Pith. sign in

REVIEW 4 major objections 4 minor 73 references

ABC: Adaptive BayesNet Structure Learning for Computational Scalable Multi-task Image Compression

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read By learning the graph structure of a neural image codec viewed as a Bayesian network, ABC makes every component computationally scalable—from 444 kMACs/pixel down to a minimal mode—while keeping rate–distortion performance competitive.

desk verdict A serious, well-built extension of the ECCV BaSIC line that deserves a referee, but only if the authors add the missing MAC and bit-cost measurements for the adaptive controller and transmitted structures. read the letter →

arxiv 2506.15228 v1 pith:43YT7OJM submitted 2025-06-18 eess.IV cs.MM

Yufeng Zhang , Wenrui Dai , Hang Yu , Shizhan Liu , Junhui Hou , Jianguo Li , Weiyao Lin This is my paper · ORCID
classification eess.IVcs.MM
keywords neuralimagecompressioncomputationalscalabilityBayesiannetworkstructurelearningslimmablenetworksautoregressivecontextmodelrate-distortion-complexityoptimizationmulti-taskcodingdynamicmaskedconvolution
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

ABC claims that the computational cost of a neural image codec can be made granularly and adaptively scalable by treating the codec as a Bayesian network and learning the network's graph structure, rather than hand-designing complexity levels. It introduces two learned structures: a heterogeneous bipartite graph that selects channel widths in the neural backbone, and a homogeneous multipartite graph that assigns latent positions to parallel stages in the autoregressive context model. An adaptive control module picks the graph according to a device budget, input difficulty, or downstream vision task. On Kodak and CLIC, the codec spans eight complexity levels from 444 kMACs/pixel downward and reaches BD-Rate at or below scalable baselines across most MAC levels, while the learned autoregressive graphs use fewer bits than hand-crafted parallel context models at equal stage counts. If these claims hold, one trained model can replace a family of separately trained codecs and adjust its computation at deployment time.

What carries the argument

The carrying object is the learned BayesNet structure $G$, decomposed into the inter-node graph and the intra-node graph. The inter-node graph is parameterized by categorical distributions $\pi$ over candidate edges per node pair; because each candidate is a slimmable network at a distinct channel width, sampling an edge samples a computational budget, and expected complexity is computed by mixing the candidate MACs with the same weights (Equation 18). The intra-node graph is parameterized by per-location categorical topological indices $T_{c,h,w}$; the number of partites, $\dim \pi_{c,h,w} = S_{\mathrm{intra}}$, sets the number of parallel stages, and the dynamic masked convolution of Equation (24) zeros out any kernel position whose topological index is not strictly smaller than the current node's, preserving acyclicity. The generative structure samples needed at decode time are claimed to be simple discrete variables, excluded from the rate loss. A two-stage optimizer first trains the rate–distortion objective for any sampled graph, then selects a finite set of near-Pareto-optimal structures and fine-tunes inside that set to stabilize the large search space.

What would settle it

Run the arithmetic coder on the decoder-side structure samples $\hat{G}$—the inter-node categorical edge selections and the intra-node topological indices—at each of the eight ABC complexity levels and compare the resulting bits with the total BPP; if that overhead reaches even a small percentage of the bitstream at ABC-L7 or at low bit-rates, the claim that structure cost is negligible and does not affect the rate–distortion trade-off is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that computational scalability in neural image compression can be achieved by learning the directed acyclic graph $G$ that defines conditional dependencies among the codec's variables, split into two learnable structures. The inter-node BayesNet is a heterogeneous bipartite graph: each adjacent pair of nodes is linked by several candidate neural-network edges of different sizes, a categorical distribution selects one edge, and each candidate is implemented as a slimmable network at a different channel width, so the selected graph directly sets a MAC budget. The intra-node BayesNet is a homogeneous multipartite graph over the latent tensor: every spatial/channel location carries a topological index drawn from a categorical distribution, dependencies point only from earlier to later partites, and the number of partites equals the number of parallel autoregressive stages; a dynamic masked convolution (Equation 24) enforces acyclicity while keeping the same parameter and MAC count as a standard convolution. The adaptive control module generates $G$ conditioned on a computation-budget controller $\Phi_C$, a task controller $\Phi_T$, and the input $x$, and the decoder-side structure samples are described as simple discrete variables with negligible storage. Experiments report BD-Rate at or below scalable baselines across most MAC levels, with learned intra-node graphs matching or beating hand-crafted checkerboard and channel-wise context models at equal parallel stages.

Load-bearing premise

The load-bearing premise, stated in Section V-C, is that the decoder-side structure samples $\hat{G}$ are simple discrete variables whose storage is negligible, so they are excluded from the rate loss; if transmitting those choices costs real bits, the reported bitrate savings shrink and the competitive-compression claim needs revision.

