RaySplats: Ray Tracing based Gaussian Splatting

Dawid Baran; Jacek Tabor; Krzysztof Byrski; Marcin K\k{a}dzio{\l}ka; Marcin Mazur; Przemys{\l}aw Spurek; Tadeusz Dziarmaga

REVIEW 3 major objections 7 minor 1 cited by

RaySplats: Ray Tracing based Gaussian Splatting

T0 review · 3 major / 7 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read RaySplats replaces rasterization in 3D Gaussian Splatting with ray tracing directly on Gaussian primitives, making shadows, transparency, mirrors, glass, and mesh integration part of the same differentiable pipeline.

desk verdict A real incremental step for ray-traced 3DGS, but Proposition 3.2's alpha proof is wrong for half-rays and the lighting claims lack quantitative evidence. read the letter →

arxiv 2501.19196 v1 pith:QVQZAUFZ submitted 2025-01-31 cs.CV

Krzysztof Byrski , Marcin Mazur , Jacek Tabor , Tadeusz Dziarmaga , Marcin Kądzio{l}ka , Dawid Baran , Przemys{l}aw Spurek This is my paper · ORCID

classification cs.CV

keywords 3DGaussianSplattingraytracingconfidenceellipsoiddifferentiablerenderingnovelviewsynthesisrelightingshadowstransparency

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

RaySplats sets out to remove a known blind spot in 3D Gaussian Splatting: because the standard pipeline rasterizes Gaussians onto the image plane, it is awkward to compose splats with meshes or to simulate light and shadow. The paper's proposal is to keep the Gaussian representation unchanged but render it with ray tracing, treating every Gaussian as a confidence ellipsoid and computing where camera rays actually cross that ellipsoid. The authors show that the whole procedure is differentiable during training as well as inference, and they report image quality on established benchmark scenes comparable to rasterized 3D Gaussian Splatting and to an earlier ray-tracing variant. If this holds, the practical payoff is that a single Gaussian scene representation can serve both conventional novel-view synthesis and physically motivated relighting, without converting splats to meshes or adding a separate shading stage.

What carries the argument

The load-bearing object is the Gaussian confidence ellipsoid, the set of points at Mahalanobis distance $Q$ from the Gaussian mean, where $Q$ is a configurable quantile of the chi-square distribution with three degrees of freedom. The first identity is the closed-form closest intersection of a ray with this ellipsoid (Proposition 3.1), obtained by whitening coordinates $o' = S^{-1}R^T(o-\mu)$ and $d' = S^{-1}R^T d$ so that the ellipsoid becomes a sphere and the intersection reduces to solving a quadratic with a numerically stable root. The second identity is the per-Gaussian $\alpha$ rule (Proposition 3.2): $\alpha_i = \hat{\alpha}_i \exp\left(-\tfrac{1}{2}\left\| o' - \frac{d'}{\|d'\|}\left\langle \frac{d'}{\|d'\|}, o' \right\rangle \right\|^2\right)$, which is the peak of the Gaussian density along the ray rather than an integral over the ray segment. A supporting mechanism is the two-phase traversal with ray-termination thresholds, plus an index buffer that records all hit Gaussians, allowing the backward pass to reuse forward traversal data and to compute the gradient of the last meaningful Gaussian.

What would settle it

Render a single strongly elongated Gaussian viewed edge-on, first with RaySplats' closed-form alpha rule and again by numerically integrating the same Gaussian density along the ray in the volumetric formula of Eq. (4); if the two pixel values diverge substantially as the confidence parameter $Q$ changes, the peak-density approximation is not capturing the primitive's contribution.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the central operation of 3D Gaussian Splatting—projecting each Gaussian to the screen—is not what makes the representation work; the same primitive can be evaluated along true camera rays. RaySplats represents each Gaussian by its confidence ellipsoid $E_{\mu,\Sigma,Q}$ defined by a Mahalanobis distance equal to a quantile of the chi-square distribution with three degrees of freedom, and assigns to a ray the first intersection point given by Proposition 3.1. Each Gaussian's $\alpha$ contribution is then the learned opacity times the maximum of the Gaussian's normalized density along the ray (Proposition 3.2), and colors are combined by ordinary $\alpha$ blending. Because the ray–ellipsoid intersection is analytic and the $\alpha$ rule is closed form, both forward and backward passes are differentiable, and an index buffer stores intersected Gaussians so gradients can be computed without re-traversing rays. The authors report quantitative results in the same range as rasterized 3DGS and 3DGRT on standard scene-collection tests, and qualitative results with glass, shadows, mirror reflections, and Gaussian splats placed inside mesh scenes.

