Pith. sign in

REVIEW 3 major objections 6 minor 70 references

ESVR: 3D Ellipsoid-based Sparse Volume Rendering via Structure-aware Primitive Learning and Per-primitive Ray Sampling

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A method that fits differentiable ellipsoids to sparse volumetric intensity fields and renders them via per-primitive ray sampling, achieving real-time interaction for hundreds of gigabytes of data at high compression.

desk verdict Solid, well-engineered method paper whose headline compression claim is unverifiable as written due to an internal contradiction about MIP-downsampled training data. read the letter →

arxiv 2608.05564 v1 pith:K3WHA4ZY submitted 2026-08-06 cs.GR

Suemin Jeon , Youjin Kim , Jungwoo Park , Kyungryun Lee , Won-Ki Jeong This is my paper · ORCID
classification cs.GR
keywords renderingsparsedataesvrvolumevolumetricaddressdatasets
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

Volume rendering turns 3D data into images, but enormous scans like whole-brain connectomes can be hundreds of gigabytes, making direct rendering impossible on normal machines. ESVR takes a different route: instead of keeping every voxel, it places a limited set of flexible 3D ellipsoids where the signal actually is, usually less than 1 percent of the space. Each ellipsoid stores a center intensity, an orientation, and a size, and the sum of these blobs approximates the original data.

The key technical pieces are three. First, the ellipsoids use a smooth sigmoid boundary, which lets the optimizer update them through standard gradients, unlike the hard-edged ellipsoids used in some prior work. Second, a structure-aware pruning step scores each ellipsoid by its volume, intensity, and how stretched it is, deleting the least informative ones while keeping elongated structures that align with fibers and vessels. Third, rendering is done by marching rays through the volume but processing each ellipsoid once and adding its contribution only to the ray segments it intersects, which is much faster than checking every ray step against every primitive.

The paper reports up to four orders of magnitude compression and frame rates above 40 FPS on volumes up to 568GB using a single RTX 3090, with reconstruction quality better than neural-network baselines on sparse data. The method is limited to sparse, anisotropic data: on dense volumes with smooth boundaries, neural representations still win.

Extended reading notes

Core claim

The abstract states: 'Across large sparse datasets, ESVR achieves up to four orders of magnitude compression and real-time rendering at 43-223 FPS while maintaining competitive reconstruction quality.' The load-bearing version is the Table 2 result: on Hemibrain (lLN2P), a 568GB volume, ESVR reaches 37.03dB 3D PSNR with a 5.8MB model at 86.76 FPS, while the strongest SRN baseline (AMGSRN++) reaches 36.55dB at 6.1MB and 0.02 FPS. If this is true, the framework delivers a 10x to 4000x speedup over neural baselines at similar or better fidelity on sparse data.

Load-bearing premise

The structure-aware pruning score in Eq. (8), score_i = |Omega_i| * I_i * FA_i, is adopted as the ranking of primitive importance. The paper validates this heuristic with ablations on three datasets (Table 8), but there is no derivation from a reconstruction objective. On a different dataset where intensity, coverage, and anisotropy are not multiplicative proxies for importance, the pruning step could remove informative primitives, breaking the reported compression-quality trade-off. The fragility is visible in the paper's own limitation statement that 'no fixed setting thus generalizes across data types,' i.e., the pruning schedule had to be engineered separately for BigBrain.

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

3 major / 6 minor

Summary. ESVR proposes a compact ellipsoidal-primitive representation for direct volume rendering of large sparse volumetric data. The method fits differentiable, normalized sigmoid-boundary ellipsoids to the raw 3D intensity field via an L1+SSIM loss, uses structure-aware pruning based on a score combining voxel coverage, center intensity, and fractional anisotropy, and renders by per-primitive ray sampling with depth-binned intra-ray parallelism. A chunk-based training scheme with ghost ellipsoids extends the method to volumes exceeding GPU memory. Experiments compare against SRN-based baselines (fV-SRN, NGP, AMGSRN++, InstantVNR) as well as DVR, VDB, and traditional compression methods, reporting model size, 3D/2D PSNR, SSIM, LPIPS, FPS, and training time on datasets from 65.5 MB to 568 GB. The headline claims are up to four orders of magnitude compression and interactive FPS (43-223) on large sparse volumes.

