Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

GSCache: Real-Time Radiance Caching for Volume Path Tracing using 3D Gaussian Splatting

T0 review · 2 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A multi-level 3D Gaussian radiance cache lets volume path tracers render less noisy images at no extra cost.

desk verdict A novel path-space Gaussian radiance cache with promising empirical results, but the training target's unbiasedness is unproven and likely biased, which casts doubt on the headline claim. read the letter →

arxiv 2507.19718 v2 pith:CILLKZJQ submitted 2025-07-25 cs.GR cs.LG

David Bauer , Qi Wu , Hamid Gadirov , Kwan-Liu Ma This is my paper
classification cs.GRcs.LG
keywords radiancecachingpathtracingvolumerendering3DGaussiansplattingpath-spacescientificvisualizationreal-timenoise2noise
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

This paper claims that the noisy, high-variance Monte Carlo images produced by volume path tracers can be cleaned up in real time by caching radiance in path space: instead of storing radiance at world-space probes, it stores, for each path length, the attenuated radiance of all paths of that length in a hierarchy of 3D Gaussian splats. The cache is trained on the fly from the very same noisy path samples the renderer produces, relying on the Noise2Noise principle that gradient descent on noisy targets converges to the clean expectation. At one sample per pixel, the cached renderer reports PSNR gains of roughly 2 to 5 dB over a next-event-estimation baseline on six volume datasets, with frame times comparable to the baseline and a cache that adapts within about 16 frames to changed lighting or transfer functions. The significance would be a drop-in, scene-agnostic way to suppress Monte Carlo noise in scientific volume visualization, where the cache's path-space design means it needs only the attenuated radiance value and path length from the host renderer.

What carries the argument

The central object is a multi-level point cloud of 3D Gaussians, one level per path length, rasterized with the differentiable 3D Gaussian splatting rasterizer to produce cache images used by the path tracer. The cache carries the argument by being trainable in real time on noisy path buffers: the HDR loss of [26] is used as the Noise2Noise objective, and an AdamW optimizer with adaptive learning rates and a size-capped, density-matched initial point cloud keeps the optimization stable under high-variance radiance targets. A throughput-based path termination heuristic, with a user coefficient $C$ and a $\beta_n$ cascade correction, decides when to cut paths short and read the cache.

What would settle it

In a simple scene with one volume and one light, render with early termination enabled and disabled at high sample counts, and compare the cached radiance at a single path-length level to a converged unbiased reference for that level; if the two diverge systematically, the early-termination training targets are biased.

Watch

Extended reading notes

Core claim

The core discovery is that the radiance integral along a volumetric ray can be reorganized by path length, with each path-length subspace represented by its own level of a Gaussian-splatting cache, and that this cache can be optimized in real time directly on noisy, single-sample path buffers. The paper argues that the expected value of the samples accumulating in each level buffer behaves like the expected value of regular path tracing, so applying the Noise2Noise training rule to these noisy targets drives the Gaussians toward the true path-space radiance of that level. Because each level stores pre-attenuated radiance at a specific path length, a path can be terminated early into the cache whenever its current throughput is low, and a product-of-cache-miss-probabilities factor ($\beta$) corrects the cached read for the chance that the path would have stopped earlier. The result is a cache that trades variance for bias but, at the reported settings, yields measurably higher PSNR at 1 SPP than both uniform-sampling and NEE baselines and than a neural radiance caching baseline, with the cache's runtime overhead constant in the sample count.

Load-bearing premise

The load-bearing premise is that path samples stored in each level buffer, including those from paths cut short by the cache's own termination rule, are unbiased estimators of that level's path-space radiance after throughput adjustment, so that Noise2Noise training converges to the correct cached radiance.

Editorial extensions

If this is right

  • At low sample counts the cache raises image quality without a frame-time penalty, and the benefit grows as sample counts rise because the cache's overhead stays constant while each sample is generated faster through early termination.
  • The cache adapts to user-driven scene changes such as transfer functions and lighting within roughly 16 frames, making it suitable for interactive exploration rather than offline rendering.
  • The cache relies only on attenuated radiance and path length from the host renderer, so it can be added to existing volume path tracers without requiring intersection metadata or sampling details.
  • Because the cache learning is self-contained, it can be trained purely on noisy Monte Carlo output, with no clean reference images needed.