Editorial extensions

If this is right

  • A single ABC model can be deployed at eight selectable complexity levels, and because the four slimmable backbone networks are controlled independently, the model offers $5^4 = 625$ configurable MAC levels for hitting a hardware budget precisely.
  • The learned intra-node graph at ten parallel stages reaches a bit-rate close to the serial MaskConv baseline while keeping the autoregressive model parallel, so learned context structure narrows the rate–distortion gap that hand-crafted parallel models leave open.
  • Adaptive control lets one codec serve multiple regimes: data-adaptive ABC improves BD-Rate on CLIC under low MAC budgets, and task-adaptive ABC supports classification and instance segmentation across a wider MAC span than the task-slimmable baseline.
  • The structure-learning layer is not tied to one backbone: ablations on three different backbone designs show the same comparative trends, so the approach can be grafted onto other hyperprior-style codecs.

Reading between the lines

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

  • Editorial inference: The paper's exclusion of $\hat{G}$ from the rate loss is an assumption, not a measurement; a direct count of the bits spent on decoder-side structure at low bit-rates and at ABC-L7 would show whether the reported BD-Rate survives in the regime where ABC claims its biggest advantage.
  • Editorial inference: The learned intra-node graphs consistently converge to interlaced channel-spatial patterns, which suggests a transferable design rule for parallel autoregressive context models; one could test it by initializing hand-crafted models with interlaced assignments and comparing final bit-rate.
  • Editorial inference: The complexity loss is defined on differentiable MAC counts, so the paper's scalability claims are about arithmetic cost rather than wall-clock time; a complete deployment comparison would need to charge the adaptive control module's own computation and memory overhead against the MAC budget, which the paper acknowledges can reduce performance at some thresholds.
  • Editorial inference: The architecture-agnostic claim rests on ablations over hyperprior-style backbones; a natural extension is to apply the same inter/intra-node structural prior to transformer-based or lossless-hybrid codecs, where context dependency patterns differ.
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

4 major / 4 minor

Summary. The paper proposes ABC, a framework for computational scalable multi-task neural image compression based on Bayesian network (BayesNet) structure learning. The BayesNet structure G is decomposed into an inter-node bipartite graph that selects among slimmable channel-width edges for the analysis, synthesis, and hyper transforms, and an intra-node multipartite graph that learns topological partitions for a dynamic masked convolution acting as a parallel autoregressive context model. An adaptive control module conditions G on device budget, input content, and downstream task. Experiments compare BD-Rate against kMACs/pixel on Kodak and CLIC, task-specific accuracy on ImageNet and MS-COCO, and processing speed against WebP, BPG, and NIC baselines; ablations cover one- vs two-stage optimization, backbone choices, and intra-node hyperparameters. Code is released.

Significance. The conceptual contribution is useful: viewing NIC modules as BayesNet edges and learning inter- and intra-node structure in a differentiable way is a coherent extension of prior slimmable-NIC work, and the learned intra-node structures (Table III, Fig. 10) offer a concrete design insight that resonates with later hand-crafted schemes such as ELIC. The two-stage optimization in Sec. III-E is sensible, and releasing code is a strength. However, the central scalability claims are currently supported by incomplete measurement: the adaptive control module's compute is not included in MAC totals, the bit cost of transmitting G is asserted rather than measured, and the reported comparisons lack repeated-run variance. These issues are fixable and should be addressed before publication.