Significance. If the reported results hold, ESVR would be a practically valuable representation for interactive visualization of massive sparse volumes such as connectome data, with a much more favorable size-accuracy-speed trade-off than neural SRN approaches. The paper's strengths include a broad evaluation with size-matched and PSNR-matched baselines, ablations of primitive type, pruning score components, boundary sharpness, depth bins, and comparisons to lossy compressors and VDB-based methods, as well as an honest limitations section. It also describes a concrete rendering algorithm (per-primitive ray sampling with depth bins) and a chunk-based training strategy that are likely to be useful beyond this specific implementation. However, the central compression claim currently rests on an unresolved ambiguity about the resolution of the input volume used for training and evaluation, so the significance cannot be fully assessed as written.

major comments (3)
  1. [Appendix C.1, Appendix Table 1, Table 2, abstract] The claimed 'up to four orders of magnitude compression' is not verifiable from the paper as written, because Appendix C.1 and Appendix Table 1 contradict each other on the resolution of the large datasets. Appendix C.1 states that 'due to the large size of full-resolution volumes, we use downsampled representations at MIP level 2 (×8 downsampling) or 3 (×64 downsampling) for analysis,' while Appendix Table 1 lists, for Hemibrain (lLN2P), 125 chunks of resolution (1226, 955, 1023), whose product matches the full-resolution 5115×4774×6128 volume split into a 5×5×5 grid. If the model was trained on a MIP2 (×8 total, or ×2 per axis) version, the actual input for lLN2P is roughly 1.2 GB and the compression ratio is about 200×, not the ≈98,000× implied by 568 GB/5.8 MB; if MIP3 (×64 total, or ×4 per axis) was used, the input is about 2.2 MB, i.e., smaller than the resulting model. The same ambiguity affects every large-dataset number in Table 2, including the 3D PSNR values, since the reconstruction metric is computed against whatever ground-truth tensor was loaded. Please report, per dataset, the exact MIP level or voxel resolution used for (i) training ESVR, (ii) training each baseline, and (iii) computing the 3D PSNR and the DVR reference images, and compute compression ratios with respect to the actual input tensor, not the full-resolution file size in Table 1.
  2. [Section 6 (Limitations), Appendix C.2, Eq. (8), Table 8] The generality of the 'structure-aware primitive learning' contribution is weakened by the paper's own statement that 'no fixed setting thus generalizes across data types,' combined with the dataset-specific training protocols in Appendix C.2: BigBrain requires three pruning passes with different ratios (0.8, 0.8, 0.85) while all other datasets use one pass, and gradient thresholds vary by two orders of magnitude (1e-10 to 6.5e-9). The importance score in Eq. (8) is a heuristic product of coverage, center intensity, and fractional anisotropy, validated by ablations on only three datasets (Table 8). The paper should either provide a systematic sensitivity analysis (e.g., PSNR as a function of pruning ratio, pruning epoch, and gradient threshold for all datasets) and identify a default configuration that works without per-dataset engineering, or explicitly reposition the contribution as requiring dataset-specific tuning. As it stands, the reported headline results may reflect hand-tuned schedules rather than a general method.
  3. [Section 5.1 (Baseline comparison), Figure 8, Table 2] The fairness of the large-volume comparison is not fully documented. The text says that for volumes >10 GB, fV-SRN and NGP use the domain decomposition implementation from AMGSRN++, and Figure 8's caption notes that Woodbranch and Hemibrain are downsampled (×8 and ×64) to enable DVR (GT) rendering without OOM. It is unclear whether the SRN baselines were trained on the same resolution as ESVR, whether the 3D PSNR in Table 2 is computed on the full-resolution or the downsampled volume, and whether the DVR reference used for 2D PSNR/SSIM/LPIPS is rendered from the same downsampled volume for all methods. Please make the per-method training and evaluation resolutions explicit and justify that the size-matched comparison is apples-to-apples with respect to input resolution.
minor comments (6)
  1. [Section 1, Introduction] The text contains the typo 'V olVis'; it should be 'VolVis'.
  2. [Figure 8 caption and Appendix C.1] The downsampling notation is inconsistent: the caption uses '(×8 and ×64)' while Appendix C.1 uses 'MIP level 2 (×8 downsampling) or 3 (×64 downsampling)'; please clarify whether these factors apply per axis or to the total voxel count.
  3. [Appendix Table 5] In the Aneurysm row under the Size-matched setting, the reported 3D PSNR values for ZFP (183.60 dB) and SZ3 (∞) at sub-MB sizes appear to be formatting errors; please verify these entries.
  4. [Table 8 versus Table 7] The 'w/o pruning' setting in Table 8 uses about 32k primitives for Aneurysm, whereas the 'before pruning' setting in Table 7 uses 93k; please specify at which stage primitive counts are taken and confirm that the comparison is controlled for primitive count as stated in Section 5.6.
  5. [Algorithm 1] Algorithm 1 uses the symbol r in the depth-range computation while Section 4.4 defines r_supp; align the notation to avoid ambiguity about the support radius factor.
  6. [Section 5.1, Rendering and evaluation] The paper reports FPS over a fixed 360° trajectory, but flexible TF editing is a claimed benefit; please state whether the reported FPS includes the cost of re-evaluating the transfer function during interactive TF changes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ESVR's reconstruction quality and compression are empirical fitting results against ground-truth volumes, with no load-bearing step that reduces to its own inputs.