Reading between the lines

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

  • The same path-space cache structure could in principle wrap surface path tracers as easily as volume ones, since pre-attenuated radiance at a given bounce depth is a scene-agnostic quantity; the paper only demonstrates it on volumes.
  • The bias-variance trade-off controlled by $C$ suggests an automatic controller could tune $C$ per pixel or per frame to a target noise level, something the paper leaves to the user.
  • The claim that early-terminated samples are unbiased targets for the cache is the one step most worth an independent derivation or Monte Carlo check, since the paper's own $\beta$ correction handles cache reads but not the training target distribution.
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

2 major / 6 minor

Summary. The manuscript introduces GSCache, a multi-level radiance cache for real-time volume path tracing. Each cache level is a set of 3D Gaussians that represents the radiance contribution of paths of a given length; the cache is rasterized into images, used to terminate and shade paths early, and optimized online with noisy path samples. The path-tracer uses a throughput-dependent termination heuristic, and cached radiance is attenuated with a cascade correction factor. The evaluation compares against a uniform-sampling path tracer, a next-event-estimation path tracer, and a re-implemented NRC baseline on six volume datasets, reporting PSNR and frame timings, with additional ablations of cache size and hyperparameters.

Significance. The paper's main novelty is the use of 3D Gaussian splatting as a trainable path-space radiance cache for volume rendering. If the training-target issue below is resolved, the practical contribution is significant: online Noise2Noise-style training of a cache from noisy radiance buffers is well motivated, the method is relatively non-invasive, and the reported PSNR gains at 1 SPP are large and consistent across the six datasets. The cache-size and hyperparameter ablations are useful, and the initialization timing table provides useful reproducibility detail. The main correctness risk is the claimed unbiasedness of the training buffers, which is the load-bearing justification for the Noise2Noise training; this is not a circularity issue but an unproven statistical claim.

major comments (2)
  1. [§3.5 and Algorithm 1] Early-terminated training samples are not shown to be unbiased estimators of path-space radiance. In Algorithm 1, when q < p the function returns True before the correction Trout <- Trout / Tr, so a sample stored at its terminating level is weighted by Trout while the probability of being stored is p = 1 - clamp(C * luminance(Trout), 0, 1), which depends on the same Trout. The expected value of a stored level-n sample is therefore E[Trout L_n | terminated] = E[Trout L_n p] / E[p], not E[Trout L_n]. Section 3.5 asserts that the behavior is "analogous to regular path tracing" without deriving this expectation. Equation (3)'s beta correction applies to cache reads, not to the training-target distribution, so it does not fix the target. Because Noise2Noise converges to the conditional mean of the noisy target, the cache would learn a termination-weighted average of radiance rather than the path-space radiance of Eq. (1). Please derive the exact target of the per-level buffers under Algorithm 1 and correct the stored weight (e.g., divide by p) or provide the independence assumptions under which the current weighting is unbiased.
  2. [§4.1, Table 2, Figure 6] Quantitative claims rest on single-run measurements without error bars or repeated seeds. Cache training is stochastic (random initialization of the Gaussian point cloud, AdamW updates, learning-rate resets on viewport change), and the reported PSNR differences and timings are presented as point values. The reader cannot distinguish systematic gains from run-to-run variation, especially for the comparison with the NRC baseline. Please report multiple runs or per-frame variance across the fly-through, specify how many frames contribute to each PSNR number, and state the warm-up schedule for all methods. The NRC baseline should also be explicitly labeled as a modified reimplementation: it uses the authors' path termination heuristic, disables self-training, and replaces the spatio-directional input with hash-grid encoding.
minor comments (6)
  1. [§3.2, Eq. (2)] The wedge symbol in the scale formula is undefined; if it denotes a component-wise minimum or clamp, please state this explicitly.
  2. [Abstract and §4.4] The phrase "without increasing rendering costs" is stronger than Table 2 supports; for example, MechanicalHand shows roughly 65.8 ms total with the cache versus 33.3 ms for NEE, and Spider shows 183.8 ms versus 148.0 ms. Suggest qualifying the claim to "comparable or lower in several tested scenes."
  3. [§4.1] The NRC description says "The model configuration is the same as NRC [34]" and then introduces hash-grid encoding; please state explicitly which parts are from the original method and which are modifications for this comparison.
  4. [Figure 7] The text says "less than 16 samples" while the caption discusses 16 SPP; clarify whether the horizontal axis is frames or accumulated samples per pixel.
  5. [Table 2] Reporting PT, ST, and OT separately is useful, but the headline comparison would be clearer if each method also had a single combined frame time; the current format makes it easy to misread the totals.
  6. [References] Reference [51] lists "Woodock"; the correct spelling is "Woodcock."