major comments (4)
  1. [Sec. V-C, Eq. (33), App. B-B3] The complexity loss in Eq. (33) sums C(G_inter) terms and omits the Adaptive Control Module, and App. B-B3 ('We utilize ptflops to tally the MAC figures') does not state whether the VGG-based g_c, the merge module, and the MLP are included. Since the ABC+DA curves in Figs. 4 and 6 are plotted against kMACs/pixel, an unmeasured per-pixel MAC cost for g_c can shift the low-complexity points to the right and narrow the claimed control span. The note in Sec. VI-B1 that the control module causes 'marginally reduced performance' is qualitative; please instrument the full encode path and report total MACs, and ideally wall-clock time, for both the w/o-DA and +DA variants at every complexity level.
  2. [Sec. V-C and Eq. (8)] The formulation in Eq. (8) includes log p_G(\hat G) in the rate loss, but the implemented loss in Eq. (31) does not contain this term, and Sec. V-C dismisses the generative structure cost as negligible without measurement. The decoder must receive G for the generative dependencies p(\hat y|\hat z, G) and p(\hat x|\hat y, G), so the bitstream necessarily contains the inter-node and intra-node structure indices. At BPP around 0.3, even a few hundred bits per image is non-negligible for BD-Rate comparisons. Please report the compressed size of the transmitted G per image and per level, or provide a measured entropy model for it, and explicitly show whether this cost is negligible.
  3. [Sec. IV-B, Eq. (33), Figs. 4/6] The MAC axis in Figs. 4 and 6 reflects inter-node backbone cost only: Eq. (33) defines C(G_intra) as constant and omits it from L_C, and Sec. IV-B/App. B show that varying S_intra changes the number of parallel stages and latency while the MACs of the dynamic masked convolution remain equal to those of a standard 2D convolution (Eq. 27). The claim of 'full computational scalability across all NIC components' therefore requires a clearer separation of metrics: the headline MAC comparisons do not exercise the intra-node autoregressive scalability, which appears only in the speed-based comparison (Fig. 8, Table III). Please state explicitly which components are traded off in each experiment and report both MAC and latency for all components.
  4. [Sec. VI-B1, VI-D, Table III] The empirical comparisons are reported as single runs without error bars or multiple seeds, even where the differences are very small (Table III, where BPP differences of 0.003-0.005 separate methods). To support the claim that the learned intra-node structures 'consistently demonstrate slightly better performance,' repeated-seed or repeated-evaluation variance should be reported. In addition, the baseline description is internally inconsistent: Sec. VI-B1 first states that the autoregressive modules are 'same as [17]' across all frameworks and then says SlimCAE uses a 'slimmable context model' according to the original paper; please specify exactly which context model each baseline uses and how λ-scheduling was disabled.
minor comments (4)
  1. [Sec. VI-D] The sentence beginning 'Moreover, the decompression speed is also reported to evaluate the efficiency of It's important to note...' is incomplete and should be rewritten.
  2. [Sec. VI-B2] The phrase 'more than 30% higher compression performance than BPG' is ambiguous; it should be phrased as 'more than 30% BD-rate savings over BPG' or similar according to the metric used.
  3. [App. B-B3] The note that reported decompression speeds are effectively quadrupled because images are loaded as 32-bit float tensors should be stated in the main text, at least in the caption of Fig. 8, to avoid misleading readers.
  4. [Table III] The table layout with 'Stages - 2 4 10' and the row of method names is hard to parse; explicit column headings for each method group would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ABC's claims are anchored by held-out rate-distortion-complexity evaluations against external baselines; the unquantified control-module MAC overhead and generative-structure bit cost are missing measurements, not circular reductions.

full rationale

ABC is an empirical system paper rather than a first-principles derivation, so the relevant check is whether the reported predictions (BD-Rate, BPP, MAC span) reduce by construction to fitted inputs. They do not. The training loss in Eq. (7) combines rate, distortion, and complexity, but the experimental claims are measured as held-out BPP and BD-Rate on Kodak and CLIC, and task metrics on ImageNet and MS-COCO, against external baselines such as SlimCAE, ELFIC, Cao2023, WebP, BPG, and the static Minnen/Cheng/He codecs. The learned structures G_inter and G_intra are optimized via Gumbel-softmax and VIMCO (Eqs. 18 and 21), and their quality is assessed by bit-rate on test images rather than read off the training objective. The self-citation to the prior BaSIC paper [15] is a normal conference-to-journal extension; the new adaptive-control experiments are externally benchmarked, so no load-bearing claim rests solely on that citation. There is no imported uniqueness theorem and no learned ansatz whose justification is merely the authors' earlier work. Two genuine gaps are quantitative, not circular: Section V-C asserts that generative structures G have a 'contribution to the overall bit-rate is negligible' and therefore 'do not factor into the bit-rate optimization process,' even though Eq. (8) formally includes log p_G(hat G); and Appendix B-B3 states 'We utilize ptflops to tally the MAC figures' without itemizing the VGG-based data-adaptive controller g_c, which Section VI-B1 concedes causes 'added computational overhead... marginally reduced performance at specific complexity thresholds.' These are unmeasured costs that could shift the low-complexity ends of Figures 4, 6, and 8, but they are missing measurements or simplifying approximations, not reductions of a prediction to its own input by construction. Correcting them would change the reported complexity numbers, not the derivation structure of the method.

Assumptions & free parameters 6 free parameters · 6 assumptions · 2 invented entities

The ledger lists the hand-chosen operating-point hyperparameters, the domain assumptions inherited from prior NIC work, and two paper-specific assumptions (negligible structure bit-rate and the two-stage Pareto-front heuristic). These are the main ingredients the central claim rests on that are not themselves derived in the paper.

