Pith. sign in

REVIEW 4 major objections 5 minor 67 references

ECoNGS: Efficient Compressive Neural Gaussian Splats for Volume Visualization

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read ECoNGS sets out to show that volume-visualization scenes can be represented as editable Gaussians predicted on the fly from compressed anchor points, and that sharing the predictors across similar scenes yields a representation that is smal

desk verdict A credible system paper that combines known components (neural anchors, entropy coding, joint learning) into a VolVis representation with consistent reported gains over iVR-GS, but the lack of error bars and reliance on the authors' own baseline make the headline margins uncertain. read the letter →

arxiv 2607.18466 v1 pith:5MCWIDCZ submitted 2026-07-20 cs.CV cs.GR

Kaiyuan Tang , Chaoli Wang This is my paper
classification cs.CVcs.GR
keywords volumevisualization3DGaussiansplattingneuralrepresentationjointlearningentropycodingeditableGaussianstransferfunctionsmodelcompression
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that editable Gaussian splatting for volume visualization can be made dramatically more compact and faster to train without sacrificing reconstruction quality. Its strategy is to stop storing every Gaussian explicitly: lightweight MLPs decode Gaussian attributes on demand from a sparse set of anchor points, geometrically similar scenes share those MLPs, and a neural entropy model compresses the remaining anchor attributes. If the approach works as reported, large volume datasets could be interactively explored and edited from models of only a few megabytes, with training times cut by several times and quality improved at the same time.

What carries the argument

The central object is the neural editable Gaussian representation: each explicit anchor point stores a 50-dimensional feature, scaling factors, and offsets for K=10 surrounding Gaussians, and four lightweight two-layer MLPs decode opacity, covariance, shading attributes, and normals on the fly from the anchor feature, view direction, and a transfer-function embedding. Joint learning shares these MLPs across scenes clustered by Chamfer distance between sparse point clouds extracted from the volume. A neural entropy model, built from binary hash grids and a shallow MLP, estimates probability distributions for the anchor attributes so they can be arithmetic-coded into a compact bitstream. Toget

What would settle it

Train ECoNGS jointly on two scenes that are close under Chamfer distance of their sparse point clouds but differ strongly in fine-scale structure or shading; if either scene's PSNR drops by more than about 1 dB relative to separate training, then geometric point-cloud similarity is not sufficient for safe parameter sharing.

Watch

Extended reading notes

Core claim

The central claim is that the explicit Gaussian primitives used in volume-visualization scene representations are redundant, and that this redundancy can be removed in three complementary ways: by predicting editable Gaussians from compact anchor points with lightweight MLPs, by sharing those MLPs across scenes that have similar geometry, and by entropy-coding the anchor attributes with a learned context model. The paper reports that, compared with the prior editable-Gaussian method iVR-GS, this design improves PSNR by up to 2.2 dB, reduces model size by up to 6.1x, and reduces training time by up to 5.9x, while keeping real-time rendering and inference-time editing of color, opacity, and li

Load-bearing premise

Joint learning assumes that scenes whose subsampled point clouds have small Chamfer distance are similar enough that one set of shared MLPs can reconstruct them all; the cross-dataset experiment shows that when this similarity assumption fails, reconstruction quality collapses.

Editorial extensions

If this is right

  • A composed volume-visualization scene can be stored in a few megabytes and rendered at hundreds of frames per second while still supporting interactive edits to color, opacity, and lighting.
  • Jointly training many similar transfer-function-specific scenes cuts both training time and model size, and the savings grow with the number of scenes: on a 41-scene dataset, training time drops by about half and anchor count by roughly 60 percent versus separate training.
  • Neural entropy coding shrinks the explicit anchor attributes by over 10x compared with no context model, and the rate-distortion results show a tunable quality-versus-size tradeoff.
  • Initializing anchors from a sparse point cloud sampled directly from the volume takes under a second and improves both convergence speed and final reconstruction accuracy relative to random initialization.
  • The representation also extends to isosurface scenes and unstructured-mesh data, producing compact high-fidelity models without pipeline changes.

