Pith. sign in

REVIEW 3 major objections 5 minor 54 references

Neural Path Guiding with Distribution Factorization

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 The paper claims that factoring a 2D directional PDF into two interpolated 1D PDFs and training against a cached-radiance target yields a neural guiding distribution that is both expressive and fast, beating prior methods in equal-time…

desk verdict A solid, incremental path guiding paper with a clever factorization and radiance-cached training, but missing error bars and code. read the letter →

arxiv 2506.00839 v2 pith:HRYOLIT2 submitted 2025-06-01 cs.GR cs.CVcs.LG

Pedro Figueiredo , Qihao He , Nima Khademi Kalantari This is my paper
classification cs.GRcs.CVcs.LG
keywords pathguidingMonteCarlorenderingneuralimportancesamplingdistributionfactorizationmarginalandconditionalPDFslinearinterpolationradiancecachingequal-timecomparison
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

Path guiding in Monte Carlo rendering needs a sampling distribution that is expressive enough to match complex lighting and cheap enough to evaluate and sample at every bounce. This paper argues that a full 2D directional distribution is unnecessary: factoring it into one 1D marginal and one 1D conditional distribution, each predicted by a small MLP over discrete bins and reconstructed by interpolation, gives both expressiveness and speed. Training minimizes a KL-divergence loss, and a cached radiance network supplies both the incoming radiance and the normalization factor, so the target distribution is far less noisy than raw Monte Carlo estimates. If the claim holds, the method provides a fast, continuous guiding distribution that outperforms existing neural and classic guides on complex light transport, with the linear-interpolation variant reported as best on all seven test scenes and the nearest-neighbor variant second.

What carries the argument

The load-bearing object is the factorization identity $\hat p_\Theta(\epsilon_1,\epsilon_2|x,\omega_o)=\hat p_{w_1}(\epsilon_1|x,\omega_o)\,\hat p_{w_2}(\epsilon_2|\epsilon_1,x,\omega_o)$ in uniform square coordinates $(\epsilon_1,\epsilon_2)\in[0,1]^2$, where the first factor is a marginal PDF over the azimuth-like coordinate and the second is a conditional PDF over the elevation-like coordinate. Two MLPs with softmax outputs predict $M_1$ and $M_2$ discrete PDF values; the continuous PDF is recovered by interpolation, with care so the interpolated function integrates to 1, and sampling runs through the inverse CDF. Around this sits the radiance-caching network $f_\Phi$, which estimates reflected radiance along a ray; evaluating it at the current and next intersection points produces the two radiance terms in the training target (Eq. 14), reducing gradient noise and supplying the normalization factor that prior methods drop.

What would settle it

Render a scene whose light transport is known but whose cache is deliberately corrupted, for example by freezing $f_\Phi$ early or adding controlled noise, and compare guiding quality with the uncorrupted cache; if relative mean-squared error does not degrade, Eq. 14's dependence on cached radiance is not the cause of the improvement. Alternatively, on the swimming-pool scene, compare the method at $32\times 16$ and $64\times 32$ resolutions: if doubling the resolution does not reduce the error, the fixed-resolution factorization is not the limiting factor.

Watch

Extended reading notes

Core claim

The paper's central claim is that the joint directional PDF over the hemisphere can be written as the product of a marginal and a conditional 1D PDF in uniform square coordinates, and that this factorization is a better building block for online neural path guiding than normalizing flows or mixture models. Each 1D factor is a tiny MLP that outputs a softmax-normalized vector of PDF values at uniformly spaced coordinates; arbitrary query points are evaluated by nearest-neighbor or linear interpolation, and sampling is done by inverse CDF transform. To train the two networks, the paper minimizes a Monte Carlo estimate of the KL divergence between the guiding distribution and a target built from the BSDF, the cosine term, and cached radiance: the cache network $f_\Phi$ provides both the incoming radiance $L_i(x,\omega_i)$ at the next intersection and the reflected radiance $L_r(x,\omega_o)$ used as the normalization factor (Eq. 14). The paper reports that this combination produces the lowest equal-time relative mean-squared error on all seven complex scenes tested, with the linear-interpolation variant ahead of the nearest-neighbor variant.