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the cache training and evaluation rest on external Noise2Noise theory and external baselines; the §3.5 buffer-expectation concern is an unproven correctness assumption, not a definitional reduction.

full rationale

The derivation chain is not circular. The cache's training targets are path-buffer samples collected from the renderer ('To train the cache, we collect unbiased samples from the renderer, attenuate them, and assign them to their appropriate cache level', Section 3.5), and the Noise2Noise principle used to justify learning from noisy targets is an external result [26], not a conclusion derived inside this paper. No equation defines the cache's prediction as its own input: Eq. (3) is an importance-sampling attenuation for cache reads, and Algorithm 1 is a termination heuristic, neither of which reduces the claimed quality gain to a fitted parameter. The self-citations ([1], [2], [9], [52], [53], [55]) appear as related work and are not load-bearing for the central claim; the evaluation compares against external baselines (uniform PT, NEE, and an NRC implementation) and a 512-SPP reference, so the PSNR improvements are empirical rather than constructed. The paper also explicitly acknowledges that the method is biased and has no guaranteed error bounds (Section 5, 'Bias'), which is a correctness limitation, not evidence of circularity. The main scientific risk—that Section 3.5's assertion 'the expected value of samples in the buffers behaves analogously to regular path tracing' is not derived, so early-terminated buffer entries may have a conditional-expectation bias—is a soundness/bias concern rather than a definitional or self-citation circularity: the training target is not defined in terms of the cache output. I therefore find no significant circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces a data structure (the multi-level Gaussian cache) and a bookkeeping variable beta, but no new physical entity. The main free parameters are tuning choices for the cache, not fitted constants in a derivation. The load-bearing assumptions are the standard rendering model, the path-space decomposition, and the unbiasedness of noisy training targets.

free parameters (5)
  • Cache termination coefficient C = 0.5
    User-defined coefficient in Algorithm 1 that sets the cache sampling probability; tuned in Figure 8 as a bias-variance trade-off.
  • Initial cache size N = 300,000
    Number of initialization points for level 0; selected by ablation in Figure 11.
  • Number of cache levels K+1 = 3
    Multi-level hierarchy depth chosen for the evaluation; each lower level is half the size of the previous.
  • Learning rates = position 1.16e-3, color 1.25e-2, rotation 1e-3, opacity 1.5e-1, scale 0
    Per-parameter learning rates listed in Section 4.1, chosen by hand for training stability.
  • Allowed throughput threshold 0.9 = 0.9
    Threshold in Algorithm 1 that guarantees some paths continue regardless of cache probability; chosen in testing.
assumptions (5)
  • domain assumption Radiative transfer equation model for volume rendering (Equation 1).
    Underlying optical model used by the baseline and cache.
  • standard math Path-space decomposition of the radiance integral (Veach, Chapter 8).
    Justifies grouping paths by length into distinct cache levels.
  • domain assumption Noise2Noise: minimizing mean squared error against noisy targets with expected value equal to clean targets recovers the clean solution.
    Applied in Section 3.5 to justify training on noisy path buffers; accepted from Lehtinen et al. [26].
  • ad hoc to paper Early-terminated, attenuated path samples have the correct expected value for their path-length buffer.
    Required for cache training to be unbiased; stated as an observation in Section 3.5 but not derived for the early-termination case.
  • domain assumption Path-space radiance at each length can be represented by view-independent isotropic 3D Gaussians with RGB color.
    Motivates cache capacity and disables SH degree 0; assumes sufficiently diffuse volume radiance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GSCache: Real-Time Radiance Caching for Volume Path Tracing using 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/CILLKZJQ

