Pith. sign in

REVIEW 3 major objections 6 minor 3 cited by

IRGS: Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read IRGS claims to be the first Gaussian splatting method to apply the full rendering equation without simplification, using a differentiable 2D Gaussian ray tracer to compute incident light on the fly.

desk verdict A useful 2D Gaussian ray tracer for inverse rendering, but the 'full rendering equation' claim is overstated because the indirect term is traced from the stage-I appearance color, not the optimized PBR materials. read the letter →

arxiv 2412.15867 v2 pith:KNZLWKXR submitted 2024-12-20 cs.CV

Chun Gu , Xiaofei Wei , Zixuan Zeng , Yuxuan Yao , Li Zhang This is my paper · ORCID
classification cs.CV
keywords inverserenderingGaussiansplattingraytracinginter-reflectionequationrelighting2Ddifferentiable
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

The paper sets out to give 3D Gaussian splatting a way to model inter-reflections in inverse rendering, something previous splatting-based methods avoided by simplifying the rendering equation or by approximating indirect light with learnable parameters. Its central proposal is a differentiable ray tracer that works directly on 2D Gaussian primitives, giving each ray a well-defined intersection with the surface so that visibility and indirect radiance can be queried on the fly. With this ray tracer in hand, the authors evaluate the full rendering equation with Monte Carlo sampling at each pixel, decompose incident light into direct environment light and ray-traced indirect light, and add a relighting pass that computes indirect radiance under a new environment map. If correct, IRGS would let Gaussian splatting recover more accurate albedo, roughness, and lighting, and produce relit images with realistic color bleeding.

What carries the argument

The load-bearing object is 2D Gaussian ray tracing (2DGRT). Each 2D Gaussian disk — defined by center, two tangent vectors, and per-axis scales — is enclosed in an icosahedron mesh, the meshes are organized in a BVH, and NVIDIA OptiX finds ray–triangle hits; the exact ray–splat plane intersection is then computed analytically, giving a well-defined surface point, normal, and Gaussian influence for $\alpha$ blending. This yields the functions $(c_{\text{rt}}, o_{\text{rt}}) \leftarrow \text{Trace}(r_o, r_d)$, which supply both the attenuation factor $1-V$ and the indirect radiance $L_{\text{ind}}$ needed in the rendering equation $L_o(\omega_o,x)=\int_\Omega f(\omega_o,\omega_i,x)L_i(\omega_i,x)(\omega_i\cdot n)\,d\omega_i$.

What would settle it

Take a synthetic scene with known geometry and albedo in which a bright red surface faces a white matte surface, train IRGS, then relight with a blue environment map. If the indirect radiance comes from the first-stage view-dependent color, the white surface will retain red-tinged indirect light from the old lighting instead of shifting toward blue; comparing IRGS relights against a ground-truth path-traced image with the same new environment would settle whether the full rendering equation is actually being solved.

Watch

Extended reading notes

Core claim

IRGS claims to be the first Gaussian splatting inverse-rendering method to apply the full rendering equation without simplification. It does so by replacing the ambiguous intersection point of a 3D Gaussian along a ray with the well-defined ray–disk intersection of a 2D Gaussian, then alpha-blending the intersected primitives along the ray to obtain both the visibility (one minus accumulated opacity) and the indirect radiance coming from that direction. The indirect radiance is taken from the view-dependent appearance color learned in the first-stage 2DGS pretraining, so gradients can flow back through the ray tracer to refine that color and the material maps. The paper reports that this explicit inter-reflection modeling improves relighting and albedo estimation on synthetic benchmarks and produces realistic relights on a real-world dataset, with ablations showing that omitting or detaching the indirect term degrades albedo and indirect radiance quality.

Load-bearing premise

The entire inter-reflection model rests on treating the first-stage view-dependent appearance color as if it were the true outgoing radiance of each surface point; if that color is not physically consistent radiance, the rendering equation is being evaluated with an indirect term that is really just re-rendered appearance, and the claimed accuracy of inter-reflection is not independently grounded.

Editorial extensions

If this is right

  • Gaussian splatting inverse rendering can now account for color bleeding and secondary shadows without baking indirect illumination into learnable parameters.
  • Material and lighting estimates improve on benchmarks with strong inter-reflection, since the optimization sees physically-grounded indirect light rather than a learned approximation.
  • Relighting can include an indirect component, so relit scenes show realistic inter-reflection consistent with the new environment, instead of dropping indirect light entirely.
  • The same differentiable 2D ray tracer could be reused for other ray-based effects in Gaussian scenes, such as reflections, shadows, or global illumination at render time.

