Pith. sign in

REVIEW 4 major objections 6 minor 42 references

GOGS: High-Fidelity Geometry and Relighting for Glossy Objects via Gaussian Surfels

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read GOGS claims a two-stage Gaussian-surfel pipeline recovers glossy-object geometry and materials from photos well enough to relight them under new illumination.

desk verdict Solid two-stage inverse-rendering system with strong relighting numbers, but the learned specular compensation term is a plausible crutch that can't be validated without code, a relighting ablation, and error bars. read the letter →

arxiv 2508.14563 v1 pith:WTWOEECU submitted 2025-08-20 cs.CV

Xingyuan Yang , Min Wei This is my paper
classification cs.CV
keywords inverserenderingglossyobjects2DGaussiansplattingrelightingmaterialdecompositionspecularreflectionsMonteCarloimportancesamplingfoundationmodelgeometricpriors
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that inverse rendering of glossy objects—recovering shape, material, and lighting from ordinary multi-view photos—can be made both accurate and practical by splitting the problem into two stages on 2D Gaussian surfels. In the first stage, geometry is reconstructed with physics-based split-sum shading plus monocular depth and normal priors from a pretrained foundation model, which suppresses the surface noise that specular highlights normally cause. In the second stage, with geometry fixed, materials are optimized against the full rendering equation using Monte Carlo importance sampling, with visibility and indirect light computed by differentiable Gaussian ray tracing and a spherical-mipmap correction for sharp highlights. The payoff, if true, is that glossy and reflective objects can be relit under new environments with plausible albedo, metalness, and roughness, at training times far below NeRF-based alternatives.

What carries the argument

The central mechanism is a two-stage decoupling on 2D Gaussian surfels: geometry first, material second. Stage I uses split-sum approximation of the specular integral—a BRDF lookup plus a prefiltered environment mipmap—as a cheap, stable shading model, and anchors the surface with monocular normal and depth priors through L1 and cosine normal losses and scale-invariant depth alignment. Stage II evaluates the full rendering equation with GGX and Lambert importance sampling combined by a balance heuristic; a differentiable 2D Gaussian ray tracer supplies visibility and indirect illumination, and a spherical-mipmap feature queried by reflected direction and roughness feeds a small MLP that adds

What would settle it

Take a textureless, highly curved mirror-like object under a known environment map and compare the reconstructed shape against a laser scan or structure-from-motion that excludes specular pixels. If the reconstructed normals and depth track the monocular prior where it disagrees with the multi-view photometric evidence, the central geometry claim fails. Alternatively, relight a scene with two facing mirrors: single-bounce ray tracing cannot capture the infinite interreflections, so relighting PSNR against ground truth should drop sharply.

Watch

Extended reading notes

Core claim

GOGS claims that the material-light ambiguity of glossy scenes is resolvable when geometry is decoupled from shading. Concretely, it proposes a two-stage pipeline built on 2D Gaussian surfels—oriented flat Gaussian disks that act as explicit surface elements. Stage I renders diffuse plus split-sum specular shading in a deferred G-buffer and regularizes the surface with foundation-model normal and depth losses and an object mask, recovering smooth geometry even under specular interference and inter-reflections. Stage II freezes that geometry and optimizes Disney BRDF parameters (albedo, metalness, roughness) plus an environment cubemap by evaluating the full rendering equation with balance-he

Load-bearing premise

That a pretrained monocular depth-and-normal model gives accurate geometry cues on glossy, specular objects, and that the per-view scale-and-shift alignment of its depth stays unique and stable; the paper's supplement notes the alignment can be sensitive and non-unique.

Editorial extensions

If this is right

  • Glossy-object inverse rendering from RGB becomes a practical two-pass pipeline: stable geometry first, then material estimation on frozen geometry.
  • Relighting quality no longer depends on the geometry stage modeling high-frequency speculars, since split-sum handles geometry and the full equation handles materials.
  • Inter-reflection regions can be recovered because the ray tracer provides visibility and indirect radiance to the material stage.
  • Decomposed albedo, metalness, roughness, and environment maps are directly usable for relighting, because the learned specular compensation is disabled under novel illumination.
  • Training cost stays near 1.3 hours without the compensation stage, versus hours for NeRF-based inverse renderers, making the approach more deployable.