Reading between the lines

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

  • Because joint learning's benefit grows with the number of basic scenes, the approach becomes more attractive for large ensemble or parameter-study visualizations where many transfer-function variants of the same volume are needed.
  • The clustering criterion is purely geometric; an editorial extension is to include appearance, shading, or transfer-function similarity in the grouping metric, which could prevent the kind of quality collapse seen when unrelated datasets are forced to share parameters.
  • The context model conditions only on anchor position; conditioning on anchor features or nearby anchors could capture spatial correlation and further reduce bitrate.
  • If view-dependent opacity prediction genuinely enables pre-rasterization culling, as the rendering breakdown suggests, then rendering cost may scale with visible structure rather than total primitive count, a promising property for very large volumes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. ECoNGS is a hybrid neural-explicit Gaussian splatting representation for volume visualization. It keeps explicit anchor points per basic TF-specific scene, decodes editable Gaussian attributes (position offsets, opacity, covariance, shading, normal) via shared lightweight MLPs, trains groups of geometrically similar scenes jointly via Chamfer-distance clustering, and compresses anchor attributes with a neural entropy model and arithmetic coding. A sparse point-cloud initialization extracted from the volume is proposed. The paper reports evaluations on four univariate and three multivariate datasets against Plenoxels, 3DGS, Scaffold-GS, CCNeRF, HAC, and iVR-GS, with PSNR/LPIPS, FPS, training time, and model size. The headline comparison to iVR-GS is up to +2.2 dB PSNR, 6.1x smaller models, and 5.9x faster training. Ablations and appendices study initialization, component removal, rate-distortion, clustering scalability, limited views, isosurface/fuzzy/unstructured scenes, and implicit-vs-explicit storage.

Significance. If the reported gains reproduce, the paper makes a useful contribution: it shows that a Scaffold-GS-style hybrid anchor/MLP representation can be adapted to editable VolVis Gaussians, that clustering based on sparse point clouds enables multi-scene parameter sharing with meaningful training-time and model-size savings, that a BiRF-style context model compresses the explicit part substantially, and that volume-derived point clouds improve convergence over random initialization. Strengths include the unusually large evaluation (7 datasets, 92 train/181 test views per scene, multiple baselines, rate-distortion, extensive appendices), the explicit ablation of each component, and the public code release. The main caveats are reproducibility (single runs), an undocumented clustering threshold for the main tables, and an overstated claim that joint learning improves reconstruction accuracy when the evidence shows a slight quality/size trade-off.

major comments (4)
  1. [Table 3; Section 4.1] The central improvement over iVR-GS is based on a single training run per method/dataset. The margins over iVR-GS in Table 3 are 0.91 dB (ionization T), 1.03 dB (supernova), 1.35 dB (combustion MF), and 2.20 dB (vortex), with LPIPS differences as small as 0.021 vs 0.023. These margins may be within run-to-run variance, particularly because the baseline is the authors' own iVR-GS and no seeds or confidence intervals are reported. Please add at least three repeated runs for ECoNGS and iVR-GS (or a justified single-run protocol) on the datasets where the margin is below ~1.5 dB, and report mean±std for PSNR, LPIPS, training time, and model size.
  2. [Section 4.1; Appendix A Tables 1-2] The Chamfer-distance clustering threshold used to produce Tables 3 and 4 is not stated in the main text. Appendix A Table 1 shows that on multivariate combustion, the threshold changes PSNR from 34.60 to 34.02 and model size from 45.75 to 14.72 MB, and Table 2 shows catastrophic failure when unrelated scenes are grouped (argon bubble PSNR 31.76→16.82). Since joint learning is a core contribution and the threshold is a free parameter, the main-experiment thresholds must be stated for every dataset, together with whether the threshold was selected on a validation set. Without this, the reported model-size/training-time gains are not reproducible.
  3. [Section 3.4; Table 7; Appendix A Table 3] Section 3.4 claims joint learning 'can achieve faster convergence and more accurate reconstruction results.' The evidence in the paper contradicts the accuracy part. Table 7 shows full ECoNGS on multivariate combustion has PSNR 34.13 vs 34.60 without joint learning, and Appendix A Table 3 shows joint learning lowers PSNR on all four datasets tested (e.g., ionization T 34.34 vs 34.53; combustion MF 34.70 vs 35.04; 24-scene ionization 35.60 vs 35.95; 41-scene combustion 34.13 vs 34.60). The demonstrated benefits of joint learning are reductions in training time, model size, and anchor count, not reconstruction fidelity. Please revise the claim and present the quality/time/size trade-off explicitly.
  4. [Section 4.1] The hyperparameters lambda_e, K, anchor feature dimension, and the clustering threshold are set 'in our experiments' without a documented validation split or selection protocol. The paper uses 181 held-out test views, but it is not stated whether any hyperparameter was chosen by inspecting those views. To support the claim of simultaneous improvement, please describe how hyperparameters were selected, or add a validation-based selection and report sensitivity on the main datasets.
minor comments (5)
  1. [Section 3.6, Eq. (11)] 'Guassians' should be 'Gaussians'.
  2. [Section 4.5] 'In addition, Moreover, increasing the number of points...' contains a redundant transition; please rephrase.
  3. [Table 7 vs Appendix A Table 1] The full ECoNGS row in Table 7 (TT 72.0 min, MS 15.99 MB) appears to describe the same configuration as the threshold-0.01 row in Appendix A Table 1 (TT 70.5 min, MS 15.72 MB); please clarify the discrepancy.
  4. [References] Reference [37] has 'Pattern Recongnition' in the venue name; please fix the typo.
  5. [Throughout] The 'V olVis' spacing artifacts from the LaTeX macro should be cleaned up in the camera-ready version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: main claims rest on held-out-view benchmarks and external baselines; self-citations are contextual, not load-bearing.

