Pith. sign in

REVIEW 3 major objections 5 minor 72 references

Visual Token Codec: Unleashing Spatial Redundancy for ViT Feature Coding

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The paper claims that VTC, a dual-path token codec, cuts the bitrate needed to preserve 90 percent of uncompressed ViT feature performance by 15.7x–37.4x compared with a VTM-based baseline.

desk verdict The dual-path split of global vs patch tokens is a clean, well-validated idea, but the headline bitrate multipliers rest on a fragile operating point chosen after BD metrics failed. read the letter →

arxiv 2608.08832 v1 pith:HZAFCOLQ submitted 2026-08-09 cs.CV

Donghui Feng , Fengxi Zhang , Changsheng Gao , Wenhan Yang , Qi Wang , Qunshan Gu , Hongwei Hu , Zhengxue Cheng
show 1 more author
Li Song
This is my paper · ORCID
classification cs.CV
keywords visualtokencodecViTfeaturecodingspatial-channelcontextentropymodelpatch-tokengridfeature-matchingsupervisionvariable-rateDINOv2SAM3
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that when a Vision Transformer is split across devices, the intermediate token features exchanged between nodes can be compressed much more aggressively than current methods do, because patch tokens still sit on their original 2D image grid and retain strong local spatial correlation. Existing ViT feature codecs flatten all tokens into an $L\times C$ pseudo image, so their entropy models mostly exploit sequence-axis dependence and leave grid redundancy untouched. VTC instead routes a few global summary tokens through a lightweight factorized 1D codec and codes the many patch tokens on their native $H\times W$ grid with a spatial–channel context entropy model. The paper reports that at 90% of uncompressed-feature (bypass) performance, VTC needs 15.7x–37.4x less bitrate than the LaMoFC-VTM baseline across SAM3 detection and instance segmentation and DINOv2 classification and semantic segmentation. A feature-matching loss and a variable-rate module make the codec usable at intermediate layers and over 65 rate points in a single model.

What carries the argument

The central object is the dual-path token layout: a short 1D path for global summary tokens and a 2D $H\times W$ grid path for patch tokens. The grid path uses a Spatial-Channel Context (SCCTX) entropy model, which partitions the latent into channel groups and checkerboard spatial positions, then predicts conditional Gaussian parameters for each position from hyperprior side information and from previously decoded spatial and channel contexts; this is the component that harvests the local 2D redundancy. Two auxiliary mechanisms carry the deployment claims: feature-matching supervision, which pushes reconstructed tokens through the next $j$ frozen ViT blocks and adds mean-squared error on those deeper features to the rate–distortion loss, and a variable-rate module bank that conditions the codec on a quantization parameter selecting one of 65 Lagrange multipliers, letting one model span multiple bitrates.

What would settle it

Run both codecs at many more rate points on DINOv2 classification and find the actual bitrate at which each reaches 90 percent of bypass accuracy; if VTC's required bitrate is above 0.013 bits per feature point, or the gap to LaMoFC-VTM is smaller than 28.4x, the headline saving does not hold. As a second check, retrain the flattened-token layout with the same nonlinear transforms and feature-matching loss; if it matches VTC's curve, the 2D grid context is not the source of the gains.

Watch

Extended reading notes

Core claim

The central claim is that ViT patch tokens carry exploitable local spatial correlation on their original 2D grid, and the right codec design is to stop treating all tokens as one flattened sequence. VTC's discovery is the dual-path layout: global tokens are coded with a factorized prior on a 1D path, while patch tokens are reshaped back to the $H\times W$ grid and coded by a spatial–channel context entropy model that predicts each quantized latent from hyperprior side information plus already-decoded spatial and channel neighbors. Because this layout matches the entropy model to the token statistics, VTC reaches 90% of bypass performance at 0.009–0.013 bits per feature point where the LaMoFC-VTM baseline needs 0.164–0.498, giving the 15.7x–37.4x savings. In the multi-task setting, one compressed bitstream serves both segmentation and reconstruction, and compressing layer 9 with feature-matching supervision at layer 12 is the best tested rate–distortion operating point for ADE20K segmentation. The paper also demonstrates that larger backbones are not automatically better under a pixel-normalized bitrate metric.

Load-bearing premise

The headline savings rest on the assumption that the comparison point—the bitrate at which a codec reaches 90 percent of the uncompressed feature's task performance—is a well-defined and fair measure on every task, and on DINOv2 classification the accuracy curve is nearly flat with bitrate, so the reported figure is an upper bound at the lowest tested rate rather than a directly measured crossover.

Editorial extensions

If this is right

  • At 90% of uncompressed-feature performance, VTC needs 0.009–0.013 bits per feature point versus 0.164–0.498 for LaMoFC-VTM, translating to 15.7x–37.4x lower bitrate across the four task-specific settings.
  • Dual-path coding beats both flattening all tokens into one sequence and padding global tokens into the 2D grid; removing the codec's nonlinear transform blocks collapses performance by 15.51 BD-mIoU in the ablation.
  • Feature-matching supervision largely determines the rate–distortion behavior: curves sharing the same supervision layer follow similar trajectories, and compressing layer 9 with supervision at layer 12 is the best ADE20K segmentation operating point among the tested configurations.
  • A single VTC model with a QP-conditioned module bank covers 65 rate points, so rate adaptation does not require retraining per bitrate.
  • A single multi-task bitstream from DINOv2-B/16-reg4 supports both semantic segmentation and subjective reconstruction; at LPIPS=0.484 VTC uses 17.7x less bitrate than LaMoFC-VTM.