Reading between the lines

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

  • If the pipeline is deployed on glossy objects outside the foundation model's training distribution, geometry quality will be bounded by the prior's bias; a testable fix is to weight the prior by multi-view stereo confidence rather than a fixed constant.
  • The need for a learned specular residual hints that the physical BRDF alone does not fit low-roughness highlights; this suggests hybrid neural-physical shading could generalize better than either component.
  • Because geometry is frozen before materials are estimated, any remaining geometry error will be absorbed into material maps; co-optimizing geometry with the full equation could trade relighting sharpness against geometric bias.
  • The single-bounce ray tracing will under-model multi-bounce specular paths in adjacent-mirror scenes, so relighting evaluations that include such scenes would likely expose the limit the authors already flag.
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

4 major / 6 minor

Summary. The paper proposes GOGS, a two-stage inverse-rendering framework based on 2D Gaussian surfels for glossy objects. Stage I reconstructs geometry using split-sum physically-based shading, regularized by monocular normal and depth priors from foundation models. Stage II fixes the geometry and decomposes materials by Monte Carlo importance sampling of the full rendering equation, with visibility and indirect illumination from differentiable 2D Gaussian ray tracing, plus a learned spherical-mipmap specular-compensation term. The authors report state-of-the-art or competitive results on Shiny Synthetic, Glossy Synthetic, and Shiny Real datasets for novel-view synthesis, normal estimation, and relighting under three environment maps.

Significance. If the claims hold, the paper makes a useful engineering contribution: it combines established components (2DGS, split-sum, MC importance sampling, 2D Gaussian ray tracing) with foundation-model geometric priors, and it reports sizable relighting improvements over Gaussian-based baselines (e.g., 25.75 vs. 22.01 PSNR on Glossy Synthetic, Table 3). The two-stage decomposition is clearly motivated and the experimental coverage is standard for the field. At the same time, the manuscript provides no code, no error bars, and no direct evidence that the learned specular-compensation term is not propping up the material decomposition. The central relighting claim therefore requires additional validation before the reported numbers can be taken at face value.

major comments (4)
  1. [Sec. 4.2, Eq. (20); Table 4] The learned compensation Lc = fc(K,h) is active during training, but the paper states it is disabled during relighting. No relighting ablation is provided: Table 4 ablate 'w/o specular compensation' only on NVS PSNR (28.58 vs. 29.38), not on relighting. If Lc absorbs a large share of the radiance in training views, then the physical BRDF/lighting parameters are not self-sufficient, and the relighting numbers in Table 3 are optimistic. Please report relighting PSNR/SSIM/LPIPS with and without Lc, and quantify the magnitude of Lc relative to Lpbr (e.g., mean ratio over training views).
  2. [Sec. 4.1, Eq. (9); Supplementary A] The supplementary material admits that the depth-alignment least-squares problem is 'highly sensitive to minimal distributional discrepancies' and 'can yield non-unique solutions for ω and b,' causing geometry fluctuations. Since geometric priors are a core contribution, this is load-bearing: the paper should either show the optimization is well-posed (e.g., by regularizing ω,b) or quantify sensitivity via repeated runs with different random seeds. Additionally, Table 4 shows that removing the geometry prior changes NVS PSNR by only 0.02 dB (29.36 vs. 29.38), which undercuts the quantitative case that the prior drives the reported gains. A more targeted geometry metric (e.g., Chamfer distance or relighting without the prior) is needed.
  3. [Table 1 and Table 3] The second stage (Ours(ir)) consistently decreases NVS metrics relative to Ours(geo) on all three datasets (e.g., Shiny Synthetic: 35.03 to 32.21 PSNR), yet the relighting evaluation reports only the second-stage model. To demonstrate that the material-decomposition stage improves relighting rather than overfitting training views or relying on Lc, the authors should report relighting results for Ours(geo) and Ours(ir) under the same three environment maps. Without this comparison, the design rationale for Stage II is not directly supported by the evidence.
  4. [Tables 1–5] All quantitative claims are point estimates with no error bars or significance tests. Some advertised improvements are small (Table 2 normal MAE 2.16 vs. 2.28; Table 1 NVS PSNR 30.83 vs. 30.77), so it is impossible to assess whether the differences are meaningful. Given the paper's stated goal of 'outperforming existing inverse rendering approaches,' the authors should provide at least error bars over multiple runs or seeds, especially for the geometry-prior and relighting claims.
minor comments (6)
  1. [Sec. 5.2] Typo: 'important sampling' should be 'importance sampling'.
  2. [Notation, Secs. 4.1/4.2] The symbol Lc is used both for the RGB reconstruction loss (Eq. 10) and for the compensation radiance (Eq. 19). In Eq. 21, Lc is the RGB loss while the compensation term from Eq. 20 is also Lc, which is confusing. Rename the compensation term (e.g., Lspeccomp).
  3. [Eq. (9)] The text says 'predicted normals ~D and rendered normals ^D' but the quantities are depth maps, not normals. Please correct the wording.
  4. [Table 2] The table caption does not state which dataset is used for the normal-map evaluation (presumably Glossy Synthetic) or the unit of MAE (degrees). Please specify.
  5. [Figure 2] The figure caption reads 'Overview of :' with a missing method name. In addition, the pipeline arrows and G-buffer terms could use a legend for readability.
  6. [Supp. Table 6 / Sec. C] The text cites 'GShader (0.48 hours)' but the table lists only ENVIDR, 3DGS-DR, Ref-GS, IRGS, and Ours. Check the correspondence or add GShader to the table.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the relighting claim is tested on held-out environment maps, and no load-bearing self-citation or equation-level reduction is present.

