Pith. sign in

REVIEW 3 major objections 5 minor 87 references

The paper claims that a single autoencoder, trained once on a diverse 6,376-volume simulation database, can compress unseen scientific volumes at compression ratios thousands of times higher than conventional lossy compressors at equal reco

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A single autoencoder trained on 6,376 simulation volumes compresses unseen scientific volumes at 2,000–10,000× compression with higher reported fidelity than prior compressors, with adjustable rate at inference.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid engineering, overblown matched-quality claim, and a real but fixable train/test overlap issue—worth refereeing with revision. the 3 major comments →

arxiv 2607.18187 v1 pith:LTU7TS2I submitted 2026-07-20 cs.GR cs.DBcs.LG

EVOLVE: Efficient Learned Volume Compression with Variable-Rate Encoding on a Cross-Domain Database

Kaiyuan Tang , Maizhe Yang , Chaoli Wang This is my paper
classification cs.GR cs.DBcs.LG
keywords volume compressionautoencoderlearned compressionvariable-rate encodingcross-domain databasescientific visualizationlossy compressionimplicit neural representation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 tries to establish that one autoencoder, trained once on a curated database of 6,376 volumes from 21 scientific simulations, can compress unseen simulation volumes far more aggressively than established lossy compressors (ZFP, TTHRESH, SZ3) at equal reconstruction quality, while being orders of magnitude faster to encode than per-volume implicit neural representations. It reports compression ratios of roughly 2,800x to 10,800x on unseen datasets at high-fidelity quality, and shows the same single model can produce a continuous range of ratios by interpolating a learned gain factor. A reader should care because scientific simulations produce data faster than storage and network bandwidth grow, and this points to a path where one pretrained model replaces both a bank of rate-specific models and costly per-volume optimization.

Core claim

On its own terms, the central discovery is that combining a context-aware entropy model — a scale hyperprior plus 3D checkerboard spatial context plus channel-wise context — with training on a large, pHash-curated cross-domain database lets an autoencoder generalize across unseen volumes in the covered simulation domains. The paper reports that at PSNR at or above 40 dB, EVOLVE achieves compression ratios of 2,846x to 10,774x on unseen datasets, exceeding conventional compressors by a factor of roughly 3-4 and INR-based methods by a factor of 5-10, with encoding times orders of magnitude shorter than INR per-volume training. A single model additionally supports continuous variable-rate encod

What carries the argument

The central object is the EVOLVE autoencoder with a context-aware entropy model: a hyperprior produces side information, a 3D checkerboard mask provides spatial context from already-decoded anchor positions, and channel-wise autoregressive slicing conditions each latent group on previously decoded groups; these are fused to predict Gaussian parameters for arithmetic coding. A learnable gain vector rescales latents before quantization and rescales back after decoding, so interpolating between gain values yields a continuous rate-distortion curve in one model.

Load-bearing premise

The load-bearing premise is that the 21-simulation, 6,376-volume curated database is representative enough of 'scientific simulation data' that one shared model transfers to unseen volumes in those domains — the paper's own foot CT result, where quality saturates at 33.7 dB below conventional compressors, shows the claim collapses when the target domain is absent from training.

What would settle it

Compress a volume from a scientific domain not represented in the 21 training simulations (e.g., a medical CT or microscopy stack) and check whether EVOLVE exceeds ZFP, TTHRESH, and SZ3 at the 40 dB high-fidelity threshold. The paper itself reports the negative case: on the foot CT, EVOLVE reaches only 33.71 dB while conventional compressors exceed 45 dB, so that single observation already delimits the claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single pretrained EVOLVE model can replace multiple rate-specific models for offline scientific volume compression.
  • At matched reconstruction quality (PSNR ≥ 40 dB), the reported CRs imply storage savings of roughly 3-4x over conventional compressors, and 5-10x over INR-based compressors.
  • Encoding is feed-forward, so per-volume compression takes seconds rather than the hours some fully connected INR methods require.
  • The variable-rate gain mechanism produces densely sampled, smooth rate-distortion curves, which is useful when scientists must trade quality against budget per dataset.

Where Pith is reading between the lines

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

  • I infer that the pHash deduplication threshold is a hidden hyperparameter: raising it would retain more near-duplicate frames and likely increase training bias toward the most common dynamical states, while lowering it would shrink training diversity; the paper gives no sensitivity analysis, so the 0.85 threshold is an unvalidated choice.
  • The failure on the foot CT scan suggests the model's 'generalization' is really interpolation over statistics typical of simulation fields; a testable extension would be to add a small amount of medical or microscopy data and see whether the same architecture reaches 40+ dB, which would tell us whether the bottleneck is data coverage or architecture.
  • Because the latent space of EVOLVE is interpretable and usable for timestep selection, compression and analysis could share one representation — for example, clustering or trajectory extraction directly in the compressed latent space, avoiding decompression — but the paper only demonstrates this for representative timestep selection, not for general analysis tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes EVOLVE, a learned autoencoder-based compressor for scientific simulation volumes. The authors construct a curated training database of 6,376 volumes from 21 simulation datasets, progressively design an AE architecture with hyperprior and context-based entropy modeling, and add a learnable gain mechanism so that one trained model can produce a continuous range of compression ratios. They evaluate on several test volumes against conventional compressors (ZFP, TTHRESH, SZ3), INR-based methods (SIREN, NeurComp, ECNR, Instant-NGP, fV-SRN, AMGSRN++), and an AE-based baseline (IDLat), reporting higher CRs at favorable quality and much faster encoding than INR methods. The paper also reports an out-of-distribution evaluation on scanned volumes and discusses deployment considerations.

