Pith. sign in

REVIEW 3 major objections 5 minor 25 references

Neural Visibility Cache for Real-Time Light Sampling

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read An online-trained neural cache of light visibility enables unbiased real-time light sampling that beats screen-space ReSTIR in occluded scenes.

desk verdict A genuinely new, practical idea: an online hash-grid MLP that caches visibility and feeds it to WRS for unbiased many-light sampling, with results that beat a standard ReSTIR config, but the paper omits the one equation that pins down the unbiasedness claim. read the letter →

arxiv 2506.05930 v2 pith:AHJLBUQM submitted 2025-06-06 cs.GR

Jakub Bokšanský , Daniel Meister This is my paper
classification cs.GR
keywords neuralvisibilitycachereal-timeraytracingdirectilluminationweightedreservoirsamplingReSTIRhash-gridencodingmany-lightrenderingonlinelearning
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

Direct illumination with many lights remains noisy in real-time rendering because visibility between a point and each light is expensive to test. This paper proposes to cache exactly that visibility in a small neural network trained online, one frame at a time, and to feed the predicted per-light visibility into weighted reservoir sampling so the renderer spends its shadow rays on likely visible lights. The reported effect is a lower FLIP error than the standard screen-space ReSTIR baseline at a similar time budget, with the largest improvements in occluded regions, and a clustered variant that extends the approach to scenes with tens of thousands of lights. Because the network only guides which light is sampled and the final shading still uses Monte Carlo integration with clamped positive weights, the method remains unbiased even while the network is still learning, producing extra noise rather than systematic error.

What carries the argument

The central object is the neural visibility cache (NVC): a compact multilayer perceptron implemented so that training and inference run entirely in GPU shared memory, with a multi-resolution hash-grid encoding that maps a 3D position to a vector of predicted visibilities, one output neuron per light or per light cluster. The mechanism that carries the argument is weighted reservoir sampling (WRS) over lights, in which each light's sampling weight is the product of the network's visibility estimate, the BRDF/cosine term approximated by linearly transformed cosines, and the light's radiance. Clamping predicted visibility to a small positive constant (0.001) makes the estimator unbiased at the cost of slightly higher variance. For scenes with many lights, a k-means clustered variant predicts the average visibility of each cluster and uses a two-step scheme combining WRS with resampled importance sampling to pick a cluster and then a light inside it.

What would settle it

Render a scene with fine, high-frequency occlusion (dense foliage or a venetian blind) using the default 8196 training rays and compare the network's per-light visibility predictions to ground-truth shadow rays; if the predicted visibility is wrong at shadow boundaries, the FLIP advantage over ReSTIR in occluded regions should shrink or reverse. A second decisive test is to teleport the camera between rooms while holding the learning rate fixed and measure whether the FLIP spike takes markedly longer to recover than the roughly 16 frames reported for training from scratch.

Watch

Extended reading notes

Core claim

The paper's central claim is that a compact online-trained hash-grid MLP can learn a nonbinary visibility field between 3D positions and light sources well enough to drive weighted reservoir sampling for real-time direct illumination. The network outputs an estimated visibility per light or per light cluster at a query point; these estimates are multiplied by an analytic BRDF/cosine term computed with linearly transformed cosines and by light radiance to form the sampling weights, and the selected light is then evaluated with a shadow ray. In the paper's experiments this neural light sampling reaches a lower FLIP error than screen-space ReSTIR at comparable frame times on scenes with 32 lights (about 20% lower in Kitchen and 45% lower in Sponza), and the clustered version, which learns average visibility per k-means light cluster, reduces ReSTIR's error after disocclusion in scenes with up to about 59K lights. Unbiasedness is preserved by clamping the network's visibility output to a small positive constant, so an undertrained network manifests as increased variance, not bias; a separate biased variant, Neural DI, uses the visibility estimates directly to produce noise-free approximate direct illumination without casting shadow rays for shading.

Load-bearing premise

The method assumes that a few thousand shadow-ray samples per frame (8196 in the default configuration) suffice for a small hash-grid MLP to learn a nonbinary visibility field that generalizes to every shaded pixel, including new camera views and moving geometry; if the network cannot generalize to complex occlusion, the reported noise reduction and fast convergence no longer follow.