Load-bearing premise

The model assumes that how much a Gaussian contributes to a pixel depends only on how close the ray passes to its center and on a fixed cut-off for including it, not on how long the ray spends inside the Gaussian or how the Gaussian's fuzziness is spread along the ray.

Editorial extensions

If this is right

A scene represented as Gaussians can be rendered either by rasterization or by ray tracing, so splats can be composited directly into path-traced mesh scenes without converting them to triangles.
Shadows, transparency, mirror reflections, and glass-like refraction become trainable features of the same optimization rather than post-processing effects.
Storing intersected Gaussians in an index buffer during the forward pass lets the backward pass reuse the same ray intersections, roughly halving the number of ray–Gaussian intersection computations.
The switch from spherical harmonics to plain RGB colors is reported to be sufficient in the ray-tracing setting, simplifying the per-Gaussian parameters.
On the tested datasets, adopting ray tracing does not appear to cost fidelity relative to rasterized baselines, making the richer rendering model usable where 3DGS already works.

Reading between the lines

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

Extension: because the alpha rule uses the density peak rather than an integral over the ray segment, physically based volumetric effects such as absorption, scattering, or participating media would likely need a segment-length correction or per-sample integration to remain accurate for elongated or overlapping Gaussians.
Extension: the same closed-form ellipsoid intersection can be reused for shadow rays and light rays, not only camera rays, which suggests a direct path to soft shadows and one-bounce global illumination that the paper's examples only begin to demonstrate.
Extension: the confidence quantile $Q$ is a smooth quality-versus-speed dial; one could schedule it during training, starting broad to catch sparse coverage and tightening it later, instead of fixing it per dataset.

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.

Desk Editor's Note

read the letter

Short version: this is a decent incremental paper that deserves referee time, but the central alpha computation has a real mathematical bug that the authors should fix before it lands. The code is public, and the contribution is concrete—exact ray-ellipsoid intersection instead of 3DGRT's polytope approximations, with a clean extension to flat 2D Gaussians, which is what makes the mesh, light, and shadow integration possible.

What is genuinely new: replacing polytope culling with the quadric intersection and using the Gaussian peak response for alpha aggregation. The derivations in Appendix A are mostly careful, and the gradient formulas are spelled out in enough detail to be reimplemented. The standard novel-view synthesis numbers in Table 1 are in the same range as 3DGRT, and the qualitative demos of glass, mirrors, and shadows are suggestive, though not conclusive.

Where it gets soft. First, Proposition 3.2 is proved incorrectly. Equation (10) defines the max over t≥0, but Appendix A.2 solves the unconstrained min over the whole line. The step 'min_{t≥0} ||td′+o′|| = min_{t≥0} ||td′−o′||' is false because the ray is not symmetric. When the perpendicular foot lies behind the ray origin (⟨o′,d′⟩>0), the true max over the ray is at t=0, and Eq. (11) overestimates α_i by exp(+½⟨o′,d′⟩²/⟨d′,d′⟩). This is not a purely theoretical corner case: secondary rays for reflections, transparency, and shadows start on surfaces and can easily point away from nearby Gaussians. The hard-ellipsoid culling rule aggravates the problem—it rejects a ray whenever the smaller root is negative, even when the ray starts inside the ellipsoid and has a positive exit root. This is an internal inconsistency with the paper's own definition, and it directly affects the lighting effects that are the stated headline.

Second, the headline lighting and mesh results are only shown as images. There is no quantitative comparison against IRGS, EnvGS, or 3DGRT on relighting scenes, and no error bars anywhere. The switch from L1 to L2 and SH to RGB, plus the per-dataset tuning of Q and epsilon, is not ablated. These gaps are addressable, but as presented the evidence for the final claim is incomplete.