Load-bearing premise

Everything rests on the radiance-cache network estimating the true light arriving at each point accurately enough; if that estimate is biased or noisy, the guiding network is trained to match the wrong target distribution.

Editorial extensions

If this is right

  • The marginal-conditional representation can be evaluated and sampled with two parallel network evaluations, avoiding the sequential coupling-layer cost of normalizing-flow guides and making online training faster.
  • Radiance caching supplies both terms of the target distribution, so training gradients are smoother and the normalization factor is no longer silently discarded; the paper credits this with the largest gains on scenes with complex light transport.
  • In equal-time comparisons on seven complex scenes, the linear-interpolation variant achieves the lowest relative mean-squared error and the nearest-neighbor variant is second, ahead of the classic guides PPG and Variance and the neural guides NIS and NPM.
  • The representation's fixed discretization is its main constraint: features smaller than one bin, such as a sun disk, are blurred, and raising the resolution from $32\times 16$ to $64\times 32$ recovers quality at added computational cost.
  • On simple scenes where many samples can be thrown cheaply, the method loses its advantage because its per-sample cost is higher than that of simpler guides.

Reading between the lines

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

  • Beyond the paper's own claims, the same marginal-conditional chain extends naturally to product sampling in higher dimensions, with network cost growing linearly in the number of dimensions rather than exponentially; this is a direct extrapolation of the 2D construction.
  • A possible extension of the paper's actor-critic analogy is to treat the cached radiance as a critic baseline and replace the KL loss with a variance-reduced policy-gradient objective, which the paper mentions only as an analogy and does not test.
  • A testable prediction: combining the factorization with adaptive spatial resolution or variable bin sizes, which the paper itself proposes for future work, should close most of the gap on strongly directional light sources like the swimming-pool sun.
  • Because the target distribution is non-negative and normalized by construction, the method is naturally biased toward stable training; however, the same property means any error in the cache shifts the learned guide systematically, so the practical gains hinge on the cache's accuracy.
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 neural path guiding method in which the 2D directional PDF is factorized into a marginal and a conditional 1D PDF, each represented by an MLP that outputs PDF values at discrete grid points; continuous evaluation and sampling are performed via nearest-neighbor or linear interpolation. The guiding networks are trained by minimizing KL divergence using a target distribution (Eq. 14) that replaces incoming radiance with a neural radiance cache and normalizes by the cached reflected radiance. Experiments on seven scenes compare equal-time relMSE against PT, PPG, Variance, NIS, and NPM, reporting DF-L as best in all tested scenes, with ablations on discretization resolution and radiance caching.

Significance. The factorization idea is simple and principled: the normalized construction via softmax and the explicit constraints in Eqs. 5 and 9 are correct, and the parallel evaluation of the two 1D networks is a genuine efficiency advantage over sequential normalizing-flow evaluation. The use of radiance caching to reduce gradient variance and to estimate the normalization factor is an interesting and potentially practical contribution. If the empirical claims hold, this would be a useful step for neural path guiding in complex indoor scenes. The paper also deserves credit for openly discussing limitations, including the SWIMMINGPOOL and CORNELLBOX failures. However, the central comparison rests on a training target that depends entirely on an unvalidated neural cache, and the quantitative evidence lacks variance estimates, so the headline claim of universal superiority is not yet fully supported.