@misc{pith2026250719718,
  author       = {Pith},
  title        = {Pith review of: GSCache: Real-Time Radiance Caching for Volume Path Tracing using 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CILLKZJQ}},
  note         = {Machine review of arXiv:2507.19718}
}
read the original abstract

Real-time path tracing is rapidly becoming the standard for rendering in entertainment and professional applications. In scientific visualization, volume rendering plays a crucial role in helping researchers analyze and interpret complex 3D data. Recently, photorealistic rendering techniques have gained popularity in scientific visualization, yet they face significant challenges. One of the most prominent issues is slow rendering performance and high pixel variance caused by Monte Carlo integration. In this work, we introduce a novel radiance caching approach for path-traced volume rendering. Our method leverages advances in volumetric scene representation and adapts 3D Gaussian splatting to function as a multi-level, path-space radiance cache. This cache is designed to be trainable on the fly, dynamically adapting to changes in scene parameters such as lighting configurations and transfer functions. By incorporating our cache, we achieve less noisy, higher-quality images without increasing rendering costs. To evaluate our approach, we compare it against a baseline path tracer that supports uniform sampling and next-event estimation and the state-of-the-art for neural radiance caching. Through both quantitative and qualitative analyses, we demonstrate that our path-space radiance cache is a robust solution that is easy to integrate and significantly enhances the rendering quality of volumetric visualization applications while maintaining comparable computational efficiency.

Figures

Figures reproduced from arXiv: 2507.19718 by the authors.

Figure 1
Figure 1. Our path-space cache improves image quality at low sample counts at comparable compute cost. We compare against a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The path tracing integral can be characterized as the collection of all possible paths of all possible lengths. When combined, these paths [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of our cache sampling and path termination mech [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (13 more)
Figure 5
Figure 5. Figure 5: Our cache is trained in real time on paths obtained during the [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Visual quality of our method compared to the baseline path tracer. We show results for images at 1 SPP and compare our method [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Comparison of cache state after a cold-start training for 64 frames. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Influence of the cache sampling probability [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Cache amortization behavior example from the [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Comparison of the visual quality of our method when using different cache initialization sizes [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Ablation of our hyperparameter choice. The image compares results from our configuration with various modified versions of those settings. [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Visual quality of our method on the Carp dataset compared to the baseline path tracer. We show results for images at 1 SPP and compare our method (GSCache) against a baseline volume path tracer with uniform sampling (Uniform) and a version that uses next-event estimat…
Figure 14
Figure 14. Figure 14: Visual quality of our method on the FullBody dataset compared to the baseline path tracer. We show results for images at 1 SPP and compare our method (GSCache) against a baseline volume path tracer with uniform sampling (Uniform) and a version that uses next-event est…
Figure 15
Figure 15. Figure 15: Visual quality of our method on the MechanicalHand dataset compared to the baseline path tracer. We show results for images at 2 SPP and compare our method (GSCache) against a baseline volume path tracer with uniform sampling (Uniform) and a version that uses next-eve…
Figure 16
Figure 16. Figure 16: Visual quality of our method on the Spider dataset compared to the baseline path tracer. We show results for images at 1 SPP and compare our method (GSCache) against a baseline volume path tracer with uniform sampling (Uniform) and a version that uses next-event estim…
Figure 17
Figure 17. Figure 17: Visual quality of our method on the Supernova dataset compared to the baseline path tracer. We show results for images at 1 SPP and compare our method (GSCache) against a baseline volume path tracer with uniform sampling (Uniform) and a version that uses next-event es…
Figure 18
Figure 18. Figure 18: Visual quality of our method on the ZebraFish dataset compared to the baseline path tracer. We show results for images at 1 SPP and compare our method (GSCache) against a baseline volume path tracer with uniform sampling (Uniform) and a version that uses next-event es…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. A Query-Efficient Stochastic Volume Rendering Framework for Time-Varying Implicit Neural Volumes

    cs.GR 2026-07 accept novelty 6.0 of 10

    Delta-tracking plus a four-stage RT/tensor-core pipeline renders unmodified time-varying INRs interactively (~30–40 FPS at 1024²) with ~1–2 ms timestep updates.

