++ and TCM-large is not conclusive. The authors should provide FLOPs measured with a profiler that includes all operations, or derive a rigorous upper/lower bound for the omitted attention cost.
  2. [Section 4.3, Table 1 (FAT row)] The reported FAT decoding time of >10 s is inconsistent with the 242 ms reported in the FAT paper, and Supp. A attributes the discrepancy to an unresolved issue in the official GitHub repository. Reporting a >10 s runtime as a scientific comparison point is misleading unless the same inference setup, including code optimizations and environment, is used for all methods. Either the authors should obtain a representative FAT runtime (for example, from the original authors or a properly optimized implementation) or they should exclude the FAT runtime from the comparison table and instead discuss the issue in the text.
  3. [Abstract and Section 5 (Conclusion)] The headline BD-rate results are internally inconsistent. The abstract reports -15.26%, -15.41%, and -17.63% BD-rate against VTM-9.1 on Kodak, CLIC, and Tecnick, while Section 5 states 'outperforming VTM-9.1 by -14.84% in BD-rate on Kodak, Tecnick and CLIC Professional validation datasets.' The paper must clarify which number is correct, whether -14.84% is an average, and correct the erroneous statement. Since these numbers are the paper's main quantitative claim, this inconsistency affects reproducibility and must be fixed.
  4. [Section 4.2, Table 1] The claim that LALIC achieves state-of-the-art performance rests on BD-rate differences from MLIC++ of 0.24 percentage points on Kodak, 0.96 on CLIC, and 0.42 on Tecnick. No error bars, standard deviations, or statistical tests are provided. In learned image compression, differences below 0.5% are typically within run-to-run and test-set variance. The authors should report confidence intervals (for example via bootstrap over images) or a paired significance test to support the SOTA claim over MLIC++.
minor comments (7)
  1. [Section 1, Contribution bullets] The first contribution bullet says 'Spatial-Mix and Time-Mix' but Section 3.2 only defines Spatial Mix and Channel Mix; please correct the terminology.
  2. [Table 2] The table header says 'BD-rate' but the text states 'we use inference loss as a proxy'; please align the metric name with the reported values.
  3. [Table 3] The note that DeltaFLOPs 'only counts the operations of attention layer or shift layer' is useful, but it should be stated in the main text that these are analytical counts, not measured FLOPs, given that the thop package cannot count these operations.
  4. [References] References [9] and [10] appear to refer to the same Vision-RWKV paper; please merge or disambiguate them.
  5. [Figures 3 and 9] The ERF visualizations lack a color bar and a description of normalization; adding these would make the comparison more interpretable.
  6. [Table 1] The table does not specify the measurement protocol for encoding/decoding times (e.g., GPU/CPU, batch size, image resolution, number of runs). Please provide this information for reproducibility.
  7. [Section 4.1] The training dataset description 'the first 400,000 images of the OpenImages dataset' is ambiguous; please specify the exact split or provide a link if this subset is public.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RD and complexity claims are direct empirical measurements anchored to external codecs and external architecture sources, not derived by construction from fitted parameters or self-citations.

full rationale

The paper's central claims are empirical and self-contained. The BD-rate figures (-15.26%, -15.41%, -17.63% relative to VTM-9.1) are direct measurements or adopted R-D curves from external baselines, not quantities reconstructed from the model's own fitted parameters. The architecture is an application of Bi-RWKV blocks and an RWKV-based context model to learned image compression; the BiWKV attention formula in Eq. (6) is credited to Vision-RWKV [9], and the Omni-Shift module to Restore-RWKV [39], both external works with no author overlap with the present paper, so no load-bearing argument reduces to a self-citation. The entropy-model comparison (Conv vs. Conv-Plus vs. RWKV SCCTX, Table 2) is an ablation, and the claim that increasing context parameters helps is supported by an external reference [34], not by the paper's own prior result. The only caveat is in Supp. A, where the authors admit that the thop FLOPs counter excludes matrix multiplications, attention operations, and CUDA-specific kernels; this is a measurement limitation that weakens the precision of the complexity comparison, but it is not circularity because the reported FLOPs and latency are still independently measured quantities rather than quantities forced by definition or by the paper's own assumptions. No fitted input is relabeled as a prediction, no uniqueness theorem is imported from the authors' prior work, and no known result is merely renamed. The derivation chain is therefore not circular.