full rationale

I walked the derivation chain and found no circular step that meets the evidentiary bar. The core claim is that ESVR fits compact ellipsoidal primitives to raw volumetric intensity fields and reconstructs them with high 3D PSNR at small model sizes (Table 2). The fitting objective, L_rec = βL1 + (1−β)L_D−SSIM (Eq. 7), is minimized against ground-truth subvolumes, and the reported PSNR is a reconstruction-quality measure against the same ground-truth data. This is standard representation-learning evaluation, not a prediction derived from fitted parameters, so it is not circular. The structure-aware importance score in Eq. (8), score_i = |Ω_i| · I_i · FA_i, is explicitly a heuristic ('we define a structure-aware importance score'), not a quantity derived from the reconstruction target; its usefulness is supported by ablations (Table 8) rather than assumed by construction. The ellipsoid primitive is inspired by EVER and 3DGS, which are external prior works, and the paper cites no author-overlapping 'uniqueness theorem' or load-bearing self-citation that forces its design choices. The per-primitive ray sampling and ghost-ellipsoid chunking are engineering contributions evaluated against DVR and neural baselines, and their superiority is an empirical result rather than an identity. The Appendix C.1 statement about MIP-level downsampling creates a potential inconsistency with the reported compression ratios, but this is a correctness/verifiability concern about how the input volume was sized, not a case where a prediction is equivalent to its input by definition. No fitted parameter is renamed as a prediction, no known result is merely re-labeled, and no ansatz is smuggled in via self-citation. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 1 invented entities

The method rests on the representational assumption that sparse anisotropic volumes are well-approximated by sums of bounded ellipsoidal primitives, plus a set of hand-tuned hyperparameters (k, cutoff, loss weight, densification thresholds, pruning schedule, subvolume size). No new physical entities are introduced; ghost ellipsoids are an algorithmic device. The empirical validation covers the tested datasets, but the heuristics are not derived from first principles.

free parameters (7)
  • boundary sharpness k = 5
    Fixed hyperparameter controlling sigmoid sharpness; ablations show a 2.2 dB PSNR drop at k=3 on Aneurysm (Appendix Table 10), so the reported numbers are sensitive to this choice.
  • spatial cutoff radius factor = 1.42
    Sets the primitive support cutoff; used in ray sampling and voxelization to avoid truncation artifacts.
  • L1/SSIM loss weight beta = 0.2
    Balances L1 and SSIM terms in Eq. (7); set by hand and used across all experiments.
  • densification gradient threshold and percent-dense parameter = 1e-10 to 6.5e-9 and 0.001 to 0.01 per dataset
    Control primitive splitting and cloning; tuned per dataset (Appendix Table 1), affecting the final primitive count and quality.
  • pruning schedule and ratio = pruning at 42-50% of training; BigBrain uses three passes at epochs 200/300/800 with ratios 0.8/0.85
    Dataset-specific schedule; BigBrain required custom multi-stage pruning to handle heterogeneous regions (Appendix C.2).
  • subvolume side length = 128 to 614 voxels depending on dataset
    Training subvolume size is set as approximately half the longest axis; affects training time and memory footprint.
  • number of depth bins = 8
    Intra-ray binning parameter; default 8 bins, with 4 or 16 as alternatives (Appendix E.5), balancing speed and memory.
assumptions (5)
  • domain assumption The volume's intensity field can be approximated by the sum of a finite set of ellipsoidal primitives with bounded support.
    The entire method is built on this representational assumption; validated only empirically on the tested datasets (Section 4.1).
  • domain assumption Sparse and anisotropic structures dominate the data of interest.
    The method is designed for and evaluated on sparse data; the paper's own limitations state ellipsoids are ill-suited for dense or diffuse-boundary volumes (Section 6).
  • standard math The standard L1+SSIM loss and alpha compositing are appropriate quality measures for volumetric reconstruction.
    Used in Eq. (7) and Eq. (2), following common practice in image and volume quality assessment.
  • standard math The covariance matrix Sigma is positive definite and factorized as R S S^T R^T.
    Standard 3DGS parametrization, used throughout for ellipsoid geometry.
  • ad hoc to paper The sigmoid boundary with normalization preserves peak intensity and is differentiable.
    Eq. (5) and the normalization are a design choice to ensure differentiability; the paper shows hard boundaries (large k) fail to converge (Appendix E.4).
invented entities (1)
  • Ghost ellipsoid
    purpose: Provide boundary context during chunk-based training by voxelizing ellipsoids from neighboring chunks without updating them.
    This is an algorithmic construct inspired by ghost-cell methods, not a physical entity; it has no falsifiable handle outside the implementation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ESVR: 3D Ellipsoid-based Sparse Volume Rendering via Structure-aware Primitive Learning and Per-primitive Ray Sampling." pith.science (2026). https://pith.science/paper/K3WHA4ZY