Reference graph

Works this paper leans on

61 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bauer, Q

    D. Bauer, Q. Wu, and K.-L. Ma. FoVolNet: Fast volume ren- dering using foveated deep neural networks.IEEE Transactions on Visualization and Computer Graphics , 29(1):515–525, 2023. doi: 10.1109/TVCG.2022.3209498 3

    arXiv 2023

  2. [2]

    Bauer, Q

    D. Bauer, Q. Wu, and K.-L. Ma. Photon field networks for dynamic real-time volumetric global illumination.IEEE Trans- actions on Visualization and Computer Graphics , 30(1):975–985,

    work page

  3. [3]

    Bekaert, P

    P. Bekaert, P. Slussalek, R. Cools, V. Havran, and H.-P. Seidel. A custom designed density estimator for light transport.Technical Report, Max Planck Institut für Informatik , (4), 2003. 2, 3

    work page 2003

  4. [4]

    Bitterli, C

    B. Bitterli, C. Wyman, M. Pharr, P. Shirley, A. Lefohn, and W. Jarosz. Spatiotemporal reservoir resampling for real-time ray tracing with dynamic direct lighting.ACM Transactions on Graphics, 39(4):148:1–148:17, article no. 148, 2020. doi: 10. 1145/3386569.3392481 1, 9

    arXiv 2020

  5. [5]

    Condor, S

    J. Condor, S. Speierer, L. Bode, A. Bozic, S. Green, P. Didyk, and A. Jarabo. Don’t splat your gaussians: Volumetric ray- traced primitives for modeling and rendering scattering and emissive media. ACM Transactions on Graphics , 44(1):10:1– 10:17, article no. 10, 2025. doi: 10.1145/3711853 9

    doi:10.1145/3711853 2025

  6. [6]

    Dappa, K

    E. Dappa, K. Higashigaito, J. Fornaro, S. Leschka, S. Wilder- muth, and H. Alkadhi. Cinematic rendering–an alternative to volume rendering for 3d computed tomography imaging.Insights Into Imaging, 7(6):849–856, 2016. doi: 10.1007/s13244-016-0518 -1 2

    work page doi:10.1007/s13244-016-0518 2016

  7. [7]

    Engel and T

    D. Engel and T. Ropinski. Deep volumetric ambient occlusion. IEEE Transactions on Visualization and Computer Graphics , 27(2):1268–1278, 2021. doi: 10.1109/TVCG.2020.3030344 3

    arXiv 2021

  8. [8]

    Facebook Inc. PyTorch. https://pytorch.org/, 2024. [Online; accessed 01-November-2024]. 6

    work page 2024

Show all 61 references
  1. [9]

    Gadirov, Q

    H. Gadirov, Q. Wu, D. Bauer, K.-L. Ma, J. B. Roerdink, and S. Frey. Hyperflint: Hypernetwork-based flow estimation and temporal interpolation for scientific ensemble visualization.Com- puter Graphics Forum , 44(3):e70134, 2025. doi: 10.1111/cgf. 70134 3

    2025 doi

  2. [10]

    Hadadan, S

    S. Hadadan, S. Chen, and M. Zwicker. Neural radiosity.ACM Transactions on Graphics, 40(6):236:1–236:11, article no. 236,

  3. [11]

    Han and C

    J. Han and C. Wang. Coordnet: Data generation and visualiza- tion generation for time-varying volumes via a coordinate-based neural network. IEEE Transactions on Visualization and Com- puter Graphics, 29(12):4951–4963, 2023. doi: 10.1109/TVCG. 2022.3197203 3

    2023

  4. [12]

    Jakob, C

    W. Jakob, C. Regg, and W. Jarosz. Progressive Expectation– Maximization for hierarchical volumetric photon mapping.Com- puter Graphics Forum (Proceedings of EGSR) , 30(4), 2011. doi: 10.1111/j.1467-8659.2011.01988.x 2

    2011

  5. [13]

    Jarosz, C

    W. Jarosz, C. Donner, M. Zwicker, and H. W. Jensen. Radi- ance caching for participating media. ACM Transactions on Graphics (Presented at SIGGRAPH) , 27(1):7:1–7:11, 2008. doi: 10/cwnw78 2, 3

