Pith. sign in

REVIEW 4 major objections 4 minor 45 references

Treating each voxel as a full time series, not scattered space-time points, lets neural compression of time-varying volumes run ~50x faster with comparable quality on most datasets.

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 ↗

Time-varying volumes are compressed by mapping each spatial coordinate directly to its full temporal sequence, using mixture-of-experts routing and low-rank decoders.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection The sequence-level INR reformulation is a real idea, but the paper's own tables contradict the quality claim and the speedup comparison isn't yet fair. the 4 major comments →

arxiv 2607.20970 v1 pith:SIFHLRXU submitted 2026-07-23 cs.AI

From Scalars to Time Series: Rethinking Implicit Neural Representations for Time-Varying Volumetric Data

Weihan Zhang , Xuan Zhao , Yenwen Peng , Yuqi Chen , Jun Tao This is my paper
classification cs.AI MSC 68T0768U10
keywords implicit neural representationtime-varying volumesequence-level supervisionmixture of expertsvolumetric compressiontemporal coherencecoordinate-based networkslow-rank adaptation
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 argues that dense spatiotemporal sampling is unnecessary when training implicit neural representations (INRs) on time-varying volumes. It reformulates the data as a spatial array of temporal signals - a mapping from each spatial location to its T time steps - and supervises each location with its entire time series rather than with individual scalar samples. The central claim is that this sequence-level supervision consistently improves reconstruction quality while cutting compression time by about an order of magnitude (42-60x) across several INR backbones, with a mixture-of-experts instantiation adding a trainable time-embedding-guided router and low-rank expert decoders for heterogeneous dynamics. The paper is careful to note that the quality advantage is not uniform - on the vortex dataset the scalar-wise MoE baseline remains ahead - so the headline gain is efficiency with competitive fidelity. If the claim holds, INR-based compression of scientific volumes shifts from a multi-day to a sub-hour workflow.

Core claim

Coordinate-wise INR training treats each (x,t) sample as an independent scalar. The paper instead defines F(x) = [V_1(x),...,V_T(x)], supervising each voxel's full temporal sequence with one squared-error loss, and argues this is why the same backbones reach comparable or better PSNR at 42-60x lower compression time. The formulation is architecture-agnostic: SIREN, CoordNet, NeurComp, and MoE baselines adapted to sequence prediction retain the speedup. The paper concedes that for high-frequency, weakly coherent dynamics (e.g., vortex), the scalar formulation can still win on PSNR.

What carries the argument

The central object is the reformulation F: Omega -> R^T, which turns each voxel into a temporal sequence and replaces scalar-wise sampling with sequence-level supervision, Loss(x) = ||s_hat(x) - s(x)||^2. The MoE instantiation consists of a spatial encoder with trainable harmonic positional encoding, a router that fuses spatial features with a trainable time-embedding table E in R^(T x M) (mean/variance-pooled temporally) and selects one expert via hard routing, and a shared expert decoder with LoRA adaptations W_k = W_0 + A_k B_k to keep parameter growth small. A clustering warm-up initializes the router with pseudo-labels. The role of this machinery is to make temporal coherence the primar

Load-bearing premise

The efficiency claim rests on the premise that a single sequence-level forward pass converges to target fidelity in far fewer optimization steps than coordinate-wise scalar training, so that the wall-clock speedup is not an artifact of the baselines' training budgets; the paper does not report FLOPs or convergence curves, and its own vortex result (scalar MoE-INR ahead by 4.2 dB) shows the quality premise is not uniform.