Who this is for: anyone working on ray-traced Gaussian splatting or splat-mesh integration. It deserves a serious referee—the math bug should be caught. A revision that fixes or clearly scopes the t≥0 issue and adds quantitative light-effect evaluation would make this a solid paper. Send it out, but point the referees at Appendix A.2.

Referee Report

3 major / 7 minor

Summary. The paper proposes RaySplats, a differentiable ray-tracing renderer for 3D Gaussian Splatting that replaces the standard splatting projection with explicit ray-Gaussian intersections computed on confidence ellipsoids. The color of each pixel is obtained by front-to-back alpha blending, where the alpha of a Gaussian is the product of a learned opacity and the maximum of its normalized Gaussian density along the ray. The method is implemented in C/C++ with OptiX and includes hand-derived gradient formulas for all trainable parameters. The experimental section reports novel-view synthesis results on Mip-NeRF360, Tanks and Temples, and Deep Blending, together with qualitative demonstrations of shadows, reflections, and transparency, and an ablation study of three hyperparameters.

Significance. If the proposed rendering procedure is correct, it would provide a useful alternative to polytope-based 3DGRT and enable secondary-ray effects such as reflections and shadows within a Gaussian Splatting framework. The paper includes detailed analytic gradient derivations in the appendices and a full from-scratch CUDA implementation, which is a substantial engineering effort. However, the central alpha computation in Proposition 3.2 contains a genuine mathematical error for rays whose perpendicular foot lies behind the ray origin, and the claimed lighting capabilities are not evaluated quantitatively. These issues currently prevent the paper from being a reliable foundation for the stated claims.

major comments (3)

[Appendix A.2, Eq. (11)] The closed form for alpha_i in Proposition 3.2 is derived by replacing the constrained maximum over t>=0 with the unconstrained distance from the point o' to the full line {td' : t in R}. This replacement is valid only when the orthogonal projection parameter t* = -<o',d'>/<d',d'> is nonnegative. When <o',d'> > 0, the constrained maximum over t>=0 occurs at t=0, giving alpha_i = sigmoid(alpha_hat_i) * exp(-1/2 ||o'||^2), whereas Eq. (11) reports exp(-1/2(||o'||^2 - <o',d'>^2/||d'||^2)), which is larger by the factor exp(+1/2 <o',d'>^2/||d'||^2). The proof's step 'min_{t>=0}||td'+o'|| = min_{t>=0}||td'-o'||' is false because the ray is a half-line, not a full line; only the full line is symmetric about the origin. Since this alpha value is used both in training and inference, and since secondary rays for reflections, transparency, and shadows may start on surfaces or inside ellipsoids, this error is load-bearing and must be corrected by clamping t* at zero and re-deriving the gradient formulas accordingly.
[Section 3, Proposition 3.1 and ray-ellipsoid intersection rule] The ray-Gaussian intersection procedure discards a Gaussian whenever the smaller root of the quadratic is negative, even if the larger root is positive. This is inconsistent with the definition of the intersection set {t>=0 : ||o'+td'||^2 <= Q} and causes rays whose origin lies inside a confidence ellipsoid (for example, secondary rays originating at a mesh surface or inside a Gaussian) to be incorrectly classified as having no intersection. The paper's own description in Section 3 states that a ray is treated as missing if the closest intersection point 'is negative and does not belong to the ray,' which is exactly the situation in which the ray starts inside the ellipsoid and exits through the positive root. Given that the central claim includes mesh integration and secondary-ray effects, the intersection rule must accept the positive root when it exists, and the implementation in OptiX should be checked against this case.
[Section 4, Table 1] The experimental section validates RaySplats only on standard novel-view synthesis metrics. The central claim of the paper—that ray tracing enables the incorporation of meshes, lights, shadows, and transparency—is supported only by qualitative figures (Figs. 3-6), with no quantitative evaluation of relighting, shadow, or reflection accuracy. Moreover, Table 1 shows that on Tanks and Temples RaySplats achieves PSNR 22.20 dB, which is 1.00 dB below 3DGRT (23.20 dB) and 0.94 dB below 3DGS-30K (23.14 dB); the statement that RaySplats attains 'outcomes that are analogous' to classical 3DGS is therefore not supported by the reported numbers. The authors should add a dedicated lighting-effects benchmark and compare with IRGS and EnvGS, which are discussed in the related work but not evaluated.