full rationale

ECoNGS is an empirical systems paper whose central claims are supported by held-out-view reconstruction benchmarks (181 unseen testing views) against multiple external baselines (3DGS, Scaffold-GS, Plenoxels, CCNeRF, HAC, VEG) in addition to the authors' own iVR-GS. The hybrid anchor/MLP representation is trained end-to-end with a reconstruction loss on training views; no fitted parameter is reused as a predicted output. Joint learning is evaluated by explicit comparison to separate per-scene optimization (Appendix A, Table 3) and by a cross-dataset failure case (Appendix A, Table 2) that exposes the similarity assumption rather than hiding it. The neural entropy model minimizes a real bitrate objective (Eq. 9) and reported model sizes correspond to actual encoded bitstreams, not assumed values. Self-citations to iVR-GS, TexGS, and NLI4VolVis are used as baselines and building blocks, but the core derivation does not reduce to any self-citation chain or imported uniqueness theorem. The paper contains no equation where a claimed prediction is equivalent by construction to its input. Concerns about missing error bars or undocumented clustering thresholds are reproducibility/correctness issues, not circularity. Therefore no significant circularity is present.

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

The central claim depends on the 3DGS rasterizer being an accurate forward model, Blinn-Phong shading matching the appearance of the ParaView/IndeX rendering, Gaussian-distributed anchor attributes for the entropy model, and the Chamfer-distance scene similarity for joint learning. No new physical entities are introduced. Several hyperparameters are hand-selected on the evaluation datasets.

free parameters (6)
  • entropy regularization weight lambda_e = 0.0001
    Balances reconstruction fidelity vs. bitrate in Eq. (12); authors fixed it at 0.0001 for main results and swept it 0.0001-0.01 for rate-distortion curves (Sec. 4.6).
  • per-attribute quantization step Q = f_a:1, l_s:0.001, l_r:0.001, delta_x_g:0.2
    Set by hand according to numerical scales of anchor attributes (Sec. 3.5); directly sets compression distortion.
  • number of neural editable Gaussians per anchor K = 10
    Chosen as a quality/size trade-off in Appendix I; 20 improves PSNR slightly but increases model size and slows rendering.
  • anchor feature dimension = 50
    Hyperparameter from Appendix I; 32 degrades PSNR, 100 gives equal PSNR at higher storage.
  • scaling regularization weight lambda_s = 0.001
    Regularizer in Eq. (11) to shrink predicted Gaussians; no sensitivity study reported.
  • clustering threshold (Chamfer distance) = not explicitly stated for main runs; swept 0-0.04 in Appendix A
    Controls scene grouping; threshold 0.01 gives 10 groups for combustion (MF), affecting training memory/time and final model size.
assumptions (6)
  • domain assumption 3DGS tile-based differentiable rasterization accurately renders the decoded neural Gaussians (Eqs. 1-2 used in Sec. 3.1 and 3.3).
    All reconstruction quality numbers depend on the fidelity of the rasterizer and its gradients during training.
  • domain assumption Blinn-Phong shading with a white specular light matches the appearance of the volume renderings produced by ParaView's NVIDIA IndeX plugin with depth enhancement.
    Eq. (5a-c) in Sec. 3.3; the training images are generated with this renderer, so editability and PSNR values rest on this equivalence.
  • domain assumption Anchor attribute values are approximately Gaussian-distributed, so the context model can model them with mean/variance outputs.
    Figure 3 in Sec. 3.5 shows histograms; entropy coding's bit estimates in Eq. (9) assume this parametric form.
  • domain assumption Chamfer distance on subsampled point clouds is a valid similarity metric for grouping scenes for joint learning.
    Sec. 3.4; Appendix A Table 2 shows joint learning across unrelated datasets degrades quality, so the metric is load-bearing.
  • domain assumption A sparse point cloud sampled from opaque TF voxels is a good geometric prior for initializing anchors.
    Sec. 3.2; ablation in Table 5 shows ~3.4 dB PSNR improvement over random init, so the initialization is a major contributor to the claimed gains.
  • standard math Uniform-noise quantization (Eq. 6) is a trainable approximation of rounding (Eq. 7), as standard in neural compression.
    Sec. 3.5, following Balle et al. [2]; not proved for this specific loss, but standard in the field.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ECoNGS: Efficient Compressive Neural Gaussian Splats for Volume Visualization." pith.science (2026). https://pith.science/paper/5MCWIDCZ