Editorial extensions

If this is right

  • On the 32-light test scenes, neural light sampling reports roughly 20% lower FLIP error than screen-space ReSTIR in Kitchen and 45% lower in Sponza at similar per-frame cost.
  • Clustered NVC can replace ReSTIR's initial-candidate generator, and in a simulated full-screen disocclusion on the 59K-light Subway scene it drops FLIP from 0.876 to 0.802; the paper recommends running it only on disoccluded pixels to keep the overhead near 5%.
  • Because the cached quantity is world-space visibility rather than screen-space radiance, the method can follow camera motion, animated geometry and lights, and can in principle sample direct illumination in participating media.
  • Neural DI, the biased variant, produces noise-free images with one sample per pixel and no shadow rays for shading, making it suitable as a fast preview or for deeper bounces in path tracing.

Reading between the lines

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

  • Beyond the paper's reported results, a natural stress test is a sudden insertion of a large occluder between camera and lights; the world-space training data is supposed to pre-train the network for such changes, but the recovery time as a function of learning rate is left uncharacterized.
  • The visibility cache could be reused as a proposal distribution for other Monte Carlo estimators, such as path guiding or bidirectional methods, not just reservoir sampling, since it supplies a cheap, world-space-aware probability over lights at any 3D point.
  • The reported training budget of 8196 binary shadow-ray samples per frame suggests an upper bound on how complex a visibility field can be learned per frame; scenes with very high-frequency occlusion may need either more training samples or a higher-resolution hash grid, which the paper does not quantify.
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 / 5 minor

Summary. The paper proposes a real-time direct light sampling method based on an online-trained neural visibility cache. A hash-grid-encoded MLP outputs per-light (or per-cluster) visibility estimates for a 3D position; these estimates are multiplied by LTC shading and light radiance to form importance weights for weighted reservoir sampling (WRS). A biased variant called Neural DI uses the predicted visibility directly as approximate direct illumination, and a clustered variant supports scenes with many lights by first sampling a cluster and then a light within it. The authors report that neural light sampling achieves lower FLIP than screen-space ReSTIR at comparable frame times on the Kitchen and Sponza scenes, and that clustered NVC improves ReSTIR initial candidates after disocclusions. The core algorithmic claim is that clamping predicted visibilities to a positive constant makes the method unbiased, because predicted visibility is used only to guide sampling.

Significance. If the estimator is stated precisely, this is a practical and timely contribution: it decouples the learned quantity (visibility) from the analytically evaluated radiance, provides an unbiased fallback for real-time light sampling, and integrates naturally with ReSTIR as an initial-candidate generator. The performance breakdown in Table 1 and the convergence progression in Figure 6 are useful evidence that the approach is competitive at 1080p on current GPUs. I see no circularity: the network is trained on binary shadow-ray visibility, not on the comparison metric, and the WRS step is standard importance sampling once positive clamped weights are used. The main weaknesses are that the unbiased estimator is never written down and the clustered two-stage source PDF appears to contain an unjustified factor; both are load-bearing for the paper's central comparisons.

