pith. sign in

arxiv: 2606.21033 · v1 · pith:IBDX4ATZnew · submitted 2026-06-19 · 📡 eess.IV · cs.AI· cs.CV

MoECodec: Image Compression for joint human and machine perception via Mixture-of-Experts

Jiancheng Zhao , Xiang Ji , Yifan Zhan , Zunian Wan , Yinqiang Zheng This is my paper

Pith reviewed 2026-06-26 13:09 UTC · model grok-4.3

classification 📡 eess.IV cs.AIcs.CV
keywords image compressionmixture of expertstransformermachine perceptionrate distortiontoken routingjoint human machine codec
0
0 comments X

The pith

MoECodec replaces FFN layers with token-wise mixture-of-experts to enable dynamic computation in a single image codec for both human viewing and machine tasks.

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

The paper introduces MoECodec as a transformer-based image compression model that swaps standard feed-forward layers for a token-wise mixture-of-experts structure. This change lets each image token receive computation tailored to its content and the active task objective instead of applying the same operations everywhere. A routing method that pairs expert-choice assignment with spatial total variation regularization keeps the experts coherent across neighboring pixels, while a compact Group Shuffle MLP limits parameter growth. The result is one model that handles conventional reconstruction alongside multiple downstream vision tasks and reports better rate-distortion numbers than static baselines.

Core claim

MoECodec replaces the FFN layers in transformer-based compression model with token-wise Mixture-of-Experts, enabling dynamic, token-level computation conditioned on the input content and task objective. A stable routing strategy combines expert-choice routing with spatial total variation regularization to encourage spatially coherent assignments, and a lightweight Group Shuffle MLP controls parameter growth. Extensive experiments show consistent improvement against baselines on both conventional image reconstruction and machine tasks.

What carries the argument

Token-wise Mixture-of-Experts replacing FFN layers, driven by expert-choice routing plus spatial total variation regularization and implemented with Group Shuffle MLP experts.

If this is right

  • A single set of parameters serves both human reconstruction and multiple machine vision tasks without task-specific retraining.
  • Computation cost varies per token according to local semantic complexity rather than remaining uniform across the image.
  • Spatially coherent expert assignments prevent visible artifacts that would otherwise appear from independent token routing.
  • Parameter growth stays modest because the Group Shuffle MLP replaces full expert networks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same token-level routing could be tested on video codecs where temporal consistency might replace or augment the spatial regularizer.
  • If the experts learn distinct frequency or semantic specializations, the model might allow selective decoding of only the experts needed for a given downstream task.
  • Deployment on edge devices could benefit from the dynamic compute pattern, as simple regions would use fewer experts at inference time.

Load-bearing premise

A stable routing strategy that combines expert-choice routing with spatial total variation regularization will keep token-wise MoE effective inside a compression model without harming rate-distortion performance or introducing spatial artifacts.

What would settle it

If disabling the spatial total variation term produces visible block artifacts or worse rate-distortion curves on Kodak or CLIC test sets while keeping the same expert count, the routing approach fails to stabilize the model.

Figures

Figures reproduced from arXiv: 2606.21033 by Jiancheng Zhao, Xiang Ji, Yifan Zhan, Yinqiang Zheng, Zunian Wan.

Figure 1
Figure 1. Figure 1: Overview of dynamic compute allocation in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a): Overall architecture of MoECodec. Starting from a pre-trained Transformer-based base codec, all transformer blocks in the encoder and decoder are replaced with MoE blocks while keeping other components fixed. (b): Structure of a MoE layer in each MoE block. Input tokens are distributed to different experts through EC routing, enabling token-adaptive computation within the block. (c): Each expert netwo… view at source ↗
Figure 3
Figure 3. Figure 3: Multi-task performance. Models applied using MoECodec is marked by †. For EC routing, we remove the spatial regularization term in Equation (7). The resulting routing maps are visualized in [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Expert specialization comparison under TC, EC, and EC+Spatial Total Variance regularization. Images drown from Kodak14, Kodak23. Routing results at en￾coder stage ga1. EC-based routing yields more balanced and spatially coherent expert allocation than TC, and Spatial regularization further refines continuity [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Expert-token allocation across tasks. Top: four samples from ImageNet validation coded for MSE (upper row) vs. classification (lower row); columns show re￾constructed image, FFT spectrum, and the stage-ga1 heatmap. Bottom: two samples from COCO2017 validation coded for object detection (upper row) vs. instance segmen￾tation (lower row); columns show reconstructed image, FFT spectrum, and stage-wise heatmap… view at source ↗
read the original abstract