Reading between the lines

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

  • Because the indirect radiance in training is the first-stage view-dependent color, not a self-consistent solution of the rendering equation under the estimated materials, the method likely does not fully close the loop; a scene with strong inter-reflection in the training views could still produce biased material estimates, and relighting could carry over appearance from the original lighting.
  • A direct test: render a synthetic scene with a colored wall casting light onto a neutral surface, train IRGS, then relight with a different environment; if the indirect term is appearance-based, the relit surface will not show the correct color bleed from the new lighting.
  • The split-sum relighting strategy is a pragmatic fallback: it keeps relighting tractable but reintroduces the approximation the training phase claims to avoid, so IRGS's relighting quality under extreme lighting changes may be bounded by that approximation.
  • The 2DGRT primitive could be extended to other tasks that need ray queries in Gaussian scenes, such as shadow mapping or caustics, since the ray–disk intersection is analytic.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes IRGS, an inverse rendering framework built on 2D Gaussian splatting. It introduces a differentiable 2D Gaussian ray tracer (2DGRT) that queries visibility and indirect radiance for incident light, and integrates this into a two-stage pipeline: first a standard 2DGS appearance model is pretrained, then per-Gaussian albedo, roughness, and an environment cubemap are optimized by Monte Carlo evaluation of a rendering equation whose direct term uses the cubemap and whose indirect term is obtained by tracing the first-stage view-dependent color. A relighting strategy is also proposed, which uses split-sum shading of ray-aggregated material properties. Experiments on Synthetic4Relight, TensoIR, and Stanford-ORB report improved albedo and relighting accuracy over prior Gaussian-based inverse rendering methods.

Significance. If validated, the 2DGRT contribution is a practical and useful advance: ray tracing on pretrained 2D Gaussians introduces only minimal quality degradation, and the full pipeline is reasonably efficient (~0.7 hours on one RTX 3090). The reported relighting and albedo gains over GS-IR and R3DG are consistent across two synthetic benchmarks, and the qualitative visualizations show plausible inter-reflection effects. However, the paper's central claim—that it applies the full rendering equation without simplification—is not supported by the implementation, which uses a simplified BRDF and an appearance-based proxy for indirect radiance; the significance therefore rests on an approximate inter-reflection model rather than on a solution of the full global-illumination problem.

major comments (3)
  1. [Abstract; Sec. 1; Sec. 3.3.2 (Eq. 14)] The abstract and Sec. 1 claim that IRGS applies the full rendering equation without simplification and computes incident radiance on the fly. However, in Sec. 3.3.2, Eq. (14) sets the indirect radiance Lind(ω_i,x) through Trace(x,ω_i), and the text immediately after states that 'the RGB values used in ray tracing correspond to the view-dependent color c from the first stage.' This c is the SH appearance color fitted to the training images, not a radiance evaluated from the estimated albedo, roughness, and environment map under the PBR model. Consequently, the indirect term is a re-rendering of the observed appearance, and Eqs. (13)–(15) actually solve a direct-illumination PBR integral with an appearance-based proxy for inter-reflection. The claim of solving the full rendering equation without simplification is therefore not accurate and should be revised, along with the associated novelty statement.
  2. [Sec. 3.1; Supp. Sec. 6 (Eq. 18)] The manuscript internally acknowledges two simplifications that are hard to reconcile with the 'without simplification' claim. In Sec. 3.1, the BRDF is 'a simplified Disney BRDF model [7] with only diffuse albedo a and roughness r,' and in Supp. Sec. 6, Eq. (18), the relighting stage uses split-sum approximation and pre-integrated cubemaps. At minimum, the abstract and Sec. 1 should qualify the claim to refer to the full integral over the hemisphere (rather than the full rendering equation), and the relighting approximation should be mentioned in the main text, since readers may otherwise expect a consistently unsimplified treatment across training and relighting.
  3. [Sec. 4.2, Table 3] The ablation in Table 3 shows only modest quantitative gains from the indirect component: relighting PSNR is 34.68 for the full model versus 34.22 with detached indirect gradients, 33.93 without indirect during training, and 33.84 without indirect during relighting. Since the paper's central claim is that accurate inter-reflection modeling is essential for material and lighting estimation, these small differences merit a per-scene breakdown and a discussion of which effects drive the gains, rather than relying primarily on selected qualitative examples. This would strengthen the evidence that the proposed indirect term, as implemented via the first-stage color c, is the source of the reported improvements.
minor comments (6)
  1. [Table 2 heading] The word 'Quantatitive' in the table caption should be spelled 'Quantitative'.
  2. [Eq. (14)] Clarify the relation between the accumulated opacity returned by Trace and the visibility function V; as written, '1 − V' is only an estimate when multiple surfaces are encountered along the ray.
  3. [Eq. (5)] The bounding icosahedron formula uses log(o/αmin), but the alpha value used in rasterization is the product of opacity and Gaussian response; a short derivation of the threshold would improve reproducibility.
  4. [Sec. 3.1 and Supp. Sec. 7] The composited scene in Fig. 1 and Supp. Sec. 7 assigns an additional metallic parameter m to Gaussians, while the main method in Sec. 3.1 assumes dielectric materials; this discrepancy should be stated in the main text.
  5. [Sec. 4 (Implementation details)] The k-buffer size k=16 and transmittance threshold 0.03 are heuristic; a sentence on their influence on indirect radiance accuracy would be useful.
  6. [Eq. (16)] The gray-world prior on diffuse incident light should be justified for scenes with strongly colored illumination, as it may bias albedo estimates.