major comments (3)
  1. [Section 3.1, Section 4.1] The unbiasedness argument is not analytically supported as written. No equation defines the sampling probability p(i|x), the normalization of the clamped weights, or the final Monte Carlo estimator used in shading. In particular, the manuscript never states explicitly whether the selected light's contribution is evaluated with an exact shadow ray. If it is not, clamping the predicted visibility cannot remove the bias caused by using an approximate visibility in the integrand; if it is, the estimator is standard importance sampling with positive weights and should be stated as such. The distinction matters for the comparison to ReSTIR and for the contrast with the biased Neural DI variant in Section 3.2. Please add the estimator, including the exact shading integrand (BRDF, cosine, geometry, and binary visibility) and the clamped weight normalization.
  2. [Section 3.6] The source probability density function for the clustered two-step sampler is stated as p(x) = m w(y)/w_sum * 1/m_y. Since the first-stage WRS selects cluster y with probability w(y)/w_sum and the second stage samples uniformly among the m_y lights in that cluster, the unconditional density of generating light x should be w(y)/(w_sum m_y). The factor m is not justified and, as written, p(x) need not be a probability. If the implementation uses the stated formula, the clustered results in Figures 8 and 9 are not supported; if the factor is a typographical artifact, the correct derivation should be given and the RIS weights should be checked.
  3. [Section 3.4 and Figure 3] The text and caption do not specify what operation changes a 'Biased NLS' result into an 'Unbiased NLS' result. The text says clamping to 0.001 yields an unbiased result, but this is only meaningful if the shading pass already evaluates exact visibility for the selected light; otherwise clamping only modifies a biased estimator. Since the final architecture uses a sigmoid output (Section 3.3), negative network outputs cannot occur, so the discussion of clamping 'zero and possibly negative values' in Section 3.1 should be reconciled with the actual output activation.
minor comments (5)
  1. [Section 3.4 and Section 4.4] The number of training points is given as 8196 in two places, but the text states 4096 world-space plus 4096 screen-space samples, i.e., 8192. Please correct the typo.
  2. [Section 3.1] The opening of Section 3.1 says the MLP predicts 'nonbinary' visibility accounting for soft shadows and semitransparent surfaces, while Section 3.4 says each training target is a single binary shadow ray and the network learns the average over the light area. This is a reasonable interpretation, but the text should state explicitly that the network output is an estimate of the average binary visibility over the light surface, not a prediction of the binary visibility of any particular point on the light.
  3. [Section 4.1] The reported Kitchen FLIP reduction is about 17% (0.361 to 0.298), not 20%; please make the percentage consistent with the displayed values.
  4. [Section 4.4, Table 1] The light sampling column mixes very different operations (RIS loop, network inference plus WRS, all ReSTIR passes), which makes direct cross-method timing comparisons hard to interpret; a sentence explaining what is included in each entry would improve reproducibility.
  5. [Section 5] The conclusion says 'with a minor modification, our method provides unbiased estimates' but the modification (clamping and the exact shadow-ray evaluation) should be stated in the conclusion as well; as written it is vague.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the neural visibility output is used only as a positive proposal weight in standard weighted reservoir sampling, so the unbiasedness claim does not reduce to the network's fitted values.

full rationale

The paper's derivation chain is self-contained. The core claim is that a hash-grid MLP trained on binary shadow-ray visibility can supply WRS weights for light selection, and that the resulting estimator remains unbiased when outputs are clamped to a small positive constant. This is standard importance sampling: for any positive weight function w_i, WRS selects light i with probability w_i / sum_j w_j, and evaluating the exact contribution f_i divided by that probability gives an unbiased estimate of the sum of f_i. The network's visibility estimate is thus a proposal distribution input, not a fitted proxy for the reported FLIP error or for the ReSTIR baseline. Clamping changes variance but not expectation, provided the normalized WRS distribution is recomputed from the clamped weights, which is what WRS does by construction. The FLIP comparisons are empirical measurements against ground truth and are not derived from the network's training objective, which is visibility only. The cited building blocks (hash-grid encoding, fully-fused MLP, WRS, LTC shading) are external standard components, and there are no author self-citations used as load-bearing uniqueness or equivalence arguments. The skeptic's observation that the final estimator is never written down in a single equation is a rigor/presentation gap, not circularity: an omitted equation does not make the derivation equivalent to its inputs.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard Monte Carlo theory and on empirically tuned components: a neural network and its training schedule, hash-grid configuration, clamp threshold, and ReSTIR baseline settings. None of these are physical constants; all are hand-chosen or learned and directly affect the reported quality/performance tradeoff. The unbiasedness itself does not depend on any fitted value, which is why the circularity burden is minimal.