    2008

  6. [14]

    Jönsson, J

    D. Jönsson, J. Kronander, T. Ropinski, and A. Ynnerman. His- torygrams: Enabling interactive global illumination in direct volume rendering using photon mapping.IEEE Transactions on Visualization and Computer Graphics , 18(12):2364–2371, 2012. doi: 10.1109/TVCG.2012.232 2

    2012 doi

  7. [15]

    Jönsson, E

    D. Jönsson, E. Sundén, A. Ynnerman, and T. Ropinski. A survey of volumetric illumination techniques for interactive volume rendering. Computer Graphics Forum, 33(1):27–51, 2014. doi: 10.1111/cgf.12252 2

    2014 doi

  8. [16]

    Kandlbinder, A

    L. Kandlbinder, A. Dittebrandt, A. Schipek, and C. Dachsbacher. Optimizing path termination for radiance caching through ex- plicit variance trading.Proceedings of the ACM on Computer Graphics and Interactive Techniques , 7(3):33:1–33:19, article no. 33, 2024. doi: 10.1145/367...

    2024 doi

  9. [17]

    Keller, K

    A. Keller, K. Dahm, and N. Binder. Path space filtering. In ACM SIGGRAPH 2014 Talks , SIGGRAPH ’14, article no. 68, pp. 68:1–68:1. ACM, New York, 2014. doi: 10.1145/2614106. 2614149 2

    2014 doi

  10. [18]

    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, article no. 139,

  11. [19]

    Khlebnikov, P

    R. Khlebnikov, P. Voglreiter, M. Steinberger, B. Kainz, and D. Schmalstieg. Parallel irradiance caching for interactive monte- carlo direct volume rendering. Computer Graphics Forum , 33(3):61–70, 2014. doi: 10.1111/cgf.12362 2

    2014 doi

  12. [20]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Y. Bengio and Y. LeCun, eds.,3rd International Conference on Learning Representations, ICLR.SanDiego, 2015. doi: 10.48550/arXiv.1412.6980 11

    2015 doi arXiv

  13. [21]

    Kniss, S

    J. Kniss, S. Premoze, C. Hansen, and D. Ebert. Interactive translucent volume rendering and procedural modeling. InIEEE Visualization, pp. 109–116. IEEE, 2002. doi: 10.1109/VISUAL. 2002.1183764 2

    2002 arXiv

  14. [22]

    Kniss, S

    J. Kniss, S. Premoze, C. Hansen, P. Shirley, and A. McPherson. A model for volume lighting and modeling.IEEE Transactions on Visualization and Computer Graphics , 9(2):150–162, 2003. doi: 10.1109/TVCG.2003.1196003 2

    2003 arXiv

  15. [23]

    Krivanek, P

    J. Krivanek, P. Gautron, S. Pattanaik, and K. Bouatouch. Ra- diance caching for efficient global illumination computation. IEEE Transactions on Visualization and Computer Graphics , 11(5):550–561, 2005. doi: 10.1109/TVCG.2005.83 1, 2, 3

    2005 doi

  16. [24]

    Kroes, F

    T. Kroes, F. H. Post, and C. P. Botha. Exposure render: An interactive photo-realistic volume rendering framework.PLOS ONE, 7(7):1–10, 2012. doi: 10.1371/journal.pone.0038586 2

    2012 doi

  17. [25]

    J. C. Lee, D. Rho, X. Sun, J. H. Ko, and E. Park. Compact 3d gaussian representation for radiance field. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 21719–21728, 2024. 9

    2024

  18. [26]

    Lehtinen, J

    J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Karras, M.Aittala, andT.Aila. Noise2Noise: Learningimagerestoration without clean data. In J. Dy and A. Krause, eds.,Proceedings of the 35th International Conference on Machine Learning , vol. 80 of Proceedings of Machine L...