@misc{pith2026260805564,
  author       = {Pith},
  title        = {Pith review of: ESVR: 3D Ellipsoid-based Sparse Volume Rendering via Structure-aware Primitive Learning and Per-primitive Ray Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K3WHA4ZY}},
  note         = {Machine review of arXiv:2608.05564}
}
read the original abstract

Efficient representation and rendering of large-scale sparse volumetric data remain challenging in scientific visualization, as meaningful structures often occupy only a small fraction of the spatial domain. While direct volume rendering (DVR) provides high-quality visualization, its computational and memory costs scale poorly with data size. Recent advances in 3D Gaussian Splatting (3DGS) address this challenge by representing volumetric scenes with compact geometric primitives, enabling efficient, high-fidelity rendering. However, existing 3DGS-based methods learn from DVR rendered images rather than raw volumes, leading to information loss and limiting flexible transfer function control for interactive exploration. To address these limitations, we propose ESVR, an ellipsoid-based sparse volume rendering framework that directly learns and renders volumetric data in 3D space. Our method combines differentiable ellipsoidal primitives with bounded support, structure-aware primitive learning with complementary pruning, and a per-primitive ray sampling strategy for fast and accurate transfer function mapping. To support large-scale datasets, we further introduce a chunk-based optimization scheme with ghost ellipsoids, providing boundary context during training. Across large sparse datasets, ESVR achieves up to four orders of magnitude compression and real-time rendering at 43-223 FPS while maintaining competitive reconstruction quality.

Figures

Figures reproduced from arXiv: 2608.05564 by the authors.

Figure 1
Figure 1. Rendering results on the Hemibrain (lLN2P) dataset under a model-size-matched setting (5.8MB). Our method reconstructs [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Rendering results on the neuron dataset after TF application, [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of our proposed framework: (a) 3D intensity field fitting pipeline, (b) chunk-based large volume training using ghost ellipsoid and (c) [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Visualization of Woodbranch dataset: (a) ground truth, (b) training [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Artifacts caused by inaccurate TF mapping in 2D rasterization. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Conceptual comparison of data access patterns between (a) [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Illustration of intra-ray parallelism. 3D intensity distribution along the ray. These segments are traversed front-to-back, where the TF maps non-zero intensities to color and opacity, and alpha blending (Eq. 2) produces the final pixel color GPU-acceleration through i…
Figure 8
Figure 8. Figure 8: Qualitative comparison of TF-mapped renderings and [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Visualization of FA scores of ellipsoid primitives learned on a [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Qualitative comparison with iVR-GS under unseen intensities. [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 1
Figure 1. Figure 1: 1D comparison between the Gaussian kernel [PITH_FULL_IMAGE:figures/full_fig_p012_1.png]
Figure 2
Figure 2. Figure 2: TF editing interface of the ESVR viewer. (a) Rendering result of [PITH_FULL_IMAGE:figures/full_fig_p012_2.png]
Figure 3
Figure 3. Figure 3: Qualitative results of traditional compression methods. Wood￾branch and Hemibrain are downsampled (×8 and ×64) to enable DVR (GT, TTHRESH, ZFP, SZ3) rendering without OOM [PITH_FULL_IMAGE:figures/full_fig_p015_3.png]
Figure 4
Figure 4. Figure 4: Rendering results of volumes converted into lossless/lossy VDB [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: 2D slice of synthetic datasets used for primitive type comparison. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 7
Figure 7. Figure 7: Ablation on boundary sharpness (k) of differentiable ellipsoids on [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Effect of depth bin count on (a) rendering performance (FPS), (b) GPU memory usage, and (c) rendering quality. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 33 canonical work pages