major comments (3)
  1. [Sec. 3.2, Eq. (14)] The training objective replaces the true target p(ω_i) with the estimate ρ(x,ω_o,ω_i) f_Φ(x',ω'_o)|cos θ_i| f_Φ(x,ω_o)^{-1}, where f_Φ is a neural radiance cache. The paper itself states in Sec. 3.2 that 'theoretically, caching radiance and using it as L_i should not help with variance reduction; the variance will just be passed from here to radiance caching optimization.' Since the reported gains all come from this cache-augmented objective, a biased or too-noisy cache would train the guiding PDFs toward a systematically wrong target. The only supporting evidence is a one-scene ablation (Fig. 9) and an appeal to NRC's relative L2 loss. Please add a quantitative evaluation of the cache's accuracy (e.g., error of f_Φ against a reference radiance solution on at least one scene), report the cache ablation on all scenes, and discuss how cache bias or noise propagates through Eq. 14 to the learned PDFs.
  2. [Sec. 5.2, Table 1] The central claim that DF-L 'produces the best results in all the scenes' is based on relMSE averaged over 10 runs, but no standard deviations, per-run values, or significance tests are reported. Equal-time comparisons are noisy, and some of the reported differences are small (e.g., BATHROOM: DF-N 0.2601 vs NIS 0.2738). Without a measure of run-to-run variation, the claim that the method is best in all scenes is not statistically supported. Please report standard deviations or individual run results, and state whether the ranking is consistent across runs.
  3. [Sec. 6, Figs. 10-11] The paper's own results show that on SWIMMINGPOOL the proposed method (DF-L 32x16, relMSE 0.0948) is substantially worse than PPG (0.0221) and Variance (0.0128), and on CORNELLBOX with flipped light NPM produces less noise than DF-L despite lower-quality learned PDFs. These failures are directly relevant to the abstract's unqualified claim that the approach 'is better than the existing methods.' The conclusion acknowledges these limitations, but the abstract and the 'best in all scenes' statements in Sec. 5.2 should be scoped to the tested indoor scenes with complex indirect transport, or supplemented with an explicit discussion of why these counterexamples do not affect the main claim.
minor comments (5)
  1. [Fig. 3 caption] The caption contains a typo: 'Note that the we usex and ω_o' should read 'Note that we use x and ω_o'.
  2. [Eq. (9)] The summation in Eq. (9) writes v[i+1] for i up to M-1, which is out of range; please clarify the boundary handling (nearest-neighbor for ε2 at the edges) so that the summation is unambiguous.
  3. [Table 1 vs Fig. 6] The spp values for the PT row on BREAKFAST differ between Table 1 (6675 spp) and Fig. 6 (4582 spp); please make the numbers consistent.
  4. [Acknowledgments] The scene name 'SALLE DEBAIN' should be written as 'SALLE DE BAIN' for consistency with Table 1 and Section 5.1.
  5. [Sec. 3.2, Eq. (12)] The sampling distribution q(ω_i) is used in Eq. (12) but is not defined before the equation; please state explicitly that q is the combined BSDF/guiding sampling distribution used to draw the MC samples.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed derivation is self-contained, and the cache-based training target is a standard self-supervised approximation, not a definitional reduction.

full rationale

The paper's derivation chain is not circular. The factorization in Eq. 3 is the exact product rule p(ε1,ε2)=p(ε1)p(ε2|ε1), and the network parameterization in Eq. 6, softmax followed by scaling by M, is an explicit construction that makes each 1D PDF integrate to one (Eqs. 5 and 9). The KL objective (Eqs. 10–12) is a standard divergence minimization whose target p(ω) is the normalized radiance integrand, not a function of the guiding parameters Θ being optimized. Eq. 14 replaces the unknown radiance terms with estimates from the cache network f_Φ; although f_Φ is trained online from paths generated with the current guide, the target distribution is not defined in terms of the guiding network parameters w1 and w2, and the guiding network is a separate function approximator. This is a self-supervised / actor-critic style loop, not a reduction of the claimed result to its inputs. The paper even flags the theoretical caveat that caching radiance 'should not help with variance reduction' (Sec. 3.2), showing the empirical comparison is not presented as a forced identity. No load-bearing self-citation chain exists: the radiance-caching component is attributed to the external NRC work [MRNK21], and the benchmark comparisons are against external methods. Any concern that a biased cache could train the guide toward a wrong target is a correctness and robustness risk, not circularity.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard Monte Carlo and rendering equations plus two empirical assumptions: the accuracy of the learned radiance cache and the sufficiency of the fixed discretization. These assumptions are load-bearing because they determine the target and capacity of the guiding distribution.

