Pith. sign in

REVIEW 3 major objections 6 minor 58 references

S2CFormer: Revisiting the RD-Latency Trade-off in Transformer-based Learned Image Compression

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Channel aggregation, not spatial attention, is the real driver of learned image compression performance, the paper argues.

desk verdict The empirical finding holds up—spatial attention is largely replaceable in LIC—but the paper overclaims the causal role of channel aggregation given its entropy-model and capacity confounds. read the letter →

arxiv 2502.00700 v3 pith:OKCDUG7L submitted 2025-02-02 cs.CV eess.IV

Yunuo Chen , Qian Li , Bing He , Donghui Feng , Ronghua Wu , Qi Wang , Li Song , Guo Lu
show 1 more author
Wenjun Zhang
This is my paper · ORCID
classification cs.CVeess.IV
keywords learnedimagecompressionchannelaggregationS2CFormerrate-distortionoptimizationdecodinglatencyfeed-forwardnetworktransformerBD-rate
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

Transformer-based learned image compression has long credited its rate-distortion gains to attention-based spatial operations, but this paper argues those operations are largely redundant. The authors strip spatial interaction down to an identity mapping, keeping only an MLP-based channel-mixing module, and this minimal model still beats VTM-21.0 by 9.70%, 10.70%, and 7.55% BD-rate on Kodak, Tecnick, and CLIC, respectively, nearly matching a much larger transformer-CNN hybrid. Replacing identity with a cheap separable convolution or vanilla window attention gives near-identical results, while removing the channel-mixing module costs far more. The paper proposes S2CFormer as a design paradigm: simplify spatial operations, strengthen channel operations, and get both better rate-distortion performance and over 30% faster decoding.

What carries the argument

The load-bearing object is the S2CFormer block, a two-part nonlinear transform made of a Spatial Interaction module followed by LayerNorm and a Feed-Forward Network that performs Channel Aggregation. The spatial branch can be identity, separable convolution, or vanilla window attention, and the paper shows the branch choice is almost immaterial to rate-distortion performance. The strongest FFN variant is Gated-FFN, which computes $[\sigma (XW_1) \odot (XW_2)] W_{out}$, and the paper shows it outperforms additive and vanilla FFNs at equal latency. The paradigm is completed by S2C-Hybrid, which assigns convolution to the first high-resolution stage and attention to later low-resolution stages, improving BD-rate further without slowing decoding.

What would settle it

Train an MLP-only compression model with the same parameter budget and the same SCCTX entropy model as S2C-Identity but replace the FFN channel mixer with a spatial-attention-only block of matched size: if its BD-rate on Kodak stays within a fraction of a percent of S2C-Identity, the claim that channel aggregation is the key factor would be falsified. A complementary check is to sweep channel width while holding spatial interaction fixed and measure whether BD-rate tracks channel capacity.

Watch

Extended reading notes

Core claim

The central claim is that efficient channel aggregation is the primary factor in the rate-distortion performance of learned image compression, and that the complex spatial operations in transformer-based codecs are largely redundant. This is demonstrated by S2C-Identity, which removes spatial interaction entirely and uses only an FFN-based channel mixer; it outperforms VTM-21.0 on all three benchmarks while nearly matching TCM-large. S2C-Conv (separable convolution) and S2C-Attention (vanilla window attention) achieve near-identical BD-rates, showing the specific spatial operator barely matters; ablating the channel aggregation module instead reduces BD-rate by roughly ten percentage points. The paper concludes that transformer LIC design should be rebalanced toward FFN-based channel aggregation, and presents S2C-Hybrid, which combines convolution at high resolution with attention at low resolution, as a new state of the art with significantly faster decoding than previous methods.

Load-bearing premise

The attribution of rate-distortion gains to channel aggregation rather than to model size rests on the assumption that the strong performance of S2C-Identity is not mostly an artifact of its 64.63 million parameters or of the powerful ELIC entropy model that every S2CFormer variant inherits.