Image compression for machines calls for a unified codec that serves multiple downstream vision tasks. Existing approaches either adopt task-specific end-to-end designs, raising parameter and deployment overhead, or rely on transfer-based adaptations that remain externally attached and heuristic task design. A key limitation shared by both lines of work is their largely static computation pattern, which applies similar transformations across tokens despite the fact that different image regions exhibit markedly different semantic importance and complexity for machine perception. We propose MoECodec, a token-aware image compression framework that supports multiple downstream tasks within a single model. MoECodec replaces the FFN layers in transformer-based compression model token-wise Mixture-of-Experts (MoE), enabling dynamic, token-level computation conditioned on the input content and task objective. To make MoE effective in compression model, we introduce a stable routing strategy that combines expert-choice routing with spatial total variation regularization to encourage spatially coherent assignments, and we propose a lightweight expert architecture, Group Shuffle MLP (GShMLP), to control parameter growth. Extensive experiments show consistent improvement against baselines on both conventional image reconstruction and machine tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes MoECodec, a token-aware image compression framework for joint human and machine perception. It replaces the FFN layers in a transformer-based codec with token-wise Mixture-of-Experts (MoE) using expert-choice routing combined with spatial total variation regularization, plus a lightweight Group Shuffle MLP (GShMLP) expert, to enable dynamic per-token computation conditioned on content and task. The abstract states that extensive experiments demonstrate consistent gains on both reconstruction and machine tasks relative to baselines.

Significance. If the experimental claims hold, the work is significant because it directly targets the static computation limitation of prior codecs by introducing content- and task-adaptive token-level routing inside the compression model itself, rather than relying on task-specific designs or post-hoc adaptations. The combination of expert-choice routing, spatial coherence regularization, and parameter-efficient experts offers a principled way to scale multi-task codecs without proportional parameter growth, which could influence future unified codecs for diverse vision pipelines.

major comments (2)
  1. [Abstract] Abstract: the central claim of 'consistent improvement' and 'extensive experiments' is presented without any quantitative results, baseline names, dataset identifiers, or ablation tables. Because the soundness of the routing strategy and GShMLP expert rests entirely on these unreported numbers, the load-bearing experimental validation cannot be assessed from the provided material.
  2. [Method] Method (routing strategy): the paper asserts that expert-choice routing plus spatial TV regularization yields stable assignments without rate-distortion loss or spatial artifacts, yet no equation, hyper-parameter schedule, or ablation isolating the TV term is referenced. This is load-bearing for the claim that MoE can be made effective inside a compression model.
minor comments (1)
  1. [Abstract] Abstract: specify the exact machine tasks (e.g., classification, detection, segmentation) and the datasets used so readers can immediately gauge the scope of the multi-task evaluation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for more concrete experimental support in the abstract and clearer methodological references. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of 'consistent improvement' and 'extensive experiments' is presented without any quantitative results, baseline names, dataset identifiers, or ablation tables. Because the soundness of the routing strategy and GShMLP expert rests entirely on these unreported numbers, the load-bearing experimental validation cannot be assessed from the provided material.

    Authors: We agree that the abstract would be strengthened by including specific quantitative support. In the revised version, we will add key results such as average BD-rate savings on the Kodak dataset for reconstruction and mAP gains on COCO for object detection, along with baseline names (VTM, Cheng2020) and a reference to the ablation studies. This keeps the abstract concise while making the claims verifiable. revision: yes

  2. Referee: [Method] Method (routing strategy): the paper asserts that expert-choice routing plus spatial TV regularization yields stable assignments without rate-distortion loss or spatial artifacts, yet no equation, hyper-parameter schedule, or ablation isolating the TV term is referenced. This is load-bearing for the claim that MoE can be made effective inside a compression model.