free parameters (7)
  • MLP and hash-grid weights = approximately 562K parameters, learned online
    Trained by supervised learning on binary shadow-ray visibility; the central claim that the cache improves sampling depends on this learned fit.
  • Network topology = 2 hidden layers of 32 neurons; 32 output neurons
    Chosen by hand in Section 3.3; the clustered variant uses the same output count over k=32 clusters.
  • Hash-grid configuration = 10 levels, base resolution 16, 4 features per level for NLS; 8 levels, base resolution 2 for clustered NVC
    Tuned per variant in Sections 3.3 and 3.6; the paper states the optimal settings differ for the clustered approach.
  • Learning-rate schedule = 0.05 decaying linearly to 0.001 over the first 200 steps
    Hand-selected in Section 3.4 to speed early training convergence.
  • Training-data budget = 8196 samples per frame (4096 world-space plus 4096 screen-space); 49152 samples for clustered scenes
    Chosen to balance quality and performance in Sections 3.4 and 3.6; the paper reports finding larger budgets necessary for many-light scenes.
  • Visibility clamp epsilon = 0.001
    Introduced in Section 3.1 to guarantee strictly positive WRS weights; the value is hand-chosen and controls the variance versus bias tradeoff.
  • ReSTIR baseline parameters = 8 initial candidates, temporal clamp at 20 times, spatial radius 32 pixels
    Chosen for the comparison in Section 4.1; they affect the reported FLIP differences and the fairness of the central empirical claim.
assumptions (6)
  • standard math WRS with strictly positive weights yields an unbiased Monte Carlo estimator for the target integral.
    Invoked in Section 3.1, where clamping to 0.001 preserves unbiasedness while avoiding zero weights; this is textbook importance sampling.
  • domain assumption A small MLP with 10-level hash encoding can represent the nonbinary visibility field of a scene well enough to guide sampling.
    Sections 3.3 and 3.4 establish capacity empirically, not by proof; the quality of the cache depends on this representational assumption.
  • domain assumption Binary shadow-ray training samples at one random point per light per 3D position yield a good average visibility, penumbra, estimate for area lights.
    Section 3.4 states that 'for area lights, the neural network will eventually learn the average visibility over the whole area of the light'; generalization over finite samples is assumed.
  • domain assumption LTC approximations of BRDF times cosine give weights that, while approximate, keep the sampling estimator unbiased when the final radiance is evaluated exactly.
    Section 3.1 uses LTC from Heitz et al. 2016 to weight WRS; approximate proposal densities do not bias a properly weighted estimator, but the paper does not spell out the exact final evaluation step.
  • domain assumption Combining 4096 world-space and 4096 screen-space training samples fixes dark-blob artifacts and adapts quickly to new views and dynamic changes.
    Sections 3.4 and 3.5 describe this as an empirical finding, not a derived guarantee; the claim that the method handles dynamic scenes rests on it.
  • domain assumption Screen-space ReSTIR with 8 candidates, 20 times temporal clamp, and 32-pixel spatial radius is a representative baseline.
    Section 4.1 defines the baseline configuration; the validity of the central comparison depends on this being a fair or typical ReSTIR setup.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Visibility Cache for Real-Time Light Sampling." pith.science (2026). https://pith.science/paper/AHJLBUQM

@misc{pith2026250605930,
  author       = {Pith},
  title        = {Pith review of: Neural Visibility Cache for Real-Time Light Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AHJLBUQM}},
  note         = {Machine review of arXiv:2506.05930}
}
read the original abstract

Direct illumination with many lights is an inherent component of physically-based rendering, remaining challenging, especially in real-time scenarios. We propose an online-trained neural cache that stores visibility between lights and 3D positions. We feed light visibility to weighted reservoir sampling (WRS) to sample a light source. The cache is implemented as a fully-fused multilayer perceptron (MLP) with multi-resolution hash-grid encoding, enabling online training and efficient inference on modern GPUs in real-time frame rates. The cache can be seamlessly integrated into existing rendering frameworks and can be used in combination with other real-time techniques such as spatiotemporal reservoir sampling (ReSTIR).

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 9 canonical work pages

  1. [7]

    Neural Importance Sampling of Many Lights

    URL: https://arxiv.org/ abs/2505.11729. 3, 4, 15 GUO, J. J., E ISEMANN , M., AND EISEMANN , E. Next event estimation++: Visibility mapping for efficient light transport simulation. Computer Graphics F orum, 39(7):205– 217,

    work page Pith review arXiv

  2. [13]

    3 LLOYD , S

    URL: https://doi.org/10.1145/3665320.3670993. 3 LLOYD , S. Least squares quantization in PCM. IEEE Transactions on Information Theory , 28(2):129–137,

    work page arXiv

  3. [14]

    8 MACQUEEN , J

    URL: https://doi.org/10.1109/TIT.1982.1056489. 8 MACQUEEN , J. Some methods for classification and analysis of multivariate observations. In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, V olume 1: Statistics, pages 281–297. University of California Press,

    arXiv 1982

  4. [17]

    2, 4 MÜLLER , T., G ROSS , M., AND NOVÁK, J

    URL: https://doi.org/10.1145/3528223.3530127. 2, 4 MÜLLER , T., G ROSS , M., AND NOVÁK, J. Practical path guiding for efficient light- transport simulation. Computer Graphics F orum, 36(4):91–100,

    arXiv

  5. [19]