Editorial extensions

If this is right

  • Transformer-based learned image codecs can drop shifted-window attention and other heavy spatial modules, gaining more than 30% decoding speed with no loss in rate-distortion performance.
  • The near-identical performance of S2C-Conv and S2C-Attention implies that the choice of spatial operator is secondary; research effort should shift to FFN and channel-mixing design.
  • S2C-Identity establishes that a pure MLP channel mixer, with no spatial mixing at all, already beats VTM-21.0, so any future spatial module must justify its cost against that baseline.
  • Stage-wise hybridization, using convolution at high resolution and attention at low resolution, improves rate-distortion further without extra latency, so spatial operators retain some value at the finest scales.
  • Scaling up the hybrid with more blocks in the later stages yields monotonic BD-rate gains, and a low-FLOP variant keeps state-of-the-art BD-rate at 2.63T FLOPs, suggesting the paradigm is capacity-friendly.

Reading between the lines

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

  • If channel aggregation is the true workhorse, other low-level vision tasks built on transformers, such as restoration, video compression, and super-resolution, may similarly find that their spatial attention modules can be replaced by cheap token mixers without losing quality; the paper does not test this extension.
  • The paper's effective-receptive-field analysis suggests channel mixing actually expands the field more than window shifting does; a testable corollary is that channel-mixing-only models should match attention models on tasks requiring long-range dependencies.
  • A controlled parameter-matched comparison between S2C-Identity and an attention-only model with a weak entropy model would isolate the channel-aggregation effect from the entropy-model and capacity confounds; the authors did not run this, so their causal attribution remains partly open.
  • The Gated-FFN result hints that dedicated channel-mixer architectures could be tuned further, possibly approaching the rate-distortion-latency frontier without transformer-style blocks at all.
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 argues that in transformer-based learned image compression (LIC), rate-distortion performance is driven primarily by channel aggregation (the FFN/MLP part of a transformer block) rather than by complex spatial interaction modules such as shifted-window attention. It introduces the S2CFormer paradigm, whose spatial interaction is deliberately simplified (identity mapping, separable convolution, or vanilla window attention) while channel aggregation is implemented with improved FFNs (Additive-FFN, Gated-FFN). Three model variants are presented: S2C-Identity, S2C-Conv, S2C-Attention, plus a staged hybrid S2C-Hybrid. The models are built on an ELIC-style VAE with the SCCTX entropy model, trained on Flickr2W, and evaluated on Kodak, Tecnick, and CLIC. Reported BD-rate numbers relative to VTM-21.0 are competitive with or better than published methods such as TCM, FTIC, and MLIC++, while decoding latency is reduced. The central claim is that channel aggregation, not spatial interaction, is the key to competitive LIC models.

Significance. If the central attribution claim is correct, the paper would provide a useful design principle for LIC: simple spatial operators plus strong channel mixing can match or beat much more elaborate transformer blocks, while improving decoding speed. The paper is strengthened by releasing code, by benchmarking against VTM-21.0 and several strong published models, and by reporting consistent internal ablation trends (identity spatial mapping retains strong R-D performance, and removing FFNs degrades BD-rate). The significance is currently limited by the lack of a clean isolation of the channel-aggregation effect from the entropy model and model capacity: the headline S2C-Identity result changes several components at once, so the paper's central causal conclusion is plausible but not yet established. The latency findings are more robust and are a useful independent contribution.