Circularity Check

1 steps flagged · score 6.0 of 10

Indirect radiance is traced from the first-stage fitted appearance color c, so the 'full rendering equation' claim reduces to re-rendering the training signal.

  1. fitted input called prediction [Sec. 3.2 Eq. (8); Sec. 3.3.2 Eqs. (13)-(14)]
    "It is important to note that the RGB values used in ray tracing correspond to the view-dependent color c from the first stage. Additionally, our 2D Gaussian ray tracing approach is fully differentiable, allowing gradients to propagate through the ray tracing process to optimize the indirect radiance of the incident ray."

    In Eq. (14) the indirect radiance Lind(ωi,x) is defined by Trace(x,ωi), and Trace in Eq. (8) alpha-blends the per-Gaussian view-dependent color c of vanilla 2DGS. This c is a free appearance field optimized in Stage I (and still optimized in Stage II) to reproduce the input training images; it is not computed from the estimated albedo, roughness, environment map, or BRDF. Therefore the indirect term in the claimed 'full rendering equation' is, by construction, a re-rendering of the fitted appearance signal, not an independently predicted incident radiance from the PBR model being optimized. The inter-reflection results are inherited from the appearance fit, so the central claim reduces to a fitted input being presented as a physical prediction.

full rationale

The only significant circularity is in the indirect-radiance construction. The paper's abstract and introduction claim the full rendering equation is applied without simplification, but Eqs. (13)-(14) and the explicit note in Sec. 3.3.2 show that Lind comes from Trace(x,ωi), whose radiance is the first-stage view-dependent color c. Since c is fitted to the training images, the indirect light is not a self-consistent evaluation of the PBR equation under the estimated materials and lighting; it is a re-rendered appearance term. This is a partial circularity: one central 'prediction' (on-the-fly incident radiance / inter-reflection) reduces by construction to a fitted input. The relighting stage additionally abandons the full equation in favor of split-sum (Sec. 6, Eq. (18)), which is a scope inconsistency rather than a circular step. Self-citations to R3DG [10] and Reflective Gaussian Splatting [37] are not load-bearing; the core ray tracer builds on external 2DGS [13] and 3DGRT [22], and the material/lighting outputs are evaluated on external relighting benchmarks with held-out environment maps. The circularity therefore does not void the whole method, but it does invalidate the 'full rendering equation without simplification' claim as stated. Score 6 reflects that one predicted quantity (indirect radiance/inter-reflection) reduces by construction to the fitted appearance field, while other outputs retain independent empirical support.

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

No new physical entities are introduced. The method introduces a ray tracing procedure, not an entity. The main uncharged assumptions are the use of stage-I color for indirect radiance and the gray-world light prior.