  1. [1]

    Ahrens, B

    J. Ahrens, B. Geveci, and C. Law. ParaView: An end-user tool for large data visualization.Visualization Handbook, (8):717–731, 2005. doi: 10. 1016/B978-012387582-2/50038-1 6

    work page 2005

  2. [2]

    K. Ai, K. Tang, and C. Wang. NLI4V olVis: Natural language interaction for volume visualization via LLM multi-agents and editable 3D Gaussian splatting.IEEE Transactions on Visualization and Computer Graphics, 32(1):46–56, 2026. doi: 10.1109/TVCG.2025.3633888 1, 2

    arXiv 2026

  3. [3]

    Amunts, C

    K. Amunts, C. Lepage, L. Borgeat, H. Mohlberg, T. Dickscheid, M.-É. Rousseau et al. BigBrain: An ultrahigh-resolution 3D human brain model. Science, 340(6139):1472–1475, 2013. doi: 10.1126/science.1235381 5, 13

    work page doi:10.1126/science.1235381 2013

  4. [4]

    Ballester-Ripoll, P

    R. Ballester-Ripoll, 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 9, 14

    arXiv 2020

  5. [5]

    Y . Cai, Y . Liang, J. Wang, A. Wang, Y . Zhang, X. Yang et al. Radiative Gaussian splatting for efficient X-ray novel view synthesis. InProceedings of European Conference on Computer Vision, pp. 283–299. Springer, 2024. doi: 10.1007/978-3-031-73232-4_16 2

    work page doi:10.1007/978-3-031-73232-4_16 2024

  6. [6]

    W. Chen, L. Ren, M. Zwicker, and H. Pfister. Hardware-accelerated adaptive EWA volume splatting. InProceedings of IEEE Visualization Conference, pp. 67–74, 2004. doi: 10.1109/VISUAL.2004.38 2

    work page doi:10.1109/visual.2004.38 2004

  7. [7]

    Chen and G

    Y . Chen and G. H. Lee. DOGS: Distributed-oriented Gaussian splatting for large-scale 3D reconstruction via Gaussian consensus. InProceedings of Advances in Neural Information Processing Systems, vol. 37, pp. 34487– 34512, 2024. doi: 10.48550/arXiv.2405.13943 4

    arXiv doi:10.48550/arxiv.2405.13943 2024

  8. [8]

    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 Computer Graphics, 32(6):4883–4893, 2026. doi: 10.1109/TVCG.2026. 3694460 2

    doi:10.1109/tvcg.2026 2026

Show all 70 references
  1. [9]

    Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang. LightGaussian: Unbounded 3D Gaussian compression with 15x reduction and 200+ fps. InProceedings of Advances in Neural Information Processing Systems, vol. 37, pp. 140138–140158, 2024. doi: 10.48550/arXiv.2311.17245 2, 4, 14

    2024 doi arXiv

  2. [10]

    Ganter and M

    D. Ganter and M. Manzke. An analysis of region clustered BVH volume rendering on GPU.Computer Graphics Forum, 38(8):13–21, 2019. doi: 10.1111/cgf.13756 1

    2019 doi

  3. [11]

    Hadwiger, A

    M. Hadwiger, A. K. Al-Awami, J. Beyer, M. Agus, and H. Pfister. Sparse- Leap: Efficient empty space skipping for large-scale volume rendering. IEEE Transactions on Visualization and Computer Graphics, 24(1):974– 983, 2018. doi: 10.1109/TVCG.2017.2744238 1

    2018

  4. [12]

    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.2022.3197203 1, 2

    2023

  5. [13]

    Hanson, A

    A. Hanson, A. Tu, G. Lin, V . Singla, M. Zwicker, and T. Goldstein. Speedy- Splat: Fast 3D Gaussian splatting with sparse pixels and sparse primitives. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21537–21546, 2025. doi: 10.48550/arXiv.2...