major comments (3)
  1. [Sec. 4.1.1, Sec. 4.2, Table 1] The claim that S2C-Identity isolates the contribution of channel aggregation is not supported by the reported comparisons. S2C-Identity does not merely replace spatial interactions by identity; it inherits the ELIC SCCTX entropy model and, as stated in Sec. 4.1.1, additionally incorporates S2C-Conv and S2C-Identity blocks inside the entropy model. It also has 64.63M parameters (Table 1), far more than ELIC's 33.29M. The comparison with TCM-large, FTIC, and MLIC++ therefore varies the transform architecture, the entropy model, and model capacity simultaneously. To substantiate the causal claim, please add a control where the same entropy model is paired with a standard Swin/TCM transform, or where the identity transform is paired with the unmodified ELIC entropy model, with comparable parameter count.
  2. [Sec. 4.3, Fig. 5] The 'w/o CA' ablation is under-specified. The text says 'removing all FFN modules for Channel Aggregation,' but it does not state whether FFNs are removed from the transform only or also from the entropy model, whose incorporated S2C-Conv/S2C-Identity blocks also contain FFNs. If entropy-model FFNs are removed, the ablation changes context-model capacity rather than isolating channel aggregation in the transform. In addition, no error bars or repeated-run statistics are reported despite BD-rate differences of 7-11%; with single runs, the conclusion that FFN removal is responsible for the drop is not statistically supported.
  3. [Sec. 4.4, Table 2] The Hybrid configuration [C,A,A] is selected by comparing BD-rate on Kodak only (Table 2), and this same configuration is then reported as a new state of the art on Kodak, Tecnick, and CLIC (Table 1). The Kodak SOTA claim is therefore partly circular, and the generalization to the other datasets is not validated by a selection procedure that uses held-out data. Please either report the configuration-selection result as exploratory and validate on a separate split, or provide a model-selection criterion that does not use the evaluation datasets.
minor comments (6)
  1. [Sec. 4.2] 'Tabel 1' is a typo for 'Table 1'.
  2. [Fig. 3 caption] The caption mentions L1-L6 and C1-C6, but only L1-L3 and C1-C4 are described in Sec. 4.1.1; please clarify the stage configuration for all models, including the entropy model.
  3. [Sec. 3.1] The rate-distortion loss equation contains typesetting errors ('entopy', 'pbz') and should be reformatted; also state explicitly that negative BD-rate means improvement over the anchor.
  4. [Table 1] FTIC's decoding latency is listed as '>10' seconds, which is an unusual value; please report the actual measurement or explain the cutoff, and specify the GPU and software versions used for all latency numbers.
  5. [References] Reference [27] is cited as 'Thorp et al.' but the FNet paper is by Lee-Thorp et al.; please correct the citation and in-text attribution.
  6. [Fig. 6] The effective receptive field visualization would be more interpretable if the authors specified which model variant and input image are used, and whether the ERF is computed before or after training convergence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central claims rest on external benchmarks and controlled ablations, not on fitted constants or self-citations.

full rationale

The paper's derivation chain is empirical rather than definitional. The key supporting result, S2C-Identity's BD-rate advantage over VTM-21.0 and its near-TCM-large performance, is a measured outcome of training a specified architecture and evaluating it on Kodak, Tecnick, and CLIC; it is not obtained by fitting the reported BD-rate values or by renaming an input. The ablation in Fig. 5 removes all FFN channel-aggregation modules while keeping the SCCTX entropy model and the same S2C instances, so the -7.51%/-8.17% and -10.57%/-11.14% BD-rate gaps are direct comparisons of identical pipelines with and without the module in question. The S2C-Hybrid stage configuration [C,A,A] is chosen by an ablation on Kodak (Table 2), but its state-of-the-art claim is then checked on Tecnick and CLIC and against external codecs, so this is model selection rather than a prediction that is forced by construction. Self-citations ([19], [45], [56]) appear only in contextual related-work sentences and are not load-bearing; no uniqueness theorem or prior result by the same authors is invoked to forbid alternative architectures. The attribution of the gain specifically to channel aggregation rather than to the inherited ELIC SCCTX entropy model or larger parameter count is an experimental-isolation concern, not a circularity, because the paper does not define channel aggregation in terms of the reported gains. No step in the paper equates a claimed prediction with a fitted input or imports a conclusion solely from a self-citation.

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

The central empirical finding rests on architectural choices (channel widths, block counts, expansion ratio, hybrid arrangement) that are chosen by hand or by ablation on the evaluation datasets. No new physical entities are introduced.