free parameters (8)
  • discretization resolution M1 = 32
    Chosen via ablation (Sec 5.3); higher resolutions improve quality but cost more, so 32x16 chosen as a trade-off.
  • discretization resolution M2 = 16
    Half of M1 because theta ranges 0 to pi; chosen in same ablation.
  • learning rate for f_Phi (radiance cache) = 1e-2
    Set in Sec 4; not derived from theory.
  • learning rate for PDF networks f_w1, f_w2 = 3e-2
    Set in Sec 4; authors note it is stable thanks to cached objective.
  • guided path ratio = 70%
    70% of paths use guiding, 30% use BSDF sampling for exploration (Sec 4).
  • training budget fraction = 30%
    Optimization runs for first 30% of time budget (Sec 4).
  • network architecture = 3 hidden layers of 64 neurons, ReLU
    Shared across all networks (Sec 4).
  • input encodings = dense grid encoding, SH degree 4, one-blob with 4 bins, triangle wave with 12 frequencies
    Input encodings chosen from prior work (Sec 4).
assumptions (6)
  • domain assumption The rendering equation (Eq. 1) accurately models outgoing radiance as emitted plus integrated incident radiance times BSDF and cosine.
    Used as the starting point for defining the target distribution for path guiding; if the physical model is wrong, the method optimizes the wrong objective.
  • standard math The product rule p(epsilon1, epsilon2) = p(epsilon1) p(epsilon2|epsilon1) holds exactly for any joint distribution.
    Eq. 3; this is a mathematical identity.
  • standard math Minimizing KL divergence to the normalized integrand yields a good importance sampling distribution.
    Eq. 10; standard in MC and learning literature.
  • standard math The MC gradient estimate in Eq. 12, using samples from q, is an unbiased (or sufficiently low-variance) estimator of the KL gradient.
    Assumed in the optimization; standard when q has a PDF.
  • ad hoc to paper The radiance caching network f_Phi provides accurate estimates of incoming and reflected radiance for the target distribution.
    Eq. 14; the method's effectiveness depends on this empirical assumption, which the paper acknowledges is not theoretically guaranteed.
  • ad hoc to paper The fixed grid resolution (32x16) is fine enough to capture the important features of the target distributions in the tested scenes.
    The representation capacity is limited; the paper shows a counterexample (SWIMMINGPOOL) where the resolution is too coarse unless increased.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Path Guiding with Distribution Factorization." pith.science (2026). https://pith.science/paper/HRYOLIT2

@misc{pith2026250600839,
  author       = {Pith},
  title        = {Pith review of: Neural Path Guiding with Distribution Factorization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HRYOLIT2}},
  note         = {Machine review of arXiv:2506.00839}
}
read the original abstract

In this paper, we present a neural path guiding method to aid with Monte Carlo (MC) integration in rendering. Existing neural methods utilize distribution representations that are either fast or expressive, but not both. We propose a simple, but effective, representation that is sufficiently expressive and reasonably fast. Specifically, we break down the 2D distribution over the directional domain into two 1D probability distribution functions (PDF). We propose to model each 1D PDF using a neural network that estimates the distribution at a set of discrete coordinates. The PDF at an arbitrary location can then be evaluated and sampled through interpolation. To train the network, we maximize the similarity of the learned and target distributions. To reduce the variance of the gradient during optimizations and estimate the normalization factor, we propose to cache the incoming radiance using an additional network. Through extensive experiments, we demonstrate that our approach is better than the existing methods, particularly in challenging scenes with complex light transport.

Figures

Figures reproduced from arXiv: 2506.00839 by the authors.