    2025 doi

  6. [14]

    Hanson, A

    A. Hanson, A. Tu, V . Singla, M. Jayawardhana, M. Zwicker, and T. Gold- stein. PUP 3D-GS: Principled uncertainty pruning for 3D Gaussian splat- ting. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5949–5958, 2025. doi: 10.48550/arXiv.2406. 10219 2

    2025 doi

  7. [15]

    Huang, Z

    B. Huang, Z. Liu, and N. Wong. Decomposing densification in Gaussian splatting for faster 3D scene reconstruction.arXiv preprint arXiv:2507.20239, 2025. doi: 10.48550/arXiv.2507.20239 2, 4

    2025 doi arXiv

  8. [16]

    Kahler, M

    R. Kahler, M. Simon, and H.-C. Hege. Interactive volume rendering of large sparse data sets using adaptive mesh refinement hierarchies.IEEE Transactions on Visualization and Computer Graphics, 9(3):341–351,

  9. [17]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkuehler, and G. Drettakis. 3D Gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4):139:1–139:14, 2023. doi: 10.1145/3592433 1, 2, 4, 6

    2023 doi

  10. [18]

    Kerbl, A

    B. Kerbl, A. Meuleman, G. Kopanas, M. Wimmer, A. Lanvin, and G. Dret- takis. A hierarchical 3D Gaussian representation for real-time rendering of very large datasets.ACM Transactions on Graphics, 43(4):62:1–62:15,

  11. [19]

    D. Kim, M. Lee, and K. Museth. NeuralVDB: High-resolution sparse vol- ume representation using hierarchical neural networks.ACM Transactions on Graphics, 43(2):20:1–20:21, 2024. doi: 10.1145/3641817 1, 15

    2024 doi

  12. [20]

    Kleinbeck, H

    C. Kleinbeck, H. Schieber, K. Engel, R. Gutjahr, and D. Roth. Multi- layer Gaussian splatting for immersive anatomy visualization.IEEE Transactions on Visualization and Computer Graphics, 31(5):2353–2363,

  13. [21]

    Labschütz, S

    M. Labschütz, S. Bruckner, M. E. Gröller, M. Hadwiger, and P. Rautek. Jittree: A just-in-time compiled sparse GPU volume data structure.IEEE Transactions on Visualization and Computer Graphics, 22(1):1025–1034,

  14. [22]

    M. Levoy. Display of surfaces from volume data.IEEE Computer Graph- ics and Applications, 8(3):29–37, 2002. doi: 10.1109/38.511 1

    2002 doi

  15. [23]

    Liang, S

    X. Liang, S. Di, D. Tao, S. Li, S. Li, H. Guo et al. Error-controlled lossy compression optimized for high compression ratios of scientific datasets. InProceedings of IEEE International Conference on Big Data, pp. 438–447, 2018. doi: 10.1109/BigData.2018.8622520 9, 14

    2018

  16. [24]

    J. Lin, Z. Li, X. Tang, J. Liu, S. Liu, J. Liu et al. VastGaussian: Vast 3D Gaussians for large scene reconstruction. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5166–5175,

  17. [25]

    Lindstrom

    P. Lindstrom. Fixed-rate compressed floating-point arrays.IEEE Transac- tions on Visualization and Computer Graphics, 20(12):2674–2683, 2014. doi: 10.1109/TVCG.2014.2346458 9, 14

    2014

  18. [26]

    Y . Liu, C. Luo, L. Fan, N. Wang, J. Peng, and Z. Zhang. CityGaussian: Real-time high-quality large-scale scene rendering with Gaussians. In Proceedings of European Conference on Computer Vision, pp. 265–282. Springer, 2024. doi: 10.1007/978-3-031-72640-8_1 4

    2024 doi

  19. [27]

    Ljung, J

    P. Ljung, J. Krüger, E. Groller, M. Hadwiger, C. D. Hansen, and A. Yn- nerman. State of the art in transfer functions for direct volume rendering. Computer Graphics Forum, 35(3):669–691, 2016. doi: 10.1111/cgf.12934 1

    2016 doi

  20. [28]

    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.14295 1, 2

    2021 doi

  21. [29]

    doi: 10.48550/arXiv.2402.17427 4

    doi arXiv

  22. [30]

    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 1, 2, 13

    2022

  23. [31]

    K. Museth. VDB: High-resolution sparse volumes with dynamic topology. ACM Transactions on Graphics, 32(3):27:1–27:22, 2013. doi: 10.1145/ 2487228.2487235 1, 9, 15

    2013

  24. [32]

    K. Museth. NanoVDB: A GPU-friendly and portable VDB data struc- ture for real-time rendering and simulation. InProceedings of ACM SIGGRAPH Talks, pp. 1–2, 2021. doi: 10.1145/3450623.3464653 9, 15

    2021

  25. [33]

    Neophytou, K

    N. Neophytou, K. Mueller, K. T. McDonnell, W. Hong, X. Guan, H. Qin et al. GPU-accelerated volume splatting with elliptical RBFs. InProceedings of Joint Eurographics - IEEE VGTC Symposium on Visualization, pp. 13– 20, 2006. doi: 10.2312/VisSym/EuroVis06/013-020 2