minor comments (7)

[Section 4, first paragraph] The phrase 'ourRaySplats' is missing a space and should read 'our RaySplats'.
[Section 3, Eq. (11)] The variables o' and d' are used in Eq. (11) before they are defined in Appendix A.1; define them in the main text for readability.
[Section 3, paragraph 'Color Aggregation'] The term '2.5D approach' is introduced without explanation; please define it or replace it with a more standard description.
[Section 4, Figure 7] The ablation study is presented as a small plot without exact numeric values; replace it with a table reporting the mean and standard deviation of PSNR for each hyperparameter setting.
[Section 4, Implementation Details] The paper does not report training time or rendering frames-per-second, despite the introduction claiming efficiency advantages; such numbers are needed for comparison with 3DGRT.
[Section 3, Eq. (1) and Eq. (3)] The symbol N is used both for the normal distribution N(m, Sigma) and for the number of Gaussians; this notation clash should be resolved.
[Appendix A.1, line after Eq. (7)] The derivation of the discriminant contains a typesetting artifact with repeated 'A4' fragments; please clean up the appendix display math.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ray-Gaussian intersection and alpha formulas are externally cited or explicitly defined, and the empirical evaluation is benchmark-based.

full rationale

The derivation chain in RaySplats is not circular. The ray-Gaussian intersection (Proposition 3.1) is credited to Hearn et al., and the alpha aggregation formula (Proposition 3.2) is explicitly attributed to 3DGRT; neither is a self-citation, and neither is used to "predict" a fitted quantity. The model's alpha is defined by Eq. 10 and then computed via Eq. 11; this is a rendering definition rather than a prediction from fitted parameters. The only self-citations (Kasymov et al. 2024; Waczyńska et al. 2024) appear in related-work remarks and are not load-bearing. Hyperparameters Q, epsilon1, epsilon2, and the maximum number of Gaussians are configurable thresholds, not fitted constants relabeled as predictions. The skeptic's observation that Eq. 11 may overestimate Eq. 10 when the perpendicular foot lies behind the ray origin is a mathematical correctness concern about the cited formula; it is not a case where the output is equivalent to the input by construction. The paper is evaluated against external benchmarks (Mip-NeRF360, Tanks and Temples, Deep Blending), so its central comparisons are independent.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard Gaussian-splatting alpha blending and on the approximation that a Gaussian's contribution to a ray equals the product of learned opacity and the peak of the Gaussian pdf along the ray. The free parameters Q, epsilon1, epsilon2, and max Gaussians are tuned per dataset in the ablation, so the final numbers are conditional on that tuning. No new physical or virtual entities are postulated.

free parameters (5)

Q (chi-squared quantile) = set individually per dataset (not given numerically)
Q defines the confidence ellipsoid used for ray-intersection culling. The ablation in Fig. 7 shows PSNR varies with Q, and the authors state they set it per dataset as a trade-off.
epsilon1 (first transmittance threshold) = set individually per dataset
epsilon1 controls ray termination in the forward color-aggregation phase. Ablated in Fig. 7 and tuned per dataset.
epsilon2 (second transmittance threshold) = not specified
epsilon2 controls the ray-termination criterion for the gradient of the last meaningful Gaussian (Eq. 13). Not ablated or given a value.
max Gaussians per ray = set individually per dataset
Upper bound on the number of Gaussians intersected by a ray. Ablated in Fig. 7 and tuned per dataset.
lambda (SSIM weight) = not specified
Lambda balances L2 and D-SSIM in the loss (Eq. 14). No value or sensitivity analysis is reported.