What would settle it

Run the same INR backbone in both scalar-wise and sequence-wise mode with equal total FLOPs (or equal numbers of network evaluations) on vortex and combustion, plotting PSNR against training time; if scalar-wise reaches equal PSNR with no more compute, the central efficiency claim collapses. A simpler check from the paper's own tables: on vortex, coordinate-wise MoE-INR reaches 52.38 dB versus 48.18 dB for the sequence formulation - if a convergence-curve comparison shows the sequence model saturates below the scalar model at any budget, the 'consistently improves quality' claim is falsified o

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

If this is right

  • Compression of time-varying volumes drops from tens of hours to about an hour (42-60x wall-clock speedup) with the same backbone and similar PSNR.
  • Decompression time becomes nearly flat in the number of reconstructed frames: decoding 16 frames costs about the same as decoding one, unlike scalar-wise models whose cost grows linearly.
  • The formulation transfers across INR architectures: SIREN, CoordNet, NeurComp, and MoE baselines can be adapted to sequence prediction and all speed up, with MoE models suffering the least quality loss.
  • Compared with traditional error-bounded compressors (ZFP, SZ3, TTHRESH), the method reaches 9.6x-102.2x higher compression at matched PSNR, at the price of slower encoding.
  • Per-time-step PSNR curves are smoother than those of scalar baselines, indicating more stable temporal behavior across the sequence.

Where Pith is reading between the lines

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

  • A FLOP-matched or sample-matched comparison is needed to separate algorithmic gains from implementation: the paper compares wall-clock hours without reporting per-iteration counts or convergence curves for the scalar baselines, so part of the speedup may reflect the fact that sequence training reuses the same spatial encoding across all T outputs in one pass.
  • The router's hard assignments effectively segment the volume into regions with similar temporal behavior; that map is a free byproduct that could be used for feature tracking, region-of-interest detection, or designing transfer functions.
  • The same reformulation should transfer to any structured output axis with strong coherence, such as spectral data or parameter-indexed fields, wherever the output dimension can be predicted jointly.
  • Enabling the in-situ encoding the paper mentions would require a partial-sequence variant, because the sequence-level loss assumes complete temporal trajectories at training time; a mini-batched time-step scheme could preserve the efficiency idea without waiting for the whole simulation.
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

4 major / 4 minor

Summary. The paper proposes reformulating INR-based compression of time-varying volumetric data from coordinate-wise scalar prediction to sequence-level prediction: each spatial coordinate is mapped to its full temporal trajectory. The authors argue this eliminates dense spatiotemporal sampling, improves reconstruction quality, and drastically reduces training cost. They instantiate the idea with a mixture-of-experts (MoE) INR containing a spatial encoder, a temporal-aware router, and LoRA-based expert decoders, and evaluate it on four scientific datasets against learning-based and traditional compressors. The paper also reports ablations on warm-up, LoRA, time embeddings, expert counts, and a controlled analysis in which existing INR backbones are adapted to the sequence setting.

Significance. If the central claims hold, the paper would make a useful contribution to time-varying volume compression: the sequence-level formulation is architecturally simple, compatible with several INR backbones, and the reported wall-clock savings are large. The paper is also careful in several respects: it includes extensive ablations, per-step PSNR curves, a temporal-coherence analysis, a controlled '-seq' comparison with existing backbones, and a discussion of practical trade-offs such as random time-step access. These are genuine strengths. However, the significance is currently limited by overclaiming: the abstract's 'consistently improves reconstruction quality' is contradicted by the paper's own tables, and the efficiency claim rests on an underspecified comparison of training budgets.