Significance. If the results are supported, the paper would be a useful step toward practical learned volume compression: one shared model that generalizes across simulation families, variable-rate encoding without retraining, and substantially faster encoding than per-volume INR optimization are all practically important. The authors provide code, model weights, and results, and the ablation-based design roadmap is a strength. The appended out-of-distribution study is also commendable for candidly documenting a failure mode. However, the central rate-distortion and cross-domain generalization claims are not yet fully supported by the evidence as presented; the main experiments compare at unequal reconstruction quality, and several test datasets overlap with training simulation families.

major comments (3)
  1. [§4.2, Tables 2 and 3] The central claim of 'substantially higher CRs at comparable reconstruction quality' is not directly supported by the reported tables. Baselines are tuned to just above 40 dB PSNR, while EVOLVE operates at 45–49 dB. For example, on ionization (H+) EVOLVE reports 47.58 dB at 7,843× versus TTHRESH at 40.11 dB and 6,334×; on combustion (MF) EVOLVE is 49.31 dB at 6,047× versus SZ3 at 41.10 dB and 2,948×. These are not comparable operating points, so the CR advantage may partly reflect the much larger quality gap. The authors should provide matched-quality CR comparisons (e.g., CR of every method at the same PSNR or at the baseline quality) or report full rate–distortion curves with explicit PSNR anchors for all methods. Figure 10 is a step in this direction, but the main tables and abstract should not claim comparable quality without such matched numbers.
  2. [§3.1, Appendix A, and Table 1] The cross-dataset generalization claim is weakened by overlap between training and test families. The training database in Appendix A explicitly includes combustion, half-cylinder, and ionization volumes, and Table 1 tests combustion (MF), half-cylinder (VLM, 6,400), and ionization (H+). Even if the specific timesteps are disjoint, these are not genuinely unseen simulation families. The abstract's phrase 'unseen scientific simulation datasets' therefore overstates the evidence. A leave-one-family-out experiment, or at least an evaluation on a simulation family not represented in training, is needed to support the claimed cross-domain generalization. Without it, the results may mostly show within-family generalization plus interpolation across correlated variables and timesteps.
  3. [Appendix H] The out-of-distribution evaluation on the foot CT volume shows EVOLVE saturating at 33.71 dB PSNR, below all conventional compressors, and the authors acknowledge that this quality cannot be improved by increasing the rate budget. This is an honest and useful limitation, but it also sharpens the load-bearing role of the training database: the claimed advantage depends on the target volume matching the curated simulation statistics. Since the paper does not provide an independent validation that the 21 simulation families and pHash threshold (0.85) produce a corpus that is representative of all 'covered scientific simulation domains', the scope of the generalization claim remains unclear. The authors should state this dependency more prominently in the abstract and contributions, and ideally provide a quantitative criterion for when the database can be expected to cover a target volume.
minor comments (5)
  1. [Title page] Author name appears as 'Maizhe Y ang' with a space; presumably should be 'Maizhe Yang'.
  2. [§4.1 and Appendix A] The text says 'the volumes in Table 1 have no overlap with the 9,921 volumes in Table 1', which is confusing because both refer to Table 1. The training composition table in Appendix A should have a distinct number (e.g., Table 8) and the sentence should refer to the appropriate table.
  3. [§3.1] The claim that all test volumes are 'drawn from unseen simulations or variables' is contradicted by Appendix A, which lists combustion, half-cylinder, and ionization as training sources. Please rephrase to say that the test volumes are unseen timesteps or ensembles from some of the same simulation families, plus some new datasets.
  4. [Figure 10] The rate–distortion curves show EVOLVE spanning a continuous range, but the PSNR axis starts around 32–35 dB and goes above 50 dB. It would be clearer to mark the 'comparable quality' region and to show the baseline operating points from Tables 2 and 3 explicitly, so the reader can see the quality gap at a glance.
  5. [§4.3] For IDLat, the paper says its CR is fixed and reports results without quality adjustment; IDLat's PSNR on magnetic is 38.84 dB, below the stated 40 dB high-fidelity threshold. This should be noted in the text or the threshold statement adjusted to 'close to 40 dB'.

Circularity Check

0 steps flagged

No significant circularity; EVOLVE's results are measured empirical outcomes, not derivations that reduce to their inputs.

full rationale

EVOLVE is an empirical systems/ML paper: its contributions are trained models and measured rate-distortion evaluations rather than theorems derived from assumptions. The variable-rate behavior is implemented directly by Eq. (2) (gain-scaled quantization) and by interpolating the learned gain vector, following prior work [14,69]; the continuous-rate claim is an operational property of the defined mechanism, not a fitted quantity renamed as a prediction. The entropy model, context model, and database curation are attributed to external or prior work, and no load-bearing step is justified by an unverified self-citation chain. The nearest concern is that some 'unseen' test volumes (e.g., ionization H+) may overlap in variable or simulation family with the training database in Appendix A, which weakens the cross-domain generalization claim; however, that is a data-separation/generalization risk rather than a logical reduction of a result to its inputs. Appendix H's acknowledged out-of-distribution failure (foot CT saturating at 33.7 dB although more rate is spent) is external evidence that EVOLVE's performance is not guaranteed by construction. The derivation chain is therefore self-contained as an empirical evaluation.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on hand-chosen architecture and training hyperparameters plus the domain assumption that the curated 21-simulation database transfers to unseen volumes. No new physical or formal entities are introduced. The learned gain vector is a fitted parameter, not an independent prediction. The most fragile item is the representativeness of the training corpus, which the paper itself partially falsifies for CT-like scanned data.