Figure 1
Figure 1. Method Overview. During path tracing, we generate samples using our guiding distribution pΘ to increase the num￾ber of paths that reach light sources. These paths are then used to train our radiance caching fΦ. We leverage the cached radiance as a smoother objective to improve pΘ, which in turn is used on the next sample generation. in an offline manner. These methods train a CNN on a large number of scenes and use … view at source ↗
Figure 2
Figure 2. We demonstrate the PDF evaluation and sampling pro￾cess for nearest neighbor and linear interpolation. To model the 1D distributions (marginal and conditional in Eq. 3) our network first predicts a vector v containing estimates of the PDF at discrete lo￾cations. To obtain the PDF at an arbitrary location, we either use the PDF estimate at the closest sample (top-left), or linearly in￾terpolate between the two closes… view at source ↗
Figure 3
Figure 3. Our multilayer perceptron (MLP) takes the condition C as the input and estimates an M dimensional vector. We then apply softmax function to this vector and multiply each element by M to obtain a vector containing the PDF estimate at discrete locations. We use this network to model the marginal and conditional distri￾butions in Eq. 3. Note that the we use x and ωo as the condition when modeling the marginal distribut… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Computing the target distribution requires obtaining Lr(x,ωo) and Li(x,ωi) (see Eq. 13). We use a neural network that takes location and direction as the input and estimates the cached reflected radiance along that particular ray. By evaluating the net￾work at the curr…
Figure 5
Figure 5. Figure 5: We show convergence plot of all the approaches on the seven scenes from 12 to 120 seconds (s). The 36 s line is where all the approaches (except the unidirectional path tracer) stop learning and use the learned distributions to sample the remaining paths. al. [RGH∗ 20]…
Figure 6
Figure 6. Figure 6: Equal-time comparison against Müller et al. [MGN17] (PPG), Rath et al. [RGH∗ 20] (Variance), Müller et al. [MMR∗ 19] (NIS), Dong et al. [DWL23] (NPM), and our method with nearest neighbor (DF-N) and linear (DF-L) interpolation. The time budget for all the methods is 12…
Figure 7
Figure 7. Figure 7: Equal-time comparison of learned distributions for the VEACH DOOR scene. The time budget for all methods is 120s. 6. Conclusion, Limitations, and Future Work In this paper, we have presented a novel neural path guiding method. We first factorize the distribution over t…
Figure 9
Figure 9. Figure 9: Effect of radiance caching during equal-time compari￾son of 120s on the VEACH DOOR. We compare the results of our approach (Lr +Li) against not using the cache (similar to existing methods) as well as using the cache only for Li in Eq. 14. Using the cache for both Li a…
Figure 10
Figure 10. Figure 10: Equal-time comparison (120s) on the SWIMMING POOL scene. The pool shows direct lighting from an environment map that models the Sun forming caustics in an outdoor swimming pool. Our approach struggles in this case as it does not sharply encode the directional light fr…
Figure 11
Figure 11. Figure 11: Equal-time comparison (30 seconds budget) on the CORNELL BOX with flipped light. Although DF-N and DF-L en￾code higher quality PDFs, Dong et al. [DWL23] (NPM) generate the result with least noise because they trace more samples. proposed by Müller et al. [MMR∗ 19], to…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 52 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry.original add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 i...

  3. [3]

    C., Anderson J

    Budge B. C., Anderson J. C., Joy K. I. : Caustic forecasting: Unbiased estimation of caustic lighting for global illumination. Computer Graphics Forum 27, 7 (2008), 1963--1970

    work page 2008

  4. [4]

    : Double hierarchies for directional importance sampling in monte carlo rendering

    Bus N., Boubekeur T. : Double hierarchies for directional importance sampling in monte carlo rendering. Journal of Computer Graphics Techniques (JCGT) 6, 3 (August 2017), 25--37

    work page 2017

  5. [5]

    : Fast path space filtering by jittered spatial hashing

    Binder N., Fricke S., Keller A. : Fast path space filtering by jittered spatial hashing. In ACM SIGGRAPH 2018 Talks (New York, NY, USA, 2018), SIGGRAPH '18, Association for Computing Machinery

    work page 2018

  6. [6]

    : Offline deep importance sampling for monte carlo path tracing

    Bako S., Meyer M., DeRose T., Sen P. : Offline deep importance sampling for monte carlo path tracing. Computer Graphics Forum 38, 7 (2019), 527--542

    work page 2019

  7. [7]

    : A significance cache for accelerating global illumination

    Bashford-Rogers T., Debattista K., Chalmers A. : A significance cache for accelerating global illumination. Computer Graphics Forum 31, 6 (2012), 1837--1851

    work page 2012

  8. [8]

    G., Sutton R

    Barto A. G., Sutton R. S., Anderson C. W. : Neuronlike adaptive elements that can solve difficult learning control problems. IEEE Transactions on Systems, Man, and Cybernetics SMC-13, 5 (1983), 834--846

    work page 1983

Show all 54 references
  1. [9]

    A., Belcour L., Nowrouzezahrai D

    Dubouchet R. A., Belcour L., Nowrouzezahrai D. : Frequency based radiance cache for rendering animations. In Proceedings of the Eurographics Symposium on Rendering: Experimental Ideas & Implementations (Goslar, DEU, 2017), EGSR '17, Eurographics Association, p. 41–53