free parameters (4)
  • Hybrid stage arrangement (C,A,A) = [C,A,A]
    Selected from Table 2 ablation on Kodak; the choice maximizes reported BD-rate and is used for the final S2C-Hybrid claim.
  • Channel widths C1-C4 = 192, 192, 192, 320
    Set by hand following prior LIC settings (Sec 4.1.1); influences capacity and FLOPs.
  • Block counts L1-L3 for Hybrid-S = 3, 3, 3
    Chosen by hand as the standard configuration; scaling variants use 3,5,5 and 3,8,8.
  • FFN expansion ratio r = 4
    Set to 4 for all FFN structures (Sec 3.2.3); affects capacity and latency.
assumptions (4)
  • domain assumption SCCTX entropy model from ELIC provides a strong and standard entropy model; transform network improvements are the cause of R-D gains.
    Adopted in Sec 3.1 and 4.1.2; the paper does not ablate the entropy model.
  • domain assumption Ablating all FFN modules isolates the channel aggregation contribution; no interacting effects between FFN and spatial modules are considered.
    Section 4.3 ablation compares models with and without channel aggregation.
  • ad hoc to paper The identity-mapping spatial interaction is a valid lower bound representing the absence of spatial operations.
    Section 3.2.1; assumes identity is the correct baseline rather than, for example, a single pointwise convolution.
  • domain assumption Reported baseline BD-rate and latency numbers reflect fair, consistent evaluation conditions.
    Section 4.1.2; the paper does not state how baseline numbers were obtained.

how reviews work

0 comments
Cite this review

Pith. "Pith review of S2CFormer: Revisiting the RD-Latency Trade-off in Transformer-based Learned Image Compression." pith.science (2026). https://pith.science/paper/OKCDUG7L

@misc{pith2026250200700,
  author       = {Pith},
  title        = {Pith review of: S2CFormer: Revisiting the RD-Latency Trade-off in Transformer-based Learned Image Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OKCDUG7L}},
  note         = {Machine review of arXiv:2502.00700}
}
read the original abstract

Transformer-based Learned Image Compression (LIC) suffers from a suboptimal trade-off between decoding latency and rate-distortion (R-D) performance. Moreover, the critical role of the FeedForward Network (FFN)-based channel aggregation module has been largely overlooked. Our research reveals that efficient channel aggregation-rather than complex and time-consuming spatial operations-is the key to achieving competitive LIC models. Based on this insight, we initiate the ``S2CFormer'' paradigm, a general architecture that simplifies spatial operations and enhances channel operations to overcome the previous trade-off. We present two instances of the S2CFormer: S2C-Conv, and S2C-Attention. Both models demonstrate state-of-the-art (SOTA) R-D performance and significantly faster decoding speed. Furthermore, we introduce S2C-Hybrid, an enhanced variant that maximizes the strengths of different S2CFormer instances to achieve a better performance-latency trade-off. This model outperforms all the existing methods on the Kodak, Tecnick, and CLIC Professional Validation datasets, setting a new benchmark for efficient and high-performance LIC. The code is at \href{https://github.com/YunuoChen/S2CFormer}{https://github.com/YunuoChen/S2CFormer}.

Figures

Figures reproduced from arXiv: 2502.00700 by the authors.

Figure 1
Figure 1. S2CFormer and the performance of S2CFormer-based models. The general structure of our S2CFormer is shown in (a). It consists of two key components: the Spatial Interaction module and the Channel Aggregation module. S2CFormer functions as nonlinear transform blocks for Learned Image Compression (LIC). Our analysis reveals that the competence of transformer-based LIC models primarily stems from channel aggregation. Bu… view at source ↗
Figure 2
Figure 2. Comparison of execution times for spatial interac￾tion and channel aggregation across different models. Previous methods show much higher spatial interaction times than channel aggregation, causing significant delays. Our S2CFormer effectively rebalances the time relationship between these two modules. models have emerged as a promising alternative to traditional codecs, offering improved rate-distortion performance… view at source ↗
Figure 3
Figure 3. Overview of S2CFormer-based LIC model. We adopt the basic VAE structure from [3, 42] and integrate the SCCTX entropy model from [21]. The hierarchical architecture consists of five stages of nonlinear transform blocks. Each stage contains Li S2CFormer blocks. The general S2CFormer architecture is shown in (a), and (b-d) illustrate three S2CFormer instances. L1-L6 and C1-C6 represent block numbers and channel numbers… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Vanilla FFN (a) and Advanced FFNs (b-c) The Additive-FFN combines outputs from two activation functions via addition, applying separate nonlinear transfor￾mations to the same input to leverage the strength of multiple nonlinearities: Additive-FFN(X) = [σ1 (XW1) + σ2 (X…
Figure 5
Figure 5. Figure 5: Ablation Study for Channel Aggregation. Experiments on Kodak dataset. “w/ FFN” refers to S2CFormer-based models with Channel Aggregation. “w/o CA” represents removing all FFN modules for Channel Aggregation. all the SOTA methods. As shown in Tabel 1, in terms of BD-rat…
Figure 7
Figure 7. Figure 7: Performance evaluation on the CLIC Professional Validation Dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Performance evaluation on the Kodak dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 10
Figure 10. Figure 10: Comparison of Different FFNs. Experiments on the Kodak dataset. We provide R-D curves for S2C-Conv and -Attention models with various FFNs for Channel Aggregation. and enhancing R-D performance, suggesting that certain complex spatial operations may be redundant. Effe…
Figure 11
Figure 11. Figure 11: Visual Comparison. This figure presents visualizations of decompressed images from the Kodak dataset using various methods. Each subfigure is labeled with “Method | Bpp | PSNR | MS-SSIM". Attention exhibit similar rate-distortion performance, their distinct capabiliti…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 44 canonical work pages