free parameters (6)
  • lambda_D (PSNR rate-distortion tradeoff) = [0.0018, 0.0035, 0.0067, 0.0130]
    Four distortion weights chosen from CompressAI to set bitrate levels; these set the operating points and are not derived by ABC.
  • lambda_D (MS-SSIM) = [2.40, 4.58, 8.73, 16.64]
    MS-SSIM distortion weights, fine-tuned from PSNR models.
  • lambda_C (complexity tradeoff) = [0.0, 0.25, 0.5, 1.0, 2.0, 4.0, 8.0, 16.0]
    Discrete complexity weights sampled in second-stage training to produce eight levels ABC-L0 through ABC-L7; hand-selected.
  • slimmable channel widths = {48, 72, 96, 144, 192}
    Channel width options for g_a, g_s, h_a, h_s; this memory-efficient scalability design comes from SlimCAE and is carried over.
  • AR parallel stages S_intra = {2, 4, 10}
    Number of partites in the intra-node BayesNet; controls decompression parallelism and is compared against hand-crafted models with the same stage counts.
  • channel groups C and kernel size K for dynamic masked convolution = C in {2, 4, 6}; K in {3, 5, 7}
    Ablated in Table V; C matched to the number of stages performs best and K=5 is standard.
assumptions (6)
  • domain assumption A multi-task NIC codec can be represented as a Bayesian network with factorization p(xhat|yhat)p(yhat|zhat)p(zhat) and conditional dependencies parameterized by neural networks.
    Section III-A and Figure 2 restate existing hyperprior and autoregressive NIC frameworks in BayesNet terms; this is a modeling view, not proved.
  • domain assumption Computational complexity of an edge is proportional to its channel width or MACs and is additive across selected edges, as in Equation 18.
    Used as the differentiable optimization target for the inter-node graph; plausible but not derived from hardware behavior.
  • domain assumption Slimmable networks trained with the sandwich rule from prior work preserve near-static performance at each width.
    ABC inherits this property from SlimCAE; it is not re-proved here.
  • ad hoc to paper The generative structure samples G can be transmitted to the decoder at negligible bit-rate and are excluded from the rate loss.
    Stated in Section V-C but never measured; if false, BD-Rate comparisons would need adjustment.
  • standard math VIMCO provides a usable gradient estimator for the discrete intra-node structure objective (Equation 21), and Gumbel-softmax works for the inter-node categorical structure (Equation 17).
    Standard estimators from the literature; no formal proof is included in this paper.
  • ad hoc to paper A two-stage strategy, first optimizing rate-distortion over all structures and then selecting a finite structure set via a complexity-weighted loss, yields structures near the rate-distortion-complexity Pareto front.
    Heuristic in Section III-E; supported only by ablations (Figure 11), not by a formal guarantee.
invented entities (2)
  • BayesNet structure variable G (inter-node and intra-node) independent evidence
    purpose: Latent control variable that selects slimmable channel-width edges and partite assignments, thereby setting computational complexity.
    G is not an unobservable postulate: it is output by q(G|X,Phi), transmitted or shared with the decoder, and visualized in Figure 10; its effect on BPP, MACs, and speed is measured.
  • Controller nodes Phi_C, Phi_T, Phi_D independent evidence
    purpose: Discrete conditioners representing device computation budget, vision task, and distortion level for q(G|X,Phi).
    They are input metadata (budget index, task label, rate index), observable by design; their effect on the produced codec is falsifiable through the reported adaptive-control experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ABC: Adaptive BayesNet Structure Learning for Computational Scalable Multi-task Image Compression." pith.science (2026). https://pith.science/paper/43YT7OJM

@misc{pith2026250615228,
  author       = {Pith},
  title        = {Pith review of: ABC: Adaptive BayesNet Structure Learning for Computational Scalable Multi-task Image Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/43YT7OJM}},
  note         = {Machine review of arXiv:2506.15228}
}
read the original abstract

Neural Image Compression (NIC) has revolutionized image compression with its superior rate-distortion performance and multi-task capabilities, supporting both human visual perception and machine vision tasks. However, its widespread adoption is hindered by substantial computational demands. While existing approaches attempt to address this challenge through module-specific optimizations or pre-defined complexity levels, they lack comprehensive control over computational complexity. We present ABC (Adaptive BayesNet structure learning for computational scalable multi-task image Compression), a novel, comprehensive framework that achieves computational scalability across all NIC components through Bayesian network (BayesNet) structure learning. ABC introduces three key innovations: (i) a heterogeneous bipartite BayesNet (inter-node structure) for managing neural backbone computations; (ii) a homogeneous multipartite BayesNet (intra-node structure) for optimizing autoregressive unit processing; and (iii) an adaptive control module that dynamically adjusts the BayesNet structure based on device capabilities, input data complexity, and downstream task requirements. Experiments demonstrate that ABC enables full computational scalability with better complexity adaptivity and broader complexity control span, while maintaining competitive compression performance. Furthermore, the framework's versatility allows integration with various NIC architectures that employ BayesNet representations, making it a robust solution for ensuring computational scalability in NIC applications. Code is available in https://github.com/worldlife123/cbench_BaSIC.

Figures

Figures reproduced from arXiv: 2506.15228 by the authors.