    Authors: The method section (3.2) formulates expert-choice routing and the spatial TV regularization term with accompanying equations, and the implementation details specify the hyper-parameter schedule for the TV coefficient. The ablation isolating the TV term (showing stable routing without RD penalty or artifacts) appears in Table 4. We will add explicit forward references to these equations and the table in the method overview to improve clarity. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents MoECodec as a new architectural proposal: replacing FFN layers in a transformer compression backbone with token-wise MoE, stabilized by expert-choice routing plus spatial TV regularization, and using a lightweight GShMLP expert. No equations, fitted parameters, or self-citation chains are shown that would reduce the claimed rate-distortion or task gains to a definition or input by construction. The central claims rest on the design choices and reported experiments rather than any load-bearing self-referential step, rendering the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no equations, training details, or explicit assumptions; free parameters, axioms, and invented entities cannot be enumerated.

pith-pipeline@v0.9.1-grok · 5737 in / 1033 out tokens · 23918 ms · 2026-06-26T13:09:58.825622+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

66 extracted references · 30 canonical work pages · 12 internal anchors

  1. [1]

    In: MoECodec15 Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Abati, D., Tomczak, J., Blankevoort, T., Calderara, S., Cucchiara, R., Bejnordi, B.E.: Conditional channel gated networks for task-aware continual learning. In: MoECodec15 Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 3931–3940 (2020) 4

    2020

  2. [2]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Agustsson,E.,Minnen,D.,Toderici,G.,Mentzer,F.:Multi-realismimagecompres- sion with a conditional generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22324–22333 (2023) 1, 4

    2023

  3. [3]

    Agustsson, E., Minnen, D., Toderici, G., Mentzer, F.: Multi-realism image com- pressionwithaconditionalgenerator(2023),https://arxiv.org/abs/2212.13824 3

    work page arXiv 2023

  4. [4]

    eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers

    Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Zhang, Q., Kreis, K., Ait- tala, M., Aila, T., Laine, S., et al.: ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  5. [5]

    Ballé, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression (2017),https://arxiv.org/abs/1611.017043, 4, 5

    work page arXiv 2017

  6. [6]

    Ballé, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior (2018),https://arxiv.org/abs/1802.01436 3, 4

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  7. [7]

    In: 2021 Data Compression Conference (DCC)

    Chamain, L.D., Racapé, F., Bégaint, J., Pushparaja, A., Feltman, S.: End-to-end optimized image compression for machines, a study. In: 2021 Data Compression Conference (DCC). pp. 163–172. IEEE (2021) 1, 4

    2021

  8. [8]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Chen, Y.H., Weng, Y.C., Kao, C.H., Chien, C., Chiu, W.C., Peng, W.H.: Transtic: Transferring transformer-based image compression from human perception to ma- chine perception. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23297–23307 (2023) 1, 4, 8, 9, 11, 13

    2023

  9. [9]

    Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules (2020),https: //arxiv.org/abs/2001.015683

    work page arXiv 2020

  10. [10]

    Exponentially Increasing the Capacity-to-Computation Ratio for Conditional Computation in Deep Learning

    Cho,K.,Bengio,Y.:Exponentiallyincreasingthecapacity-to-computationratiofor conditional computation in deep learning. arXiv preprint arXiv:1406.7362 (2014) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  11. [11]

    IEEE Trans- actions on Image Processing31, 2739–2754 (2022) 4

    Choi, H., Bajić, I.V.: Scalable image coding for humans and machines. IEEE Trans- actions on Image Processing31, 2739–2754 (2022) 4

    2022

  12. [12]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Choi, Y., El-Khamy, M., Lee, J.: Variable rate deep image compression with a con- ditional autoencoder. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3146–3154 (2019) 3

    2019

  13. [13]

    Cui, Z., Wang, J., Gao, S., Bai, B., Guo, T., Feng, Y.: Asymmetric gained deep image compression with continuous rate adaptation (2022),https://arxiv.org/ abs/2003.020123

    work page arXiv 2022

  14. [14]

    In: 2009 IEEE conference on computer vision and pattern recognition

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248–255. Ieee (2009) 8

    2009

  15. [15]

    In: International conference on machine learning

    Du, N., Huang, Y., Dai, A.M., Tong, S., Lepikhin, D., Xu, Y., Krikun, M., Zhou, Y., Yu, A.W., Firat, O., et al.: Glam: Efficient scaling of language models with mixture-of-experts. In: International conference on machine learning. pp. 5547–

  16. [16]

    IEEE Transactions on Circuits and Systems for Video Tech- nology33(8), 4108–4121 (2023) 1, 4