@misc{pith2026260718466,
  author       = {Pith},
  title        = {Pith review of: ECoNGS: Efficient Compressive Neural Gaussian Splats for Volume Visualization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5MCWIDCZ}},
  note         = {Machine review of arXiv:2607.18466}
}
read the original abstract

Recent advances in differentiable Gaussian splatting have highlighted the potential of primitive-based approaches as alternative scene representations for interactive, high-quality, volume visualization (VolVis) of large datasets. However, the explicit nature of current primitive-based methods, combined with isolated optimization for each VolVis scene, results in redundant, non-compact representations. We present ECoNGS, an efficient compressive neural Gaussian splatting framework for VolVis scene representation. ECoNGS employs lightweight neural networks to dynamically predict implicit, editable Gaussian splats from explicit anchor points, effectively combining model compactness and parameter efficiency of implicit representations with high-performance rendering of explicit primitives. We explore a joint learning strategy that clusters geometrically similar scenes and shares parameters across them, significantly reducing overall training time and model size while maintaining reconstruction fidelity. To achieve a more compact scene representation, we further compress the explicit anchor attributes using a neural entropy model that estimates their probability distributions, enabling compact storage via entropy coding. We systematically investigate Gaussian initialization strategies and propose a simple yet effective scheme tailored for VolVis scenes, improving reconstruction accuracy and accelerating convergence. We evaluate ECoNGS qualitatively and quantitatively across various univariate and multivariate VolVis scenes, highlighting its superior performance over prior methods in training time, reconstruction quality, and model size. In particular, compared with the prior method iVR-GS, ECoNGS improves reconstruction quality by up to 2.2 dB in PSNR while reducing the model size by up to 6.1x and the training time by up to 5.9x. The code is available at https://github.com/TouKaienn/ECoNGS.

Figures

Figures reproduced from arXiv: 2607.18466 by the authors.

Figure 1
Figure 1. Overview of the ECoNGS framework. (a) Rendering large-scale volumetric data through DVR is computationally expensive. (b) The dataset [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Statistical analysis for the distribution of the anchor attribute [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Visualization of inter-scene anchor feature similarity. (a) Three [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Comparing scene composing results of four methods w.r.t. GT. While Plenoxels, Scaffold-GS, and 3DGS focus on reconstruction, only [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparing scene composing results of four methods w.r.t. GT. CCNeRF and HAC bake TF information and lighting into their representations, [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Examples of iterative scene editing results with composed ECoNGS models on the vortex and ionization (T) datasets. The light source [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparing reconstruction results of ECoNGS and iVR-GS [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Rate-distortion curves of different GS-based methods on the [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 8
Figure 8. Figure 8: Comparing different initialization strategies on the beetle [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 1
Figure 1. Figure 1: Comparing ECoNGS with triangular meshes on the isosurface extracted from the combustion (MF) volume. (a)–(c) the three basic scenes extracted with different isovalues, and (d) is the composed scene. ECoNGS preserves high visual quality but is more compact than the orig…
Figure 2
Figure 2. Figure 2: Novel view synthesis results of ECoNGS on fuzzy VolVis scene of the vortex dataset. The difference image in the bottom-right corner shows the pixel-wise perceptible difference (blue to red indicates low to high) in the CIELUV color space. the large-scale volume data. O…
Figure 3
Figure 3. Figure 3: Comparing ECoNGS and iVR-GS under varying training views [PITH_FULL_IMAGE:figures/full_fig_p011_3.png]
Figure 4
Figure 4. Figure 4: Novel view synthesis results of ECoNGS and VEG on the ionization (T) dataset [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 11 canonical work pages