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

The paper's contribution is an empirical architecture; its central claims rest on hand-selected hyperparameters, standard LIC assumptions (conditional Gaussian entropy, hyperprior), and a FLOPs measurement tool that the authors admit undercounts certain operations. No new physical entities are introduced.

free parameters (4)
  • Channel allocation in RWKV-SCCTX = {16, 16, 32, 64, M-128}
    Hand-set slice sizes in Sec 3.3; they determine the conditional modeling precision and are central to the entropy model's bitrate.
  • Bi-RWKV layer configuration = {2, 4, 6, 6}
    Chosen in Sec 4.1; ablation in Table 2 shows BD-rate depends strongly on this.
  • Latent channel dimensions M and N = M=320, N=192
    Set in Sec 4.1; direct impact on rate-distortion and complexity.
  • Rate-distortion Lagrange multipliers = MSE: {0.0025, 0.0035, 0.0067, 0.0130, 0.0250, 0.0483}; MS-SSIM: {2.40, 4.58, 8.73, 16.64, 31.73, 60.50}
    Used to train six points on the R-D curve; the reported BD-rate numbers depend on these operating points.
assumptions (4)
  • domain assumption Conditional Gaussian entropy model for latent y
    Standard in learned compression (Sec 3.1, after Eq. 2).
  • domain assumption OpenImages first 400k images provide a representative training distribution
    Training set chosen in Sec 4.1; RD performance on Kodak/CLIC/Tecnick depends on it.
  • ad hoc to paper The thop FLOPs counter provides a fair approximation of true computational cost
    Authors acknowledge it omits attention and CUDA ops (Supp. A); the low-complexity framing relies on these numbers.
  • ad hoc to paper The BiWKV implementation achieves the theoretical linear complexity of Eq. 6
    No proof or benchmark of actual scaling is provided; Table 5 cites Vision-RWKV's GitHub count.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Linear Attention Modeling for Learned Image Compression." pith.science (2026). https://pith.science/paper/ZXPAXYND

@misc{pith2026250205741,
  author       = {Pith},
  title        = {Pith review of: Linear Attention Modeling for Learned Image Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZXPAXYND}},
  note         = {Machine review of arXiv:2502.05741}
}
read the original abstract

Recent years, learned image compression has made tremendous progress to achieve impressive coding efficiency. Its coding gain mainly comes from non-linear neural network-based transform and learnable entropy modeling. However, most studies focus on a strong backbone, and few studies consider a low complexity design. In this paper, we propose LALIC, a linear attention modeling for learned image compression. Specially, we propose to use Bi-RWKV blocks, by utilizing the Spatial Mix and Channel Mix modules to achieve more compact feature extraction, and apply the Conv based Omni-Shift module to adapt to two-dimensional latent representation. Furthermore, we propose a RWKV-based Spatial-Channel ConTeXt model (RWKV-SCCTX), that leverages the Bi-RWKV to modeling the correlation between neighboring features effectively. To our knowledge, our work is the first work to utilize efficient Bi-RWKV models with linear attention for learned image compression. Experimental results demonstrate that our method achieves competitive RD performances by outperforming VTM-9.1 by -15.26%, -15.41%, -17.63% in BD-rate on Kodak, CLIC and Tecnick datasets. The code is available at https://github.com/sjtu-medialab/RwkvCompress .

Figures

Figures reproduced from arXiv: 2502.05741 by the authors.

Figure 1
Figure 1. BD-rate vs Decoding Latency on Kodak dataset, where [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) Overview of proposed Linear Attention based Learned Image Compression (LALIC). Conv [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The effective receptive field (ERF) [23] visualization for the forward pass (ga & ha) of different models. A more exten￾sively distributed dark area indicates a larger ERF. WKV (BiWKV) attention [9] for the t-th token is defined as: wkvt = PT i=1,i̸=t e −(|t−i|−1)/T·w+ki vi + e u+kt vt PT i=1,i̸=t e−(|t−i|−1)/T·w+ki + e u+kt (6) where ki and vi represent the key and value for the i-th to￾ken. The introduced paramete… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Diagram of the RWKV Spatial-Channel Context Model. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Rate-distortion performance on the Kodak dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Rate-distortion performance on the CLIC dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: Effectiveness of the RWKV-based entropy model in improving compression efficiency. The middle two columns present the [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Visualization of different attention mechanisms. The [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Linear scaling trends of FLOPs (a) and GPU memory [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Network architecture of Conv SCCTX and Conv Plus [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Subjective quality comparison on the kodim01 image from Kodak. (a) Original (b) TCM-large 0.245 bpp / 34.20 dB (c) LALIC (Ours) 0.243 bpp / 34.30 dB (d) Original crop (e) TCM-large crop (f) LALIC (Ours) crop [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Subjective quality comparison on the kodim02 image from Kodak. of [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 39 canonical work pages

  1. [1]

    Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici

    Johannes Balle, Philip A. Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear Transform Coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339– 353, 2021. 1

    work page 2021

  2. [2]

    Simoncelli

    Johannes Ball ´e, Valero Laparra, and Eero P. Simoncelli. End-to-end optimized image compression. In 5th Interna- tional Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Pro- ceedings. OpenReview.net, 2017. 2

    work page 2017

  3. [3]

    Variational image compression with a scale hyperprior, 2018

    Johannes Ball ´e, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior, 2018. 2, 5

    work page 2018

  4. [4]

    Sullivan, and Jens-Rainer Ohm

    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 applica- tions. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736–3764, 2021. 1

    work page 2021

  5. [5]

    CompressAI: A PyTorch library and evaluation platform for end-to-end compression research, 2020

    Jean B ´egaint, Fabien Racap ´e, Simon Feltman, and Akshay Pushparaja. CompressAI: A PyTorch library and evaluation platform for end-to-end compression research, 2020. 1

    work page 2020

  6. [6]

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

    Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. ArXiv, abs/2403.09626,

    arXiv

  7. [7]

    Learned Image Compression With Discretized Gaus- sian Mixture Likelihoods and Attention Modules

    Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned Image Compression With Discretized Gaus- sian Mixture Likelihoods and Attention Modules. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7939–7948, 2020. 1, 2, 5

    work page 2020

  8. [8]

    The evolution of rwkv: Advancements in effi- cient language modeling, 2024

    Akul Datta. The evolution of rwkv: Advancements in effi- cient language modeling, 2024. 2

    work page 2024

Show all 51 references
  1. [9]

    Vision-RWKV: Efficient and Scalable Visual Per- ception with RWKV-Like Architectures, 2024

    Yuchen Duan, Weiyun Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Hongsheng Li, Jifeng Dai, and Wen- hai Wang. Vision-RWKV: Efficient and Scalable Visual Per- ception with RWKV-Like Architectures, 2024. 1, 3, 4, 2

    2024

  2. [10]

    Vision-rwkv: Efficient and scalable visual percep- tion with rwkv-like architectures

    Yuchen Duan, Weiyun Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Hongsheng Li, Jifeng Dai, and Wen- hai Wang. Vision-rwkv: Efficient and scalable visual percep- tion with rwkv-like architectures. ArXiv, abs/2403.02308,

    arXiv

  3. [11]

    Diffusion-rwkv: Scaling rwkv-like archi- tectures for diffusion models

    Zhengcong Fei, Mingyuan Fan, Changqian Yu, Debang Li, and Junshi Huang. Diffusion-rwkv: Scaling rwkv-like archi- tectures for diffusion models. ArXiv, abs/2404.04478, 2024. 3

    2024 arXiv

  4. [12]

    Mamba: Linear-time sequence mod- eling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence mod- eling with selective state spaces. ArXiv, abs/2312.00752,

    arXiv

  5. [13]

    Mambair: A simple baseline for im- age restoration with state-space model

    Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for im- age restoration with state-space model. In European Confer- ence on Computer Vision, 2024. 2

    2024

  6. [14]

    ELIC: Efficient Learned Image Com- pression 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 Com- pression with Unevenly Grouped Space-Channel Contextual Adaptive Coding. In 2022 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 5708– 5717...

    2022

  7. [15]

    Deform-mamba network for mri super-resolution

    Zexin Ji, Beiji Zou, Xiaoyan Kui, Pierre Vera, and Su Ruan. Deform-mamba network for mri super-resolution. ArXiv, abs/2407.05969, 2024. 2

    2024 arXiv

  8. [16]

    MLIC++: Linear Complexity Multi-Reference Entropy Modeling for Learned Image Compression

    Wei Jiang, Jiayu Yang, Yongqi Zhai, Feng Gao, and Rong- gang Wang. MLIC++: Linear Complexity Multi-Reference Entropy Modeling for Learned Image Compression. In Pro- ceedings of the 31st ACM International Conference on Mul- timedia, pages 7618–7627, 2023. 5, 7, 2

    2023

  9. [17]

    Burakhan Koyuncu, Han Gao, Atanas Boev, Georgii Gaikov, Elena Alshina, and Eckehard Steinbach

    A. Burakhan Koyuncu, Han Gao, Atanas Boev, Georgii Gaikov, Elena Alshina, and Eckehard Steinbach. Con- textformer: A Transformer with Spatio-Channel Attention for Context Modeling in Learned Image Compression. In Computer Vision – ECCV 2022 , pages 447–463. Springer Nature Swi...