free parameters (5)
  • λ schedule for rate-distortion tradeoff = [100, 200, 400, 800, 1600, 3200, 6400, 12800]
    Eight discrete quality levels chosen by hand in Section 4.1; these define the operating points and the initialization of the gain vector.
  • Network hyperparameters = C1=96, C2=192, C3=384; L1..L5=[2,2,4,1,3]; M=320; N=256; channel groups=[16,16,32,64,192]
    Set after the ablation roadmap in Section 3.2 on a 303-volume subset; these directly determine the reported CR and reconstruction quality.
  • Block size = 128×128×128
    Chosen as the largest block affordable on the available GPU (Section 3.2.4); affects receptive field and memory, and hence CR/quality.
  • pHash deduplication threshold = NNS ≤ 0.85
    Hand-chosen threshold in Section 3.1 that controls the composition and diversity of the training database; no independent validation of this value.
  • Learnable gain vector γ = 8 values, initialized sqrt(λ/λ0), optimized during training
    Central to the variable-rate claim in Section 3.3. The continuous-rate behavior is obtained by interpolating between these fitted gains, so it is a fitted property rather than a prediction.
axioms (4)
  • standard math Arithmetic coding is near-optimal, so the cross-entropy rate R(y) equals the actual bitstream length.
    Invoked in Section 3.2.2 to justify using the rate-distortion loss R + λD as a training objective. This is standard information theory.
  • domain assumption The 21 curated simulations are representative of the target distribution over unseen volumes in the covered domains.
    This is the load-bearing premise for cross-dataset generalization, stated in Sections 1 and 3.1. The out-of-distribution foot result in Appendix H shows the assumption fails outside the covered domains.
  • ad hoc to paper The pHash-based deduplication with the 0.85 NNS threshold preserves diversity while removing redundancy.
    There is no independent evidence that 0.85 is the right threshold; it is a heuristic chosen by the authors to make the training set more diverse (Section 3.1).
  • domain assumption Random cropping, flipping, and rotation are distribution-preserving augmentations for scientific volumes.
    Used in Section 4.1 to improve generalization. This assumes scientific volumes are roughly invariant to these transforms, which is plausible but not verified per dataset.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of EVOLVE: Efficient Learned Volume Compression with Variable-Rate Encoding on a Cross-Domain Database." pith.science (2026). https://pith.science/paper/LTU7TS2I

@misc{pith2026260718187,
  author       = {Pith},
  title        = {Pith review of: EVOLVE: Efficient Learned Volume Compression with Variable-Rate Encoding on a Cross-Domain Database},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LTU7TS2I}},
  note         = {Machine review of arXiv:2607.18187}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large-scale scientific simulations generate volumetric data at rates that far outpace advances in storage and network bandwidth, making effective lossy compression increasingly critical. However, conventional compressors often struggle to preserve fine structural details at high compression ratios (CRs), and implicit neural representations (INRs) require costly per-volume optimization and produce models with fixed CRs. To respond, we present EVOLVE, an autoencoder (AE)-based volume-compression framework that targets high CRs for offline compression, with three key contributions. First, we construct a large-scale cross-domain database of 6,376 volumes from 21 scientific simulations, curated via perceptual hashing to ensure diversity, enabling the optimized model to extract features that generalize across volumes within the covered scientific simulation domains. Second, we reexamine the design space of AE-based compressors and incorporate several macro- and micro-designs into a vanilla AE to develop EVOLVE, which substantially improves the expressive power and compression capability. Third, we develop a learnable gain mechanism with a three-stage training strategy to enable variable-rate encoding, allowing a single model to support continuous CR adjustment at inference time. Experiments on multiple unseen scientific simulation datasets demonstrate that EVOLVE achieves substantially higher CRs than conventional compressors at comparable reconstruction quality, while delivering compression speeds that are orders of magnitude faster than INR-based methods, highlighting its promise as a strong alternative for compressing scientific data. The code, model weights, and results are available on our project page at https://evolve-vis.github.io.

Figures

Figures reproduced from arXiv: 2607.18187 by Chaoli Wang, Kaiyuan Tang, Maizhe Yang.