major comments (4)
  1. [Abstract; Sec. 4.2, Table 2] The abstract claims the reformulation 'consistently improves reconstruction quality,' but Table 2 shows the opposite on two of four datasets: on combustion (CHI), ours attains 43.85 dB versus CoordNet's 44.42 dB, and on vortex, ours attains 48.18 dB versus MoE-INR's 52.38 dB. The paper itself later acknowledges in Sec. 6 that the method 'does not consistently achieve the highest reconstruction quality.' The abstract and introduction need to be recalibrated. Also, no error bars or repeated-run statistics are reported, so small PSNR differences are hard to interpret.
  2. [Sec. 4.1; Appendix A] The efficiency claim is load-bearing and currently under-specified. Sec. 4.1 states that all methods use 'the same voxel coordinate sampling strategy and sampling ratio,' but the ratio is never defined. Appendix A says our model trains with 1,600 spatial coordinates per batch, each producing a T-dimensional output, hence 1,600×T scalar supervisions per step. If a scalar baseline uses 1,600 random (x,y,z,t) samples per step, the comparison supervises T times fewer values per step and requires T times more steps per epoch. Without per-iteration sample counts, FLOPs, or convergence curves, the reported wall-clock speedup may reflect unequal optimization budgets rather than the formulation. At minimum, the manuscript should report these quantities and compare methods under matched supervision volume or matched epochs, and should re-derive the speedup ranges.
  3. [Sec. 4.2; Sec. 5.2, Table 7] The controlled '-seq' comparison in Sec. 5.2 does not rescue the quality claim. On vortex, at CR=2,300, our sequence model reaches 42.41 dB, while the paper reports that scalar CoordNet reaches 44.09 dB at the same compression ratio. Thus even when compression ratio is matched, the sequence formulation is 1.7 dB worse than a scalar baseline. This is acknowledged indirectly by the drop for CoordNet-seq (36.10 dB), but it contradicts the general claim that sequence-level supervision improves fidelity. The paper should either restrict the quality claim to specific datasets/conditions or explain why the efficiency gain justifies a fidelity loss in these cases.
  4. [Sec. 4.2, Table 2; Sec. 5.2, Table 7] The numerical speedup ranges in the text are not reproducible from the reported tables. The paper states compression speedups of 42.6–59.9× and decompression speedups of 25.7–38.0×. From Table 2, for example, vortex CoordNet compression is 18.63/0.54 ≈ 34.5×, and combustion Neural Experts is 45.86/0.64 ≈ 71.7×; decompression ratios similarly span wider ranges depending on the baseline and dataset. The ranges should be recomputed or the basis for them stated explicitly. This matters because the efficiency advantage is the paper's central quantitative contribution.
minor comments (4)
  1. [Throughout] Several table headers appear as 'T able' instead of 'Table'; please fix formatting across the manuscript.
  2. [Sec. 4.1] The sentence 'all baseline methods are uniformly configured under the scalar prediction setting' is not enough to establish fairness. Please specify the exact batch sizes, number of scalar samples per iteration, and number of epochs for each baseline, including whether they use the same total number of scalar observations as our method.
  3. [Sec. 5.2] The sentence reporting that 'Neural Experts and Switch-NeRF achieve 32.20 and 35.68 PSNR, respectively, in their original formulations' is confusing because those numbers do not appear in Table 7. Either add a table entry or clarify that these are results from the full comparison at a different CR.
  4. [Sec. 4.4, Fig. 8] The x-axis compression-ratio values are not evenly spaced and the legend is shared across four panels; consider using log-scale x-axes or separate legends to improve readability.

Circularity Check

0 steps flagged

No significant circularity: the central claims are empirical comparisons and the reformulation is a genuine modeling change, not a fitted parameter renamed as a prediction.

full rationale

The paper contains no derivation that reduces to its own inputs. The central contribution is a reformulation: instead of coordinate-wise scalar supervision, the network maps spatial coordinates to full temporal sequences and is trained with the sequence loss in Eq. (3). This is a change of output representation and loss, not a circular construction. The warm-up step in Sec. 3.3 clusters the actual temporal sequences and uses the cluster labels as pseudo-ground-truth router targets for the first 10% of training; this is a standard supervised initialization of a routing network, and the final model is trained end-to-end without clustering supervision. The paper's ablations (Appendix B, Tabs. 3-5) compare variants against each other, and the main tables compare against external baselines, including adapting existing INR backbones to the sequence setting (Tab. 7). No fitted parameter is relabeled as a prediction, no 'uniqueness theorem' is invoked, and the cited prior work is used for standard components (positional encoding, sinusoidal initialization, LoRA), not as load-bearing evidence for the paper's own conclusions. The paper itself acknowledges that the method does not consistently achieve the highest reconstruction quality (Sec. 6, Limitations), which is an overclaim concern rather than a circularity concern. The possible ambiguity in the training budget comparison is an experimental-control issue, not a circularity issue. No self-citation chain is load-bearing, and no equation in the paper is equivalent by construction to the claimed result. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper is an empirical method paper; no derivation is presented. The central claims rely on hyperparameters chosen by ablation and on the domain assumption that temporal trajectories are structured enough for vector-output regression. No new physical or latent entities are postulated beyond learned network components.