    Duan, Z., Ma, Z., Zhu, F.: Unified architecture adaptation for compressed domain semantic inference. IEEE Transactions on Circuits and Systems for Video Tech- nology33(8), 4108–4121 (2023) 1, 4

    2023

  17. [17]

    Journal of Machine Learning Research 23(120), 1–39 (2022) 4 16 J

    Fedus,W.,Zoph,B.,Shazeer,N.:Switchtransformers:Scalingtotrillionparameter models with simple and efficient sparsity. Journal of Machine Learning Research 23(120), 1–39 (2022) 4 16 J. Zhao et al

    2022

  18. [18]

    Fedus,W.,Zoph,B.,Shazeer,N.:Switchtransformers:Scalingtotrillionparameter models with simple and efficient sparsity (2022),https://arxiv.org/abs/2101.0 39619

    2022

  19. [19]

    arXiv preprint arXiv:2407.11633 (2024)

    Fei, Z., Fan, M., Yu, C., Li, D., Huang, J.: Scaling diffusion transformers to 16 billion parameters. arXiv preprint arXiv:2407.11633 (2024) 4

    work page arXiv 2024

  20. [20]

    In: European Conference on Computer Vision

    Feng, R., Jin, X., Guo, Z., Feng, R., Gao, Y., He, T., Zhang, Z., Sun, S., Chen, Z.: Image coding for machines with omnipotent feature learning. In: European Conference on Computer Vision. pp. 510–528. Springer (2022) 1, 4

    2022

  21. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Feng, Z., Zhang, Z., Yu, X., Fang, Y., Li, L., Chen, X., Lu, Y., Liu, J., Yin, W., Feng, S., et al.: Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10135– 10145 (2023) 4

    2023

  22. [22]

    In: 2022 Picture Coding Symposium (PCS)

    Harell, A., De Andrade, A., Bajić, I.V.: Rate-distortion in image coding for ma- chines. In: 2022 Picture Coding Symposium (PCS). pp. 199–203. IEEE (2022) 4

    2022

  23. [23]

    He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression (2021),https://arxiv.org/abs/2103.15306 4

    work page arXiv 2021

  24. [24]

    Hu, Y., Yang, S., Yang, W., Duan, L.Y., Liu, J.: Towards coding for human and machinevision:Ascalableimagecodingapproach.arXivpreprintarXiv:2001.02915 (2020) 4

    work page arXiv 2001

  25. [25]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Iwai, S., Miyazaki, T., Omachi, S.: Controlling rate, distortion, and realism: To- wards a single comprehensive neural image compression model. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2900–2909 (2024) 1, 4

    2024

  26. [26]

    In: ICASSP 2021-2021 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Le, N., Zhang, H., Cricri, F., Ghaznavi-Youvalari, R., Rahtu, E.: Image coding for machines: an end-to-end learned approach. In: ICASSP 2021-2021 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1590–1594. IEEE (2021) 1, 4

    2021

  27. [27]

    Lee, J., Cho, S., Beack, S.K.: Context-adaptive entropy model for end-to-end op- timized image compression (2019),https://arxiv.org/abs/1809.104524

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  28. [28]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Lee, Y., Kim, J., Go, H., Jeong, M., Oh, S., Choi, S.: Multi-architecture multi- expert diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 13427–13436 (2024) 4

    2024

  29. [29]

    GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding

    Lepikhin, D., Lee, H., Xu, Y., Chen, D., Firat, O., Huang, Y., Krikun, M., Shazeer, N., Chen, Z.: Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668 (2020) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  30. [30]

    Li, A., Li, F., Liu, Y., Cong, R., Zhao, Y., Bai, H.: Once-for-all: Controllable generative image compression with dynamic granularity adaption (2024),https: //arxiv.org/abs/2406.007583

    work page arXiv 2024

  31. [31]

    MiniMax-01: Scaling Foundation Models with Lightning Attention

    Li, A., Gong, B., Yang, B., Shan, B., Liu, C., Zhu, C., Zhang, C., Guo, C., Chen, D., Li, D., et al.: Minimax-01: Scaling foundation models with lightning attention. arXiv preprint arXiv:2501.08313 (2025) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Li, H., Li, S., Dai, W., Li, C., Zou, J., Xiong, H.: Frequency-aware transformer for learned image compression (2024),https://arxiv.org/abs/2310.163873

    work page arXiv 2024

  33. [33]