Figure 1
Figure 1. Figure 1: After training, EVOLVE, as an offline compression method, can achieve a competitive compression ratio (CR) on the unseen ionization (H+) dataset (141 MB). Unlike prior learning-based methods that require storing one model per CR setting, a single trained EVOLVE model supports a wide range of variable-rate encoding, as shown in (a). The encoding/decoding time plot in (b) uses a log scale to improve the visi… view at source ↗
Figure 2
Figure 2. Figure 2: We modernize a vanilla AE towards EVOLVE through pro [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of different entropy models and our proposed context model. (a) Factorized model with independent Gaussian assumptions. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) EVOLVE architecture. Q, EE, and ED denote quantization, entropy encoding, and entropy decoding. Blue and red lines indicate the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of volume rendering results between EVOLVE and conventional lossy compressors. Top and Bottom: combustion (MF) and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of isosurface rendering results between EVOLVE and conventional lossy compressors. Top and bottom: asteroids and [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of volume rendering results between EVOLVE and fully-connected INRs using the half-cylinder (VLM, 6,400) dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of isosurface rendering results between EVOLVE and grid-based INRs using the magnetic dataset. The chosen isovalue is [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Top: Comparison of volume rendering results between EVOLVE [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 1
Figure 1. Figure 1: Comparison of representative timestep selection results us [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Volume rendering of EVOLVE decompression results on asteroids-T (top) and Nyx-E (bottom), showing selected timesteps and ensemble [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of volume rendering results between EVOLVE and conventional lossy compressors. Top and bottom: asteroids and isotropic. [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of volume rendering results between EVOLVE and [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of isosurface rendering results between EVOLVE and conventional lossy compressors. Top and bottom: combustion (MF) and [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of isosurface rendering results between EVOLVE and [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Volume rendering comparison on scanned data (top to bottom: chameleon, engine, foot, and stag-beetle). For each method, the bottom-left inset shows the pixel-wise difference from GT, and the bottom-right inset shows the zoom-in of the red box. For foot, EVOLVE (33.71 dB) visibly smooths the trabecular bone texture, reflecting its quality saturation on out-of-distribution CT data, while for the other datase… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

87 extracted references · 17 canonical work pages · 1 internal anchor

  1. [1]

    Abbas, K

    A. Abbas, K. Tirumala, D. Simig, S. Ganguli, and A. S. Morcos. SemD- eDup: Data-efficientlearningatweb-scalethroughsemanticdeduplication. arXiv preprint arXiv:2303.09540, 2023. doi:10.48550/arXiv.2303.095403

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.09540 2023

  2. [3]

    J. Ballé. Efficient nonlinear transforms for lossy image compression. In Proc. Picture Coding Symposium, pp. 248–252, 2018. 3, 5

    2018

  3. [4]

    Ballé, V

    J. Ballé, V. Laparra, and E. P. Simoncelli. End-to-end optimized image compression. InProc. ICLR, 2017. 2, 4

    2017

  4. [5]

    Ballé, D

    J. Ballé, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston. Variational image compression with a scale hyperprior. InProc. ICLR, 2018. 2

    2018

  5. [6]

    Ballester-Ripoll, P

    R. Ballester-Ripoll, P. Lindstrom, and R. Pajarola. TTHRESH: Tensor compression for multidimensional visual data.IEEE Trans. Vis. Comput. Graph., 26(9):2891–2903, 2020. doi:10.1109/TVCG.2019.2904063 1, 2, 5, 6

    arXiv 2020

  6. [7]

    H. G. Barrow, J. M. Tenenbaum, R. C. Bolles, and H. C. Wolf. Parametric correspondence and chamfer matching: Two new techniques for image matching. InProc. IJCAI, pp. 659–663, 1977. 6

    1977

  7. [8]

    Beyer, M

    J. Beyer, M. Hadwiger, and H. Pfister. State-of-the-art in GPU-based large-scale volume visualization.Comput. Graph. Forum, 34(8):13–37,

  8. [9]

    Burkhart, S

    B. Burkhart, S. Appel, S. Bialy, J. Cho, A. Christensen, D. Collins et al. The catalogue for astrophysical turbulence simulations (CATS).Astrophys. J., 905(1):14, 2020. doi:10.3847/1538-4357/abc48410

    work page doi:10.3847/1538-4357/abc48410 2020

  9. [10]

    Chen, S.-H

    J. Chen, S.-H. Kao, H. He, W. Zhuo, S. Wen, C.-H. Lee et al. Run, don’t walk: Chasing higher FLOPS for faster neural networks. InProc. IEEE/CVF CVPR, pp. 12021–12031, 2023. doi:10.1109/CVPR52729.2023. 011575

    arXiv 2023

  10. [11]

    Y. Chen, Q. Wu, W. Lin, M. Harandi, and J. Cai. HAC: Hash-grid assisted context for 3D Gaussian splatting compression. InProc. ECCV, pp. 422–438, 2024. doi:10.1007/978-3-031-72667-5_243, 5

    work page doi:10.1007/978-3-031-72667-5_243 2024

  11. [12]

    Cheng, H

    Z. Cheng, H. Sun, M. Takeuchi, and J. Katto. Learned image compression with discretized Gaussian mixture likelihoods and attention modules. In Proc. IEEE/CVF CVPR, pp. 7936–7945, 2020. doi:10.1109/CVPR42600. 2020.007962

    arXiv 2020

  12. [13]

    Scientific visualization datasets

    Computer Graphics Laboratory, ETH Zürich. Scientific visualization datasets. https://cgl.ethz.ch/research/visualization/data. php. 3, 10

  13. [15]

    Interactivespatio-temporalexploration of massive time-varying rectilinear scalar volumes based on a variable bit-rate sparse representation over learned dictionaries.Comput

    J.Díaz,F.Marton,andE.Gobbetti. Interactivespatio-temporalexploration of massive time-varying rectilinear scalar volumes based on a variable bit-rate sparse representation over learned dictionaries.Comput. Graph., 88:45–56, 2020. doi:10.1016/j.cag.2020.03.0021, 2, 8

    work page doi:10.1016/j.cag.2020.03.0021 2020

  14. [16]