free parameters (6)
  • Number of experts = 7
    Selected by ablation on vortex (Tab. B.5) and following prior MoE-INR [7]; PSNR gains saturate beyond 12 experts.
  • LoRA rank = 8
    Selected by ablation on combustion (MF) and vortex (Tab. B.4); quality is relatively insensitive to rank.
  • Warm-up ratio = 10% of epochs
    Selected by ablation on ionization (PD) (Tab. B.3); optimal in a 5–30% sweep.
  • Embedding dimension M = 352 for vortex; varies by dataset
    Set per dataset; part of model capacity.
  • Learning rate / batch / epochs = 1e-5, 1600, 300
    Fixed training protocol following SIREN-style initialization.
  • Time embedding table E = T x M trainable parameters
    Learned latent temporal context for routing; not derived.
axioms (5)
  • domain assumption Complete temporal sequences are available at every spatial location during training
    Sec. 3.1 defines F(x) using the full time series; 'dense' temporal sampling is still required in the data, only optimization-time sampling changes.
  • domain assumption Temporal evolution is smooth enough to be learned as a vector output
    Sec. 3 and Sec. 6 Discussion assume time is regularly sampled and evolves smoothly; the paper's own ablation on combustion CHI shows this fails for high-frequency, weakly coherent dynamics.
  • domain assumption A small number of experts (7) can cover the diversity of temporal patterns
    Warm-up clustering (Sec. 3.3) assumes cluster structure exists; ablation shows quality saturates with more experts.
  • domain assumption Low-rank adaptation preserves enough capacity for expert specialization
    LoRA reduces expert variance (App. B Tab. 4); on combustion (MF) PSNR drops from 42.35 to 40.87, showing the assumption is not universally satisfied.
  • standard math SIREN initialization and sinusoidal activations stabilize training
    Standard INR practice cited to Sitzmann et al. [30]; no proof needed.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of From Scalars to Time Series: Rethinking Implicit Neural Representations for Time-Varying Volumetric Data." pith.science (2026). https://pith.science/paper/SIFHLRXU

@misc{pith2026260720970,
  author       = {Pith},
  title        = {Pith review of: From Scalars to Time Series: Rethinking Implicit Neural Representations for Time-Varying Volumetric Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SIFHLRXU}},
  note         = {Machine review of arXiv:2607.20970}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Implicit neural representations (INRs) for time-varying volumetric data are typically trained using dense sampling over spatiotemporal coordinates, where each observation corresponds to a single point in space and time. This coordinate-wise formulation requires extensive sampling during optimization, leading to high computational cost and inefficient use of temporal structure. In this work, we revisit this design choice and show that dense spatiotemporal sampling is not necessary for learning time-varying fields. Instead, we represent the data as a collection of spatially indexed time series and train INRs using sequence-level supervision over each spatial location, rather than coordinate-wise scalar samples. This reformulation eliminates the need for dense spatiotemporal sampling and instead learns each spatial location from its full temporal evolution in a structured manner. We demonstrate that this representation is compatible with a range of existing INR architectures and consistently improves reconstruction quality, while significantly reducing training cost. Furthermore, we show that this formulation can be combined with mixture-of-experts architectures, and that our MoE instantiation further improves reconstruction quality compared to both the base reformulation and existing MoE-based INR methods, providing a stronger capacity allocation under heterogeneous temporal dynamics.

Figures

Figures reproduced from arXiv: 2607.20970 by Jun Tao, Weihan Zhang, Xuan Zhao, Yenwen Peng, Yuqi Chen.