Reading between the lines

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

  • If the patch-grid prior is the main source of savings, similar dual-path layouts should transfer to other dense-grid ViT features beyond DINOv2 and SAM3; that is testable by running VTC's token split on such features without retraining the codec.
  • The flat DINOv2 classification rate–accuracy curve hints that a classification-only codec could spend nearly all bits on global tokens and almost nothing on patch detail, which would make dense tasks the real constraint on bitrate.
  • Because the supervision layer shapes the rate–distortion trend more than the compressed layer does, a practical design rule would be to pick the feature-matching layer according to the downstream task head rather than the transmission point.
  • The reported multipliers are measured against a VTM-based baseline; an independent learned baseline that also flattens tokens but uses the same feature-matching supervision would isolate how much of the gain is the 2D layout itself.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Visual Token Codec (VTC), a learned codec for compressing intermediate ViT features in split-inference and feature-storage settings. The key design is a dual-path architecture that codes the few global tokens with a lightweight factorized prior and codes the patch tokens on their native 2D grid with a spatial-channel context entropy model (SCCTX from ELIC), instead of flattening all tokens into a pseudo-image. VTC also adds feature-matching supervision on deeper ViT blocks and a variable-rate mechanism. The authors evaluate on DINOv2 (classification, semantic segmentation, reconstruction) and SAM3 (object detection, instance segmentation) and report that at 90% of uncompressed-feature performance, VTC reduces bitrate by 15.7x–37.4x relative to LaMoFC-VTM. They additionally provide intermediate-layer rate-distortion trade-offs, a backbone-scale study, and ablations showing that the dual-path layout outperforms all-1D and all-2D layouts under fixed transforms.

Significance. If the quantitative claims hold, the paper makes a useful contribution to ViT feature coding: it identifies a structural prior (patch-token spatial locality) that prior pseudo-image codecs miss, and it validates the design with extensive experiments on two modern backbones and five tasks, using substantially larger test sets than earlier feature-coding studies. The ablation in Fig. 9 is well designed and supports the qualitative central claim that layout matters independently of transform capacity. The intermediate-layer deployment analysis in Sec. IV-C and Table III is a valuable addition for practical bandwidth/computation trade-offs. The manuscript is also transparent about where BD metrics could not be computed. However, the headline bitrate-reduction magnitudes rest on a post-hoc operating point and on single-point ratios without uncertainty quantification, so the exact multipliers are less secure than the qualitative ranking.

major comments (3)
  1. [Sec. IV-B1, Table II] The headline 15.7x–37.4x multipliers are computed at a 90%-of-bypass operating point adopted only after BD metrics were abandoned because the rate curves have limited overlap. For the DINOv2 Cls row, Table II reports VTC as '≤0.013 bpfp', which is the lowest tested rate on a nearly horizontal accuracy curve rather than a measured crossing at the 90% target. This makes the claimed multipliers protocol-dependent: the DINOv2 Cls entry is an inequality, and if a 95% or 99% threshold were used, or if denser low-rate points were added, the reported gains could shift substantially. Please report the DINOv2 Cls entry explicitly as an inequality, either exclude it from the headline range or provide denser low-rate measurements, and state that the remaining multipliers are single-point estimates rather than BD-derived averages.
  2. [Sec. IV-A4 and Sec. IV-B1] All compared baselines (LaMoFC-VTM, VQFC, and the uncompressed bypass) come from the same research group or the authors' previous work, and no uncertainty is reported for any rate-performance point. Since the quantitative magnitude is the main contribution beyond the qualitative ablation, the paper should report error bars or multiple evaluation runs for the Table II bitrate estimates, and ideally include an independent off-the-shelf learned image codec (for example, applying ELIC to the flattened pseudo-image layout) as an additional baseline. Such a baseline would also directly test whether the gains come from the layout prior rather than from codec choice.
  3. [Sec. IV-C, Table III] The intermediate-layer analysis is informative, but the BD metrics in Table III are computed relative to a single hand-selected operating point (Input=9, Spv=12, fa=fs=2), and rows with '–' in BD-rate are not explained. Because the choice of reference row substantially determines the reported BD-Rate percentages, the paper should clarify how sensitive the conclusions are to this reference choice and should report the full rate-performance curves for the key rows, not only in the appendix text but as accessible data, so that readers can compare operating points directly.
minor comments (5)
  1. [Sec. III-E] The 1:1 weighting between global-token and patch-token distortion is described as an empirical choice, but no ablation or sensitivity study is provided; a small experiment varying this weighting would strengthen the claim that the choice is robust.
  2. [Table II caption] The notation '≤0.013' in the DINOv2 Cls row should be explained in the caption, since it denotes an upper bound from the lowest tested rate rather than a measured crossing point.
  3. [Table III] Rows 4–5 are first discussed in Sec. IV-C but are later used as codec-transform ablations in Sec. IV-E; adding a forward pointer in the Table III discussion would avoid confusion.
  4. [Table IV] The text states that ImageNet BD-rate is omitted because the curves are nearly horizontal; please clarify whether BD-rate is undefined or simply not reported, and whether the flatness is due to the accuracy range or the rate range.
  5. [Fig. 9] The all-1D, all-2D, and dual-path curves are visually close at low bitrates; consider reporting the BD-mIoU or BD-rate differences between the three layouts numerically to make the ablation comparison easier to assess.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: VTC's gains are measured against external task metrics; self-cited baselines and the 90%-of-bypass threshold are evaluation choices, not circular inputs.