full rationale

The paper's two-stage derivation is self-contained against external benchmarks. Stage I reconstructs geometry using split-sum shading regularized by Marigold depth/normal priors; Stage II optimizes materials/lighting with the full rendering equation and Monte Carlo sampling; relighting is then evaluated under unseen environment maps (Table 3, Table 5). None of these steps defines its output in terms of the target metric. The only non-circular weakness is the specular compensation Lc: Eq. 20 trains with Lfinal = Lpbr + Lc, and the paper states Lc is 'explicitly disabled during relighting.' This means the physical model Lpbr alone is evaluated on novel illuminations after being optimized jointly with a learned residual, so the material/lighting parameters are not independently validated without Lc. However, this is a potential confound or correctness risk, not a circularity: the relighting test uses novel light maps not seen in training, so the reported PSNR is not forced by construction. The supplementary's admission that the depth alignment (Eq. 9) can yield non-unique solutions is also a robustness limitation, not a circular step. There are no load-bearing self-citations; the cited prior works (IRGS, Ref-Gaussian, Ref-GS, 2DGS) are external and are used as standard building blocks. Therefore, no circular step meeting the required evidence threshold is present.

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

The paper introduces no new physical entities. Its free parameters are training hyperparameters and per-scene fitted quantities (including the specular compensation network). The main axioms are domain assumptions about the adequacy of the Disney BRDF, split-sum approximation, Marigold priors, and 2D Gaussian ray tracing, plus the ad hoc assumption that the compensation term can be dropped for relighting.

free parameters (4)
  • Stage I loss weights (λn, λo, λsmooth, λgeo-n, λgeo-d) = 0.05, λo unstated, 0.01, 0.005, 0.005
    Hand-chosen hyperparameters; the paper does not justify these values or show sensitivity to them.
  • Stage II loss weights (λsmooth, λlight) and learning rates = λsmooth=2.0, λlight=0.01; lr albedo 0.0075, roughness 0.005, metallic 0.005, cubemap 0.01
    Hand-selected training hyperparameters.
  • Spherical mipmap specular compensation network = 8x512x512 mipmap with 16 channels; MLP with two 256-unit layers; up to 80k iterations
    This learned residual term is fitted to the training views and absorbs high-frequency appearance; it is a per-scene fitted module.
  • Per-scene environment cubemap and per-Gaussian BRDF fields = 6x128x128 cubemap; albedo, roughness, metallic per Gaussian
    Optimized to match training images. These are the intended outputs of inverse rendering, but they are fitted to data rather than derived.
assumptions (5)
  • domain assumption Disney BRDF with split-sum approximation is sufficiently accurate for geometry reconstruction of glossy objects.
    Invoked in Section 4.1; split-sum loses high-frequency specular detail, but the paper asserts it is adequate for Stage I.
  • domain assumption Marigold monocular depth and normal predictions provide reliable pseudo-ground truth on glossy synthetic objects.
    Stage I supervises reconstructions with these priors (Eq. 8, 9); the supplementary text admits the depth alignment can be unstable.
  • domain assumption 2D Gaussian ray tracing approximates visibility and indirect illumination well enough for decomposition, effectively a single-bounce approximation.
    Used in Stage II (Eq. 17); the paper's own limitation section says multi-bounce inter-reflections are not captured.
  • ad hoc to paper The learnable specular compensation term can be removed during relighting without invalidating the optimized materials.
    Eq. 20 adds Lc during training, then it is disabled for relighting; the physical parameters were optimized while Lc was active, so the decomposition may absorb errors that Lc would otherwise fix.
  • ad hoc to paper Geometry can be fixed after Stage I without significantly degrading material decomposition in Stage II.
    The two-stage decoupling is a design choice; if geometry errors remain in inter-reflection regions, material estimates may be biased.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GOGS: High-Fidelity Geometry and Relighting for Glossy Objects via Gaussian Surfels." pith.science (2026). https://pith.science/paper/WTWOEECU