free parameters (7)
  • per-Gaussian albedo a = learned per Gaussian
    Material parameter optimized in stage II to match input images; the core output of inverse rendering.
  • per-Gaussian roughness r = learned per Gaussian
    Material parameter optimized in stage II.
  • environment cubemap Ldir = 32x32 cubemap
    Global lighting optimized in stage II; the direct lighting model.
  • white light prior weight lambda_light = 0.01
    Hyperparameter that enforces the gray-world assumption in Eq. (16); chosen by the authors.
  • MC sample count Nr = 256
    Number of incident directions per pixel; ablation shows it affects quality (Table 3).
  • rays per iteration Nrays = 2^18
    Computational budget for rendering equation evaluation; an efficiency choice.
  • k-buffer size k = 16
    Per-ray sorting buffer size, borrowed from 3DGRT; affects ray tracing accuracy.
assumptions (6)
  • standard math Rendering equation (Eq. 3) is the correct physical model of light transport.
    The paper evaluates this integral with uniform hemisphere sampling (Eq. 15).
  • domain assumption Disney BRDF with only diffuse albedo and roughness, dielectric material assumption.
    Restricts the material model; no metals or transmission, which limits the generality of the 'full' rendering equation.
  • ad hoc to paper Indirect radiance can be represented by the view-dependent color c from the first-stage 2DGS, via ray tracing (Eq. 14).
    This is the hidden approximation that undermines the 'without simplification' claim.
  • ad hoc to paper Diffuse incident light is white (gray-world prior, Eq. 16).
    A heuristic regularization to aid material-light disentanglement.
  • domain assumption 2D Gaussian disks have well-defined ray-splat intersections (Eqs. 6-7).
    Borrowed from 2DGS; the basis for 2DGRT.
  • domain assumption Split-sum approximation is adequate for relighting indirect radiance (Supp. Eq. 18).
    Standard approximation, but not the 'full rendering equation'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IRGS: Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing." pith.science (2026). https://pith.science/paper/KNZLWKXR

@misc{pith2026241215867,
  author       = {Pith},
  title        = {Pith review of: IRGS: Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KNZLWKXR}},
  note         = {Machine review of arXiv:2412.15867}
}
read the original abstract

In inverse rendering, accurately modeling visibility and indirect radiance for incident light is essential for capturing secondary effects. Due to the absence of a powerful Gaussian ray tracer, previous 3DGS-based methods have either adopted a simplified rendering equation or used learnable parameters to approximate incident light, resulting in inaccurate material and lighting estimations. To this end, we introduce inter-reflective Gaussian splatting (IRGS) for inverse rendering. To capture inter-reflection, we apply the full rendering equation without simplification and compute incident radiance on the fly using the proposed differentiable 2D Gaussian ray tracing. Additionally, we present an efficient optimization scheme to handle the computational demands of Monte Carlo sampling for rendering equation evaluation. Furthermore, we introduce a novel strategy for querying the indirect radiance of incident light when relighting the optimized scenes. Extensive experiments on multiple standard benchmarks validate the effectiveness of IRGS, demonstrating its capability to accurately model complex inter-reflection effects.

Figures

Figures reproduced from arXiv: 2412.15867 by the authors.

Figure 1
Figure 1. Global and indirect illumination in a Gaussian-based [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Schematic illustration of the proposed IRGS. Starting from a set of 2D Gaussians equipped with material properties, we apply rasterization to generate albedo, roughness, position, and normal maps. We then evaluate the rendering equation using stratified sampling at the corresponding position, drawing geometry and material values from these feature maps. The radiance of incident light is decomposed into direct radian… view at source ↗
Figure 3
Figure 3. Performance of directly applying Gaussian ray tracing [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of NVS, material and lighting estimation, and relighting results on the Synthetic4Relight dataset [ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of estimated components in incident light, including the averaged direct radiance [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of relighting results on TensoIR dataset [ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Relighting results on Stanford-ORB dataset [ [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Ablation studies on various components of [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Global, direct, and indirect illumination in a Gaussian-based scene using our [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Qualitative comparison of NVS, material and lighting estimation, and relighting results on the Synthetic4Relight dataset [ [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparison of NVS, material and lighting estimation, and relighting results on the Synthetic4Relight dataset [ [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison of NVS, material and lighting estimation, and relighting results on the Synthetic4Relight dataset [ [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Visualization of estimated components in novel view, including the averaged direct radiance [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: Qualitative comparison of rendered normal maps on the Synthetic4Relight dataset [ [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]
Figure 15
Figure 15. Figure 15: Qualitative comparison of rendered normal maps on the TensoIR dataset [ [PITH_FULL_IMAGE:figures/full_fig_p014_15.png]
Figure 16
Figure 16. Figure 16: Qualitative comparison of estimated albedo maps on the TensoIR dataset [ [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. MV-CoLight: Efficient Object Compositing with Consistent Lighting and Shadow Generation

    cs.CV 2025-05 conditional novelty 6.0 of 10

    A feed-forward two-stage compositing framework that harmonizes inserted objects across views using a Hilbert-ordered Gaussian color mapping, trained and evaluated on a new 480k-scene synthetic dataset.