    Importance Sampling of Many Lights with Reinforcement Lightcuts Learning

    URL: https://arxiv.org/abs/ 1911.10217. 3 REN, H., H UO, Y., P ENG , Y., S HENG , H., X UE, W., H UANG , H., L AN, J., W ANG , R., AND BAO, H. LightFormer: Light-oriented global neural rendering in dynamic scene. ACM Transactions on Graphics , 43(4):75:1–75:14, July

    work page Pith review arXiv 1911

  6. [20]

    org/10.1145/3658229

    URL: https://doi. org/10.1145/3658229. 3 17 Journal of Computer Graphics Techniques Neural Visibility Cache for Real-Time Light Sampling V ol. 14, No. 2, 2025 http://jcgt.org SHIRLEY , P., WANG , C., AND ZIMMERMAN , K. Monte Carlo techniques for direct lighting calculations. ACM Transactions on Graphics , 15(1):1–36, January

    doi:10.1145/3658229 2025

  7. [22]

    15 TOKUYOSHI , Y., I KEDA , S., K ULKARNI , P., AND HARADA , T

    URL: http://jcgt.org/ published/0005/01/02/. 15 TOKUYOSHI , Y., I KEDA , S., K ULKARNI , P., AND HARADA , T. Hierarchical light sampling with accurate spherical gaussian lighting. In SIGGRAPH Asia 2024 Conference Papers , pages 82:1–82:11. Association for Computing Machinery,

    work page 2024

  8. [23]

    org/10.1145/3680528.3687647

    URL: https://doi. org/10.1145/3680528.3687647. 3 VÉVODA , P., K ONDAPANENI , I., AND K ˇRIVÁNEK , J. Bayesian online regression for adap- tive direct illumination sampling. ACM Transactions on Graphics , 37(4):125:1–125:12, July

    arXiv

Show all 25 references
  1. [24]

    3, 15 VORBA , J., K ARLÍK , O., Š IK, M., R ITSCHEL , T., AND K ˇRIVÁNEK , J

    URL: https://doi.org/10.1145/3197517.3201340. 3, 15 VORBA , J., K ARLÍK , O., Š IK, M., R ITSCHEL , T., AND K ˇRIVÁNEK , J. On-line learning of parametric mixture models for light transport simulation. ACM Transactions on Graph- ics, 33(4):101:1–101:11, July

  2. [25]

    URL:https://doi.org/10.1145/2601097. 2601203. 3 WALTER , B., F ERNANDEZ , S., A RBREE , A., B ALA , K., D ONIKIAN , M., AND GREEN - BERG , D. P. Lightcuts: A scalable approach to illumination. ACM Transactions on Graph- ics, 24(3):1098–1107, July

    doi

  3. [26]

    URL: https://doi.org/10.1145/1073204. 1073318. 3 WARD , G. J. Adaptive shadow testing for ray tracing. In Photorealistic Render- ing in Computer Graphics: Proceedings of the Second Eurographics Workshop on Rendering, pages 11–20. Springer,

    doi

  4. [27]

    URL: https://doi.org/10.1007/978-1-4842-7185-8_22 . 2, 4 Index of Supplemental Materials • Video of Sponza walkthrough jcgt.org/published/0014/02/01/NVC_Sponza_Walkthrough.mp4 • Video of Sponza with dynamic lighting jcgt.org/published/0014/02/01/NVC_Sponza_Dynamic_Light.avi 18...

    doi

  5. [1982]

    2, 4 CONTY ESTEVEZ , A

    URL: https://doi.org/10.1093/biomet/69.3.653. 2, 4 CONTY ESTEVEZ , A. AND KULLA , C. Importance sampling of many lights with adaptive tree splitting. Proceedings of the ACM on Computer Graphics and Interactive Techniques , 1(2):25:1–25:17, August

    doi

  6. [1996]

    15 TALBOT , J

    URL: https: //doi.org/10.1145/226150.226151. 15 TALBOT , J. F., C LINE , D., AND EGBERT , P. Importance resampling for global illumination. In Proceedings of the Sixteenth Eurographics Conference on Rendering Techniques, EGSR ’05, page 139–146. Eurographics Association,

  7. [2005]

    URL: https://doi.org/10.1109/TVCG. 2005.83. 3 LI, Y. K., Z HU, C., N ICHOLS , G., K UTZ , P., H UANG , W.-F. W., A DLER , D., B UR- LEY, B., AND TEECE , D. Cache points for production-scale occlusion-aware many-lights sampling and volumetric. In DigiPro ’24: Proceedings of the...