assumptions (4)

domain assumption A 3D scene can be represented by a set of 3D Gaussian primitives with means, covariances, opacities, and colors.
This is the standard 3DGS representation (Eq. 1), inherited from Kerbl et al., and the paper builds on it without questioning it.
ad hoc to paper The color of a pixel is computed by front-to-back alpha blending with the alpha of each Gaussian being the product of learned opacity and the peak of its probability density along the ray.
Eqs. 3, 10, and 11 define the core rendering equation. This peak-response approximation comes from 3DGRT but is a modeling choice that the paper adopts without independent derivation or validation.
ad hoc to paper Gaussians whose confidence ellipsoid does not intersect the ray contribute zero color, so the ray-Gaussian intersection can be safely computed using the ellipsoid defined by quantile Q.
Eq. 6 defines the hard culling criterion. The choice of Q is a free parameter and the hard cutoff is an approximation of the unbounded Gaussian tail.
domain assumption OptiX-provided ray traversal correctly handles the ray-ellipsoid intersections and mesh intersections used in training and inference.
The implementation relies on NVIDIA OptiX 8.0.0 as the ray-tracing engine; the paper treats its correctness as given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RaySplats: Ray Tracing based Gaussian Splatting." pith.science (2026). https://pith.science/paper/QVQZAUFZ

@misc{pith2026250119196,
  author       = {Pith},
  title        = {Pith review of: RaySplats: Ray Tracing based Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QVQZAUFZ}},
  note         = {Machine review of arXiv:2501.19196}
}

read the original abstract

3D Gaussian Splatting (3DGS) is a process that enables the direct creation of 3D objects from 2D images. This representation offers numerous advantages, including rapid training and rendering. However, a significant limitation of 3DGS is the challenge of incorporating light and shadow reflections, primarily due to the utilization of rasterization rather than ray tracing for rendering. This paper introduces RaySplats, a model that employs ray-tracing based Gaussian Splatting. Rather than utilizing the projection of Gaussians, our method employs a ray-tracing mechanism, operating directly on Gaussian primitives represented by confidence ellipses with RGB colors. In practice, we compute the intersection between ellipses and rays to construct ray-tracing algorithms, facilitating the incorporation of meshes with Gaussian Splatting models and the addition of lights, shadows, and other related effects.

Figures

Figures reproduced from arXiv: 2501.19196 by the authors.

**Figure 1.** RaySplats (our) incorporates ray tracing into the 3D Gaussian Splatting framework. This allows us to integrate meshes with lighting conditions and mirror effects. of 3D Gaussian Splatting with the ray-tracing model, leveraging bounding primitives (polytopes) for each Gaussian. This approach utilizes iterations and is efficient, producing high-quality renders thanks to the NVIDIA OptiX programming interface (Parker… view at source ↗

**Figure 2.** RaySplats (our) uses ray-tracing based solutions. In practice, we need two important points on rays passing through Gaussian distributions. Then, the maximum response point is utilized for aggregating colors along each ray. On the other hand, the intersection of Gaussian confidence ellipses is used to efficiently detect Gaussians with non-empty intersection with the ray. This technique utilizes a tile-based raster… view at source ↗

**Figure 3.** RaySplats (our) allows us to combine 3D Gaussian splatting with a mesh-based rendering using lighting effects such as shadows and transparency. RaySplats for Mirror Reflections [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: RaySplats (our) allows us to combine 3D Gaussian splatting with a mesh-based rendering using lighting effects such as mirror reflections. restricts its applicability to 3D scene reconstruction or integration with mesh-based models. Another approach, EnvGS (Xie et al.,…

**Figure 5.** Figure 5: RaySplats (our) is capable of modeling glass elements in the 3D Gaussian Splatting environment, thereby facilitating the accurate visualization of glass reflections and the distortion of light due to refraction. by indices i ∈ {1, . . . , N}). This procedure is analogo…

**Figure 6.** Figure 6: RaySplats (our) combines meshes with 3DGS-based representations with different material structures. 3DGRT, which do not utilize the theoretical maximum for the number of intersectable Gaussians, imposing a threshold on the transmittance value Ti , RaySplats applies a …

**Figure 7.** Figure 7: Ablation study investigating the effect of three key parameters of the RaySplats model (our): the upper limit of Gaussians that can be hit by the ray, the ray termination threshold ε1 used throughout the forward phase, and the quantile Q of order α of the χ 2 (3) distr…

**Figure 8.** Figure 8: Examples of renderings of different types: the first column shows the ground truth image, the second column shows the rendering of the optimized RaySplats, and the third column shows the rendering of 3D Gaussian splatting with RGB colors. The first two rows consist of…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Floating Radiance Networks
cs.CV 2026-08 conditional novelty 6.0 of 10

FlaRe combines per-primitive latent radiance descriptors on planar Gaussians with a shared decoder and hardware ray tracing, making rendering, secondary rays, editing, and mesh extraction work in one scene model.