    2022

  10. [18]

    Frequency-Aware Transformer for Learned Image Compression

    Han Li, Shaohui Li, Wenrui Dai, Chenglin Li, Junni Zou, and Hongkai Xiong. Frequency-Aware Transformer for Learned Image Compression. In The Twelfth International Confer- ence on Learning Representations . arXiv, 2023. 1, 2, 4, 5, 7

    2023

  11. [19]

    Learned Image Compression with Mixed Transformer-CNN Architectures

    Jinming Liu, Heming Sun, and Jiro Katto. Learned Image Compression with Mixed Transformer-CNN Architectures. In 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 14388–14397. IEEE, 2023. 1, 2, 4, 5, 7

    2023

  12. [20]

    Vmamba: Visual state space model

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. ArXiv, abs/2401.10166, 2024. 1, 2

    2024 arXiv

  13. [21]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In 2021 IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 9992–10002. IEEE, 2021. 1, 2

    2021

  14. [22]

    Transformer-based Image Compression

    Ming Lu, Peiyao Guo, Huiqing Shi, Chuntong Cao, and Zhan Ma. Transformer-based Image Compression. In 2022 Data Compression Conference (DCC), pages 469–469, Snowbird, UT, USA, 2022. IEEE. 2

    2022

  15. [23]

    Understanding the effective receptive field in deep convolu- tional neural networks

    Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolu- tional neural networks. Advances in neural information pro- cessing systems, 29, 2016. 4, 8

    2016

  16. [24]

    U-mamba: Enhancing long-range dependency for biomedical image segmentation

    Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. ArXiv, abs/2401.04722, 2024. 2

    2024 arXiv

  17. [25]

    Channel-Wise Autore- gressive Entropy Models for Learned Image Compression

    David Minnen and Saurabh Singh. Channel-Wise Autore- gressive Entropy Models for Learned Image Compression. In 2020 IEEE International Conference on Image Process- ing (ICIP), pages 3339–3343. IEEE, 2020. 1, 2, 7

    2020

  18. [26]

    Joint Autoregressive and Hierarchical Priors for Learned Image Compression

    David Minnen, Johannes Ball ´e, and George Toderici. Joint Autoregressive and Hierarchical Priors for Learned Image Compression. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Pro- cessing Systems 2018, NeurIPS 2018, 3-8 Decemb...

    2018

  19. [27]

    Wind, Stanislaw Wozniak, Ruichong Zhang, Zhenyuan Zhang, Qihang Zhao, Peng Zhou, Qinghua Zhou, Jian Zhu, and Rui-Jie Zhu

    Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, Kranthi Kiran GV , Xuzheng He, Haowen Hou, Jiaju Lin, Przemyslaw Kazienko, Jan Kocon, Jiaming Kong, Bartlomiej Kop- tyra, Hayden Lau, K...

    2023

  20. [28]

    Rwkv: Rein- venting rnns for the transformer era, 2023

    Bo Peng, Eric Alcaide, Quentin Anthony, et al. Rwkv: Rein- venting rnns for the transformer era, 2023. 2

    2023

  21. [29]

    Eagle and finch: Rwkv with matrix-valued states and dynamic recur- rence

    Bo Peng, Daniel Goldstein, Quentin Anthony, Alon Albalak, Eric Alcaide, Stella Biderman, Eugene Cheah, Teddy Fer- dinan, Haowen Hou, Przemys l aw Kazienko, G Kranthiki- ran, Jan Koco’n, Bartlomiej Koptyra, Satyapriya Krishna, Ronald McClelland, Niklas Muennighoff, Fares Obeid,...

    2024 arXiv

  22. [30]

    Entroformer: A Transformer-based Entropy Model for Learned Image Compression, 2022

    Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A Transformer-based Entropy Model for Learned Image Compression, 2022. 2

    2022

  23. [31]

    Vl-mamba: Ex- ploring state space models for multimodal learning

    Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba: Ex- ploring state space models for multimodal learning. ArXiv, abs/2403.13600, 2024. 2

    2024 arXiv

  24. [32]

    Mambavc: Learned visual compression with selective state spaces

    Shiyu Qin, Jinpeng Wang, Yimin Zhou, Bin Chen, Tianci Luo, Baoyi An, Tao Dai, Shu-Tao Xia, and Yaowei Wang. Mambavc: Learned visual compression with selective state spaces. ArXiv, abs/2405.15413, 2024. 1, 2, 7

    2024 arXiv

  25. [33]

    Temporal Context Mining for Learned Video Compression

    Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal Context Mining for Learned Video Compression. IEEE Transactions on Multimedia, 25:7311–7322, 2023. 2

    2023

  26. [34]