Figure 1
Figure 1. Figure 1: Overview of our framework. Time-varying volumetric data are reformulated as spatially indexed temporal sequences, where each 3D coordinate is mapped to its corresponding scalar values across all time steps. This formulation is network-agnostic and can be instantiated with different INR backbones. Such a spatial assignment explicitly reveals heterogeneous temporal patterns across locations and motivates ada… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of volume rendering results on learning-based INR methods using the argon bubble data set, and the CR is 3,140. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of volume rendering results on learning-based INR methods using the vortex data set, and the CR is 1,120. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of volume rendering results with MoE-based compression INR methods on the combustion (CHI) data set, and the CR is [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of isosurface rendering results with MoE-based compression INR methods on the Tangaroa data set. The CR is 2,314, and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of volume rendering results on the ionization (PD) dataset with traditional lossy compression methods under the same reconstruction quality, and the PSNR is 57.89 dB. (a) TTHRESH (b) SZ3 (c) ZFP (d) ours (e) GT [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of volume rendering results on the combustion (YOH) dataset with traditional lossy compression methods under the same reconstruction quality, and the PSNR is 42.25 dB. 2256 4431 9294 16444 30613 Compression Ratio 36 38 40 42 44 PSNR (dB) 2550 4393 8595 19776 36755 Compression Ratio 45.0 47.5 50.0 52.5 55.0 57.5 PSNR (dB) 2256 4431 9294 16444 30613 Compression Ratio 0.04 0.06 0.08 0.10 0.12 LPIPS… view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of PSNR (top) and LPIPS (bottom) under varying [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Impact of output sequence length on reconstruction quality [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of voxel assignments across seven experts on [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of voxel-level routing results on three orthogo [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 1
Figure 1. Figure 1: Per-time-step PSNR comparison among different methods on four datasets: Argon Bubble, combustion (CHI), ionization (PD), and vortex. [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Analysis of temporal complexity and spatial coherence of tem [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] 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

45 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Ainsworth, O

    M. Ainsworth, O. Tugluk, B. Whitney, and S. Klasky. Multilevel tech- niques for compression and reduction of scientific data-quantitative con- trol of accuracy in derived quantities. SIAM Journal on Scientific Com- puting, 41(4):A2146–A2171, 2019. 2

    2019

  2. [2]

    BallesterRipoll, P

    R. BallesterRipoll, P . Lindstrom, and R. Pajarola. TTHRESH: Tensor compression for multidimensional visual data. IEEE Transactions on Visualization and Computer Graphics , 26(9):2891–2903, 2020. doi: 10 .1109/TVCG.2019.2904063 2, 4

    arXiv 2020

  3. [3]

    Ben-Shabat, C

    Y . Ben-Shabat, C. Hewa Koneputugodage, S. Ramasinghe, and S. Gould. Neural experts: Mixture of experts for implicit neural representations. Ad- vances in Neural Information Processing Systems , 37:101641–101670,

  4. [4]

    S. Fu, N. Tamir, S. Sundaram, L. Chai, R. Zhang, T. Dekel et al. Dream- Sim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023. doi: 10.48550/arXiv.2306. 09344 4

    Pith/arXiv arXiv doi:10.48550/arxiv.2306 2023

  5. [5]

    Gadirov, Q

    H. Gadirov, Q. Wu, D. Bauer, K.-L. Ma, J. B. Roerdink, and S. Frey. HyperFLINT: Hypernetwork-based flow estimation and temporal interpo- lation for scientific ensemble visualization. Computer Graphics F orum, 44(3):e70134, 2025. doi: 10.1111/cgf.70134 2

    doi:10.1111/cgf.70134 2025

  6. [6]

    Gobbetti, J

    E. Gobbetti, J. A. Iglesias Guitián, and F. Marton. COVRA: A compression-domain output-sensitive volume rendering architecture based on a sparse representation of voxel blocks. Computer Graphics F orum, 31:1315–1324, 2012. doi: 10.1111/j.1467-8659.2012.03124.x 2

    arXiv 2012

  7. [7]

    J. Han, K. Tang, and C. Wang. MoE-INR: Implicit neural representa- tion with mixture-of-experts for time-varying volumetric data compres- sion. IEEE Transactions on Visualization and Computer Graphics , 2025. doi: 10.1109/TVCG.2025.3633893 1, 2, 3, 4, 9

    arXiv 2025

  8. [8]