full rationale

I walked the paper's derivation chain from the dual-path codec (Sec. III-B) through training losses (Sec. III-E), variable-rate conditioning (Sec. III-F), and the main experimental claims (Sec. IV-B). No step reduces a prediction to its input by construction. VTC is an empirical codec: its rate-utility claims are evaluated on frozen DINOv2/SAM3 backbones with external task metrics (top-1 accuracy, mIoU, mAP, LPIPS), not on quantities fitted from the same data. The dual-path design is justified by a measured correlation observation (Fig. 1) and by an ablation (Fig. 9) that fixes the ViT transform blocks and varies only the token layout, which is the correct control for the claimed mechanism. The paper does adopt existing entropy models (factorized prior [12], SCCTX [13]) and existing variable-rate ideas [32,66,67], but these are standard external building blocks, not self-citations carrying the argument. The primary baseline LaMoFC-VTM [7] is from the same research group as several current authors, and VQFC [10] is also from that group; however, these are empirical comparison points, not evidence invoked to justify the codec design. The text explicitly says BD metrics could not be computed due to limited curve overlap and that the paper therefore compares bitrate at 90% of bypass performance: 'Because the curves of different methods have limited overlap, BD metrics cannot be computed reliably here; we therefore compare the bitrate required to reach 90% of the bypass (uncompressed-feature) performance.' This is a protocol fallback, not a circular reduction; the 90% target is anchored to the uncompressed-feature performance of the frozen backbone, independent of VTC's parameters. The DINOv2 Cls entry '0.375 ≤0.013' is an inequality, meaning the reported 28.4x is a lower bound on the bitrate reduction, not a fitted or inflated number; if anything, denser low-rate points could strengthen it. The 1:1 global/patch weighting is stated openly as an empirical choice ('The 1:1 weighting between MSE1d and MSE2d is an empirical choice'), so it is a selection choice rather than a concealed fit. The intermediate-layer analysis (Sec. IV-C) is also empirical and internally consistent. In sum, the paper is self-contained against external benchmarks, and the cited self-works are not load-bearing premises. Any concern about threshold sensitivity or baseline representativeness is a correctness/robustness issue, not circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim is empirical, so the ledger lists the hand-chosen training and architecture parameters that shape the reported rate-utility curves, plus the domain assumptions that the patch-grid prior, the borrowed SCCTX model, the equal global/patch weighting, the baseline choice, and the feature-matching loss are all appropriate. No new physical or conceptual entities are introduced.

free parameters (6)
  • Lagrange multiplier set (lambda) = 65 logarithmically spaced values in [1,256], mapped to QP in [0,64]
    Controls the rate-distortion trade-off in Eq. (18); the QP-to-lambda mapping is hand-defined and sampled uniformly in VBR training (Sec. IV-A3).
  • Global/patch distortion weighting = 1:1 weighting of MSE1d and MSE2d
    Sec. III-E explicitly calls the 1:1 weighting an empirical choice to support both global-heavy classification and patch-heavy segmentation; it is not derived from token counts.
  • Patch latent dimension = 48 (DINOv2-S/B, SAM3), 120 (DINOv2-L)
    Hand-chosen bottleneck of the analysis transform; larger for DINOv2-L to avoid an overly strict information bottleneck (Sec. IV-A2).
  • Codec transform depth (f_a/f_s) = 2/2 ViT blocks for default DINOv2 multi-task point
    Table III shows removing these blocks costs 2.49 BD-mIoU per block and removing all collapses to -15.51 BD-mIoU; depth is a manual architecture choice.
  • Compressed layer and supervision layer = Input=9, Spv=12 for DINOv2-B-Reg4 multi-task
    Selected as the default operating point in Table III (anchored at 0.00% BD-rate); a validation-set choice among tested configurations that affects the intermediate-layer analysis.
  • SAM3 scale-loss balancing scalar alpha = not reported
    Balances multi-scale supervision losses on SAM3 (Sec. IV-A3); the value is undisclosed, so the exact training objective is underspecified.
assumptions (5)
  • domain assumption ViT patch tokens at the compressed layers retain strong local 2D spatial correlation.
    Fig. 1 reports sliding-window Pearson correlations; the entire patch-grid entropy model assumes this prior holds for DINOv2 and SAM3 at the selected layers.
  • domain assumption The SCCTX entropy model, designed for image latents, is an effective entropy model for patch-token latents after the learned transforms.
    Sec. III-D adopts ELIC's SCCTX without modification; if latent statistics differ sharply from natural-image latents, the context model could be suboptimal, and only indirect empirical support is given.
  • ad hoc to paper The 1:1 weighting between global-token and patch-token distortion is a reasonable and generalizable choice.
    Sec. III-E labels this empirical; it is not derived from rate-distortion theory and may not transfer to other backbones or tasks.
  • domain assumption LaMoFC-VTM is a representative strong baseline for ViT feature coding.
    Sec. IV-A4 chooses LaMoFC-VTM as primary baseline and excludes LaMoFC-Hyperprior based on the original study; all baselines (LaMoFC, VQFC) are from the same research group, so external SOTA is not measured.
  • domain assumption Feature-matching supervision on deeper ViT blocks reduces quantization drift and preserves downstream utility.
    The method assumes minimizing distortion at layer n+j, not only at the compressed layer, transfers to frozen task heads; Fig. 7 supports it empirically but it is not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Visual Token Codec: Unleashing Spatial Redundancy for ViT Feature Coding." pith.science (2026). https://pith.science/paper/HZAFCOLQ