@misc{pith2026250814563,
  author       = {Pith},
  title        = {Pith review of: GOGS: High-Fidelity Geometry and Relighting for Glossy Objects via Gaussian Surfels},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WTWOEECU}},
  note         = {Machine review of arXiv:2508.14563}
}
read the original abstract

Inverse rendering of glossy objects from RGB imagery remains fundamentally limited by inherent ambiguity. Although NeRF-based methods achieve high-fidelity reconstruction via dense-ray sampling, their computational cost is prohibitive. Recent 3D Gaussian Splatting achieves high reconstruction efficiency but exhibits limitations under specular reflections. Multi-view inconsistencies introduce high-frequency surface noise and structural artifacts, while simplified rendering equations obscure material properties, leading to implausible relighting results. To address these issues, we propose GOGS, a novel two-stage framework based on 2D Gaussian surfels. First, we establish robust surface reconstruction through physics-based rendering with split-sum approximation, enhanced by geometric priors from foundation models. Second, we perform material decomposition by leveraging Monte Carlo importance sampling of the full rendering equation, modeling indirect illumination via differentiable 2D Gaussian ray tracing and refining high-frequency specular details through spherical mipmap-based directional encoding that captures anisotropic highlights. Extensive experiments demonstrate state-of-the-art performance in geometry reconstruction, material separation, and photorealistic relighting under novel illuminations, outperforming existing inverse rendering approaches.

Figures

Figures reproduced from arXiv: 2508.14563 by the authors.

Figure 1
Figure 1. Our method achieves accurate geometry recovery and photorealistic rendering under novel illumination. Our rendering closely [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of : Stage I (Sec. 4.1) reconstructs geometry with 2DGS using split-sum shading, supervised by foundation model priors with geometric curvature losses. Stage II (Sec. 4.2) performs inverse rendering on fixed geometry via Monte Carlo sampling, utilizing 2D Gaussian ray tracing for visibility and indirect illumination and spherical mipmap-based compensation for high-frequency specular details. The pipeline’s … view at source ↗
Figure 3
Figure 3. Per-scene qualitative comparisons of normals. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Qualitative decomposition results of our model. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparisons of NVS renderings and [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison of relighting results on the [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of normal, material, and lighting estimation, and relighting results (using the Corri [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Relighting results on the Glossy Synthetic [ [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 11
Figure 11. Figure 11: Estimated environment maps for reflective object [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Outputs of our model including ground truth (GT), novel view synthesis (NVS), surface normals (visualized with [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Our GOGS framework generates high-fidelity [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 38 canonical work pages

  1. [1]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM , 65(1):99–106, 2021. 1, 2

    work page 2021

  2. [2]

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

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume render- ing for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2

    arXiv 2021

  3. [3]

    Mip-nerf: A multiscale representation for anti-aliasing neural 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 neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2

    work page 2021

  4. [4]

    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. InPro- ceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 5470–5479, 2022. 2

    work page 2022

  5. [5]

    Instant neural graphics primitives with a multiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transac- tions on graphics (TOG), 41(4):1–15, 2022. 2

    work page 2022

  6. [6]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333–

    work page

  7. [7]