  1. [1]

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

    Nicola Asuni, Andrea Giachetti, et al. Testimages: a large- scale archive for testing visual devices and basic image pro- cessing algorithms. In STAG, pages 63–70, 2014. 5

    work page 2014

  2. [2]

    End-to-end optimized image compression

    Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 2, 8

    arXiv 2016

  3. [3]

    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,

    arXiv

  4. [4]

    Compressai: a pytorch library and evaluation platform for end-to-end compression research

    Jean Bégaint, Fabien Racapé, Simon Feltman, and Akshay Pushparaja. Compressai: a pytorch library and evaluation platform for end-to-end compression research. arXiv e-prints, pages arXiv–2011, 2020. 3

    work page 2011

  5. [5]

    Calculation of average psnr differences between rd-curves

    Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. In VCEG-M33, 2001. 5

    work page 2001

  6. [6]

    Overview of the versatile video coding (vvc) standard and its applica- tions

    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. 5

    work page 2021

  7. [7]

    Browne, Y

    A. Browne, Y . Ye, and S. Kim. Algorithm description for versatile video coding and test model 21 (vtm 21), docu- ment jvet-af2002. In Joint Video Experts Team (JVET) of ITU-T SG 16 WP 3 and ISO/IEC JTC 1/SC 29/WG 11, 32nd Meeting, Hannover. 5, 6

    work page

  8. [8]

    Two-stage octave residual network for end-to-end image compres- sion

    Fangdong Chen, Yumeng Xu, and Li Wang. Two-stage octave residual network for end-to-end image compres- sion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3922–3929, 2022. 2, 3

    work page 2022

Show all 58 references
  1. [9]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17–33. Springer, 2022. 3, 5

    2022

  2. [10]

    Learned image compression with discretized gaussian mixture likelihoods and attention modules

    Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939–7948, 2020. 2, 3, 8

    2020

  3. [11]

    Xception: Deep learning with depth- wise separable convolutions

    François Chollet. Xception: Deep learning with depth- wise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251–1258, 2017. 4

    2017

  4. [12]

    Workshop and challenge on learned image com- pression

    CLIC. Workshop and challenge on learned image com- pression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 5

    2021

  5. [13]

    Language modeling with gated convolutional networks

    Yann N Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In International conference on machine learning, pages 933–941. PMLR, 2017. 5

    2017

  6. [14]

    Cswin transformer: A general vision transformer backbone with cross-shaped windows

    Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pa...

    2022

  7. [15]

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

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3

    2010 arXiv

  8. [16]

    Asymmetric learned image com- pression with multi-scale residual block, importance scal- ing, and post-quantization filtering

    Haisheng Fu, Feng Liang, Jie Liang, Binglin Li, Guohe Zhang, and Jingning Han. Asymmetric learned image com- pression with multi-scale residual block, importance scal- ing, and post-quantization filtering. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):4...