    Han and C

    J. Han and C. Wang. TSRTVD: Temporal superresolution for timevarying data analysis and visualization. IEEE Transactions on Visualization and Computer Graphics , 26(1):205–215, 2019. doi: 10.1109/TVCG.2019. 2934255 2

    doi:10.1109/tvcg.2019 2019

  9. [9]

    Han and C

    J. Han and C. Wang. SSRTVD: Spatial superresolution for timevarying data analysis and visualization. IEEE Transactions on Visualization and Computer Graphics, 28(6):2445–2456, 2020. doi: 10.1109/TVCG.2020. 3032123 2

    doi:10.1109/tvcg.2020 2020

  10. [10]

    Han and C

    J. Han and C. Wang. VCNet: A generative model for volume completion. Visual Informatics, 6(2):62–73, 2022. doi: 10.1016/j.visinf.2022.04.004 2

    work page doi:10.1016/j.visinf.2022.04.004 2022

  11. [11]

    Han and C

    J. Han and C. Wang. CoordNet: Data generation and visualization genera- tion for timevarying volumes via a coordinatebased neural network. IEEE Transactions on Visualization and Computer Graphics , 29(12):4951– 4963, 2023. doi: 10.1109/TVCG.2022.3197203 1, 2, 4

    arXiv 2023

  12. [12]

    Han and F

    J. Han and F. Y ang. DCINR: A divideandconquer implicit neural repre- sentation for compressing timevarying volumetric data in hours. IEEE Transactions on Visualization and Computer Graphics , 31(10):8116– 8128, 2025. doi: 10.1109/TVCG.2025.3564255 1, 2, 3

    arXiv 2025

  13. [13]

    J. Han, H. Zheng, and C. Bi. KDINR: Timevarying volumetric data com- pression via knowledge distillationbased implicit neural representation. IEEE Transactions on Visualization and Computer Graphics , 2023. doi: 10.1109/TVCG.2023.3345373 1, 2

    arXiv 2023

  14. [14]

    J. Han, H. Zheng, D. Z. Chen, and C. Wang. STNet: An endtoend genera- tive framework for synthesizing spatiotemporal superresolution volumes. IEEE Transactions on Visualization and Computer Graphics , 28(1):270– 280, 2021. doi: 10.1109/TVCG.2021.3114815 2

    arXiv 2021

  15. [15]

    Hoang, B

    D. Hoang, B. Summa, H. Bhatia, P . Lindstrom, P . Klacansky, W. Usher et al. Efficient and flexible hierarchical data layouts for a unified encoding of scalar field precision and resolution. IEEE Transactions on Visualiza- tion and Computer Graphics , 27(2):603–613, 2020. doi: 10.1109/TVCG .2020.3030381 2

    arXiv 2020

  16. [16]

    E. J. Hu, Y . Shen, P . Wallis, Z. AllenZhu, Y . Li, S. Wang et al. LoRA: Lowrank adaptation of large language models. In Proceedings of the International Conference on Learning Representations , 2022. Preprint. doi: 10.48550/arXiv.2106.09685 1, 4

    Pith/arXiv arXiv doi:10.48550/arxiv.2106.09685 2022

  17. [17]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of the Confer- ence on Neural Information Processing Systems , pp. 1097–1105. Curran Associates, 2012. doi: 10.1145/3065386 4

    doi:10.1145/3065386 2012

  18. [18]

    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 Transactions on Big Data , 9(2):485–498, 2022. doi: 10.48550/arXiv.2111.02925 2, 4

    Pith/arXiv arXiv doi:10.48550/arxiv.2111.02925 2022

  19. [19]

    Lindstrom

    P . Lindstrom. Fixedrate compressed floatingpoint arrays. IEEE Transac- tions on Visualization and Computer Graphics , 20(12):2674–2683, 2014. doi: 10.1109/TVCG.2014.2346458 2, 4

    arXiv 2014

  20. [20]

    Y . Liu, Y . Wang, L. Deng, F. Wang, F. Liu, Y . Lu et al. A novel in situ compression method for CFD data based on generative adversarial network. Journal of Visualization , 22(1):95–108, 2019. doi: 10.1007/ s12650-018-0519-x 2