Figure 1
Figure 1. Computational Scalable Multi-task Image Compression. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Bayesian network diagrams for various NIC frameworks. Our proposed framework ABC generalizes all concepts in (a-f). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The architecture of the implemented example based on the joint autoregressive framework [17]. Comp/Decomp means Compression / Decompression. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Comparative results for different computational scalable frameworks under MAC computation metric on Kodak dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Rate/Distortion plots for different computational scalable frameworks [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Comparative results for different computational scalable frameworks under MAC computation metric on CLIC dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Comparative results for different computational scalable frameworks under MAC computation metric for image classification on ImageNet dataset [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Comparison for computational scalable frameworks subject to total [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Comparative results subject to total MACs constraints. The parenthetical value following each method in the legend denotes its MACs value in [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Partite topological indices visualization for intra-node BayesNet in the 10-stage case. Each image patch indicates topological indices of nodes in [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Ablation study results for different inter-node BayesNet implemen [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 59 canonical work pages

  1. [17]

    Joint autoregressive and hierarchical priors for learned image compression,

    D. C. Minnen, J. Ball ´e, and G. Toderici, “Joint autoregressive and hierarchical priors for learned image compression,”ArXiv, vol. abs/1809.02736, 2018

    arXiv 2018

  2. [1]

    Variational image compression with a scale hyperprior,

    J. Ball ´e, D. C. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,”ArXiv, vol. abs/1802.01436, 2018

    arXiv 2018

  3. [2]

    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,”2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7936–7945, 2020

    work page 2020

  4. [3]

    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,”2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5708–5717, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:247594672

    2022

  5. [4]

    Transtic: Transferring transformer-based image compression from human perception to machine perception,

    Y .-H. Chen, Y . Weng, C.-H. Kao, C. Chien, W.-C. Chiu, and W. Peng, “Transtic: Transferring transformer-based image compression from human perception to machine perception,”2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 23 240–23 250, 2023. [On- line]. Available: https://api.semanticscholar.org/CorpusID:259108311

    work page 2023

  6. [5]

    The jpeg ai standard: Providing efficient human and machine visual data consumption,

    J. Ascenso, E. Alshina, and T. Ebrahimi, “The jpeg ai standard: Providing efficient human and machine visual data consumption,” IEEE MultiMedia, vol. 30, pp. 100–111, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:258641408

    work page 2023

  7. [6]

    Slimmable multi-task image compression for human and machine vision,

    J. Cao, X. Yao, H. Zhang, J. Jin, Y . Zhang, and B. W.-K. Ling, “Slimmable multi-task image compression for human and machine vision,”IEEE Access, vol. 11, pp. 29 946–29 958, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:257744357 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 16

    work page 2023

  8. [7]

    Image Compression for Machine and Human Vision with Spatial-Frequency Adaptation

    H. Li, S. Li, S. Ding, W. Dai, M. Cao, C. Li, J. Zou, and H. Xiong, “Image compression for machine and human vision with spatial-frequency adaptation,” vol. abs/2407.09853, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:271212363

    work page Pith review arXiv 2024

Show all 73 references
  1. [8]

    libwebp,

    webmproject, “libwebp,” https://github.com/webmproject/libwebp, 2023, accessed: 2023-01-09

    2023

  2. [9]

    Slimmable compressive autoencoders for practical neural image compression,

    F. Yang, L. Herranz, Y . Cheng, and M. G. Mozerov, “Slimmable compressive autoencoders for practical neural image compression,” 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4996–5005, 2021. [Online]. Available: https://api.semanticscholar.org...

    2021

  3. [10]

    Slimmable neural networks,

    J. Yu, L. Yang, N. Xu, J. Yang, and T. S. Huang, “Slimmable neural networks,”ArXiv, vol. abs/1812.08928, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:56657799

    2018 arXiv

  4. [11]

    Adanic: Towards practical neural image compression via dynamic transform routing,

    L. Tao, W. Gao, G. Li, and C. Zhang, “Adanic: Towards practical neural image compression via dynamic transform routing,”2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 16 833–16 842, 2023. [Online]. Available: https://api.semanticscholar. org/CorpusID:267026318

    2023

  5. [12]

    Elfic: A learning-based flexible image codec with rate-distortion-complexity optimization,

    Z. Zhang, B. Chen, H. Lin, J. Lin, X. Wang, and T. Zhao, “Elfic: A learning-based flexible image codec with rate-distortion-complexity optimization,”Proceedings of the 31st ACM International Conference on Multimedia, 2023. [Online]. Available: https://api.semanticscholar. org/...