    NVC-1B: A Large Neural Video Coding Model, 2024

    Xihua Sheng, Chuanbo Tang, Li Li, Dong Liu, and Feng Wu. NVC-1B: A Large Neural Video Coding Model, 2024. 7

    2024

  27. [35]

    The JPEG 2000 still image compression standard

    Athanassios Skodras, Charilaos Christopoulos, and Touradj Ebrahimi. The JPEG 2000 still image compression standard. IEEE Signal processing magazine, 18(5):36–58, 2001. 1

    2000

  28. [36]

    Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand

    Gary J. Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (HEVC) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649–1668, 2012. 1

    2012

  29. [37]

    Gregory K. Wallace. The JPEG still picture compression standard. IEEE transactions on consumer electronics, 38(1): xviii–xxxiv, 1992. 1

    1992

  30. [38]

    Enhanced Invertible Encoding for Learned Image Compression

    Yueqi Xie, Ka Leong Cheng, and Qifeng Chen. Enhanced Invertible Encoding for Learned Image Compression. Pro- ceedings of the 29th ACM International Conference on Mul- timedia, pages 162–170, 2021. 7

    2021

  31. [39]

    Restore-RWKV: Efficient and Effective Medical Image Restoration with RWKV, 2024

    Zhiwen Yang, Hui Zhang, Dan Zhao, Bingzheng Wei, and Yan Xu. Restore-RWKV: Efficient and Effective Medical Image Restoration with RWKV, 2024. 2, 3

    2024

  32. [40]

    Restore-rwkv: Efficient and effective medical image restoration with rwkv

    Zhiwen Yang, Hui Zhang, Dan Zhao, Bingzheng Wei, and Yan Xu. Restore-rwkv: Efficient and effective medical image restoration with rwkv. ArXiv, abs/2407.11087, 2024. 3

    2024 arXiv

  33. [41]

    Mamba or rwkv: Exploring high-quality and high-efficiency segment anything model

    Haobo Yuan, Xiangtai Li, Lu Qi, Tao Zhang, Ming-Hsuan Yang, Shuicheng Yan, and Chen Change Loy. Mamba or rwkv: Exploring high-quality and high-efficiency segment anything model. ArXiv, abs/2406.19369, 2024. 3

    2024 arXiv

  34. [42]

    An Attention Free Transformer, 2021

    Shuangfei Zhai, Walter Talbott, Nitish Srivastava, Chen Huang, Hanlin Goh, Ruixiang Zhang, and Josh Susskind. An Attention Free Transformer, 2021. 4, 1

    2021

  35. [43]

    End-to-end Optimized Image Compression with Attention Mechanism

    Lei Zhou, Zhenhong Sun, Xiangjian Wu, and Junmin Wu. End-to-end Optimized Image Compression with Attention Mechanism. In CVPR workshops, 2019. 2

    2019

  36. [44]

    Bsbp-rwkv: Background suppression with boundary preservation for efficient medical image segmentation

    Xudong Zhou and Tianxiang Chen. Bsbp-rwkv: Background suppression with boundary preservation for efficient medical image segmentation. In ACM Multimedia, 2024. 3

    2024

  37. [45]

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

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. ArXiv, abs/2401.09417, 2024. 2

    2024 arXiv

  38. [46]

    Transformer- based Transform Coding

    Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer- based Transform Coding. In International Conference on Learning Representations, 2021. 2

    2021

  39. [47]

    The Devil Is in the Details: Window-based Attention for Image Compression

    Renjie Zou, Chunfeng Song, and Zhaoxiang Zhang. The Devil Is in the Details: Window-based Attention for Image Compression. 2022 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 17471–17480,

    2022

  40. [48]

    The Open Images Dataset V4: Unified image classifica- tion, object detection, and visual relationship detection at scale

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Ui- jlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, et al. The Open Images Dataset V4: Unified image classifica- tion, object detection, and visual relationship detection at scale. International Journal of Computer Visio...

    1956

  41. [49]

    Kodak lossless true color image suite (pho- tocd pcd0992)

    Eastman Kodak. Kodak lossless true color image suite (pho- tocd pcd0992). 1993. 5

    1993

  42. [50]

    Testimages: A large- scale archive for testing visual devices and basic image pro- cessing algorithms

    Nicola Asuni and Andrea Giachetti. Testimages: A large- scale archive for testing visual devices and basic image pro- cessing algorithms. In STAG: Smart Tools and Applications in Computer Graphics, pages 63–70, 2014. 5

    2014

  43. [51]

    Workshop and challenge on learned image compres- sion

    CLIC. Workshop and challenge on learned image compres- sion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2021. 5 Linear Attention Modeling for Learned Image Compression Supplementary Material A. Performance Details This section provi...

    2021

Pith tools

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