    Li, H., Li, S., Ding, S., Dai, W., Cao, M., Li, C., Zou, J., Xiong, H.: Image compres- sionformachineandhumanvisionwithspatial-frequencyadaptation.In:European Conference on Computer Vision. pp. 382–399. Springer (2024) 1, 4

    2024

  34. [34]

    IEEE Journal on Emerging and Selected Topics in Circuits and Systems14(2), 198–208 (2024) 4 MoECodec17

    Li, H., Zhang, X.: Human–machine collaborative image compression method based on implicit neural representations. IEEE Journal on Emerging and Selected Topics in Circuits and Systems14(2), 198–208 (2024) 4 MoECodec17

    2024

  35. [35]

    Conditional Computation for Continual Learning

    Lin, M., Fu, J., Bengio, Y.: Conditional computation for continual learning. arXiv preprint arXiv:1906.06635 (2019) 4

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  36. [36]

    In: European conference on computer vision

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740–755. Springer (2014) 8

    2014

  37. [37]

    DeepSeek-V3 Technical Report

    Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al.: Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  38. [38]

    In: 2022 26th International Conference on Pattern Recognition (ICPR)

    Liu, J., Sun, H., Katto, J.: Improving multiple machine vision tasks in the com- pressed domain. In: 2022 26th International Conference on Pattern Recognition (ICPR). pp. 331–337. IEEE (2022) 1, 4

    2022

  39. [39]

    Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures (2023),https://arxiv.org/abs/2303.149783, 11

    work page arXiv 2023

  40. [40]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer- cnn architectures. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14388–14397 (2023) 8, 9

    2023

  41. [41]

    In: Proceedings of the 31st ACM International Conference on Multimedia

    Liu,L.,Hu,Z.,Chen,Z.,Xu,D.:Icmh-net:Neuralimagecompressiontowardsboth machine vision and human vision. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8047–8056 (2023) 1, 4, 8, 9, 11, 13

    2023

  42. [42]

    Lu, M., Guo, P., Shi, H., Cao, C., Ma, Z.: Transformer-based image compression (2021),https://arxiv.org/abs/2111.067073, 8, 9, 11, 13

    work page arXiv 2021

  43. [43]

    Ma, G., Chai, Y., Jiang, T., Lu, M., Chen, T.: Tinylic-high efficiency lossy image compression method (2024),https://arxiv.org/abs/2402.111648, 9, 11

    work page arXiv 2024

  44. [44]

    Minnen, D., Ballé, J., Toderici, G.: Joint autoregressive and hierarchical priors for learned image compression (2018),https://arxiv.org/abs/1809.027364

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  45. [45]

    Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression (2020),https://arxiv.org/abs/2007.087394

    work page arXiv 2020

  46. [46]

    In: European Conference on Computer Vision

    Park, B., Go, H., Kim, J.Y., Woo, S., Ham, S., Kim, C.: Switch diffusion trans- former: Synergizing denoising tasks with sparse mixture-of-experts. In: European Conference on Computer Vision. pp. 461–477. Springer (2024) 4

    2024

  47. [47]

    arXiv preprint arXiv:2310.07138 (2023) 4

    Park, B., Woo, S., Go, H., Kim, J.Y., Kim, C.: Denoising task routing for diffusion models. arXiv preprint arXiv:2310.07138 (2023) 4

    work page arXiv 2023

  48. [48]

    arXiv preprint arXiv:2009.13239 (2020) 4

    Puigcerver, J., Riquelme, C., Mustafa, B., Renggli, C., Pinto, A.S., Gelly, S., Key- sers, D., Houlsby, N.: Scalable transfer learning with expert models. arXiv preprint arXiv:2009.13239 (2020) 4

    work page arXiv 2009

  49. [49]

    Qian, Y., Lin, M., Sun, X., Tan, Z., Jin, R.: Entroformer: A transformer-based entropy model for learned image compression (2022),https://arxiv.org/abs/22 02.054924

    2022

  50. [50]

    Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer

    Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., Dean, J.: Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538 (2017) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  51. [51]

    arXiv preprint arXiv:2503.14487 (2025)

    Shi, M., Yuan, Z., Yang, H., Wang, X., Zheng, M., Tao, X., Zhao, W., Zheng, W., Zhou, J., Lu, J., et al.: Diffmoe: Dynamic token selection for scalable diffusion transformers. arXiv preprint arXiv:2503.14487 (2025) 4

    work page arXiv 2025

  52. [52]