    2023

  6. [13]

    Dag-gnn: Dag structure learning with graph neural networks,

    Y . Yu, J. Chen, T. Gao, and M. Yu, “Dag-gnn: Dag structure learning with graph neural networks,” inInternational Conference on Machine Learning, 2019. [Online]. Available: https://api.semanticscholar.org/ CorpusID:128358697

    2019

  7. [14]

    Dibs: Differentiable bayesian structure learning,

    L. Lorch, J. Rothfuss, B. Scholkopf, and A. Krause, “Dibs: Differentiable bayesian structure learning,”ArXiv, vol. abs/2105.11839, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:235187432

    2021 arXiv

  8. [15]

    Basic: Bayesnet structure learning for computational scalable neural image compression,

    Y . Zhang, H. Yu, S. Liu, W. Dai, and W. Lin, “Basic: Bayesnet structure learning for computational scalable neural image compression,” inComputer Vision – ECCV 2024, 2025, pp. 269–285

    2024

  9. [16]

    End-to-end optimized image compression,

    J. Ball ´e, V . Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,”ArXiv, vol. abs/1611.01704, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:6775391

    2016 arXiv

  10. [18]

    Context-adaptive entropy model for end- to-end optimized image compression,

    J. Lee, S. Cho, and S. Beack, “Context-adaptive entropy model for end- to-end optimized image compression,” inInternational Conference on Learning Representations, 2018

    2018

  11. [19]

    Density modeling of images using a generalized normalization transformation,

    J. Ball ´e, V . Laparra, and E. P. Simoncelli, “Density modeling of images using a generalized normalization transformation,”CoRR, vol. abs/1511.06281, 2015. [Online]. Available: https://api.semanticscholar. org/CorpusID:2684987

    2015 arXiv

  12. [20]

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

    T. Chen, H. Liu, Z. Ma, Q. Shen, X. Cao, and Y . Wang, “End- to-end learnt image compression via non-local attention optimization and improved context modeling,”IEEE Transactions on Image Processing, vol. 30, pp. 3179–3191, 2019. [Online]. Available: https://api.semanticschola...

    2019

  13. [21]

    Transformer-based image compression,

    M.-T. Lu, P. Guo, H. Shi, C. Cao, and Z. Ma, “Transformer-based image compression,”2022 Data Compression Conference (DCC), pp. 469–469,

    2022

  14. [22]

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

    L. yu 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. [Online]. Available: https://api.semanticscholar. org/CorpusID...

    2020

  15. [23]

    Learned image compression for machine perception,

    F. Codevilla, J.-G. Simard, R. Goroshin, and C. J. Pal, “Learned image compression for machine perception,”ArXiv, vol. abs/2111.02249,

    arXiv

  16. [24]

    End-to-end optimized image compression for multiple machine tasks,

    L. D. Chamain, F. Racap’e, J. B ´egaint, A. Pushparaja, and S. Feltman, “End-to-end optimized image compression for multiple machine tasks,”ArXiv, vol. abs/2103.04178, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:232148015

    2021 arXiv

  17. [25]

    Available: https://api.semanticscholar.org/CorpusID: 241033392

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 241033392

  18. [26]

    Channel-wise autoregressive entropy models for learned image compression,

    D. C. Minnen and S. Singh, “Channel-wise autoregressive entropy models for learned image compression,”2020 IEEE International Conference on Image Processing (ICIP), pp. 3339–3343, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:220633468

    2020

  19. [27]

    Improving multiple machine vision tasks in the compressed domain,

    J. Liu, H. Sun, and J. Katto, “Improving multiple machine vision tasks in the compressed domain,”2022 26th International Conference on Pattern Recognition (ICPR), pp. 331–337, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:254100969

    2022

  20. [28]

    Multistage spatial context models for learned image compression,

    F. Lin, H. Sun, J. Liu, and J. Katto, “Multistage spatial context models for learned image compression,”ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5, 2023. [Online]. Available: https: //api.semanticscholar.org/Co...

    2023

  21. [29]

    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,”2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 14 766–14 775, 2021. [Online]. Available: https://api.semanticscholar. org/Corpu...

    2021

  22. [30]

    Finite-state autoregressive entropy coding for efficient learned lossless compression,

    Y . Zhang, H. Yu, J. Li, and W. Lin, “Finite-state autoregressive entropy coding for efficient learned lossless compression,” inInternational Conference on Learning Representations, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:271532743

    2024

  23. [31]

    Pilc: Practical image lossless compression with an end-to-end gpu oriented neural framework,

    N. Kang, S. Qiu, S. Zhang, Z. Li, and S. Xia, “Pilc: Practical image lossless compression with an end-to-end gpu oriented neural framework,”2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3729–3738, 2022. [Online]. Available: https://api.semanti...