@misc{pith2026260808832,
  author       = {Pith},
  title        = {Pith review of: Visual Token Codec: Unleashing Spatial Redundancy for ViT Feature Coding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HZAFCOLQ}},
  note         = {Machine review of arXiv:2608.08832}
}
read the original abstract

Distributed deployment of large vision foundation models often partitions a ViT backbone and exchanges intermediate token features between computing nodes, making efficient feature compression critical under bandwidth and computation constraints. Existing ViT feature codecs typically flatten heterogeneous global and patch tokens into an L x C pseudo image, causing entropy models to mainly capture sequence-axis dependencies while overlooking the native two-dimensional patch-grid structure. In this paper, we show that ViT patch tokens retain strong local spatial correlations on the original grid. To exploit this structural prior, we propose the Visual Token Codec (VTC), a dual-path learned codec that separates global and patch tokens into dedicated coding paths. Global tokens are compressed with a lightweight factorized prior, whereas patch tokens are encoded on the patch-token grid using a spatial-channel context entropy model. To support intermediate-layer compression and practical rate adaptation, VTC further incorporates feature-matching supervision after subsequent ViT blocks and variable-rate modules within a single codec. Experiments on DINOv2 and SAM3 show that VTC consistently outperforms representative ViT feature coding baselines on classification, segmentation, and detection tasks. At 90% of uncompressed-feature performance, VTC reduces bitrate by 15.7x-37.4x across these tasks. We further provide intermediate-layer rate-utility analyses for practical transmission- and storage-oriented deployment scenarios.

Figures

Figures reproduced from arXiv: 2608.08832 by the authors.

Figure 1
Figure 1. Sliding-window 2D Pearson correlations in ViT tokens. Existing [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of ViT feature coding and its two deployment paradigms. The upper part illustrates the generic feature coding pipeline, where an input image [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview and key components of the proposed Visual Token Codec (VTC). (a) Overall architecture. The input feature [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Hyperprior and Spatial-Channel Context (SCCTX) entropy model for [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Rate–performance under task-specific feature coding for two foundation models and four tasks. Bypass: uncompressed features; LaMoFC VTM: VTM [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Rate–performance under the multi-task-compatible setting using DINOv2-Base-Reg4 features on ADE20K val. Bypass: uncompressed features; [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Rate–performance curves under different compressed-layer choices on ADE20K semantic segmentation. Left: varying the compressed input layer. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Rate–performance curves for VTC with different DINOv2-Reg4 backbone scales. Horizontal lines indicate uncompressed baselines. Top: ImageNet [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Rate–performance comparison of dual-path vs. all-1D or all-2D token [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

72 extracted references · 57 canonical work pages

  1. [1]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” inICLR, 2021

    2021

  2. [2]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . B. Huang, S.-W. Li, I. Misra, M. G. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. J ´egou, J. Mairal, P. Labatut, A. Joulin, and P. Bojanowski, “Dinov2: Learning robust visual feature...

    arXiv 2023

  3. [3]

    Siglip 2: Multilin- gual vision-language encoders with improved semantic understanding, localization, and dense features,

    M. Tschannen, A. Gritsenko, X. Wang, M. F. Naeem, I. M. Alab- dulmohsin, N. Parthasarathy, T. Evans, L. Beyer, Y . Xia, B. Mustafa, O. H’enaff, J. Harmsen, A. Steiner, and X.-Q. Zhai, “Siglip 2: Multilin- gual vision-language encoders with improved semantic understanding, localization, and dense features,”ArXiv, vol. abs/2502.14786, 2025

    arXiv 2025

  4. [4]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. K. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. B. Girshick, P. Doll’ar, and C. Feichten- hofer, “Sam 2: Segment anything in images and videos,”ArXiv, vol. abs/2408.00714, 2024

    arXiv 2024

  5. [5]