    Ref-nerf: Structured view-dependent appearance for neural radiance fields

    Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022. 2, 5, 6, 8, 13

    work page 2022

  8. [8]

    Nero: Neural geometry and brdf recon- struction of reflective objects from multiview images

    Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wen- ping Wang. Nero: Neural geometry and brdf recon- struction of reflective objects from multiview images. ACM Transactions on Graphics (ToG) , 42(4):1–22,

    work page

Show all 42 references
  1. [9]

    En- vidr: Implicit differentiable renderer with neural en- vironment lighting

    Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. En- vidr: Implicit differentiable renderer with neural en- vironment lighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 79–89, 2023. 2, 5, 13

    2023

  2. [10]

    Tensoir: Tensorial inverse rendering

    Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 165–174,

  3. [11]

    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. ACM Trans. Graph., 42(4):139–1, 2023. 1, 2

    2023

  4. [12]

    Sugar: Surface- aligned gaussian splatting for efficient 3d mesh recon- struction and high-quality mesh rendering

    Antoine Gu ´edon and Vincent Lepetit. Sugar: Surface- aligned gaussian splatting for efficient 3d mesh recon- struction and high-quality mesh rendering. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5354–5363,

  5. [13]

    Mip-splatting: Alias-free 3d gaussian splatting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sat- tler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 19447–19456, 2024. 2

    2024

  6. [14]

    Gaus- sian opacity fields: Efficient adaptive surface recon- struction in unbounded scenes

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaus- sian opacity fields: Efficient adaptive surface recon- struction in unbounded scenes. ACM Transactions on Graphics (ToG), 43(6):1–13, 2024. 2

    2024

  7. [15]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity sur- face reconstruction

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Wei- jian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hu- jun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity sur- face reconstruction. IEEE Transactions on Visualiza- tion and Computer Gr...

    2024

  8. [16]

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 20654– 20664, 2024. 2

    2024

  9. [17]

    2d gaussian splatting for geometrically accurate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIG- GRAPH 2024 conference papers , pages 1–11, 2024. 1, 2, 4, 5, 6

    2024

  10. [18]

    3igs: Factorised ten- sorial illumination for 3d gaussian splatting

    Zhe Jun Tang and Tat-Jen Cham. 3igs: Factorised ten- sorial illumination for 3d gaussian splatting. In Euro- pean Conference on Computer Vision, pages 143–159. Springer, 2024. 1

    2024

  11. [19]

    Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting

    Ziyi Yang, Xinyu Gao, Yang-Tian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. Advances in Neural Information Processing Systems , 37:61192–61216, 2024

    2024

  12. [20]

    Gs3: Efficient re- lighting with triple gaussian splatting

    Zoubin Bi, Yixin Zeng, Chong Zeng, Fan Pei, Xiang Feng, Kun Zhou, and Hongzhi Wu. Gs3: Efficient re- lighting with triple gaussian splatting. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024

    2024

  13. [21]

    Ref-gs: Direc- tional factorization for 2d gaussian splatting

    Youjia Zhang, Anpei Chen, Yumin Wan, Zikai Song, Junqing Yu, Yawei Luo, and Wei Yang. Ref-gs: Direc- tional factorization for 2d gaussian splatting. In Pro- ceedings of the Computer Vision and Pattern Recog- nition Conference, pages 26483–26492, 2025. 1, 2, 5

    2025

  14. [22]

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

    Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 53...

    2024

  15. [23]

    Gs-ir: 3d gaussian splatting for inverse ren- dering

    Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse ren- dering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 21644–21653, 2024. 2

    2024

  16. [24]

    Relightable 3d gaussians: Realistic point cloud relighting with brdf decomposition and ray tracing

    Jian Gao, Chun Gu, Youtian Lin, Zhihao Li, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussians: Realistic point cloud relighting with brdf decomposition and ray tracing. In European Con- ference on Computer Vision , pages 73–89. Springer,

  17. [25]

    3d gaussian splatting with deferred reflection

    Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splatting with deferred reflection. InACM SIGGRAPH 2024 Conference Papers, pages 1–10, 2024. 2, 3, 5, 6, 13

    2024

  18. [26]

