Pith. sign in

REVIEW 3 major objections 4 minor 4 cited by

Reflective Gaussian Splatting

T0 review · 3 major / 4 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 This paper presents Ref-Gaussian, a real-time Gaussian-splatting renderer for reflective surfaces that models inter-reflection by combining split-sum deferred shading with ray-traced binary visibility.

desk verdict Solid engineering with credible NVS results; the inter-reflection claim is overclaimed and needs revalidation. read the letter →

arxiv 2412.19282 v2 pith:HBWHFMHY submitted 2024-12-26 cs.CV

Yuxuan Yao , Zixuan Zeng , Chun Gu , Xiatian Zhu , Li Zhang This is my paper · ORCID
classification cs.CV
keywords reflectiveobjectreconstructionGaussiansplattinginverserenderinginter-reflectionphysicallybaseddeferredshadingsplit-sumapproximationrelighting
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 aims to make reflective and glossy surfaces renderable in real time, including the complicated effect of objects reflecting one another. It proposes Ref-Gaussian, a Gaussian-splatting framework that combines physically based deferred shading with a new inter-reflection term: each reflective point's indirect light is modeled by a per-Gaussian spherical-harmonic color evaluated along the mirror direction, gated by a single binary ray-traced visibility check. If the central claim is right, one system delivers real-time novel views, relighting, and material editing for both reflective and ordinary scenes, whereas previous methods either sacrificed inter-reflection or paid for it with Monte Carlo sampling.

What carries the argument

The core is a split-sum approximation of the rendering equation evaluated after $\alpha$-blending, so that material properties (albedo, metallic, roughness, normal) are aggregated per-pixel; the specular integral is factorized into a precomputed BRDF term and an environment-map convolution. The inter-reflection extension (Eqs. 9–10) adds an indirect term $L_{\mathrm{ind}}$: each Gaussian carries spherical-harmonic coefficients, evaluated in the reflected direction $R = 2(\omega_o \cdot N)N - \omega_o$, and the aggregate is multiplied by $(1 - V)$ where $V$ is a per-pixel binary visibility from ray tracing against a TSDF mesh. Geometry support comes from 2D Gaussian disks, a per-Gaussian shading initialization stage, and a material-aware normal propagation that enlarges high-metallic, low-roughness Gaussians.

What would settle it

Render a scene with two parallel mirrors facing each other and compare against a path-traced reference; because Ref-Gaussian uses only single-bounce binary visibility and an SH-based indirect term evaluated in the reflected direction, it cannot represent the series of bounces that such a scene produces, so the multiply-reflected pattern should be missing or blurred.

Watch

Extended reading notes

Core claim

The paper's central claim is that inter-reflection can be incorporated into Gaussian splatting without leaving the real-time regime. The method separates the specular term into a direct part from a pre-integrated environment map and an indirect part computed from per-Gaussian spherical-harmonic colors evaluated along the reflected direction; the indirect part is gated by a binary visibility flag obtained by tracing one ray per pixel against a mesh that is periodically extracted during training. With 2D Gaussian primitives, a per-Gaussian shading pre-stage, and material-aware normal propagation, Ref-Gaussian reports higher PSNR and SSIM and lower LPIPS than prior reflective-object methods on the Shiny Blender, Glossy Synthetic, and Ref-Real datasets, at 122 FPS and roughly 0.58 hours of training per scene.

Load-bearing premise

The inter-reflection model rests on a single binary ray test along the exact mirror direction deciding whether a surface is occluded from its environment.

Editorial extensions

If this is right

  • Real-time rendering of reflective objects becomes feasible on a single A6000 GPU (122 FPS), making interactive relighting and editing possible.
  • The method works as a unified solution for both glossy and matte scenes, so users do not need to switch representations by surface type.
  • The split-sum approximation avoids Monte Carlo noise, keeping training fast (about 0.58 hours) while still modeling inter-reflection.
  • Downstream applications such as relighting and editing follow directly from the decomposed materials and environment map.