    2019

  21. [21]

    Y . Lu, K. Jiang, J. A. Levine, and M. Berger. Compressive neural representations of volumetric scalar fields. Computer Graphics F orum, 40(3):135–146, 2021. doi: 10.1111/cgf.14295 1, 2, 4

    doi:10.1111/cgf.14295 2021

  22. [22]

    J. N. P . Martel, D. B. Lindell, C. Z. Lin, E. R. Chan, M. Monteiro, and G. Wetzstein. ACORN: adaptive coordinate networks for neural scene representation. ACM Trans. Graph., 40(4), art. no. 58, 13 pp., July 2021. doi: 10.1145/3450626.3459785 2

    arXiv 2021

  23. [23]

    Müller, A

    T. Müller, A. Evans, C. Schied, and A. Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):102:1–102:15, 2022. doi: 10.1145/3528223.3530127 2, 1

    arXiv 2022

  24. [24]

    Nowlan and G

    S. Nowlan and G. E. Hinton. Evaluation of adaptive mixtures of compet- ing experts. Advances in neural information processing systems , 3, 1990. 1

    1990

  25. [25]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pp. 8748–8763, 2021. doi: 10.48550/arXiv.2103.00020 4

    Pith/arXiv arXiv doi:10.48550/arxiv.2103.00020 2021

  26. [26]

    Reiser, S

    C. Reiser, S. Peng, Y . Liao, and A. Geiger. KiloNeRF: Speeding up neural radiance fields with thousands of tiny MLPs. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pp. 14335– 14345, October 2021. doi: 10.48550/arXiv.2103.13744 2

    Pith/arXiv arXiv doi:10.48550/arxiv.2103.13744 2021

  27. [27]

    Sahoo, Y

    S. Sahoo, Y . Lu, and M. Berger. Neural flow map reconstruction. Com- puter Graphics F orum, 41(3):391–402, 2022. doi: 10.1111/cgf.14549 2

    work page doi:10.1111/cgf.14549 2022

  28. [28]

    MINER: Multiscale Implicit Neural Representations

    V . Saragadam, J. Tan, G. Balakrishnan, R. G. Baraniuk, and A. V eer- araghavan. MINER: Multiscale implicit neural representation. In Euro- pean Conference on Computer Vision , pp. 318–333. Springer, 2022. doi: 10.48550/arXiv.2202.03532 2

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2202.03532 2022

  29. [29]

    Sitzmann, E

    V . Sitzmann, E. Chan, R. Tucker, N. Snavely, and G. Wetzstein. MetaSDF: Meta-learning signed distance functions. In Proceedings of the Conference on Neural Information Processing Systems , pp. 10136– 10147. Curran Associates, Red Hook, NY , 2020. doi: 10.5555/3495724. 3496574 2

    doi:10.5555/3495724 2020

  30. [30]

    Sitzmann, J

    V . Sitzmann, J. N. P . Martel, A. W. Bergman, D. B. Lindell, and G. Wet- zstein. Implicit neural representations with periodic activation func- tions. Advances in Neural Information Processing Systems , 33:7462– 7473, 2020. doi: 10.48550/arXiv.2006.09661 1, 2, 3, 4

    Pith/arXiv arXiv doi:10.48550/arxiv.2006.09661 2020

  31. [31]

    S. K. Suter, M. Makhynia, and R. Pajarola. TAMRESH tensor ap- proximation multiresolution hierarchy for interactive volume visualiza- tion. Computer Graphics F orum, 32(3):151–160, 2013. doi: 10.1111/cgf .12102 2

    doi:10.1111/cgf 2013

  32. [32]

    Tang and C

    K. Tang and C. Wang. ECNR: Efficient compressive neural represen- tation of timevarying volumetric datasets. In Proceedings of the IEEE Pacific Visualization Conference (PacificVis), pp. 72–81, 2024. doi: 10. 1109/PacificVis52379.2024.00014 1, 2

    arXiv 2024

  33. [33]