    Reflective gaussian splatting.arXiv preprint arXiv:2412.19282, 2024

    Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting.arXiv preprint arXiv:2412.19282, 2024. 2, 5, 6, 7, 12, 13

    2024 arXiv

  19. [27]

    Irgs: Inter-reflective gaussian splat- ting with 2d gaussian ray tracing

    Chun Gu, Xiaofei Wei, Zixuan Zeng, Yuxuan Yao, and Li Zhang. Irgs: Inter-reflective gaussian splat- ting with 2d gaussian ray tracing. In Proceedings of the Computer Vision and Pattern Recognition Confer- ence, pages 10943–10952, 2025. 1, 2, 5, 6, 7, 13

    2025

  20. [28]

    Glossygs: Inverse rendering of glossy objects with 3d gaussian splatting

    Shuichang Lai, Letian Huang, Jie Guo, Kai Cheng, Bowen Pan, Xiaoxiao Long, Jiangjing Lyu, Chengfei Lv, and Yanwen Guo. Glossygs: Inverse rendering of glossy objects with 3d gaussian splatting. IEEE Transactions on Visualization and Computer Graph- ics, 2025. 2

    2025

  21. [29]

    Gs-2dgs: Geometrically supervised 2dgs for reflective object reconstruction

    Jinguang Tong, Xuesong Li, Fahira Afzal Maken, Sundaram Muthu, Lars Petersson, Chuong Nguyen, and Hongdong Li. Gs-2dgs: Geometrically supervised 2dgs for reflective object reconstruction. In Proceed- ings of the Computer Vision and Pattern Recognition Conference, pages 21547–2...

    2025

  22. [30]

    Gs-id: Illu- mination decomposition on gaussian splatting via dif- fusion prior and parametric light source optimization

    Kang Du, Zhihao Liang, and Zeyu Wang. Gs-id: Illu- mination decomposition on gaussian splatting via dif- fusion prior and parametric light source optimization. arXiv preprint arXiv:2408.08524, 2024. 2

    2024 arXiv

  23. [31]

    3d gaussian ray tracing: Fast tracing of particle scenes

    Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Riccardode Lutio, JanickMartinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. Oct 2024. 2

    2024

  24. [32]

    Rgs-dr: Reflective gaussian surfels with de- ferred rendering for shiny objects

    Georgios Kouros, Minye Wu, and Tinne Tuyte- laars. Rgs-dr: Reflective gaussian surfels with de- ferred rendering for shiny objects. arXiv preprint arXiv:2504.18468, 2025. 2

    2025

  25. [33]

    Physically-based shading at disney

    Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In Acm siggraph, volume 2012, pages 1–7. vol. 2012, 2012. 3

    2012

  26. [34]

    James T. Kajiya. The rendering equation , page 157–164. Jul 1998. 3

    1998

  27. [35]

    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 Fidler. Extracting triangular 3d models, materials, and lighting from images. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , ...

  28. [36]

    Repurposing diffusion-based image generators for monocular depth estimation

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, RodrigoCaye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. Dec 2023. 4, 11

    2023

  29. [37]

    Fine-tuning image-conditional diffusion models is easier than you think

    Gonzalo Martin Garcia, Karim Abou Zeid, Chris- tian Schmidt, Daan De Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages 753–762. IEEE,

    2025

  30. [38]

    Towards robust monocular depth estimation: Mixing datasets for zero- shot cross-dataset transfer

    Ren ´e Ranftl, Katrin Lasinger, David Hafner, Kon- rad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero- shot cross-dataset transfer. IEEE transactions on pat- tern analysis and machine intelligence , 44(3):1623– 1637, 2020. 4

    2020

  31. [39]

    Torrance

    Robert L Cook and Kenneth E. Torrance. A re- flectance model for computer graphics. ACM Trans- actions on Graphics (ToG), 1(1):7–24, 1982. 5

    1982

  32. [40]

    Robust Monte Carlo methods for light transport simulation

    Eric Veach. Robust Monte Carlo methods for light transport simulation. Stanford University, 1998. 5

    1998

  33. [41]

    Wang, A.C

    Z. Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simon- celli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, page 600–612, Apr 2004. 6, 13

    2004

  34. [42]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable ef- fectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun 2018. 6, 13 GOGS: High-Fidelity Geometry and Rel...

    2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.