Reference graph

Works this paper leans on

29 extracted references · 19 canonical work pages · cited by 1 Pith paper

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., and Srinivasan, P

Barron, J. T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., and Srinivasan, P. P. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 5855--5864, 2021

work page 2021
[3]

T., Mildenhall, B., Verbin, D., Srinivasan, P

Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., and Hedman, P. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 5470--5479, 2022

work page 2022
[4]

Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering

Chen, Y., Gu, C., Jiang, J., Zhu, X., and Zhang, L. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023

arXiv 2023
[5]

Don't splat your gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media

Condor, J., Speierer, S., Bode, L., Bozic, A., Green, S., Didyk, P., and Jarabo, A. Don't splat your gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Transactions on Graphics, 2025

work page 2025
[6]

Plenoxels: Radiance fields without neural networks

Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., and Kanazawa, A. Plenoxels: Radiance fields without neural networks. In CVPR, pp.\ 5501--5510, 2022

work page 2022
[7]

Glassner, A. S. An introduction to ray tracing. Morgan Kaufmann, 1989

work page 1989
[8]

Irgs: Inter-reflective gaussian splatting with 2d gaussian ray tracing

Gu, C., Wei, X., Zeng, Z., Yao, Y., and Zhang, L. Irgs: Inter-reflective gaussian splatting with 2d gaussian ray tracing. arXiv preprint arXiv:2412.15867, 2024

arXiv 2024

Show all 29 references

[9]

Ges: Generalized exponential splatting for efficient radiance field rendering

Hamdi, A., Melas-Kyriazi, L., Mai, J., Qian, G., Liu, R., Vondrick, C., Ghanem, B., and Vedaldi, A. Ges: Generalized exponential splatting for efficient radiance field rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 19812--...

2024
[10]

D., Baker, M

Hearn, D. D., Baker, M. P., and Carithers, W. Computer graphics with open GL. Prentice Hall Press, 2010

2010
[11]

Deep blending for free-viewpoint image-based rendering

Hedman, P., Philip, J., Price, T., Frahm, J.-M., Drettakis, G., and Brostow, G. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37 0 (6): 0 1--15, 2018

2018
[12]

2d gaussian splatting for geometrically accurate radiance fields

Huang, B., Yu, Z., Chen, A., Geiger, A., and Gao, S. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 conference papers, pp.\ 1--11, 2024

2024
[13]

Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces

Jiang, Y., Tu, J., Liu, Y., Gao, X., Long, X., Wang, W., and Ma, Y. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 5322--5332, 2024

2024
[14]

Neggs: Negative gaussian splatting

Kasymov, A., Czekaj, B., Mazur, M., Tabor, J., and Spurek, P. Neggs: Negative gaussian splatting. arXiv preprint arXiv:2405.18163, 2024

2024 arXiv
[15]

3d gaussian splatting for real-time radiance field rendering

Kerbl, B., Kopanas, G., Leimk \"u hler, T., and Drettakis, G. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42 0 (4): 0 139--1, 2023

2023
[16]

Tanks and temples: Benchmarking large-scale scene reconstruction

Knapitsch, A., Park, J., Zhou, Q.-Y., and Koltun, V. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 0 (4): 0 1--13, 2017

2017
[17]

Point-based neural rendering with per-view optimization

Kopanas, G., Philip, J., Leimk \"u hler, T., and Drettakis, G. Point-based neural rendering with per-view optimization. In Computer Graphics Forum, volume 40, pp.\ 29--43. Wiley Online Library, 2021

2021
[18]

Neural point catacaustics for novel-view synthesis of reflections

Kopanas, G., Leimk \"u hler, T., Rainer, G., Jambon, C., and Drettakis, G. Neural point catacaustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41 0 (6): 0 1--15, 2022