    2017

  2. [10]

    : Learning light transport the reinforced way

    Dahm K., Keller A. : Learning light transport the reinforced way. In ACM SIGGRAPH 2017 Talks (New York, NY, USA, 2017), SIGGRAPH '17, Association for Computing Machinery

    2017

  3. [11]

    O ztireli C., M \

    Dodik A., Papas M., \" O ztireli C., M \" u ller T. : Path guiding using spatio-directional mixture models. Computer Graphics Forum 41, 1 (2022), 172--189

    2022

  4. [12]

    : Density estimation using real NVP

    Dinh L., Sohl - Dickstein J., Bengio S. : Density estimation using real NVP . In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings (2017), OpenReview.net

    2017

  5. [13]

    : Neural parametric mixtures for path guiding

    Dong H., Wang G., Li S. : Neural parametric mixtures for path guiding. In ACM SIGGRAPH 2023 Conference Proceedings (New York, NY, USA, 2023), SIGGRAPH '23, Association for Computing Machinery

    2023

  6. [14]

    : Manifold path guiding for importance sampling specular chains

    Fan Z., Hong P., Guo J., Zou C., Guo Y., Yan L.-Q. : Manifold path guiding for importance sampling specular chains. ACM Trans. Graph. 42, 6 (dec 2023)

    2023

  7. [15]

    J., Bauszat P., Bikker J., Eisemann E

    Guo J. J., Bauszat P., Bikker J., Eisemann E. : Primary sample space path guiding. In Proceedings of the Eurographics Symposium on Rendering: Experimental Ideas & Implementations (Goslar, DEU, 2018), SR '18, Eurographics Association, p. 73–82

    2018

  8. [16]

    : Spatial directional radiance caching

    Gassenbauer V., Křivánek J., Bouatouch K. : Spatial directional radiance caching. Computer Graphics Forum 28, 4 (2009), 1189--1198

    2009

  9. [17]

    : Product importance sampling for light transport path guiding

    Herholz S., Elek O., Vorba J., Lensch H., Křivánek J. : Product importance sampling for light transport path guiding. Computer Graphics Forum 35, 4 (2016), 67--77

    2016

  10. [18]

    : Online neural path guiding with normalized anisotropic spherical gaussians

    Huang J., Iizuka A., Tanaka H., Komura T., Kitamura Y. : Online neural path guiding with normalized anisotropic spherical gaussians. ACM Trans. Graph. 43, 3 (Apr. 2024)

    2024

  11. [19]

    : Importance sampling with hemispherical particle footprints

    Hey H., Purgathofer W. : Importance sampling with hemispherical particle footprints. In Proceedings of the 18th Spring Conference on Computer Graphics (New York, NY, USA, 2002), SCCG '02, Association for Computing Machinery, p. 107–114

    2002

  12. [20]

    : Adaptive incident radiance field sampling and reconstruction using deep reinforcement learning

    Huo Y., Wang R., Zheng R., Xu H., Bao H., Yoon S.-E. : Adaptive incident radiance field sampling and reconstruction using deep reinforcement learning. ACM Trans. Graph. 39, 1 (jan 2020)

    2020

  13. [21]

    Herholz S., Zhao Y., Elek O., Nowrouzezahrai D., Lensch H. P. A., K r iv\' a nek J. : Volume path guiding based on zero-variance random walk theory. ACM Trans. Graph. 38, 3 (jun 2019)

    2019

  14. [22]

    Jensen H. W. : Importance driven path tracing using the photon map. In Rendering Techniques '95 (Vienna, 1995), Hanrahan P. M., Purgathofer W., (Eds.), Springer Vienna, pp. 326--335

    1995

  15. [23]

    Kajiya J. T. : The rendering equation. 143–150

  16. [24]

    : Adam: A method for stochastic optimization

    Kingma D., Ba J. : Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR) (2015)

    2015

  17. [25]

    : Neural product importance sampling via warp composition

    Litalien J., Ha s an M., Luan F., Mullia K., Georgiev I. : Neural product importance sampling via warp composition. In SIGGRAPH Asia 2024 Conference Papers (New York, NY, USA, 2024), SA '24, Association for Computing Machinery