    SAM 3: Segment anything with concepts,

    N. Carion, L. Gustafson, Y .-T. Hu, R. Hu, D. Suris, C. Ryali, T. Ma et al., “SAM 3: Segment anything with concepts,”arXiv preprint arXiv:2511.16719, 2025

    arXiv 2025

  6. [6]

    OpenFedLLM: Training Large Language Models on Decentralized Private Data via Federated Learning,

    R. Ye, W. Wang, J. Chai, D. Li, Z. Li, Y . Xu, Y . Du, Y . Wang, and S. Chen, “OpenFedLLM: Training Large Language Models on Decentralized Private Data via Federated Learning,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, ser. KDD ’24. Association for Computing Machinery, 2024, pp. 6137–6147

    work page 2024

  7. [7]

    Feature coding in the era of large models: Dataset, test conditions, and benchmark,

    C. Gao, Y . Ma, Q. Chen, Y . Xu, D. Liu, and W. Lin, “Feature coding in the era of large models: Dataset, test conditions, and benchmark,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Oct. 2025, pp. 1068–1077. [Online]. Available: https://arxiv.org/abs/2412.04307

    arXiv 2025

  8. [8]

    Vision transformers need registers,

    T. Darcet, M. Oquab, J. Mairal, and P. Bojanowski, “Vision transformers need registers,” inInternational Conference on Learning Representations (ICLR), 2024. [Online]. Available: https://openreview. net/forum?id=2dnO3LLiJ1

    work page 2024

Show all 72 references
  1. [9]

    DT-UFC: Universal Large Model Feature Coding via Peaky-to-Balanced Distribution Trans- formation,

    C. Gao, Z. Liu, L. Li, D. Liu, X. Sun, and W. Lin, “DT-UFC: Universal Large Model Feature Coding via Peaky-to-Balanced Distribution Trans- formation,”Proceedings of the 33rd ACM International Conference on Multimedia, pp. 5198–5207, Oct. 2025

    2025

  2. [10]

    Transform-Free Feature Coding via Entropy-Constrained Vector Quantization,

    Q. Chen, C. Gao, L. Li, and D. Liu, “Transform-Free Feature Coding via Entropy-Constrained Vector Quantization,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 1. AAAI Press, 2026

    2026

  3. [11]

    VVC and VTM 11.0: The Versatile Video Coding standard and its reference software,

    B. Bross, Y . Wang, J. Ye, Y . Chen, H. Chen, G. J. Sullivan, and J.- R. Ohm, “VVC and VTM 11.0: The Versatile Video Coding standard and its reference software,” inProceedings of the ACM International Conference on Multimedia, 2021

    2021

  4. [12]

    Variational image compression with a scale hyperprior,

    J. Ball ´e, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,” inInternational Conference on Learning Representations, 2018

    2018

  5. [13]

    ELIC: Efficient Learned Image Compression with Unevenly Grouped Space- Channel Contextual Adaptive Coding,

    D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y . Wang, “ELIC: Efficient Learned Image Compression with Unevenly Grouped Space- Channel Contextual Adaptive Coding,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 5708– 5717

    2022

  6. [14]

    Nonlinear Transform Coding,

    J. Balle, P. A. Chou, D. Minnen, S. Singh, N. Johnston, E. Agustsson, S. J. Hwang, and G. Toderici, “Nonlinear Transform Coding,”IEEE Journal of Selected Topics in Signal Processing, vol. 15, no. 2, pp. 339–353, Feb. 2021

    2021

  7. [15]

    End-to-end optimized image compression,

    J. Ball ´e, V . Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,” in5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. [Online]. Available: https://o...

    2017

  8. [16]

    Joint Autoregressive and Hierar- chical Priors for Learned Image Compression,

    D. Minnen, J. Ball ´e, and G. Toderici, “Joint Autoregressive and Hierar- chical Priors for Learned Image Compression,” inAdvances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, 3-8 December 2018, Mon...

    2018

  9. [17]

    Learned Image Com- pression With Discretized Gaussian Mixture Likelihoods and Attention Modules,

    Z. Cheng, H. Sun, M. Takeuchi, and J. Katto, “Learned Image Com- pression With Discretized Gaussian Mixture Likelihoods and Attention Modules,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 7939–7948

    2020

  10. [18]

    Enhanced Invertible Encoding for Learned Image Compression,

    Y . Xie, K. L. Cheng, and Q. Chen, “Enhanced Invertible Encoding for Learned Image Compression,” pp. 162–170, 2021. [Online]. Available: https://dl.acm.org/doi/10.1145/3474085.3475213

    2021

  11. [19]

    H. Tu, S. Wu, L. Li, W. Zhou, and H. Li. Multi-Scale Invertible Neural Network for Wide-Range Variable-Rate Learned Image Compression

  12. [20]

    End-to-End Optimized Versatile Image Compression With Wavelet-Like Transform,

    H. Ma, D. Liu, N. Yan, H. Li, and F. Wu, “End-to-End Optimized Versatile Image Compression With Wavelet-Like Transform,” pp. 1–1

  13. [21]

    Transformer-based Transform Coding,

    Y . Zhu, Y . Yang, and T. Cohen, “Transformer-based Transform Coding,” inProceedings of the International Conference on Learning Represen- tations (ICLR), 2021