    2022

  24. [32]

    Get the best of the three worlds: Real-time neural image compression in a non-gpu environment,

    Z. Zheng, X. Wang, X. Lin, and S. Lv, “Get the best of the three worlds: Real-time neural image compression in a non-gpu environment,” Proceedings of the 29th ACM International Conference on Multimedia,

  25. [33]

    Computationally efficient neural image compression,

    N. Johnston, E. Eban, A. Gordon, and J. Ball’e, “Computationally efficient neural image compression,”ArXiv, vol. abs/1912.08771,

    1912 arXiv

  26. [34]

    Exploring structural sparsity in neural image compression,

    S. Yin, F. Meng, W.-T. Tan, C. Li, Y . Bao, Y . Liang, and W. Liu, “Exploring structural sparsity in neural image compression,”2022 IEEE International Conference on Image Processing (ICIP), pp. 471–475,

    2022

  27. [35]

    Content adaptive optimization for neural image compression,

    J. Campos, S. Meierhans, A. Djelouah, and C. Schroers, “Content adaptive optimization for neural image compression,”ArXiv, vol. abs/1906.01223, 2019. [Online]. Available: https://api.semanticscholar. org/CorpusID:174797811

    1906 arXiv

  28. [36]

    Available: https://api.semanticscholar.org/CorpusID: 239011955

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 239011955

  29. [37]

    Universal efficient variable-rate neural image compression,

    S. Yin, C. Li, Y . Bao, and Y . Liang, “Universal efficient variable-rate neural image compression,”ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2025–2029, 2021. [Online]. Available: https://api.semanticscholar.org/ ...

    2022

  30. [38]

    Optimal structure identification with greedy search,

    D. M. Chickering, “Optimal structure identification with greedy search,” J. Mach. Learn. Res., vol. 3, pp. 507–554, 2003. [Online]. Available: https://api.semanticscholar.org/CorpusID:1191614

    2003

  31. [39]

    Learning optimal bayesian networks: A shortest path perspective,

    C. Yuan and B. M. Malone, “Learning optimal bayesian networks: A shortest path perspective,”J. Artif. Intell. Res., vol. 48, pp. 23–65, 2013. [Online]. Available: https://api.semanticscholar.org/CorpusID:5951724

    2013

  32. [40]

    Learning bayesian networks from data: An information-theory based approach,

    J. Cheng, R. Greiner, J. Kelly, D. A. Bell, and W. Liu, “Learning bayesian networks from data: An information-theory based approach,” Artif. Intell., vol. 137, pp. 43–90, 2002. [Online]. Available: https://api.semanticscholar.org/CorpusID:7821347

    2002

  33. [41]

    Content adaptive latents and decoder for neural image compression,

    G. Pan, G. Lu, Z. Hu, and D. Xu, “Content adaptive latents and decoder for neural image compression,”ArXiv, vol. abs/2212.10132,

    arXiv

  34. [42]

    Available: https://api.semanticscholar.org/CorpusID: 253448404

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 253448404

  35. [43]

    Content adaptive checkerboard context model for learned image compression,

    Y . Zhang, G. Lu, D. Feng, C. Zhu, and L. Song, “Content adaptive checkerboard context model for learned image compression,”2023 IEEE International Symposium on Circuits and Systems (ISCAS), pp. 1– 5, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID: 260002843

    2023

  36. [44]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”CoRR, vol. abs/1312.6114, 2013. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 17

    2013 arXiv

  37. [45]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” inNeural Information Processing Systems, 2017. [Online]. Available: https://api.semanticscholar.org/CorpusID:326772

    2017

  38. [46]

    Darts: Differentiable architecture search,

    H. Liu, K. Simonyan, and Y . Yang, “Darts: Differentiable architecture search,”ArXiv, vol. abs/1806.09055, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:49411844

    2018 arXiv

  39. [47]

    Exact bayesian structure discovery in bayesian networks,

    M. Koivisto and K. Sood, “Exact bayesian structure discovery in bayesian networks,”J. Mach. Learn. Res., vol. 5, pp. 549–573, 2004. [Online]. Available: https://api.semanticscholar.org/CorpusID:12412451

    2004

  40. [48]

    Dags with no tears: Continuous optimization for structure learning,

    X. Zheng, B. Aragam, P. Ravikumar, and E. P. Xing, “Dags with no tears: Continuous optimization for structure learning,” in Neural Information Processing Systems, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:53217974

    2018

  41. [49]

    Bayesian structure learning with generative flow networks,

    T. Deleu, A. G’ois, C. C. Emezue, M. Rankawat, S. Lacoste-Julien, S. Bauer, and Y . Bengio, “Bayesian structure learning with generative flow networks,”ArXiv, vol. abs/2202.13903, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:247158659

    2022 arXiv

  42. [50]

    Variational inference for monte carlo objectives,

    A. Mnih and D. J. Rezende, “Variational inference for monte carlo objectives,” inInternational Conference on Machine Learning, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:5859948

    2016

  43. [51]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,”2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255, 2009

    2009

  44. [52]

    Workshop and challenge on learned image compression,

    “Workshop and challenge on learned image compression,” https://clic. compression.cc/, 2023, accessed: 2023-01-09

    2023

  45. [53]