    2018

  19. [27]

    M. Levoy. Display of surfaces from volume data.IEEE Computer Graphics and Applications, 8(3):29–37, 1988. doi: 10.1109/38. 511 2

    1988 doi

  20. [28]

    D. Lin*, M. Kettunen*, B. Bitterli, J. Pantaleoni, C. Yuksel, and C. Wyman. Generalized resampled importance sampling: Foundations of restir. ACM Transactions on Graphics (Pro- ceedings of SIGGRAPH 2022) , 41(4):75:1–75:23, article no. 75,

    2022

  21. [29]

    D. Lin, C. Wyman, and C. Yuksel. Fast volume rendering with spatiotemporal reservoir resampling. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia 2021) , 40(6):278:1– 278:18, article no. 278, 2021. doi: 10.1145/3478513.3480499 1, 9

    2021

  22. [30]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regulariza- tion. In International Conference on Learning Representations ,

  23. [31]

    S. S. Mallick, R. Goel, B. Kerbl, M. Steinberger, F. V. Carrasco, and F. De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. InSIGGRAPH Asia 2024 Conference Papers, SA ’24, article no. 2, pp. 2:1–2:11. ACM, New York,

    2024

  24. [32]

    N. Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics , 1(2):99– 108, 1995. doi: 10.1109/2945.468400 2

    1995

  25. [33]

    Moenne-Loccoz, A

    N. Moenne-Loccoz, A. Mirzaei, O. Perel, R. de Lutio, J. Mar- tinez Esturo, G. State, S. Fidler, N. Sharp, and Z. Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics, 43(6):232:1–232:19, article no. 232,

  26. [34]

    Müller, F

    T. Müller, F. Rousselle, J. Novák, and A. Keller. Real-time neural radiance caching for path tracing.ACM Trans. Graph., 40(4), article no. 36, 2021. doi: 10.1145/3450626.3459812 1, 2, 3, 6, 7, 8, 9

    2021

  27. [35]

    Niedermayr, J

    S. Niedermayr, J. Stumpfegger, and R. Westermann. Com- pressed 3d gaussian splatting for accelerated novel view synthe- sis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 10349–10358, 2024. 9

    2024

  28. [36]

    doi: 10.1145/3680528.3687694 5, 8, 9

  29. [37]

    Supernova dataset

    Oak Ridge National Lab. Supernova dataset. https://www.ornl.gov/. 7

  30. [38]

    Ouyang, S

    Y. Ouyang, S. Liu, M. Kettunen, M. Pharr, and J. Pantaleoni. Restir gi: Path resampling for real-time path tracing.Computer Graphics Forum, 40(8):17–29, 2021. doi: 10.1111/cgf.14378 1

    2021 doi

  31. [39]

    doi: 10.1145/3687934 9

    doi

  32. [40]

    Papantonakis, G

    P. Papantonakis, G. Kopanas, B. Kerbl, A. Lanvin, and G. Dret- takis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interac- tive Techniques, 7(1):16:1–16:17, article no. 16, 2024. doi: 10. 1145/3651282 9

    2024

  33. [41]

    M. Ruiz, I. Boada, I. Viola, S. Bruckner, M. Feixas, and M. Sbert. Obscurance-based volume rendering framework. In H.-C. Hege, D. Laidlaw, R. Pajarola, and O. Staadt, eds.,IEEE/ EG Sym- posium on Volume and Point-Based Graphics , pp. 113–120. The Eurographics Association, 2008...

    2008 doi

  34. [42]

    OptiX.https://developer.nvidia.com/ rtx/ray-tracing/optix, 2024

    NVIDIA Corporation. OptiX.https://developer.nvidia.com/ rtx/ray-tracing/optix, 2024. [Online; accessed 01-November- 2024]. 6

    2024

  35. [43]

    Šoltészová, D

    V. Šoltészová, D. Patel, S. Bruckner, and I. Viola. A multidi- rectional occlusion shading model for direct volume rendering. Computer Graphics Forum, 29(3):883–891, 2010. doi: 10.1111/j. 1467-8659.2009.01695.x 2

    2010

  36. [44]

    Stadlbauer, W

    P. Stadlbauer, W. Tatzgern, J. Mueller, M. Winter, R. Sto- janovic, A. Weinrauch, and M. Steinberger. Adaptive multi-view radiance caching for heterogeneous participating media.Com- puter Graphics Forum, p. e70051, 2025. doi: 10.1111/cgf.70051 2

    2025 doi

  37. [45]

    B. Pan, J. Lu, H. Li, W. Chen, Y. Wang, M. Zhu, C. Yu, and W. Chen. Differentiable design galleries: A differentiable approach to explore the design space of transfer functions. IEEE Transactions on Visualization and Computer Graphics , 30(1):1369–1379, 2024. doi: 10.1109/TVCG...