    2024

  18. [26]

    P., Willems Y

    Lafortune E. P., Willems Y. D. : A 5d tree to reduce the variance of monte carlo ray tracing. In Rendering Techniques '95 (Vienna, 1995), Hanrahan P. M., Purgathofer W., (Eds.), Springer Vienna, pp. 11--20

    1995

  19. [27]

    : Unbiased caustics rendering guided by representative specular paths

    Li H., Wang B., Tu C., Xu K., Holzschuch N., Yan L.-Q. : Unbiased caustics rendering guided by representative specular paths. In SIGGRAPH Asia 2022 Conference Papers (New York, NY, USA, 2022), SA '22, Association for Computing Machinery

    2022

  20. [28]

    : tiny-cuda-nn , 4 2021

    M\"uller T. : tiny-cuda-nn , 4 2021. URL: https://github.com/NVlabs/tiny-cuda-nn

    2021

  21. [29]

    : Practical path guiding for efficient light-transport simulation

    Müller T., Gross M., Novák J. : Practical path guiding for efficient light-transport simulation. Computer Graphics Forum 36, 4 (2017), 91--100

    2017

  22. [30]

    : Second-order occlusion-aware volumetric radiance caching

    Marco J., Jarabo A., Jarosz W., Gutierrez D. : Second-order occlusion-aware volumetric radiance caching. ACM Trans. Graph. 37, 2 (jul 2018)

    2018

  23. [31]

    : Neural importance sampling

    M\" u ller T., Mcwilliams B., Rousselle F., Gross M., Nov\' a k J. : Neural importance sampling. ACM Trans. Graph. 38, 5 (oct 2019)

    2019

  24. [32]

    : Real-time neural radiance caching for path tracing

    M\" u ller T., Rousselle F., Nov\' a k J., Keller A. : Real-time neural radiance caching for path tracing. ACM Trans. Graph. 40, 4 (jul 2021)

    2021

  25. [33]

    : Online path sampling control with progressive spatio-temporal filtering, 2020

    Pantaleoni J. : Online path sampling control with progressive spatio-temporal filtering, 2020. http://arxiv.org/abs/2005.07547 arXiv:2005.07547

    2020 arXiv

  26. [34]

    Pegoraro V., Wald I., Parker S. G. : Sequential monte carlo adaptation in low-anisotropy participating media. Computer Graphics Forum 27, 4 (2008), 1097--1104

    2008

  27. [35]

    : Variance-aware path guiding

    Rath A., Grittmann P., Herholz S., V\' e voda P., Slusallek P., K r iv\' a nek J. : Variance-aware path guiding. ACM Trans. Graph. 39, 4 (aug 2020)

    2020

  28. [36]

    Ruppert L., Herholz S., Lensch H. P. A. : Robust fitting of parallax-aware mixtures for path guiding. ACM Trans. Graph. 39, 4 (aug 2020)

    2020

  29. [37]

    : Adaptive sampling and reconstruction using greedy error minimization

    Rousselle F., Knaus C., Zwicker M. : Adaptive sampling and reconstruction using greedy error minimization. ACM Transactions on Graphics (TOG) 30, 6 (2011), 1--12

    2011

  30. [38]

    : Clustered pre-convolved radiance caching

    Rehfeld H., Zirr T., Dachsbacher C. : Clustered pre-convolved radiance caching. In Proceedings of the 14th Eurographics Symposium on Parallel Graphics and Visualization (Goslar, DEU, 2014), PGV '14, Eurographics Association, p. 25–32

    2014

  31. [39]

    S., Barto A

    Sutton R. S., Barto A. G. : Reinforcement Learning: An Introduction, second ed. The MIT Press, 2018

    2018

  32. [40]

    : Global importance sampling of glossy surfaces using the photon map

    Steinhurst J., Lastra A. : Global importance sampling of glossy surfaces using the photon map. 2006 IEEE Symposium on Interactive Ray Tracing (2006), 133--138

    2006

  33. [41]

    : Real-time global illumination by precomputed local reconstruction from sparse radiance probes