    2006 doi

  26. [34]

    A. Mai, P. Hedman, G. Kopanas, D. Verbin, D. Futschik, Q. Xu et al. EVER: Exact volumetric ellipsoid rendering for real-time view synthe- sis. InProceedings of IEEE/CVF International Conference on Computer Vision, pp. 4930–4939, 2025. doi: 10.1109/ICCV51701.2025.00469 2, 3

    2025

  27. [35]

    Niedermayr, J

    S. Niedermayr, J. Stumpfegger, and R. Westermann. Compressed 3D Gaussian splatting for accelerated novel view synthesis. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10349–10358, 2024. doi: 10.1109/CVPR52733.2024.00985 2 10 © 2026 IEEE. ...

    2024

  28. [36]

    Papantonakis, G

    P. Papantonakis, G. Kopanas, B. Kerbl, A. Lanvin, and G. Drettakis. Reducing the memory footprint of 3D Gaussian splatting.Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):16:1– 16:17, 2024. doi: 10.1145/3651282 2

    2024 doi

  29. [37]

    S. M. Plaza, J. Clements, T. Dolafi, L. Umayam, N. N. Neubarth, L. K. Scheffer et al. neuPrint: An open access tool for EM connectomics. Frontiers in Neuroinformatics, 16, 2022. doi: 10.3389/fninf.2022.896292 13

    2022

  30. [38]

    L. K. Scheffer, C. S. Xu, M. Januszewski, Z. Lu, S.-y. Takemura, K. J. Hayworth et al. A connectome and analysis of the adult Drosophila central brain.eLife, 9:e57443, 2020. doi: 10.7554/eLife.57443 5, 13

    2020 doi

  31. [39]

    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.arXiv preprint arXiv:2404.11285, 2024. doi: 10.48550/ arXiv.2404.11285 1, 2

    2024 doi arXiv

  32. [40]

    Silversmith, A

    W. Silversmith, A. Zlateski, J. A. Bae, I. Tartavull, N. Kemnitz, J. Wu et al. Igneous: Distributed dense 3D segmentation meshing, neuron skeletonization, and hierarchical downsampling.Frontiers in Neural Circuits, 16, 2022. doi: 10.3389/fncir.2022.977700 13

    2022

  33. [41]

    Talegaonkar, Y

    C. Talegaonkar, Y . Belhe, R. Ramamoorthi, and N. Antipa. V olumetri- cally consistent 3D Gaussian rasterization. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10953–10963, 2025. doi: 10.48550/arXiv.2412.03378 2, 3

    2025 doi arXiv

  34. [42]

    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,

  35. [43]

    Tang and C

    K. Tang and C. Wang. STSR-INR: Spatiotemporal super-resolution for multivariate time-varying volumetric data via implicit neural representa- tion.Computers & Graphics, 119:103874, 2024. doi: 10.1016/j.cag.2024. 01.001 2

    2024 doi

  36. [44]

    Sicat, J

    R. Sicat, J. Krüger, T. Möller, and M. Hadwiger. Sparse PDF volumes for consistent multi-resolution volume rendering.IEEE Transactions on Visualization and Computer Graphics, 20(12):2417–2426, 2014. doi: 10. 1109/TVCG.2014.2346324 1

    2014

  37. [45]

    Tseng and J

    Y .-H. Tseng and J. H. Ferziger. A ghost-cell immersed boundary method for flow in complex geometry.Journal of Computational Physics, 192(2):593–623, 2003. doi: 10.1016/j.jcp.2003.07.024 4

    2003 doi

  38. [46]

    Vega-Higuera, P

    F. Vega-Higuera, P. Hastreiter, R. Fahlbusch, and G. Greiner. High per- formance volume splatting for visualization of neurovascular data. In Proceedings of IEEE Visualization Conference, pp. 271–278, 2005. doi: 10.1109/VISUAL.2005.1532805 1, 2

    2005

  39. [47]

    I. Wald, S. Zellmann, N. Morrical, M. Larsen, and F. Sadlo. Faster rtx- accelerated empty space skipping using triangulated active region bound- ary geometry. InProceedings of Eurographics Symposium on Parallel Graphics and Visualization, pp. 37–44, 2021. doi: 10.2312/pgv.2021...