  2. BEAM: Bridging Physically-based Rendering and Gaussian Modeling for Relightable Volumetric Video

    cs.GR 2025-02 conditional novelty 5.0 of 10

    A pipeline that adds physically-based surface materials to dynamic 4D Gaussians, producing relightable volumetric video from multi-view RGB footage.

  3. RaySplats: Ray Tracing based Gaussian Splatting

    cs.CV 2025-01 conditional novelty 4.0 of 10

    RaySplats renders 3D Gaussian Splatting scenes by computing exact ray-ellipsoid intersections instead of using rasterization or bounding polytopes.

Reference graph

Works this paper leans on

46 extracted references · 39 canonical work pages · cited by 3 Pith papers

  1. [7]

    Physically-based shading at disney

    Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In Acm Siggraph ,

    work page

  2. [1]

    Flash cache: Reducing bias in radiance cache based inverse rendering

    Benjamin Attal, Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Matthew O’Toole, and Pratul P Srini- vasan. Flash cache: Reducing bias in radiance cache based inverse rendering. In ECCV, 2025. 2

    work page 2025

  3. [2]

    Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields. In ICCV, 2021. 2

    work page 2021

  4. [3]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022

    2022

  5. [4]

    Zip-nerf: Anti-aliased grid- based neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid- based neural radiance fields. In ICCV, 2023. 2

    work page 2023

  6. [5]

    Bar- ron, Ce Liu, and Hendrik P.A

    Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Bar- ron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural re- flectance decomposition from image collections. In ICCV,

  7. [6]

    Samurai: Shape and material from uncon- strained real-world arbitrary image collections

    Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. Samurai: Shape and material from uncon- strained real-world arbitrary image collections. NeurIPS,

    work page

  8. [8]

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

    Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint, 2023. 2

    work page 2023

Show all 46 references
  1. [9]

    Subsurface scat- tering for 3d gaussian splatting

    Jan-Niklas Dihlmann, Arjun Majumdar, Andreas Engel- hardt, Raphael Braun, and Hendrik Lensch. Subsurface scat- tering for 3d gaussian splatting. arXiv preprint, 2024. 2

    2024

  2. [10]

    Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray trac- ing

    Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray trac- ing. arXiv preprint, 2023. 1, 2, 3, 5, 6, 7, 8

    2023

  3. [11]

    Prtgs: Precomputed radiance transfer of gaussian splats for real-time high-quality relighting

    Yijia Guo, Yuanxi Bai, Liwen Hu, Ziyi Guo, Mianzhi Liu, Yu Cai, Tiejun Huang, and Lei Ma. Prtgs: Precomputed radiance transfer of gaussian splats for real-time high-quality relighting. In ACMMM, 2024. 2

    2024

  4. [12]

    Nerfren: Neural radiance fields with reflections

    Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song- Hai Zhang. Nerfren: Neural radiance fields with reflections. In CVPR, 2022. 5

    2022

  5. [13]

    2d gaussian splatting for geometrically ac- curate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac- curate radiance fields. In SIGGRAPH, 2024. 2, 4, 5, 7

    2024

  6. [14]

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

    Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaox- iao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint, 2023. 1, 2

    2023

  7. [15]

    Tensoir: Tensorial inverse rendering

    Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Song- fang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 1, 2, 6, 7, 8, 4, 5

    2023

  8. [16]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 2023. 1, 2, 3, 5

    2023

  9. [17]

    Stanford-orb: a real-world 3d object inverse rendering benchmark

    Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agar- wala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. NeurIPS, 2024. 6, 8

    2024

  10. [18]

    Neuralangelo: High-fidelity neural surface reconstruction

    Zhaoshuo Li, Thomas M ¨uller, Alex Evans, Russell H Tay- lor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2

    2023

  11. [19]

    Gs-ir: 3d gaussian splatting for inverse rendering

    Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In CVPR,

  12. [20]

    Nero: Neural geometry and brdf reconstruction of reflective objects from multiview images

    Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nero: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 1, 2, 5

    2023

  13. [21]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. In ECCV, 2020. 1, 2

    2020

  14. [22]

    3d gaussian ray trac- ing: Fast tracing of particle scenes

    Nicolas Mo ¨enne-Loccoz, Ashkan Mirzaei, Or Perel, Ric- cardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray trac- ing: Fast tracing of particle scenes. arXiv preprint, 2024. 2, 4, 7