    2021

  14. [22]

    Learned Image Compression with Mixed Transformer-CNN Architectures,

    J. Liu, H. Sun, and J. Katto, “Learned Image Compression with Mixed Transformer-CNN Architectures,” in2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023, pp. 14 388–14 397

    2023

  15. [23]

    Frequency-Aware Transformer for Learned Image Compression,

    H. Li, S. Li, W. Dai, C. Li, J. Zou, and H. Xiong, “Frequency-Aware Transformer for Learned Image Compression,” inProceedings of the International Conference on Learning Representations (ICLR), 2023

    2023

  16. [24]

    M. Lu, F. Chen, S. Pu, and Z. Ma. High-Efficiency Lossy Image Coding Through Adaptive Neighborhood Information Aggregation

  17. [25]

    S. Qin, J. Wang, Y . Zhou, B. Chen, T. Luo, B. An, T. Dai, S. Xia, and Y . Wang. MambaVC: Learned Visual Compression with Selective State Spaces

  18. [26]

    MambaIC: State Space Models for High-Performance Learned Image Compres- sion,

    F. Zeng, H. Tang, Y . Shao, S. Chen, L. Shao, and Y . Wang, “MambaIC: State Space Models for High-Performance Learned Image Compres- sion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 18 041–18 050

    2025

  19. [27]

    Z. Wu, H. Du, S. Wang, M. Lu, H. Sun, Y . Guo, and X. Yu. CMamba: Learned Image Compression with State Space Models

  20. [28]

    Y . Chen, Z. Lyu, B. He, H. Hu, Q. Wang, Y . Tian, L. Song, W. Zhang, and G. Lu. CMIC: Content-Adaptive Mamba for Learned Image Com- pression

  21. [29]

    Linear Attention Modeling for Learned Image Compression,

    D. Feng, Z. Cheng, S. Wang, R. Wu, H. Hu, G. Lu, and L. Song, “Linear Attention Modeling for Learned Image Compression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 7623–7632

    2025

  22. [30]

    Distilling complexity-scalable learned image compression models via neural architecture search,

    S. Wang, Z. Cheng, D. Feng, Q. Wang, G. Lu, L. Song, and W. Zhang, “Distilling complexity-scalable learned image compression models via neural architecture search,”IEEE Transactions on Circuits and Systems for Video Technology, 2026. [Online]. Available: https://ieeexplore.iee...

    2026

  23. [31]

    EVC: Towards Real-Time Neural Image Compression with Mask Decay,

    G.-H. Wang, J. Li, B. Li, and Y . Lu, “EVC: Towards Real-Time Neural Image Compression with Mask Decay,” inInternational Conference on Learning Representations (ICLR), 2023. [Online]. Available: https://openreview.net/forum?id=XUxad2Gj40n

    2023

  24. [32]

    Towards practical real-time neural video compression,

    Z. Jia, B. Li, J. Li, W. Xie, L. Qi, H. Li, and Y . Lu, “Towards practical real-time neural video compression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 12 543–12 552

    2025

  25. [33]

    Checkerboard context model for efficient learned image compression,

    D. He, Y . Zheng, B. Sun, Y . Wang, and H. Qin, “Checkerboard context model for efficient learned image compression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 14 771–14 780

    2021

  26. [34]

    Channel-Wise Autoregressive Entropy Models for Learned Image Compression,

    D. Minnen and S. Singh, “Channel-Wise Autoregressive Entropy Models for Learned Image Compression,” in2020 IEEE International Confer- ence on Image Processing (ICIP). IEEE, 2020, pp. 3339–3343

    2020

  27. [35]

    MLIC: Multi-Reference Entropy Model for Learned Image Compression,

    W. Jiang, J. Yang, Y . Zhai, P. Ning, F. Gao, and R. Wang, “MLIC: Multi-Reference Entropy Model for Learned Image Compression,” pp. 7618–7627

  28. [36]

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

    W. Jiang, J. Yang, Y . Zhai, F. Gao, and R. Wang, “MLIC++: Linear Complexity Multi-Reference Entropy Modeling for Learned Image Compression.”

  29. [37]

    Y . Qian, M. Lin, X. Sun, Z. Tan, and R. Jin. Entroformer: A Transformer- based Entropy Model for Learned Image Compression

  30. [38]

    Contextformer: A Transformer with Spatio-Channel Attention for Context Modeling in Learned Image Compression,

    A. B. Koyuncu, H. Gao, A. Boev, G. Gaikov, E. Alshina, and E. Stein- bach, “Contextformer: A Transformer with Spatio-Channel Attention for Context Modeling in Learned Image Compression,” inComputer Vision – ECCV 2022, S. Avidan, G. Brostow, M. Ciss ´e, G. M. Farinella, and T. ...