  1. [2]

    Ballé, D

    J. Ballé, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston. Vari- ational image compression with a scale hyperprior.arXiv preprint arXiv:1802.01436, 2018. doi:10.48550/arXiv.1802.014364

    arXiv doi:10.48550/arxiv.1802.01436 2018

  2. [3]

    Bauer, Q

    D. Bauer, Q. Wu, H. Gadirov, and K.-L. Ma. GSCache: Real-time radiance caching for volume path tracing using 3D Gaussian splatting.IEEE Transactions on Visualization and Computer Graphics, 32(1):901–911,

  3. [4]

    Berger, J

    M. Berger, J. Li, and J. A. Levine. A generative model for volume rendering.IEEE Transactions on Visualization and Computer Graphics, 25(4):1636–1650, 2019. doi:10.1109/TVCG.2018.28160592

    arXiv 2019

  4. [5]

    C. Bi, X. Gao, J. Deng, G. Li, and J. Han. CD-TVD: Contrastive diffusion for 3D super-resolution with scarce high-resolution time-varying data. IEEE Transactions on Visualization and Computer Graphics, 32(1):506– 516, 2026. doi:10.1109/TVCG.2025.36347872

    arXiv 2026

  5. [6]

    J. F. Blinn. Models of light reflection for computer synthesized pictures. InProceedings of ACM SIGGRAPH Conference, pp. 192–198, 1977. doi: 10.1145/965141.5638933

    arXiv 1977

  6. [8]

    Y . Chen, Q. Wu, W. Lin, M. Harandi, and J. Cai. HAC: Hash-grid as- sisted context for 3D Gaussian splatting compression. InProceedings of European Conference on Computer Vision, pp. 422–438, 2024. doi: 10. 1007/978-3-031-72667-5_242, 4, 5

    2024

  7. [9]

    Y .-T. Chen, H. Li, N. Shi, X. Luo, W. Xu, and H.-W. Shen. Explorable INR: An implicit neural representation for ensemble simulation enabling efficient spatial and parameter exploration.IEEE Transactions on Vi- sualization and Computer Graphics, 31(6):3758–3770, 2025. doi: 10. 1109/TVCG.2025.35670522

    arXiv 2025

  8. [10]

    Cheng, H

    Z. Cheng, H. Sun, M. Takeuchi, and J. Katto. Learned image compression with discretized Gaussian mixture likelihoods and attention modules. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7936–7945, 2020. doi: 10.1109/CVPR42600.2020.00796 2, 4

    arXiv 2020

Show all 67 references
  1. [11]

    Y . J. Cho and H. Y . Yoon. Numerical analysis of the ROCOM boron dilu- tion benchmark experiment using the CUPID code.Nuclear Engineering and Design, 341:167–175, 2019. doi:10.1016/j.nucengdes.2018.11.00512

    2019 doi

  2. [12]

    Dyken, A

    L. Dyken, A. Sewell, W. Usher, N. Debardeleben, S. Petruzza, and S. Ku- mar. V olume Encoding Gaussians: Transfer function-agnostic 3D Gaus- sians for volume rendering.IEEE Transactions on Visualization and Com- puter Graphics, 32(6):4883–4893, 2026. doi: 10.1109/TVCG.2026.369...

    2026

  3. [13]

    Engel and T

    D. Engel and T. Ropinski. Deep volumetric ambient occlusion.IEEE Transactions on Visualization and Computer Graphics, 27(2):1268–1278,

  4. [15]

    J. Gao, C. Gu, Y . Lin, Z. Li, H. Zhu, X. Cao et al. Relightable 3D Gaussians: Realistic point cloud relighting with BRDF decomposition and ray tracing. InProceedings of European Conference on Computer Vision, pp. 73–89, 2024. doi:10.1007/978-3-031-72995-9_52

    2024 doi

  5. [16]

    J. Han, K. Tang, and C. Wang. MoE-INR: Implicit neural representation with mixture-of-experts for time-varying volumetric data compression. IEEE Transactions on Visualization and Computer Graphics, 32(1):254– 264, 2026. doi:10.1109/TVCG.2025.36338932