    Autoslim: Towards one-shot architecture search for channel numbers,

    J. Yu and T. Huang, “Autoslim: Towards one-shot architecture search for channel numbers,”arXiv: Computer Vision and Pattern Recognition,

  46. [54]

    Available: https://api.semanticscholar.org/CorpusID: 173990379

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 173990379

  47. [55]

    Categorical reparameterization with gumbel-softmax,

    E. Jang, S. S. Gu, and B. Poole, “Categorical reparameterization with gumbel-softmax,”ArXiv, vol. abs/1611.01144, 2016

    2016 arXiv

  48. [56]

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

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

  49. [57]

    Available: https://api.semanticscholar.org/CorpusID: 195345342

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 195345342

  50. [61]

    Kodak lossless true color image suite,

    “Kodak lossless true color image suite,” https://r0k.us/graphics/kodak/, 2023, accessed: 2023-11-12

    2023

  51. [62]

    Microsoft coco: Common objects in context,

    T.-Y . Linet al., “Microsoft coco: Common objects in context,” in European Conference on Computer Vision, 2014

    2014

  52. [63]

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

    J. B ´egaint, F. Racap ´e, S. Feltman, and A. Pushparaja, “Compressai: a pytorch library and evaluation platform for end-to-end compression research,”arXiv preprint arXiv:2011.03029, 2020

    2011 arXiv

  53. [64]

    Bpg image format,

    F. Bellard, “Bpg image format,” https://bellard.org/bpg/, 2023, accessed: 2023-01-09

    2023

  54. [65]

    Neural video compression with diverse contexts,

    J. Li, B. Li, and Y . Lu, “Neural video compression with diverse contexts,”2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 22 616–22 626, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:257232512 Yufeng Zhangreceived the B.E....

    2023

  55. [68]

    Here we further provide a PyTorch4-style pseudo-code implementation in Algorithm 1 to clarify the implementation details

    Computational Complexity of Multipartite-based Dy- namic Masked Convolution:Building on the technique outlined in Section IV-B, Equation 23 and Equation 24 provide an instantiation of intra-node BayesNet via multipartite-based dynamic masked convolution. Here we further provid...

  56. [69]

    Merging Predictions from Intra-node and Inter-node BayesNet:In the original framework, the complete prior for predicting ˆygiven ˆz, denoted asp( ˆy|ˆz), fuses the output fromh s(ˆz)and the context modelAR( ˆy). This is achieved using a 3-layer MLP, consisting of1×1convolution...

  57. [70]

    This model is a straight- forward 2-layer network comprising Linear and ReLU layers

    Learning the Intra-node BayesNet Generator:Earlier in Section IV-B, we outlined the use of a generative model g(N)in Equation 22 to model the latent correlations among nodes in the intra-node BayesNet. This model is a straight- forward 2-layer network comprising Linear and ReL...

  58. [71]

    Optimization:Recall that we use the loss function Equation 26: L=L R +λ DLD +λ CLC +L Gintra (30) Specifically, the rate loss,L R, integrates all latent variables withinLas: LR = Eˆy∼q(y|x,G y,x inter),ˆz∼q(z|y,Gz,y inter) log2p(ˆy|ˆz,G ˆy,ˆz inter,G ˆy intra)p(ˆz). (31) Disto...

  59. [72]

    Training Schedule:The training stage spans ap- proximately 2M iterations (1.6M in the first stage and 0.4M in the second stage), employing the Adam opti- mizer with a learning rate of 0.0001. We follow Com- pressAI by employing four different distortion weights, λD = [0.0018,0...

  60. [73]

    Complexity Metrics:We utilize ptflops 5 to tally the MAC figures. For decompression speed, we measure by cal- culating the ratio of the total size of the decompressed image data (in MegaBytes) to the aggregate time taken (in seconds) to decompress the entire image dataset. It’...

  61. [2014]

    Before joining in SJTU, he was with the faculty of the University of Texas Health Science Center at Houston from 2018 to 2019

    He is currently an associate professor at the Department of Computer Science and Engineering, Shanghai Jiao Tong University (SJTU). Before joining in SJTU, he was with the faculty of the University of Texas Health Science Center at Houston from 2018 to 2019. He was a postdocto...

    2018

  62. [2019]

    Available: https://api.semanticscholar.org/CorpusID: 209404937

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 209404937

  63. [2020]

    His research interests include time series modeling, graph machine learning, natural language processing, and generative models

    Before joining Ant Group, he was a senior research fellow with NTU, Singapore. His research interests include time series modeling, graph machine learning, natural language processing, and generative models. Shizhan LiuShizhan Liu received the B.E. and M.E. degrees in electron...

    2021

  64. [2021]

    Available: https://api.semanticscholar.org/CorpusID: 244102730

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 244102730

  65. [2022]

    Available: https://api.semanticscholar.org/CorpusID: 246680031

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 246680031

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.