    D. B. Fielding, E. C. Ostriker, G. L. Bryan, and A. S. Jermyn. Multiphase gas and the fractal nature of radiative turbulent mixing layers.Astrophys. J. Lett., 894(2):L24, 2020. doi:10.3847/2041-8213/ab8d2c10

    work page doi:10.3847/2041-8213/ab8d2c10 2020

  15. [17]

    Gobbetti, J

    E. Gobbetti, J. A. I. Guitián, and F. Marton. COVRA: A compression- domain output-sensitive volume rendering architecture based on a sparse representation of voxel blocks.Comput. Graph. Forum, 31(3pt4):1315– 1324, 2012. doi:10.1111/J.1467-8659.2012.03124.X1, 2, 8

    arXiv 2012

  16. [18]

    Accurate, large minibatch SGD: Training ImageNet in 1 hour.arXiv preprint arXiv:1706.02677, 2017

    P.Goyal,P.Dollár,R.B.Girshick,P.Noordhuis,L.Wesolowski,A.Kyrola et al. Accurate, large minibatch SGD: Training ImageNet in 1 hour.arXiv preprint arXiv:1706.02677, 2017. doi:10.48550/arXiv.1706.026773

    Pith/arXiv arXiv doi:10.48550/arxiv.1706.02677 2017

  17. [19]

    J. Han, K. Tang, and C. Wang. MoE-INR: Implicit neural representation with mixture-of-experts for time-varying volumetric data compression. IEEETrans.Vis.Comput.Graph.,32(1):254–264,2026.doi: 10.1109/TVCG .2025.36338936

    arXiv 2026

  18. [20]

    J. Han, J. Tao, and C. Wang. FlowNet: A deep learning framework for clustering and selection of streamlines and stream surfaces.IEEE Trans. Vis. Comput. Graph., 26(4):1732–1744, 2020. doi:10.1109/TVCG.2018. 28802071, 2, 11

    doi:10.1109/tvcg.2018 2020

  19. [21]

    Han and C

    J. Han and C. Wang. SSR-TVD: Spatial super-resolution for time- varying data analysis and visualization.IEEE Trans. Vis. Comput. Graph., 28(6):2445–2456, 2022. doi:10.1109/TVCG.2020.30321233, 9

    arXiv 2022

  20. [22]

    J. Han, H. Zheng, and C. Bi. KD-INR: Time-varying volumetric data compressionviaknowledgedistillation-basedimplicitneuralrepresentation. IEEE Trans. Vis. Comput. Graph., 30(10):6826–6838, 2024. doi:10.1109/ TVCG.2023.33453731, 2, 6

    arXiv 2024

  21. [23]

    J. Han, H. Zheng, D. Z. Chen, and C. Wang. STNet: An end-to-end generative framework for synthesizing spatiotemporal super-resolution volumes.IEEE Trans. Vis. Comput. Graph., 28(1):270–280, 2022. doi:10. 1109/TVCG.2021.31148152, 3

    arXiv 2022

  22. [24]

    Scalarmixing in direct numerical simulations of temporally evolving plane jet flames with skeletal CO/H2 kinetics.Proc

    E.R.Hawkes,R.Sankaran,J.C.Sutherland,andJ.H.Chen. Scalarmixing in direct numerical simulations of temporally evolving plane jet flames with skeletal CO/H2 kinetics.Proc. Combust. Inst., 31(1):1633–1640,

  23. [25]

    D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y. Wang. ELIC: Efficient learned image compression with unevenly grouped space-channel contex- tual adaptive coding. InProc. IEEE/CVF CVPR, pp. 5708–5717, 2022. doi:10.1109/CVPR52688.2022.005632

    arXiv 2022

  24. [26]

    D. He, Y. Zheng, B. Sun, Y. Wang, and H. Qin. Checkerboard context modelfor efficientlearnedimagecompression. InProc.IEEE/CVFCVPR, pp. 14771–14780, 2021. doi:10.1109/CVPR46437.2021.014532, 5

    arXiv 2021

  25. [27]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. InProc. IEEE/CVF CVPR, pp. 770–778, 2016. doi: 10. 1109/CVPR.2016.905

    2016

  26. [29]

    3D-Spatiotemporal Forecasting the Expansion of Supernova Shells Using Deep Learning toward High-Resolution Galaxy Simulations

    K. Hirashima, K. Moriwaki, M. S. Fujii, Y. Hirai, T. R. Saitoh, and J. Makino. 3D-spatiotemporal forecasting the expansion of supernova shells using deep learning towards high-resolution galaxy simulations. Mon. Not. R. Astron. Soc., 526(3):4054–4066, 2023. doi:10.48550/arXiv. 2302.0002610

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2302.00026 2023

  27. [30]

    Hirashima, K

    K. Hirashima, K. Moriwaki, M. S. Fujii, Y. Hirai, T. R. Saitoh, J. Makino et al. Surrogate modeling for computationally expensive simulations of supernovae in high-resolution galaxy simulations.arXiv preprint arXiv:2311.08460, 2023. doi:10.48550/arXiv.2311.0846010

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.08460 2023

  28. [31]

    IEEE SciVis contests.https://sciviscontest.ieeevis

    IEEE VIS. IEEE SciVis contests.https://sciviscontest.ieeevis. org/. 3, 10

  29. [32]