    2021 doi

  40. [48]

    Weiss, P

    S. Weiss, P. Hermüller, and R. Westermann. Fast neural representations for direct volume rendering.Computer Graphics Forum, 41(6):196–211,

  41. [49]

    Q. Wu, D. Bauer, M. J. Doyle, and K.-L. Ma. Interactive volume visualiza- tion via multi-resolution hash encoding based neural representation.IEEE Transactions on Visualization and Computer Graphics, 30(8):5404–5418,

  42. [50]

    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.2025.3567121 1, 2, 13

    2025

  43. [51]

    S. W. Wurster and H.-W. Shen. AMGSRN++: Improved adaptive SRN for scientific visualization. InProceedings of IEEE Pacific Visualization Sym- posium, pp. 182–191, 2025. doi: 10.1109/PacificVis64226.2025.00024 2, 13

    2025

  44. [52]

    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 Transactions on Visualization and Computer Graphics, 30(1):965–974, 2024. doi: 10.1109/TVCG.2023.3327194 1, 2, 13

    2024

  45. [53]

    Yao and C

    S. Yao and C. Wang. V olSegGS: Segmentation and tracking in dynamic volumetric scenes via deformable 3D Gaussians.IEEE Transactions on Visualization and Computer Graphics, 32(1):407–417, 2026. doi: 10. 1109/TVCG.2025.3642516 2

    2026

  46. [54]

    Zellmann, M

    S. Zellmann, M. Hellmann, and U. Lang. A linear time bvh construction algorithm for sparse volumes. InProceedings of IEEE Pacific Visualization Symposium, pp. 222–226, 2019. doi: 10.1109/PacificVis.2019.00033 1

    2019

  47. [55]

    Zellmann, M

    S. Zellmann, M. Jaros, J. Amstutz, and I. Wald. GPU volume ren- dering with hierarchical compression using VDB.arXiv preprint arXiv:2504.04564, 2025. doi: 10.48550/arXiv.2504.04564 15

    2025 doi arXiv

  48. [56]

    Zellmann, D

    S. Zellmann, D. Meurer, and U. Lang. Hybrid grids for sparse volume rendering. InProceedings of IEEE Visualization Conference, pp. 1–5,

  49. [57]

    doi: 10.1109/TVCG.2023.3293121 1, 2, 13

    2023

  50. [58]

    S. W. Wurster, H. Guo, H.-W. Shen, T. Peterka, and J. Xu. Deep hierarchi- cal super resolution for scientific data.IEEE Transactions on Visualization and Computer Graphics, 29(12):5483–5495, 2023. doi: 10.1109/TVCG. 2022.3214420 1, 2

    2023

  51. [59]

    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.2018.00068 6

    2018

  52. [60]

    Zhang, T

    Z. Zhang, T. Song, Y . Lee, L. Yang, C. Peng, R. Chellappa et al. LP-3DGS: Learning to prune 3D Gaussian splatting.Advances in Neural Information Processing Systems, 37:122434–122457, 2024. doi: 10.48550/arXiv.2405. 18784 2

    2024 doi

  53. [61]

    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.2001.964490 2 11 A METHODSUPPLEMENTARY Algorithm 1Structure-aware primitive learning V : ground-truth volume E: ellip...

    2001

  54. [66]

    Zellmann, J

    S. Zellmann, J. P. Schulze, and U. Lang. Binned k-d tree construction for sparse volume data on multi-core and GPU systems.IEEE Transactions on Visualization and Computer Graphics, 27(3):1904–1915, 2021. doi: 10 .1109/TVCG.2019.2938957 1

    1904

  55. [67]

    R. Zha, T. J. Lin, Y . Cai, J. Cao, Y . Zhang, and H. Li. R 2-Gaussian: rectifying radiative Gaussian splatting for tomographic reconstruction. In Proceedings of Advances in Neural Information Processing Systems, pp. 44907–44934, 2024. doi: 10.48550/arXiv.2405.20693 6, 14

    2024 doi arXiv

  56. [2003]

    doi: 10.1109/TVCG.2003.1207442 1

    2003 arXiv

  57. [2016]

    doi: 10.1109/TVCG.2015.2467331 1

    2015

  58. [2019]

    doi: 10.1109/VISUAL.2019.8933631 1

    2019

  59. [2022]

    doi: 10.1111/cgf.14578 1, 2, 13, 14

    doi

  60. [2024]

    doi: 10.1145/3658160 4

    doi

  61. [2025]

    doi: 10.1109/TVCG.2025.3549882 1, 2

    2025

  62. [2026]

    doi: 10.1109/TVCG.2025.3634643 1, 2

    2025

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.