    2026

  6. [17]

    Han and C

    J. Han and C. Wang. TSR-TVD: Temporal super-resolution for time- varying data analysis and visualization.IEEE Transactions on Visual- ization and Computer Graphics, 26(1):205–215, 2020. doi: 10.1109/TVCG. 2019.29342552

    2020

  7. [18]

    Han and C

    J. Han and C. Wang. SSR-TVD: Spatial super-resolution for time-varying data analysis and visualization.IEEE Transactions on Visualization and Computer Graphics, 28(6):2445–2456, 2022. doi: 10.1109/TVCG.2020. 30321232

    2022 doi

  8. [19]

    Han and C

    J. Han and C. Wang. CoordNet: Data generation and visualization gen- eration for time-varying volumes via a coordinate-based neural network. IEEE Transactions on Visualization and Computer Graphics, 29(12):4951– 4963, 2023. doi:10.1109/TVCG.2022.31972032

    2023

  9. [20]

    J. Han, H. Zheng, D. Z. Chen, and C. Wang. STNet: An end-to-end generative framework for synthesizing spatiotemporal super-resolution volumes.IEEE Transactions on Visualization and Computer Graphics, 28(1):270–280, 2022. doi:10.1109/TVCG.2021.31148152

    2022

  10. [21]

    D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y . Wang. ELIC: Efficient learned image compression with unevenly grouped space-channel con- textual adaptive coding. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5708–5717, 2022. doi: 10. 1109...

    2022

  11. [22]

    D. He, Y . Zheng, B. Sun, Y . Wang, and H. Qin. Checkerboard con- text model for efficient learned image compression. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14771–14780, 2021. doi:10.1109/CVPR46437.2021.014532

    2021

  12. [23]

    W. He, J. Wang, H. Guo, K.-C. Wang, H.-W. Shen, M. Raj et al. InSi- tuNet: Deep image synthesis for parameter space exploration of ensemble simulations.IEEE Transactions on Visualization and Computer Graphics, 26(1):23–33, 2020. doi:10.1109/TVCG.2019.29343122

    2020

  13. [24]

    S. Jeon, K. Tang, C. Wang, and W.-K. Jeong. Super-Gaussian: Interactive scene editing for 3D Gaussian splatting and NLI-based volume visualiza- tion in virtual reality.IEEE Transactions on Visualization and Computer Graphics, 33(1), 2027. Accepted. 9

    2027

  14. [26]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis. 3D Gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4):139:1–139:14, 2023. doi:10.1145/35924331, 2, 3, 5, 12

    2023 doi

  15. [28]

    G. Li, Y . Liu, G. Shan, S. Cheng, W. Cao, J. Wang et al. ParamsDrag: Interactive parameter space exploration via image-space dragging.IEEE Transactions on Visualization and Computer Graphics, 31(1):624–634,

  16. [29]

    J. Li, B. Li, and Y . Lu. Deep contextual video compression. InProceedings of Advances in Neural Information Processing Systems, pp. 18114–18125,

  17. [30]

    L. Liu, Z. Chen, and D. Xu. HEMGS: A hybrid entropy model for 3D Gaussian splatting data compression.arXiv preprint arXiv:2411.18473,

    arXiv

  18. [31]

    S. Liu, J. Ding, J. Yang, and W. Liu. Mixed Gaussian splatting for high- quality rendering and reconstruction. InProceedings of IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 1–5, 2025. doi:10.1109/ICASSP49660.2025.108904022, 3

    2025

  19. [33]

    Y . Lu, K. Jiang, J. A. Levine, and M. Berger. Compressive neural represen- tations of volumetric scalar fields.Computer Graphics Forum, 40(3):135– 146, 2021. doi:10.1111/cgf.142952

    2021 doi

  20. [34]

    Mihajlovic, S

    M. Mihajlovic, S. Prokudin, S. Tang, R. Maier, F. Bogo, T. Tung et al. SplatFields: Neural Gaussian splats for sparse 3D and 4D reconstruction. InProceedings of European Conference on Computer Vision, pp. 313–332,

  21. [35]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. InProceedings of European Conference on Computer Vision, pp. 405–421, 2020. doi:10.1007/978-3-030-58452-8_241, 3

    2020 doi

  22. [36]

    Niedermayr, C

    S. Niedermayr, C. Neuhauser, K. Petkov, K. Engel, and R. Westermann. Application of 3D Gaussian splatting for cinematic anatomy on consumer class devices. InProceedings of International Symposium on Vision, Modeling, and Visualization, pp. 1–13, 2024. doi: 10.2312/VMV.20241195 1, 2