    2024

  38. [46]

    Šmajdek, v

    U. Šmajdek, v. Lesar, M. Marolt, and C. Bohak. Combined volume and surface rendering with global illumination caching. The Visual Computer , 40(4):2491–2503, 2023. doi: 10.1007/ s00371-023-02932-9 2

    2023

  39. [47]

    G. J. Ward, F. M. Rubinstein, and R. D. Clear. A ray tracing solution for diffuse interreflection.ACM SIGGRAPH Computer Graphics, 22(4):85–92, 1988. doi: 10.1145/378456.378490 1

    1988

  40. [49]

    Weiss, M

    S. Weiss, M. IşIk, J. Thies, and R. Westermann. Learning adaptive sampling and reconstruction for volume visualization. IEEE Transactions on Visualization and Computer Graphics , 28(7):2654–2667, 2022. doi: 10.1109/TVCG.2020.3039340 3

    2022

  41. [50]

    Differentiabledirectvolumerender- ing

    S.WeissandR.Westermann. Differentiabledirectvolumerender- ing. IEEE Transactions on Visualization and Computer Graph- ics, 28(1):562–572, 2022. doi: 10.1109/TVCG.2021.3114769 2

    2022

  42. [51]

    E. Veach. Robust Monte Carlo methods for light transport simulation. Stanford University, 1998. 3

    1998

  43. [52]

    Q. Wu, D. Bauer, Y. Chen, and K.-L. Ma. Hyperinr: A fast and predictive hypernetwork for implicit neural representations via knowledge distillation, 2023. 3

    2023

  44. [53]

    Q. Wu, D. Bauer, M. J. Doyle, and K.-L. Ma. Interactive volume visualization via multi-resolution hash encoding based neural representation. IEEE Transactions on Visualization and Computer Graphics, 30(8):5404–5418, 2024. doi: 10.1109/TVCG .2023.3293121 3

    2024

  45. [54]

    Weiss, P

    S. Weiss, P. Hermüller, and R. Westermann. Fast neural rep- resentations for direct volume rendering.Computer Graphics Forum, 41(6):196–211, 2022. doi: 10.1111/cgf.14578 3

    2022 doi

  46. [55]

    Zavorotny, Q

    D. Zavorotny, Q. Wu, D. Bauer, and K.-L. Ma. From Cluster to Desktop: A Cache-Accelerated INR framework for Interactive Visualization of Tera-Scale Data. In G. Reina, S. Rizzi, and C. Gueunet, eds.,Eurographics Symposium on Parallel Graphics and Visualization. The Eurographics...

    2025

  47. [56]

    Zhang, Z

    Y. Zhang, Z. Dong, and K.-L. Ma. Real-time volume rendering in dynamic lighting environments using precomputed photon mapping. IEEE Transactions on Visualization and Computer Graphics, 19(8):1317–1330, 2013. doi: 10.1109/TVCG.2013.17 2 A Ablation Study We provide an ablation s...

    2013 doi

  48. [57]

    E. R. Woodock, T. Murphy, P. J. Hemmings, and T. C. Long- worth. Techniques used in the GEM code for Monte Carlo neutronics calculation in reactors and other systems of complex geometry. In Proceedings of the conference on the application of computing methods to reactor proble...

    1965

  49. [60]

    Q. Wu, J. M. Esturo, A. Mirzaei, N. Moënne-Loccoz, and Z. Go- jcic. 3dgut: Enabling distorted cameras and secondary rays in gaussian splatting. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , pp. 26036–26046, 2025. 9

    2025

  50. [2021]

    doi: 10.1145/3478513.3480569 1, 2, 3

  51. [2022]

    doi: 10.1145/3528223.3530158 1

    (*Joint First Authors). doi: 10.1145/3528223.3530158 1

  52. [2023]

    doi: 10.1145/3592433 3, 5, 6, 11

    doi

  53. [2024]

    doi: 10.1109/TVCG.2023.3327107 2, 3

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.