    2022

  31. [39]

    Efficient Contextformer: Spatio-Channel Window Attention for Fast Context Modeling in Learned Image Compression,

    A. B. Koyuncu, H. Gao, A. Boev, G. Gaikov, E. Alshina, E. Steinbach, A. B. Koyuncu, P. Jia, A. Boev, E. Alshina, and E. Steinbach, “Efficient Contextformer: Spatio-Channel Window Attention for Fast Context Modeling in Learned Image Compression,” vol. 34, no. 8, pp. 7498– 7511

  32. [40]

    M. Li, K. Zhang, W. Zuo, R. Timofte, and D. Zhang. Learning Context-Based Non-local Entropy Modeling for Image Compression. [Online]. Available: http://arxiv.org/abs/2005.04661 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY 13

    2005 arXiv

  33. [41]

    Video coding for machines: A paradigm of collaborative compression and intelligent analytics,

    L. Duan, J. Liu, W. Yang, T. Huang, and W. Gao, “Video coding for machines: A paradigm of collaborative compression and intelligent analytics,”IEEE Transactions on Image Processing, vol. 29, pp. 8680– 8695, 2020

    2020

  34. [42]

    Video coding for machines: Compact visual representation compression for intelligent collaborative analytics,

    W. Yang, H. Huang, Y . Hu, L.-Y . Duan, and J. Liu, “Video coding for machines: Compact visual representation compression for intelligent collaborative analytics,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 7, pp. 5174–5191, 2024

    2024

  35. [43]

    Human-Machine Collaborative Image and Video Compression: A Survey,

    H. Li, X. Zhang, S. Wang, S. Wang, and J. Pan, “Human-Machine Collaborative Image and Video Compression: A Survey,” vol. 13, no. 6

  36. [44]

    Call for proposals on feature compression for video coding for machines,

    W. April, “Call for proposals on feature compression for video coding for machines,”ISO/IEC JTC, vol. 1, 2023. [Online]. Available: https://www.mpeg.org/standards/Explorations/34/

    2023

  37. [45]

    Toward Intelligent Sensing: Intermediate Deep Feature Compression,

    Z. Chen, K. Fan, S. Wang, L. Duan, W. Lin, and A. C. Kot, “Toward Intelligent Sensing: Intermediate Deep Feature Compression,”IEEE Transactions on Image Processing, vol. 29, pp. 2230–2243, 2020

    2020

  38. [46]

    AlphaVC: High-Performance and Efficient Learned Video Compression,

    Y . Shi, Y . Ge, J. Wang, and J. Mao, “AlphaVC: High-Performance and Efficient Learned Video Compression,” inComputer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XIX. Springer-Verlag, pp. 616–631

    2022

  39. [47]

    MIMT: Masked Image Modeling Transformer for Video Compression,

    J. Xiang, K. Tian, and J. Zhang, “MIMT: Masked Image Modeling Transformer for Video Compression,” 2023. [Online]. Available: https://openreview.net/forum?id=j9m-mVnndbm

    2023

  40. [48]

    FLA VC: Learned Video Compression with Feature Level Attention,

    C. Zhang, H. Sun, and J. Katto, “FLA VC: Learned Video Compression with Feature Level Attention,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 28 019–28 028. [Online]. Available: https://openaccess. thecvf.com/content/CVPR2025...

    2025

  41. [49]

    End-to-End Learnable Multi-Scale Feature Compression for VCM,

    Y . Kim, H. Jeong, J. Yu, Y . Kim, J. Lee, S. Y . Jeong, and H. Y . Kim, “End-to-End Learnable Multi-Scale Feature Compression for VCM,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 5, pp. 3156–3167, 2024. [Online]. Available: https://arxiv.org/a...

    2024 arXiv

  42. [50]

    Learnt Mutual Feature Compression for Machine Vision,

    T. Liu, M. Xu, S. Li, C. Chen, L. Yang, and Z. Lv, “Learnt Mutual Feature Compression for Machine Vision,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5. [Online]. Available: https://doi.org/10.1109/ICASSP...

    2023

  43. [51]

    Sensitivity-aware bit allocation for intermediate deep feature compression,

    Y . Hu, S. Xia, W. Yang, and J. Liu, “Sensitivity-aware bit allocation for intermediate deep feature compression,” in2020 IEEE International Conference on Visual Communications and Image Processing (VCIP). IEEE, 2020

    2020

  44. [52]

    Feature compression with 3d sparse convolution,

    Y . Ma, C. Gao, Q. Chen, L. Li, D. Liu, and X. Sun, “Feature compression with 3d sparse convolution,” in2024 IEEE International Conference on Visual Communications and Image Processing (VCIP). IEEE, 2024

    2024

  45. [53]

    Latent-Space Scalability for Multi-Task Collaborative Intelligence,

    H. Choi and I. V . Baji ´c, “Latent-Space Scalability for Multi-Task Collaborative Intelligence,” in2021 IEEE International Conference on Image Processing (ICIP), 2021, pp. 3562–3566. [Online]. Available: https://arxiv.org/abs/2105.10089

    2021 arXiv

  46. [54]

    DMOFC: Discrimination Metric-Optimized Feature Compression,

    C. Gao, Y . Jiang, L. Li, D. Liu, and F. Wu, “DMOFC: Discrimination Metric-Optimized Feature Compression,” in2024 Picture Coding Symposium (PCS). IEEE, 2024, pp. 1–5. [Online]. Available: https://arxiv.org/abs/2405.04044

    2024 arXiv

  47. [55]

    IMOFC: Identity-Level Metric Optimized Feature Compression for Identification Tasks,

    C. Gao, Y . Jiang, S. Wu, Y . Ma, L. Li, and D. Liu, “IMOFC: Identity-Level Metric Optimized Feature Compression for Identification Tasks,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 2, pp. 1855–1869, 2025. [Online]. Available: https://ieeexplo...