2022
[19]

Gs-ir: 3d gaussian splatting for inverse rendering

Liang, Z., Zhang, Q., Feng, Y., Shan, Y., and Jia, K. Gs-ir: 3d gaussian splatting for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 21644--21653, 2024

2024
[20]

P., Tancik, M., Barron, J

Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R., and Ng, R. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65 0 (1): 0 99--106, 2021

2021
[21]

3d gaussian ray tracing: Fast tracing of particle scenes

Moenne-Loccoz, N., Mirzaei, A., Perel, O., de Lutio, R., Martinez Esturo, J., State, G., Fidler, S., Sharp, N., and Gojcic, Z. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics (TOG), 43 0 (6): 0 1--19, 2024

2024
[22]

Instant neural graphics primitives with a multiresolution hash encoding

M \"u ller, T., Evans, A., Schied, C., and Keller, A. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41 0 (4): 0 1--15, 2022

2022
[23]

G., Bigler, J., Dietrich, A., Friedrich, H., Hoberock, J., Luebke, D., McAllister, D., McGuire, M., Morley, K., Robison, A., et al

Parker, S. G., Bigler, J., Dietrich, A., Friedrich, H., Hoberock, J., Luebke, D., McAllister, D., McGuire, M., Morley, K., Robison, A., et al. Optix: a general purpose ray tracing engine. Acm transactions on graphics (tog), 29 0 (4): 0 1--13, 2010

2010
[24]

D-miso: Editing dynamic 3d scenes using multi-gaussians soup

Waczy \'n ska, J., Borycki, P., Kaleta, J., Tadeja, S., and Spurek, P. D-miso: Editing dynamic 3d scenes using multi-gaussians soup. arXiv preprint arXiv:2405.14276, 2024

2024 arXiv
[25]

C., Sheikh, H

Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13 0 (4): 0 600--612, 2004

2004
[26]

Envgs: Modeling view-dependent appearance with environment gaussian

Xie, T., Chen, X., Xu, Z., Xie, Y., Jin, Y., Shen, Y., Peng, S., Bao, H., and Zhou, X. Envgs: Modeling view-dependent appearance with environment gaussian. arXiv preprint arXiv:2412.15215, 2024

2024 arXiv
[27]

Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting

Yang, Z., Yang, H., Pan, Z., and Zhang, L. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023

2023 arXiv
[28]

Differentiable surface splatting for point-based geometry processing

Yifan, W., Serena, F., Wu, S., \"O ztireli, C., and Sorkine-Hornung, O. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38 0 (6): 0 1--14, 2019

2019
[29]

A., Shechtman, E., and Wang, O

Zhang, R., Isola, P., Efros, A. A., Shechtman, E., and Wang, O. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 586--595, 2018

2018

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., and Srinivasan, P

Barron, J. T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., and Srinivasan, P. P. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 5855--5864, 2021

work page 2021

[3] [3]

T., Mildenhall, B., Verbin, D., Srinivasan, P

Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., and Hedman, P. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 5470--5479, 2022

work page 2022

[4] [4]

Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering

Chen, Y., Gu, C., Jiang, J., Zhu, X., and Zhang, L. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023

arXiv 2023

[5] [5]

Don't splat your gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media

Condor, J., Speierer, S., Bode, L., Bozic, A., Green, S., Didyk, P., and Jarabo, A. Don't splat your gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Transactions on Graphics, 2025

work page 2025

[6] [6]

Plenoxels: Radiance fields without neural networks

Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., and Kanazawa, A. Plenoxels: Radiance fields without neural networks. In CVPR, pp.\ 5501--5510, 2022

work page 2022

[7] [7]

Glassner, A. S. An introduction to ray tracing. Morgan Kaufmann, 1989

work page 1989

[8] [8]

Irgs: Inter-reflective gaussian splatting with 2d gaussian ray tracing

Gu, C., Wei, X., Zeng, Z., Yao, Y., and Zhang, L. Irgs: Inter-reflective gaussian splatting with 2d gaussian ray tracing. arXiv preprint arXiv:2412.15867, 2024

arXiv 2024