    2005 doi

  8. [2007]

    3 HE, K., Z HANG , X., R EN, S., AND SUN, J

    URL: https: //doi.org/10.1145/1276377.1276410. 3 HE, K., Z HANG , X., R EN, S., AND SUN, J. Delving deep into rectifiers: Surpassing human- level performance on ImageNet classification. In 2015 IEEE International Conference on Computer Vision (ICCV) , pages 1026–1034. IEEE,

    2015

  9. [2014]

    3 DATTA, S., N OWROUZEZAHRAI , D., S CHIED , C., AND DONG , Z

    URL: https://doi.org/10.1111/cgf.12256. 3 DATTA, S., N OWROUZEZAHRAI , D., S CHIED , C., AND DONG , Z. Neural shadow map- ping. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH ’22, pages 8:1–8:9. Association for Computing Machinery,

    2022 doi

  10. [2015]

    7 HEITZ , E., D UPUY, J., H ILL , S., AND NEUBELT , D

    URL: https://doi.org/ 10.1109/ICCV.2015.123. 7 HEITZ , E., D UPUY, J., H ILL , S., AND NEUBELT , D. Real-time polygonal-light shading with linearly transformed cosines. ACM Transactions on Graphics, 35(4):41:1–41:8, July

    2015 doi

  11. [2016]

    4 HUANG , J., I IZUKA , A., T ANAKA , H., K OMURA , T., AND KITAMURA , Y

    URL: https://doi.org/10.1145/2897824.2925895. 4 HUANG , J., I IZUKA , A., T ANAKA , H., K OMURA , T., AND KITAMURA , Y. Online neu- ral path guiding with normalized anisotropic spherical gaussians. ACM Transactions on Graphics , 43(3):26:1–26:18, April

  12. [2018]

    3 DACHSBACHER , C., K ˇRIVÁNEK , J., H AŠAN , M., A RBREE , A., WALTER , B., AND NOVÁK, J

    URL: https://doi.org/10.1145/3233305. 3 DACHSBACHER , C., K ˇRIVÁNEK , J., H AŠAN , M., A RBREE , A., WALTER , B., AND NOVÁK, J. Scalable realistic rendering with many-light methods. Computer Graphics F orum, 33 (1):88–104, February

    doi

  13. [2019]

    3 MÜLLER , T., R OUSSELLE , F., N OVÁK, J., AND KELLER , A

    URL: https://doi.org/10.1145/3341156. 3 MÜLLER , T., R OUSSELLE , F., N OVÁK, J., AND KELLER , A. Real-time neural radiance caching for path tracing. ACM Transactions on Graphics, 40(4), July

    doi

  14. [2020]

    2, 3, 8, 9 15 Journal of Computer Graphics Techniques Neural Visibility Cache for Real-Time Light Sampling V ol

    URL: https:// doi.org/10.1145/3386569.3392481. 2, 3, 8, 9 15 Journal of Computer Graphics Techniques Neural Visibility Cache for Real-Time Light Sampling V ol. 14, No. 2, 2025 http://jcgt.org CHAO, M. T. A general purpose unequal probability sampling plan. Biometrika, 69(3):65...

    2025

  15. [2021]

    2, 3, 9, 11 MÜLLER , T., E VANS, A., S CHIED , C., AND KELLER , A

    URL: https: //doi.org/10.1145/3450626.3459812. 2, 3, 9, 11 MÜLLER , T., E VANS, A., S CHIED , C., AND KELLER , A. Instant neural graphics primi- tives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):102:1– 102:15, July

  16. [2022]

    3 DEREVIANNYKH , M., K LEPIKOV , D., H ANIKA , J., AND DACHSBACHER , C

    URL: https://doi.org/10.1145/ 3528233.3530700. 3 DEREVIANNYKH , M., K LEPIKOV , D., H ANIKA , J., AND DACHSBACHER , C. Neu- ral two-level Monte Carlo real-time rendering. Computer Graphics F orum , 44(2): e70050,

  17. [2024]

    activision.com/publications/2024/08/Neural_Light_Grid

    URL: https://research. activision.com/publications/2024/08/Neural_Light_Grid. 3 16 Journal of Computer Graphics Techniques Neural Visibility Cache for Real-Time Light Sampling V ol. 14, No. 2, 2025 http://jcgt.org KINGMA , D. P. AND BA, J. Adam: A method for stochastic optimization,

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.