    2024 doi

  23. [37]

    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. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recongnition, pp. 12162–12172, 2026. doi:10.48550/arXiv.2603.077892

    2026 doi

  24. [38]

    Sainz and R

    M. Sainz and R. Pajarola. Point-based rendering techniques.Computers & Graphics, 28(6):869–879, 2004. doi:10.1016/j.cag.2004.08.0142

    2004 doi

  25. [39]

    doi:10.1007/978-3-031-72627-9_182, 3

    doi

  26. [40]

    Sheng, J

    X. Sheng, J. Li, B. Li, L. Li, D. Liu, and Y . Lu. Temporal context mining for learned video compression.IEEE Transactions on Multimedia, 25:7311–7322, 2023. doi:10.1109/TMM.2022.32204212

    2023

  27. [41]

    N. Shi, J. Xu, H. Li, H. Guo, J. Woodring, and H.-W. Shen. VDL-Surrogate: A view-dependent latent-based model for parameter space exploration of ensemble simulations.IEEE Transactions on Visualization and Computer Graphics, 29(1):820–830, 2023. doi:10.1109/TVCG.2022.32094132

    2023

  28. [42]

    Shin and J

    S. Shin and J. Park. Binary radiance fields. InProceedings of Advances in Neural Information Processing Systems, pp. 1–26, 2023. doi: 10.48550/ arXiv.2306.075814

    2023 doi arXiv

  29. [43]

    J. Sun, D. Lenz, H. Yu, and T. Peterka. F-Hash: Feature-based hash de- sign for time-varying volume visualization via multi-resolution Tesseract encoding.IEEE Transactions on Visualization and Computer Graphics, 32(1):396–406, 2026. doi:10.1109/TVCG.2025.36348122

    2026

  30. [44]

    J. L. Schönberger and J.-M. Frahm. Structure-from-motion revisited. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4104–4113, 2016. doi:10.1109/CVPR.2016.4451, 3

    2016

  31. [45]

    K. Tang, K. Ai, , J. Han, and C. Wang. TexGS-V olVis: Expressive scene editing for volume visualization via textured Gaussian splatting.IEEE Transactions on Visualization and Computer Graphics, 32(1):933–943,

  32. [46]

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

    2026

  33. [47]

    Tang and C

    K. Tang and C. Wang. ECNR: Efficient compressive neural representation of time-varying volumetric datasets. InProceedings of IEEE PacificVis Visualization Conference, pp. 72–81, 2024. doi: 10.1109/PACIFICVIS60374. 2024.000172

    2024

  34. [48]

    Tang and C

    K. Tang and C. Wang. STSR-INR: Spatiotemporal super-resolution for time-varying multivariate volumetric data via implicit neural representa- tion.Computers & Graphics, 119:103874, 2024. doi: 10.1016/j.cag.2024. 01.0012

    2024 doi

  35. [49]

    J. Tang, X. Chen, J. Wang, and G. Zeng. Compressible-composable NeRF via rank-residual decomposition. InProceedings of Advances in Neural Information Processing Systems, pp. 14798–14809, 2022. 5

    2022

  36. [50]

    K. Tang, M. Yang, and C. Wang. EVOLVE: Efficient learned volume compression with variable-rate encoding on a cross-domain database. IEEE Transactions on Visualization and Computer Graphics, 33(1), 2027. Accepted. 2

    2027

  37. [51]

    doi:10.1109/TVCG.2025.36346431, 2, 3

    2025

  38. [52]

    Wang and J

    C. Wang and J. Han. DL4SciVis: A state-of-the-art survey on deep learning for scientific visualization.IEEE Transactions on Visualization and Computer Graphics, 29(8):3714–3733, 2023. doi: 10.1109/TVCG.2022. 31678961

    2023 doi

  39. [53]

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotný. VGGT: Visual geometry grounded transformer. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5294–5306, 2025. doi:10.1109/CVPR52734.2025.004992

    2025

  40. [54]

    Y . Wang, Z. Li, L. Guo, W. Yang, A. C. Kot, and B. Wen. ContextGS: Compact 3D Gaussian splatting with anchor level context model. In Proceedings of Advances in Neural Information Processing Systems, 2024. doi:10.48550/arXiv.2450.207212