    S. Jeon, K. Tang, C. Wang, and W.-K. Jeong. Super-Gaussian: Interactive sceneeditingfor3DGaussiansplattingandNLI-basedvolumevisualization invirtualreality.IEEETrans.Vis.Comput.Graph.,33(1),2027. Accepted. 2

    2027

  30. [33]

    Deep Fluids: A generative network for parameterized fluid simulations

    B.Kim,V.C.Azevedo,N.Thuerey,T.Kim,M.H.Gross,andB.Solenthaler. Deep Fluids: A generative network for parameterized fluid simulations. Comput. Graph. Forum, 38(2):59–70, 2019. doi:10.1111/CGF.136192

    work page doi:10.1111/cgf.136192 2019

  31. [34]

    Openscientificvisualizationdatasets

    P.Klacansky. Openscientificvisualizationdatasets. http://klacansky. com/open-scivis-datasets/, 2017. 3, 10, 11

    2017

  32. [35]

    ImageNetclassificationwith deep convolutional neural networks

    A.Krizhevsky,I.Sutskever,andG.E.Hinton. ImageNetclassificationwith deep convolutional neural networks. InProc. NeurIPS, pp. 1106–1114,

  33. [36]

    K. Lee, D. Ippolito, A. Nystrom, C. Zhang, D. Eck, C. Callison-Burch et al. Deduplicating training data makes language models better. InProc. ACL, pp. 8424–8445, 2022. doi:10.18653/V1/2022.ACL-LONG.5773

    work page doi:10.18653/v1/2022.acl-long.5773 2022

  34. [37]

    Liang, K

    X. Liang, K. Zhao, S. Di, S. Li, R. Underwood, A. M. Gok et al. SZ3: A modular framework for composing prediction-based error-bounded lossy compressors.IEEE Trans. Big Data, 9(2):485–498, 2023. doi:10. 1109/TBDATA.2022.32011761, 2, 6

    arXiv 2023

  35. [38]

    Lindstrom

    P. Lindstrom. Fixed-rate compressed floating-point arrays.IEEE Trans. Vis. Comput. Graph., 20(12):2674–2683, 2014. doi:10.1109/TVCG.2014. 23464581, 2, 6

    doi:10.1109/tvcg.2014 2014

  36. [39]

    Exploringautoencoder- based error-bounded compression for scientific data

    J.Liu,S.Di,K.Zhao,S.Jin,D.Tao,X.Liangetal. Exploringautoencoder- based error-bounded compression for scientific data. InProc. IEEE CLUSTER, pp. 294–306, 2021. doi:10.1109/CLUSTER48925.2021.00034 1, 2, 3, 5

    arXiv 2021

  37. [40]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In Proc. ICLR, 2019. 4, 6

    2019

  38. [41]

    G. Lu, W. Ouyang, D. Xu, X. Zhang, C. Cai, and Z. Gao. DVC: An end-to- end deep video compression framework. InProc. IEEE/CVF CVPR, pp. 11006–11015, 2019. doi:10.1109/CVPR.2019.011263

    arXiv 2019

  39. [42]

    Compressiveneuralrepresen- tationsofvolumetricscalarfields.Comput.Graph.Forum,40(3):135–146,

    Y.Lu,K.Jiang,J.A.Levine,andM.Berger. Compressiveneuralrepresen- tationsofvolumetricscalarfields.Comput.Graph.Forum,40(3):135–146,

  40. [43]

    Marton, M

    F. Marton, M. Agus, and E. Gobbetti. A framework for GPU-accelerated exploration of massive time-varying rectilinear scalar volumes.Comput. Graph. Forum, 38(3):53–66, 2019. doi:10.1111/cgf.136711, 2, 8

    work page doi:10.1111/cgf.136711 2019

  41. [44]

    Minnen, J

    D. Minnen, J. Ballé, and G. Toderici. Joint autoregressive and hierarchical priorsforlearnedimagecompression. InProc.NeurIPS,pp.10794–10803,

  42. [45]

    Minnen and S

    D. Minnen and S. Singh. Channel-wise autoregressive entropy models for learned image compression. InProc. IEEE ICIP, pp. 3339–3343, 2020. doi:10.1109/ICIP40778.2020.91909352, 5

    arXiv 2020

  43. [46]

    Müller, A

    T. Müller, A. Evans, C. Schied, and A. Keller. Instant neural graphics primitives with a multiresolution hash encoding.ACM Trans. Graph., 41(4):102:1–102:15, 2022. doi:10.1145/3528223.35301272, 6, 8

    arXiv 2022

  44. [47]

    Nystad, A

    J. Nystad, A. Lassen, A. Pomianowski, S. Ellis, and T. Olson. Adaptive scalable texture compression. InProc. ACM SIGGRAPH/Eurographics HPG, pp. 105–114, 2012. doi:10.2312/EGGH/HPG12/105-1142

    work page doi:10.2312/eggh/hpg12/105-1142 2012

  45. [48]

    Ohana, M

    R. Ohana, M. McCabe, L. Meyer, R. Morel, F. Agocs, M. Beneitez et al. The Well: A large-scale collection of diverse physics simulations for machine learning. InProc. NeurIPS, pp. 44989–45037, 2024. 3, 10

    2024

  46. [49]

    DINOv2: Learningrobustvisualfeatureswithoutsupervision.Trans

    M.Oquab,T.Darcet,T.Moutakanni,H.V.Vo,M.Szafraniec,V.Khalidov etal. DINOv2: Learningrobustvisualfeatureswithoutsupervision.Trans. Mach. Learn. Res., 2024, 2024. 3