    2023

  9. [17]

    Neural image compression via attentional multi-scale back projection and frequency de- composition

    Ge Gao, Pei You, Rong Pan, Shunyuan Han, Yuanyuan Zhang, Yuchao Dai, and Hojae Lee. Neural image compression via attentional multi-scale back projection and frequency de- composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14677–14686, 2021

    2021

  10. [18]

    Causal context adjustment loss for learned image compression

    Minghao Han, Shiyin Jiang, Shengxi Li, Xin Deng, Mai Xu, Ce Zhu, and Shuhang Gu. Causal context adjustment loss for learned image compression. In The Thirty-eighth Annual Conference on Neural Information Processing Systems,

  11. [19]

    S4d: Streaming 4d real-world reconstruction with gaussians and 3d control points

    Bing He, Yunuo Chen, Guo Lu, Li Song, and Wenjun Zhang. S4d: Streaming 4d real-world reconstruction with gaussians and 3d control points. arXiv preprint arXiv:2408.13036,

  12. [20]

    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 Computer Vision and Pattern Recognition, pages 14771–14780, 2021. 3

    2021

  13. [21]

    Elic: Efficient learned image compres- sion with unevenly grouped space-channel contextual adap- tive coding

    Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compres- sion with unevenly grouped space-channel contextual adap- tive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718–5727,

  14. [22]

    MLIC$^{++}$: Linear com- plexity multi-reference entropy modeling for learned image compression

    Wei Jiang and Ronggang Wang. MLIC$^{++}$: Linear com- plexity multi-reference entropy modeling for learned image compression. In ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 3

    2023

  15. [23]

    Mlic: Multi-reference entropy model for learned image compression

    Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi-reference entropy model for learned image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7618–7627, 2023. 3, 5

    2023

  16. [24]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 5

    2014 arXiv

  17. [25]

    Kodak lossless true color im- age suite (photocd pcd0992), 1993

    Eastman Kodak. Kodak lossless true color im- age suite (photocd pcd0992), 1993. Available from http://r0k.us/graphics/kodak/. 5

    1993

  18. [26]

    Con- textformer: A transformer with spatio-channel attention for context modeling in learned image compression

    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 European Conference on Computer Vision, pages 447–463. Springer,

  19. [27]

    Fnet: Mixing tokens with fourier transforms

    James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, and Santiago Ontanon. Fnet: Mixing tokens with fourier transforms. arXiv preprint arXiv:2105.03824, 2021. 3

    2021 arXiv

  20. [28]

    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. arXiv preprint arXiv:2310.16387, 2023. 2, 3, 5, 8

    2023 arXiv

  21. [29]

    Hybrid spatial-temporal en- tropy modelling for neural video compression

    Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal en- tropy modelling for neural video compression. InProceedings of the 30th ACM International Conference on Multimedia, pages 1503–1511, 2022. 3, 5

    2022

  22. [30]

    Neural video compression with diverse contexts

    Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616–22626, 2023. 3, 5

    2023

  23. [31]

    Swinir: Image restoration using swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833– 1844, 2021. 3

    2021

  24. [32]

    A unified end-to-end framework for efficient deep image compression

    Jiaheng Liu, Guo Lu, Zhihao Hu, and Dong Xu. A unified end-to-end framework for efficient deep image compression. arXiv preprint arXiv:2002.03370, 2020. 3, 5

    2002 arXiv

  25. [33]

    Learned im- age compression with mixed transformer-cnn architectures

    Jinming Liu, Heming Sun, and Jiro Katto. Learned im- age compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14388–14397, 2023. 2, 3, 5

    2023

  26. [34]

    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 Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 2, 3

    2021

  27. [35]

    Swin transformer v2: Scaling up capacity and resolution

    Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009–12019, 2...