    Silvennoinen A., Lehtinen J. : Real-time global illumination by precomputed local reconstruction from sparse radiance probes. ACM Trans. Graph. 36, 6 (nov 2017)

    2017

  34. [42]

    H., Ritschel T., Seidel H.-P

    Scherzer D., Nguyen C. H., Ritschel T., Seidel H.-P. : Pre-convolved radiance caching. Comput. Graph. Forum 31, 4 (jun 2012), 1391–1397

    2012

  35. [43]

    : Discretizing continuous action space for on-policy optimization

    Tang Y., Agrawal S. : Discretizing continuous action space for on-policy optimization. Proceedings of the AAAI Conference on Artificial Intelligence 34, 04 (Apr. 2020), 5981--5988

    2020

  36. [44]

    : Path guiding in production

    Vorba J., Hanika J., Herholz S., M\" u ller T., K r iv\' a nek J., Keller A. : Path guiding in production. In ACM SIGGRAPH 2019 Courses (New York, NY, USA, 2019), SIGGRAPH '19, ACM, pp. 18:1--18:77

    2019

  37. [45]

    : Adjoint-driven russian roulette and splitting in light transport simulation

    Vorba J., K r iv\' a nek J. : Adjoint-driven russian roulette and splitting in light transport simulation. ACM Trans. Graph. 35, 4 (jul 2016)

    2016

  38. [46]

    : On-line learning of parametric mixture models for light transport simulation

    Vorba J., Karl\' k O., S ik M., Ritschel T., K r iv\' a nek J. : On-line learning of parametric mixture models for light transport simulation. ACM Trans. Graph. 33, 4 (jul 2014)

    2014

  39. [47]

    : Real-time radiance caching using chrominance compression

    Vardis K., Papaioannou G., Gkaravelis A. : Real-time radiance caching using chrominance compression. Journal of Computer Graphics Techniques (JCGT) 3, 4 (December 2014), 111--131

    2014

  40. [48]

    J., Rubinstein F

    Ward G. J., Rubinstein F. M., Clear R. D. : A ray tracing solution for diffuse interreflection. In Proceedings of the 15th Annual Conference on Computer Graphics and Interactive Techniques (New York, NY, USA, 1988), SIGGRAPH '88, Association for Computing Machinery, p. 85–92

    1988

  41. [49]

    : Neusample: Importance sampling for neural materials

    Xu B., Wu L., Hasan M., Luan F., Georgiev I., Xu Z., Ramamoorthi R. : Neusample: Importance sampling for neural materials. In ACM SIGGRAPH 2023 Conference Proceedings (2023), pp. 1--10

    2023

  42. [50]

    : View-dependent radiance caching

    Zhao Y., Belcour L., Nowrouzezahrai D. : View-dependent radiance caching. In Proceedings of the 45th Graphics Interface Conference on Proceedings of Graphics Interface 2019 (Waterloo, CAN, 2019), GI'19, Canadian Human-Computer Communications Society

    2019

  43. [51]

    : Neural complex luminaires: representation and rendering

    Zhu J., Bai Y., Xu Z., Bako S., Vel \'a zquez-Armend \'a riz E., Wang L., Sen P., Hasan M., Yan L.-Q. : Neural complex luminaires: representation and rendering. ACM Trans. Graph. 40, 4 (2021), 57--1

    2021

  44. [52]

    W., Su H., Ramamoorthi R

    Zhu S., Xu Z., Sun T., Kuznetsov A., Meyer M., Jensen H. W., Su H., Ramamoorthi R. : Hierarchical neural reconstruction for path guiding using hybrid path and photon samples. ACM Trans. Graph. 40, 4 (jul 2021)

    2021

  45. [53]

    W., Su H., Ramamoorthi R

    Zhu S., Xu Z., Sun T., Kuznetsov A., Meyer M., Jensen H. W., Su H., Ramamoorthi R. : Photon-driven neural reconstruction for path guiding. ACM Trans. Graph. 41, 1 (Nov. 2021)

    2021

  46. [54]

    : Learning to importance sample in primary sample space

    Zheng Q., Zwicker M. : Learning to importance sample in primary sample space. Computer Graphics Forum 38, 2 (2019), 169--179

    2019

Pith tools

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