    2024 doi

  41. [55]

    Tang and C

    K. Tang and C. Wang. StyleRF-V olVis: Style transfer of neural radiance fields for expressive volume visualization.IEEE Transactions on Visual- ization and Computer Graphics, 31(1):613–623, 2025. doi: 10.1109/TVCG. 2024.34563421

    2025

  42. [56]

    Q. Wu, J. M. Esturo, A. Mirzaei, N. Moënne-Loccoz, and Z. Gojcic. 3DGUT: Enabling distorted cameras and secondary rays in Gaussian splatting. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 26036–26046, 2025. doi: 10.1109/CVPR52734. 2025.024252

    2025

  43. [57]

    K. Tang, S. Yao, and C. Wang. iVR-GS: Inverse volume rendering for explorable visualization via editable 3D Gaussian splatting.IEEE Trans- actions on Visualization and Computer Graphics, 31(6):3783–3795, 2025. doi:10.1109/TVCG.2025.35671211, 2, 3, 5

    2025

  44. [59]

    Z. Yang, X. Gao, W. Zhou, S. Jiao, Y . Zhang, and X. Jin. Deformable 3D Gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20331–20341, 2024. doi: 10.1109/CVPR52733.2024.01922 9

    2024

  45. [60]

    Z. Yang, H. Yang, Z. Pan, and L. Zhang. Real-time photorealistic dynamic scene representation and rendering with 4D Gaussian splatting. InPro- ceedings of International Conference on Learning Representations, pp. 1–18, 2024. doi:10.48550/arXiv.2310.106429

    2024 doi arXiv

  46. [61]

    I. H. Witten, R. M. Neal, and J. G. Cleary. Arithmetic coding for data compression.Communications of the ACM, 30(6):520–540, 1987. doi: 10. 1145/214762.2147714

    1987

  47. [62]

    Yao and C

    S. Yao and C. Wang. ReV olVE: Neural reconstruction of volumes for visualization enhancement of direct volume rendering.Computers & Graphics, 131:104295, 2025. doi:10.1016/j.cag.2025.1042951

    2025

  48. [63]

    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,

  49. [64]

    doi:10.1109/TVCG.2024.34327102

    2024

  50. [65]

    Zhang, P

    Z. Zhang, P. Luo, C. C. Loy, and X. Tang. Facial landmark detection by deep multi-task learning. InProceedings of European Conference on Computer Vision, pp. 94–108, 2014. doi:10.1007/978-3-319-10599-4_72

    2014 doi

  51. [66]

    Z. Zhao, L. Hong, L. Wei, J. Chen, A. Nath, S. Andrews et al. Rec- ommending what video to watch next: A multitask ranking system. In Proceedings of ACM Conference on Recommender Systems, pp. 43–51,

  52. [67]

    Zheng, Y

    L. Zheng, Y . Wu, and K.-L. Ma. Perceptually-based depth-ordering en- hancement for direct volume rendering.IEEE Transactions on Visual- ization and Computer Graphics, 19(3):446–459, 2013. doi: 10.1109/TVCG. 2012.1445, 11

    2013

  53. [68]

    S. Yao, Y . Lu, and C. Wang. ViSNeRF: Efficient multidimensional neural radiance field representation for visualization synthesis of dynamic volu- metric scenes. InProceedings of IEEE Pacific Visualization Conference, pp. 235–245, 2025. doi:10.1109/PacificVis64226.2025.000291

    2025

  54. [70]

    Yao and C

    S. Yao and C. Wang. V olSegGS: V olumetric segmentation of dynamic visualization scenes using deformable 3D Gaussians.IEEE Transactions on Visualization and Computer Graphics, 32(1):407–417, 2026. doi: 10. 1109/TVCG.2025.36425161, 2

    2026

  55. [71]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreason- able effectiveness of deep features as a perceptual metric. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 586–595, 2018. doi:10.1109/CVPR.2018.000685

    2018

  56. [76]

    Zwicker, H

    M. Zwicker, H. Pfister, J. van Baar, and M. Gross. EW A volume splatting. InProceedings of IEEE Visualization Conference, pp. 29–538, 2001. doi: 10.1109/VISUAL.2001.9644902, 3

    2001

  57. [2019]

    doi:10.1145/3298689.33469972

  58. [2020]

    doi:10.1109/TVCG.2020.30303442

    2020

  59. [2024]

    doi:10.48550/arXiv.2411.184732

    doi arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.