    2022

  28. [36]

    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. IEEE,

    2022

  29. [37]

    Understanding the effective receptive field in deep convo- lutional neural networks

    Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convo- lutional neural networks. Advances in neural information processing systems, 29, 2016. 6

    2016

  30. [38]

    iwave: Cnn-based wavelet-like transform for image compression

    Haichuan Ma, Dong Liu, Ruiqin Xiong, and Feng Wu. iwave: Cnn-based wavelet-like transform for image compression. IEEE Transactions on Multimedia, 22(7):1667–1679, 2019. 3

    2019

  31. [39]

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

    Haichuan Ma, Dong Liu, Ning Yan, Houqiang Li, and Feng Wu. End-to-end optimized versatile image compression with wavelet-like transform. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1247–1263, 2020. 2

    2020

  32. [40]

    Vct: A video compression transformer

    Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agusts- son. Vct: A video compression transformer. arXiv preprint arXiv:2206.07307, 2022. 3

    2022 arXiv

  33. [41]

    Channel-wise autoregres- sive entropy models for learned image compression

    David Minnen and Saurabh Singh. Channel-wise autoregres- sive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339–3343. IEEE, 2020. 2, 3

    2020

  34. [42]

    Joint autoregressive and hierarchical priors for learned im- age compression

    David Minnen, Johannes Ballé, and George D Toderici. Joint autoregressive and hierarchical priors for learned im- age compression. Advances in neural information processing systems, 31, 2018. 2, 3, 4, 8

    2018

  35. [43]

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

    Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492, 2022. 2, 3

    2022 arXiv

  36. [44]

    Mambavc: Learned visual compression with selective state spaces

    Shiyu Qin, Jinpeng Wang, Yimin Zhou, Bin Chen, Tianci Luo, Baoyi An, Tao Dai, Shutao Xia, and Yaowei Wang. Mambavc: Learned visual compression with selective state spaces. arXiv preprint arXiv:2405.15413, 2024. 5, 8

    2024 arXiv

  37. [45]

    Bayesian neural networks avoid encoding com- plex and perturbation-sensitive concepts

    Qihan Ren, Huiqi Deng, Yunuo Chen, Siyu Lou, and Quan- shi Zhang. Bayesian neural networks avoid encoding com- plex and perturbation-sensitive concepts. In International Conference on Machine Learning, pages 28889–28913. PMLR, 2023. 2

    2023

  38. [46]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh- moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510–4520, 2018. 4

    2018

  39. [47]

    Training data-efficient image transformers & distillation through atten- tion

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through atten- tion. In International conference on machine learning, pages 10347–10357. PMLR, 2021. 3

    2021

  40. [48]

    Attention is all you need

    A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3

    2017

  41. [49]

    Vision transformer with deformable attention

    Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794–4803, 2022. 3

    2022

  42. [50]

    Enhanced invertible encoding for learned image compression

    Yueqi Xie, Ka Leong Cheng, and Qifeng Chen. Enhanced invertible encoding for learned image compression. In Proceedings of the 29th ACM international conference on multimedia, pages 162–170, 2021. 2, 3

    2021

  43. [51]

    Metaformer is actually what you need for vision

    Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10819–10829, 2022. 3, 4

    2022

  44. [52]

    Metaformer baselines for vision

    Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer baselines for vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3

    2023

  45. [53]

    Frequency disentangled features in 10 neural image compression

    Ali Zafari, Atefeh Khoshkhahtinat, Piyush Mehta, Moham- mad Saeed Ebrahimi Saadabadi, Mohammad Akyash, and Nasser M Nasrabadi. Frequency disentangled features in 10 neural image compression. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2815–2819. IEE...

    2023

  46. [54]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728–5739,

  47. [55]

    Practical blind denoising via swin-conv-unet and data synthesis

    Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yulun Zhang, Hao Tang, Radu Timofte, and Luc Van Gool. Practical blind denoising via swin-conv-unet and data synthesis. arXiv e-prints, pages arXiv–2203, 2022. 3

    2022

  48. [56]

    Neural rate control for learned video compression

    Yiwei Zhang, Guo Lu, Yunuo Chen, Shen Wang, Yibo Shi, Jing Wang, and Li Song. Neural rate control for learned video compression. In The Twelfth International Conference on Learning Representations, 2023. 2

    2023

  49. [57]

    Transformer-based transform coding

    Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In International Conference on Learning Representations, 2022. 2, 3

    2022

  50. [58]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17492– 17501, 2022. 3 11

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.