Reading between the lines

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

  • Editorial inference: the single-ray binary visibility restricts $L_{\mathrm{ind}}$ to a single mirror bounce; scenes with two facing mirrors or strong multi-bounce reflections are a stress test the paper does not run.
  • Editorial inference: because $L_{\mathrm{ind}}$ is a per-Gaussian spherical-harmonic color in the reflected direction, it can also absorb view-dependent residuals that are not true inter-reflection; an evaluation that isolates a known inter-reflection component would separate the two.
  • Editorial inference: the material-aware normal propagation (growing Gaussians with high metallic and low roughness) suggests a cheap regularizer for other deferred-shading radiance fields, independent of the inter-reflection machinery.
  • Editorial inference: substituting a learned or precomputed multi-ray visibility for the binary flag could extend the same framework to global illumination without giving up the real-time property.
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 / 4 minor

Summary. The paper proposes Ref-Gaussian, a 2D Gaussian Splatting framework for inverse rendering of reflective scenes. It combines pixel-level deferred shading with a simplified Disney BRDF and split-sum environment lighting, and adds an inter-reflection term Lind modeled as per-Gaussian spherical harmonics evaluated in the reflected direction, gated by a binary visibility computed by ray tracing on a TSDF-extracted mesh. Geometry optimization is supported by 2D Gaussian primitives, an initial per-Gaussian shading stage, and material-aware normal propagation. The authors report strong novel-view-synthesis results on Shiny Blender, Glossy Synthetic, and Ref-Real datasets, with training times and frame rates competitive with or better than NeRF-based and several Gaussian-based baselines, and they demonstrate relighting and editing applications.

Significance. If the inter-reflection claim is established, the paper would be a meaningful contribution: it would be the first Gaussian-splatting formulation to explicitly model indirect specular inter-reflection while retaining real-time rendering, and the material/lighting decomposition enables downstream relighting and editing. The strengths of the paper are its extensive external benchmarks (Table 1 and appendix Tables 5-8), a thorough ablation suite (Tables 3-4), and concrete, reproducible efficiency numbers. The main risk is that the physical interpretation of Lind is not independently validated: the ablation effect is small and could be explained by added view-dependent fitting capacity rather than true indirect illumination.

major comments (3)
  1. [Section 3.2, Eq. (9)-(10), Appendix A.1, Table 4] The central novelty claim that Lind 'realizes the inter-reflection function' is not supported by the evidence presented. Eq. (10) defines Lind as per-Gaussian spherical harmonics evaluated in the reflected direction and alpha-blended, which is the same functional form as the ordinary 3DGS view-dependent color in Eq. (2); the only physical coupling is the binary visibility V in Eq. (9), computed from a single ray along the mirror direction. Appendix A.1 explicitly concedes that this ignores the rest of the specular lobe. Since Lind is fitted from the same RGB reconstruction loss as all other parameters, it can absorb residual view-dependent signal in pixels flagged as occluded, regardless of physical origin. The ablation in Table 4 measures only 0.19 dB PSNR, 0.001 SSIM, and 0.001 LPIPS when removing inter-reflection, with no statistical significance, and there is no comparison against path-traced indirect illumination on the Glossy Synthetic dataset where ground truth is available. I request a quantitative validation of Lind against reference indirect radiance (e.g., rendered with Blender Cycles) and/or an isolation experiment where only indirect light is rendered and compared to ground truth, preferably with a capacity-matched baseline that has the same SH residual but without visibility gating.
  2. [Abstract, Conclusion, Table 2] The abstract and conclusion state that Ref-Gaussian 'surpasses existing approaches' in compute efficiency, but Table 2 shows that 3DGS-DR trains in 0.35h versus 0.58h for Ref-Gaussian and renders at 251 FPS versus 122 FPS. The efficiency claim is defensible only against NeRF-based and slower Gaussian baselines. Please qualify the claim or provide an efficiency-equitable comparison, for example quality-versus-time curves and repeated runs with error bars.
  3. [Appendix A.3, Table 8] A.3 states that Ref-Gaussian 'excels over both alternatives' on the NeRF-Synthetic dataset, but Table 8 shows that Ref-Gaussian is not uniformly superior to 3DGS: chair PSNR is 34.71 versus 35.03, lego PSNR is 33.46 versus 33.71, and ship SSIM/LPIPS are 0.898/0.125 versus 0.906/0.111. The unified-solution claim should be supported by averages and appropriate significance testing, or softened to reflect the mixed per-scene results.
minor comments (4)
  1. [Section 4, Implementation Details] The sentence 'retaining only the the geometry of Gaussians' contains a duplicated article; Appendix A.3 also contains the typo 'soluton'.
  2. [Table 2] Training time and FPS are reported as single scalars with no indication of whether these are single-run measurements or averages across scenes; please state the measurement protocol and, if possible, report variance.
  3. [Figure 9 caption] The caption 'w/w/o' is ambiguous; please spell out 'with/without' for readability.
  4. [Section 3.1, Eq. (8)] The split-sum approximation as written could be clarified by stating explicitly how the GGX distribution is normalized in the pre-integrated environment map lookups, to avoid ambiguity about energy conservation.