    2024

  47. [50]

    Z. Pan, K. Tang, J. Xia, Y. Qin, L. Gu, C. Wang et al. SGI: Structured 2D Gaussians for efficient and compact large image representation. InProc. IEEE/CVF CVPR, pp. 12162–12172, 2026. doi:10.48550/arXiv.2603.07789 3

    doi:10.48550/arxiv.2603.07789 2026

  48. [51]

    Popinet, M

    S. Popinet, M. Smith, and C. Stevens. Experimental and numerical study of the turbulence characteristics of airflow around a research vessel.J. Atmos. Ocean. Technol., 21(10):1575–1589, 2004. doi:10.1175/1520-0426 (2004)021<1575:EANSOT>2.0.CO;210

    work page doi:10.1175/1520-0426 2004

  49. [52]

    W. P. Porter, Y. Xing, B. R. von Ohlen, J. Han, and C. Wang. A deep learning approach to selecting representative time steps for time-varying multivariate data. InProc. IEEE VIS (Short Papers), pp. 131–135, 2019. doi:10.1109/VISUAL.2019.893375911

    arXiv 2019

  50. [53]

    Universalmodelingandcoding.IEEE Trans

    J.RissanenandJ.Langdon,GlenG. Universalmodelingandcoding.IEEE Trans. Inf. Theory, 27(1):12–22, 1981. doi:10.1109/TIT.1981.10562824

    arXiv 1981

  51. [54]

    M. B. Rodríguez, E. Gobbetti, J. A. I. Guitián, M. Makhinya, F. Marton, R. Pajarola et al. State-of-the-art in compressed GPU-based direct volume rendering.Comput. Graph. Forum, 33(6):77–100, 2014. doi:10.1111/cgf. 122802

    doi:10.1111/cgf 2014

  52. [55]

    Vectorfieldtopologyoftime-dependentflowsin asteadyreferenceframe.IEEETrans.Vis.Comput.Graph.,26(1):280–290,

    I.B.RojoandT.Günther. Vectorfieldtopologyoftime-dependentflowsin asteadyreferenceframe.IEEETrans.Vis.Comput.Graph.,26(1):280–290,

  53. [56]

    Schneider and R

    J. Schneider and R. Westermann. Compression domain volume rendering. InProc. IEEE VIS, pp. 293–300, 2003. doi:10.1109/VISUAL.2003.1250385 1, 2, 8

    Pith/arXiv arXiv 2003

  54. [57]

    J. Shen, H. Li, J. Xu, A. Biswas, and H.-W. Shen. IDLat: An importance- driven latent generation method for scientific data.IEEE Trans. Vis. Comput. Graph., 29(1):679–689, 2023. doi:10.1109/TVCG.2022.3209419 1, 2, 3, 4, 5, 6, 11

    arXiv 2023

  55. [58]

    Sheng, J

    X. Sheng, J. Li, B. Li, L. Li, D. Liu, and Y. Lu. Temporal context mining forlearnedvideocompression.IEEETrans.Multim., 25:7311–7322, 2023. doi:10.1109/TMM.2022.32204213, 5

    arXiv 2023

  56. [59]

    Sitzmann, J

    V. Sitzmann, J. N. P. Martel, A. W. Bergman, D. B. Lindell, and G. Wet- zstein. Implicit neural representations with periodic activation functions. InProc. NeurIPS, pp. 7462–7473, 2020. 1, 2, 6

    2020

  57. [60]

    H. Son, J. Noh, S. Jeon, C. Wang, and W.-K. Jeong. MC-INR: Efficient encoding of multivariate scientific simulation data using meta-learning and clustered implicit neural representations. InProc. IEEE VIS (Short Papers), pp. 206–210, 2025. doi:10.1109/VIS60296.2025.000472

    arXiv 2025

  58. [61]

    J. Sun, D. Lenz, H. Yu, and T. Peterka. F-Hash: Feature-based hash designfortime-varyingvolumevisualizationviamulti-resolutiontesseract encoding.IEEE Trans. Vis. Comput. Graph., 32(1):396–406, 2026. doi: 10.1109/TVCG.2025.36348122

    arXiv 2026

  59. [63]

    K. Tang, D. Burke, and C. Wang. Lossless-INR: Lossless volumetric implicit neural representations. InProc. IEEE VIS (Short Papers), 2026. Accepted. 2

    2026

  60. [64]

    Tang and C

    K. Tang and C. Wang. ECNR: Efficient compressive neural representation of time-varying volumetric datasets. InProc. IEEE PacificVis, pp. 72–81,

  61. [65]

    Tang and C

    K. Tang and C. Wang. STSR-INR: Spatiotemporal super-resolution for time-varyingmultivariatevolumetricdataviaimplicitneuralrepresentation. Comput. Graph., 119:103874, 2024. doi:10.1016/j.cag.2024.01.0012

    work page doi:10.1016/j.cag.2024.01.0012 2024

  62. [66]

    Tang and C

    K. Tang and C. Wang. StyleRF-VolVis: Style transfer of neural radiance fields for expressive volume visualization.IEEE Trans. Vis. Comput. Graph., 31(1):613–623, 2025. doi:10.1109/TVCG.2024.34563422

    arXiv 2025

  63. [67]