    2025

  48. [56]

    Rethinking joint optimization in feature compression: Insights from person re- identification,

    C. Gao, Z. Li, L. Li, D. Liu, F. Wu, and W. Lin, “Rethinking joint optimization in feature compression: Insights from person re- identification,” in2025 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2025, pp. 1–6

    2025

  49. [57]

    Compressed feature quality assessment: Dataset and baselines,

    C. Gao, W. Zhou, G. Lin, and W. Lin, “Compressed feature quality assessment: Dataset and baselines,” inACMMM 2025, 2025

    2025

  50. [58]

    Scalable Facial Image Compression with Deep Feature Reconstruction,

    S. Wang, S. Wang, X. Zhang, S. Wang, S. Ma, and W. Gao, “Scalable Facial Image Compression with Deep Feature Reconstruction,” in2019 IEEE International Conference on Image Processing (ICIP), 2019, pp. 2691–2695

    2019

  51. [59]

    Towards coding for human and machine vision: Scalable face image coding,

    S. Yang, Y . Hu, W. Yang, L. Duan, and J. Liu, “Towards coding for human and machine vision: Scalable face image coding,”IEEE Transactions on Multimedia, vol. 23, pp. 2957–2971, 2021

    2021

  52. [60]

    SSSIC: Semantics-to- Signal Scalable Image Coding With Learned Structural Representations,

    N. Yan, C. Gao, D. Liu, H. Li, L. Li, and F. Wu, “SSSIC: Semantics-to- Signal Scalable Image Coding With Learned Structural Representations,” IEEE Transactions on Image Processing, vol. 30, pp. 8939–8954, 2021

    2021

  53. [61]

    Semantics-to-Signal Scalable Image Compression with Learned Revertible Representations,

    K. Liu, D. Liu, L. Li, N. Yan, and H. Li, “Semantics-to-Signal Scalable Image Compression with Learned Revertible Representations,”Interna- tional Journal of Computer Vision, vol. 129, no. 9, pp. 2605–2621, 2021

    2021

  54. [62]

    End-to-End Learned Scalable Multilayer Feature Compression for Machine Vision Tasks,

    Q. Chen, C. Gao, and D. Liu, “End-to-End Learned Scalable Multilayer Feature Compression for Machine Vision Tasks,” in2024 IEEE International Conference on Image Processing (ICIP), 2024, pp. 1781–1787. [Online]. Available: https://cmsworkshops.com/ICIP2024/ view paper.php?Pape...

    2024

  55. [63]

    Towards large model feature coding,

    Y . Pang, C. Gao, D. Liu, H. Lu, and W. Lin, “Towards large model feature coding,”arXiv preprint arXiv:2605.24025, 2026

    2026 arXiv

  56. [64]

    Compression of Self-Supervised Representa- tions for Machine Vision,

    Z. Duan and F. M. Zhu, “Compression of Self-Supervised Representa- tions for Machine Vision,” in2024 IEEE 26th International Workshop on Multimedia Signal Processing (MMSP), 2024, pp. 1–6

    2024

  57. [65]

    Cross-architecture universal feature coding via distribution alignment,

    C. Gao, S. Liu, F. Wu, and W. Lin, “Cross-architecture universal feature coding via distribution alignment,” inICIP 2025, 2025

    2025

  58. [66]

    Multirate Neural Image Compression with Adaptive Lattice Vector Quantization,

    H. Xu, X. Wu, and X. Zhang, “Multirate Neural Image Compression with Adaptive Lattice Vector Quantization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7633–7642. [Online]. Available: https://openaccess.thecvf.com/content/CV...

    2025

  59. [67]

    Zhang, Y

    H. Zhang, Y . Li, L. Li, and D. Liu. Learning Switchable Priors for Neural Image Compression

  60. [68]

    ImageNet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei- Fei, “ImageNet large scale visual recognition challenge,”International Journal of Computer Vision, vol. 115, no. 3, pp. 211–252, 2015

    2015

  61. [69]

    Microsoft COCO: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft COCO: Common objects in context,” inECCV, 2014

    2014

  62. [70]

    Scene parsing through ADE20K dataset,

    B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba, “Scene parsing through ADE20K dataset,” inCVPR, 2017

    2017

  63. [71]

    Calculation of average psnr differences between rd- curves,

    G. Bjontegaard, “Calculation of average psnr differences between rd- curves,”ITU-T SG16, Doc. VCEG-M33, 2001

    2001

  64. [72]

    Diffusion transformers with representation autoencoders,

    B. Zheng, N. Ma, S. Tong, and S. Xie, “Diffusion transformers with representation autoencoders,”arXiv preprint arXiv:2510.11690, 2025

    2025 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.