    Weiss, M

    S. Weiss, M. Chu, N. Thuerey, and R. Westermann. V olumetric isosurface rendering with deep learningbased superresolution. IEEE Transactions on Visualization and Computer Graphics , 27(6):3064–3078, 2019. doi: 10.1109/TVCG.2019.2956697 2

    arXiv 2019

  34. [34]

    Weiss, M

    S. Weiss, M. I¸ sIk, J. Thies, and R. Westermann. Learning adaptive sam- pling and reconstruction for volume visualization. IEEE Transactions on Visualization and Computer Graphics , 28(7):2654–2667, 2022. doi: 10. 1109/TVCG.2020.3039340 2

    arXiv 2022

  35. [35]

    Q. Wu, D. Bauer, Y . Chen, and K. Ma. HyperINR: A fast and predictive hypernetwork for implicit neural representations via knowledge distilla- tion. arXiv preprint arXiv:2304.04188, 2023. doi: 10.48550/arXiv.2304. 04188 2

    Pith/arXiv arXiv doi:10.48550/arxiv.2304 2023

  36. [36]

    Q. Wu, J. A. Insley, V . A. Mateevitsi, S. Rizzi, M. E. Papka, and K.-L. Ma. Distributed neural representation for reactive in situ visualization. IEEE Transactions on Visualization and Computer Graphics, 31(9):5199–5214,

  37. [37]

    S. W. Wurster, H. Guo, H.-W. Shen, T. Peterka, and J. Xu. Deep hier- archical super resolution for scientific data. IEEE Transactions on Vi- sualization and Computer Graphics , 29(12):5483–5495, 2022. doi: 10. 1109/TVCG.2022.3214420 2

    arXiv 2022

  38. [38]

    L. Y an, X. Liang, H. Guo, and B. Wang. TopoSZ: Preserving topology in error-bounded lossy compression. IEEE Transactions on Visualization and Computer Graphics , 30(1):1302–1312, 2023. doi: 10.1109/TVCG. 2023.3326920 2

    arXiv 2023

  39. [39]

    Y ang, K

    M. Y ang, K. Tang, and C. Wang. Meta-INR: Efficient encoding of volumetric data via meta-learning implicit neural representation. In IEEE Pacific Visualization Conference, pp. 246–251, 2025. doi: 10.1109/ PacificVis64226.2025.00030 2

    arXiv 2025

  40. [40]

    Zhang, P

    R. Zhang, P . Isola, A. A. Efros, E. Shechtman, and O. Wang. The un- reasonable effectiveness of deep features as a perceptual metric. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 586–595, 2018. doi: 10.1109/CVPR.2018.00068 4

    arXiv 2018

  41. [41]

    Zhang, Z

    Y . Zhang, Z. Lin, X. Y ao, J. Hu, F. Meng, C. Liu et al. Kimi lin- ear: An expressive, efficient attention architecture. arXiv preprint arXiv:2510.26692, 2025. doi: 10.48550/arXiv.2510.26692 1, 4

    Pith/arXiv arXiv doi:10.48550/arxiv.2510.26692 2025

  42. [42]

    Zhao, C.-C

    J. Zhao, C.-C. Tseng, M. Lu, R. An, X. Wei, H. Sun et al. MoEC: Mixture of experts implicit neural compression, 2023. doi: 10.48550/arXiv.2312. 01361 1, 2, 3

    doi:10.48550/arxiv.2312 2023

  43. [43]

    Zhenxing and D

    M. Zhenxing and D. Xu. Switch-NeRF: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In The In- ternational Conference on Learning Representations , 2022. 2, 4

    2022

  44. [44]

    Z. Zhou, Y . Hou, Q. Wang, G. Chen, J. Lu, Y . Tao et al. V olume upscaling with convolutional neural networks. In Proceedings of the Computer Graphics International Conference, pp. 38:1–38:6. ACM, New Y ork, 2017.doi: 10.1145/3095140.3095178 2 A N ETWORK CONFIGURATION In this section, we provide the network architecture and training of our framework. Net...

    arXiv 2017

  45. [2025]

    doi: 10.1109/TVCG.2024.3432710 2

    arXiv 2024

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