    Tang and C

    K. Tang and C. Wang. ECoNGS: Efficient compressive neural Gaussian splats for volume visualization.IEEE Trans. Vis. Comput. Graph., 33(1),

  64. [68]

    K. Tang, S. Yao, and C. Wang. iVR-GS: Inverse volume rendering for explorable visualization via editable 3D Gaussian splatting.IEEE Trans. Vis. Comput. Graph., 31(6):3783–3795, 2025. doi:10.1109/TVCG.2025. 35671212

    doi:10.1109/tvcg.2025 2025

  65. [69]

    K. Tong, Y. Wu, Y. Li, K. Zhang, L. Zhang, and X. Jin. QVRF: A quantization-error-aware variable rate framework for learned image com- pression. InProc.IEEEICIP,pp.1310–1314,2023.doi: 10.1109/ICIP49359 .2023.102227176

    arXiv 2023

  66. [70]

    Treib, K

    M. Treib, K. Bürger, F. Reichl, C. Meneveau, A. S. Szalay, and R. West- ermann. Turbulence visualization at the terascale on desktop PCs.IEEE Trans. Vis. Comput. Graph., 18(12):2169–2177, 2012. doi:10.1109/TVCG. 2012.2742, 8

    arXiv 2012

  67. [71]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez et al. Attention is all you need. InProc. NeurIPS, pp. 5998–6008, 2017. doi: 10.48550/arXiv.1706.037625

    Pith/arXiv arXiv doi:10.48550/arxiv.1706.03762 2017

  68. [72]

    G. K. Wallace. The JPEG still picture compression standard.Commun. ACM, 34(4):30–44, 1991. doi:10.1145/103085.1030892

    arXiv 1991

  69. [73]

    Wang and J

    C. Wang and J. Han. DL4SciVis: A state-of-the-art survey on deep learning for scientific visualization.IEEE Trans. Vis. Comput. Graph., 29(8):3714–3733, 2023. doi:10.1109/TVCG.2022.31678962

    arXiv 2023

  70. [74]

    C. Wang, H. Yu, and K.-L. Ma. Importance-driven time-varying data visualization.IEEE Trans. Vis. Comput. Graph., 14(6):1547–1554, 2008. doi:10.1109/TVCG.2008.14011

    arXiv 2008

  71. [75]

    Y. Wang, Z. Li, L. Guo, W. Yang, A. C. Kot, and B. Wen. ContextGS: Compact 3D Gaussian splatting with anchor level context model. InProc. NeurIPS, pp. 51532–51551, 2024. 3, 5

    2024

  72. [76]

    Webster, J

    R. Webster, J. Rabin, L. Simon, and F. Jurie. On the de-duplication of LAION-2B.arXiv preprint arXiv:2303.12733, 2023. doi:10.48550/arXiv. 2303.127333

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.12733 2023

  73. [77]

    Fastneuralrepresentationsfor direct volume rendering.Comput

    S.Weiss,P.Hermüller,andR.Westermann. Fastneuralrepresentationsfor direct volume rendering.Comput. Graph. Forum, 41(6):196–211, 2022. doi:10.1111/CGF.145781, 2, 6, 8

    work page doi:10.1111/cgf.145781 2022

  74. [78]

    Whalen and M

    D. Whalen and M. L. Norman. Ionization front instabilities in primordial HIIregions.Astrophys.J., 673(2):664–675, 2008.doi: 10.1086/524400 10

    doi:10.1086/524400 2008

  75. [79]

    Q. Wu, D. Bauer, M. J. Doyle, and K.-L. Ma. Interactive volume visualiza- tion via multi-resolution hash encoding based neural representation.IEEE Trans. Vis. Comput. Graph., 30(8):5404–5418, 2024. doi:10.1109/TVCG. 2023.32931211, 8

    arXiv 2024

  76. [81]

    S. W. Wurster, T. Xiong, H.-W. Shen, H. Guo, and T. Peterka. Adaptively placed multi-grid scene representation networks for large-scale data visu- alization.IEEE Trans. Vis. Comput. Graph., 30(1):965–974, 2024. doi:10. 1109/TVCG.2023.33271941, 2, 5, 8

    arXiv 2024

  77. [82]

    L. Yan, X. Liang, H. Guo, and B. Wang. TopoSZ: Preserving topology in error-bounded lossy compression.IEEE Trans. Vis. Comput. Graph., 30(1):1302–1312, 2024. doi:10.1109/TVCG.2023.33269202

    arXiv 2024

  78. [83]

    M. Yang, K. Tang, and C. Wang. Meta-INR: Efficient encoding of volumetric data via meta-learning implicit neural representation. InProc. IEEE PacificVis (Visualization Notes), pp. 246–251, 2025. doi:10.1109/ PACIFICVIS64226.2025.000302

    arXiv 2025

  79. [84]

    P. Yin, J. Lyu, S. Zhang, S. J. Osher, Y. Qi, and J. Xin. Understanding straight-through estimator in training activation quantized neural nets. In Proc. ICLR, 2019. 6

    2019

  80. [85]

    Zauner, M

    C. Zauner, M. Steinebach, and E. Hermann. Rihamark: Perceptual image hash benchmarking. InProc. IS&T/SPIE MWSF, pp. 343–357, 2011. doi: 10.1117/12.8766173

    work page doi:10.1117/12.8766173 2011

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.