    2024

  15. [23]

    Instant neural graphics primitives with a multires- olution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a multires- olution hash encoding. ACM Trans. Graph., 2022. 2

    2022

  16. [24]

    Extracting triangular 3d models, materials, and lighting from images

    Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas M¨uller, and Sanja Fi- dler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2022. 1, 5, 6

    2022

  17. [25]

    Optix: a general purpose ray tracing engine

    Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllis- ter, Morgan McGuire, Keith Morley, Austin Robison, et al. Optix: a general purpose ray tracing engine. TOG, 2010. 2, 4, 7

    2010

  18. [26]

    Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps

    Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In ICCV, 2021. 2

    2021

  19. [27]

    Gir: 3d gaussian inverse rendering for relightable scene factorization

    Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint, 2023. 1, 2

    2023

  20. [28]

    Nerv: Neural reflectance and visibility fields for relighting and view synthesis

    Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 1, 2

    2021

  21. [29]

    Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint, 2021. 2, 5

    2021

  22. [30]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 2004. 6

    2004

  23. [31]

    Deferredgs: Decoupled and editable gaussian splatting with deferred shading

    Tong Wu, Jia-Mu Sun, Yu-Kun Lai, Yuewen Ma, Leif Kobbelt, and Lin Gao. Deferredgs: Decoupled and editable gaussian splatting with deferred shading. arXiv preprint ,

  24. [32]

    Street gaussians for modeling dynamic ur- ban scenes

    Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic ur- ban scenes. arXiv preprint, 2024. 2

    2024

  25. [33]

    Sire-ir: Inverse render- ing for brdf reconstruction with shadow and illumination re- moval in high-illuminance scenes

    Ziyi Yang, Yanzhen Chen, Xinyu Gao, Yazhen Yuan, Yu Wu, Xiaowei Zhou, and Xiaogang Jin. Sire-ir: Inverse render- ing for brdf reconstruction with shadow and illumination re- moval in high-illuminance scenes. arXiv preprint, 2023. 2

    2023

  26. [34]

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

    Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real- time photorealistic dynamic scene representation and render- ing with 4d gaussian splatting. arXiv preprint, 2023. 2

    2023

  27. [35]

    Deformable 3d gaussians for high- fidelity monocular dynamic scene reconstruction

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high- fidelity monocular dynamic scene reconstruction. In CVPR,

  28. [36]

    Neilf: Neural incident light field for physically-based ma- terial estimation

    Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based ma- terial estimation. In ECCV, 2022. 2

    2022

  29. [37]

    Reflective gaussian splatting

    Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting. In ICLR, 2025. 2

    2025

  30. [38]

    V ol- ume rendering of neural implicit surfaces

    Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. V ol- ume rendering of neural implicit surfaces. NeurIPS, 2021. 2

    2021

  31. [39]

    Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. TOG, 2024. 2

    2024

  32. [40]

    Neilf++: Inter-reflectable light fields for geometry and material esti- mation

    Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material esti- mation. In ICCV, 2023. 1, 2

    2023

  33. [41]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6

    2018

  34. [42]

    Nerfac- tor: Neural factorization of shape and reflectance under an unknown illumination

    Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul De- bevec, William T Freeman, and Jonathan T Barron. Nerfac- tor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 6, 7, 8

    2021

  35. [43]

    Modeling indirect illumination for inverse rendering

    Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 2, 6, 7, 8, 1, 3, 4 IRGS: Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing Supplementary Material Global illumin...

    2022

  36. [44]

    Relighting details Given a sampled incident direction, we conduct 2D Gaus- sian ray tracing to obtain the intersected Gaussians along the ray, and aggregate the albedo, roughness, and normal by alpha-blending: {Ar, Rr, Nr} = P i ωi{ai, ri, ni}, where ωi = TiαiP i Tiαi . Then, ...

  37. [45]

    Mic,” “Ficus,

    Composited scene details In Fig. 1 and Fig. 9, we relight a scene composed of four Gaussian-based objects: “Mic,” “Ficus,” “Lego,” and “Ground.” The objects “Mic,” “Ficus,” and “Lego” are reconstructed using the proposed IRGS framework, while “Ground” is manually designed usin...

  38. [46]

    air ballons

    More results 8.1. Results on Synthetic4Relight We further provide a qualitative comparison on three addi- tional scenes from the Synthetic4Relight dataset [43], in- cluding “air ballons” (Fig. 10), “hotdog” (Fig. 11), and “jugs” (Fig. 12). It is evident that our estimated mate...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.