Circularity Check

1 steps flagged · score 4.0 of 10

Inter-reflection novelty reduces to a relabeled per-Gaussian SH color: Eq. 10 is the standard 3DGS alpha-blending formula fit to training views, not a physically validated indirect-light prediction.

  1. renaming known result [Section 3.2, Eqs. (9)-(10); compare Eq. (2) in Section 3 and Table 4]
    "We represent the indirect light from occluded part as Lind: L′s(ωo) ≈ ( Z Ω fs(ωi, ωo)(ωi · N )dωi) · [Ldir · V + Lind · (1 − V )]. (9) ... For the indirect lighting component, each Gaussian is assigned an additional view-dependent color lind, modeled by spherical harmonics. During the rendering process,lind is evaluated in the reflected direction at the Gaussian level, and alpha blending is applied to aggregate the indirect lighting map as follows: Lind = NX i=1 lindαi i−1Y j=1 (1 − αj). (10)"

    Equation (10) is the same alpha-blending formula as the standard 3DGS view-dependent color in Eq. (2), with the per-Gaussian SH color renamed lind and evaluated at the reflected direction; the only physical coupling is the binary mask (1−V) from a single mirror ray. Since lind is optimized solely by the RGB reconstruction loss against the training views and is never supervised by ground-truth or path-traced indirect radiance, the 'indirect light' term is, by construction, whatever residual color minimizes the masked-pixel reconstruction error. Calling this 'Gaussian-grounded inter-reflection' is therefore a relabeling of ordinary SH view-dependent fitting capacity.

full rationale

The paper's central novel-view-synthesis claim is benchmarked against external datasets (Shiny Blender, Glossy Synthetic, Ref-Real) and does not depend on a self-citation chain, so most of the method is not circular. The load-bearing concern is the Gaussian-grounded inter-reflection component: Eq. (10) defines Lind as alpha-blended per-Gaussian spherical-harmonic colors, which is the same functional form as the standard 3DGS view-dependent color in Eq. (2), with only the evaluation direction changed to the reflected direction and a binary visibility mask applied. Because lind is fitted purely through the reconstruction loss and is never compared with measured or simulated indirect radiance, the quantity called 'indirect light' is defined as whatever residual fits the masked pixels. The claim of realizing the inter-reflection function for the first time is therefore partly a renaming of standard SH view-dependent fitting capacity rather than a physically derived and independently verified prediction. The small ablation difference in Table 4 further supports the view that the term adds fitting capacity. This is a partial circularity affecting the physical-interpretation novelty, but the quantitative rendering comparisons and the PBR deferred-shading pipeline retain independent content, so the score is 4 rather than 6.