    In: Proceedings of the IEEE/CVF interna- tional conference on computer vision

    Song, M., Choi, J., Han, B.: Variable-rate deep image compression through spatially-adaptive feature transform. In: Proceedings of the IEEE/CVF interna- tional conference on computer vision. pp. 2380–2389 (2021) 1, 4

    2021

  53. [53]

    arXiv preprint arXiv:2410.02098 (2024)

    Sun, H., Lei, T., Zhang, B., Li, Y., Huang, H., Pang, R., Dai, B., Du, N.: Ec-dit: Scaling diffusion transformers with adaptive expert-choice routing. arXiv preprint arXiv:2410.02098 (2024) 4 18 J. Zhao et al

    work page arXiv 2024

  54. [54]

    In: CVPR (2020) 9

    Toderici, G., Shi, W., Timofte, R., Theis, L., Balle, J., Agustsson, E., Johnston, N., Mentzer, F.: Workshop and challenge on learned image compression (clic2020). In: CVPR (2020) 9

    2020

  55. [55]

    IEEE Transactions on Multimedia24, 3169–3181 (2021) 4

    Wang, S., Wang, S., Yang, W., Zhang, X., Wang, S., Ma, S., Gao, W.: Towards analysis-friendly face representation with scalable feature and texture compression. IEEE Transactions on Multimedia24, 3169–3181 (2021) 4

    2021

  56. [56]

    IEEE Transactions on Circuits and Sys- tems for Video Technology33(6), 2979–2989 (2022) 1, 4

    Wang, S., Wang, Z., Wang, S., Ye, Y.: Deep image compression toward machine vision: A unified optimization framework. IEEE Transactions on Circuits and Sys- tems for Video Technology33(6), 2979–2989 (2022) 1, 4

    2022

  57. [57]

    Advances in Neural Infor- mation Processing Systems36, 41693–41706 (2023) 4

    Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to- image generation via large mixture of diffusion paths. Advances in Neural Infor- mation Processing Systems36, 41693–41706 (2023) 4

    2023

  58. [58]

    IEEE Transactions on Image Processing30, 8939–8954 (2021) 4

    Yan, N., Gao, C., Liu, D., Li, H., Li, L., Wu, F.: Sssic: Semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing30, 8939–8954 (2021) 4

    2021

  59. [59]

    Yang, F., Herranz, L., Cheng, Y., Mozerov, M.G.: Slimmable compressive autoen- coders for practical neural image compression (2022),https://arxiv.org/abs/21 03.157263

    2022

  60. [60]

    Scaling Autoregressive Models for Content-Rich Text-to-Image Generation

    Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.107892(3), 5 (2022) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  61. [61]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolu- tional neural network for mobile devices. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6848–6856 (2018) 3, 8

    2018

  62. [62]

    Advances in Neural Information Pro- cessing Systems37, 71465–71503 (2024) 1, 4

    Zhang, X., Guo, P., Lu, M., Ma, Z.: All-in-one image coding for joint human- machine vision with multi-path aggregation. Advances in Neural Information Pro- cessing Systems37, 71465–71503 (2024) 1, 4

    2024

  63. [63]

    In: The Eleventh International Conference on Learning Representations (2022) 4

    Zhenxing, M., Xu, D.: Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In: The Eleventh International Conference on Learning Representations (2022) 4

    2022

  64. [64]

    Advances in Neu- ral Information Processing Systems35, 7103–7114 (2022) 2, 4, 6

    Zhou, Y., Lei, T., Liu, H., Du, N., Huang, Y., Zhao, V., Dai, A.M., Le, Q.V., Laudon, J., et al.: Mixture-of-experts with expert choice routing. Advances in Neu- ral Information Processing Systems35, 7103–7114 (2022) 2, 4, 6

    2022

  65. [65]

    ST-MoE: Designing Stable and Transferable Sparse Expert Models

    Zoph, B., Bello, I., Kumar, S., Du, N., Huang, Y., Dean, J., Shazeer, N., Fedus, W.: St-moe: Designing stable and transferable sparse expert models. arxiv 2022. arXiv preprint arXiv:2202.08906 4

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  66. [66]

    Zou, R., Song, C., Zhang, Z.: The devil is in the details: Window-based attention for image compression (2022),https://arxiv.org/abs/2203.084503

    work page arXiv 2022