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

The method relies on several domain assumptions and hand-chosen hyperparameters listed above. No new physical entities are introduced; the indirect-light term is a learned appearance channel rather than a new force, particle, or conserved quantity. The central NVS claim is tested against external datasets, so the ledger contains no invented physical mechanisms.

free parameters (4)
  • Loss balancing weights (lambda=0.2, lambda_n=0.05, lambda_smooth=1.0) = 0.2 / 0.05 / 1.0
    Chosen by hand in Section 3.3; they weight RGB, normal consistency, and normal smoothness and directly affect the geometry and material split.
  • Material-aware normal propagation thresholds = metallic >= 0.02, roughness <= 0.1
    Chosen in implementation details to decide which Gaussians get enlarged for normal propagation; no sensitivity analysis is given.
  • Training stage lengths = 18,000 per-Gaussian steps; about 40,000 deferred steps
    Chosen in implementation details; the two-stage schedule is central to the geometry initialization claim.
  • Initial roughness and mesh extraction interval = roughness 0.1; mesh every 3,000 steps
    Fixed in implementation details; visibility and indirect-light estimation depend on the mesh extraction cadence.
assumptions (5)
  • domain assumption Split-sum approximation of the specular rendering equation is accurate enough for material and lighting decomposition.
    Used in Eq. 8 to replace Monte Carlo integration; the approximation's error is not quantified for the tested scenes.
  • domain assumption Disney BRDF with albedo, metallic, and roughness is an adequate material model for these scenes.
    Assumed in Section 3.1; no validation that the BRDF form captures the real reflection behavior.
  • domain assumption TSDF-fused mesh extracted from 2D Gaussians is an accurate proxy for visibility ray tracing.
    Section 3.2 uses mesh ray tracing to compute binary visibility; mesh errors directly corrupt V and hence Lind.
  • ad hoc to paper A per-Gaussian spherical harmonic directivity can represent true indirect inter-reflection.
    Eq. 10 defines Lind as a fitted SH color; nothing ties it to actual secondary bounces, so it can absorb view-dependent residuals instead.
  • domain assumption 2D Gaussian primitives are a valid surface representation for reflective object reconstruction.
    Adopted from 2DGS in Section 3; the paper's geometry optimization relies on this representation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reflective Gaussian Splatting." pith.science (2026). https://pith.science/paper/HBWHFMHY

@misc{pith2026241219282,
  author       = {Pith},
  title        = {Pith review of: Reflective Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HBWHFMHY}},
  note         = {Machine review of arXiv:2412.19282}
}
read the original abstract

Novel view synthesis has experienced significant advancements owing to increasingly capable NeRF- and 3DGS-based methods. However, reflective object reconstruction remains challenging, lacking a proper solution to achieve real-time, high-quality rendering while accommodating inter-reflection. To fill this gap, we introduce a Reflective Gaussian splatting (Ref-Gaussian) framework characterized with two components: (I) Physically based deferred rendering that empowers the rendering equation with pixel-level material properties via formulating split-sum approximation; (II) Gaussian-grounded inter-reflection that realizes the desired inter-reflection function within a Gaussian splatting paradigm for the first time. To enhance geometry modeling, we further introduce material-aware normal propagation and an initial per-Gaussian shading stage, along with 2D Gaussian primitives. Extensive experiments on standard datasets demonstrate that Ref-Gaussian surpasses existing approaches in terms of quantitative metrics, visual quality, and compute efficiency. Further, we show that our method serves as a unified solution for both reflective and non-reflective scenes, going beyond the previous alternatives focusing on only reflective scenes. Also, we illustrate that Ref-Gaussian supports more applications such as relighting and editing.

Figures

Figures reproduced from arXiv: 2412.19282 by the authors.

Figure 1
Figure 1. Our Ref-Gaussian achieves superior rendering quality in novel view synthesis while en￾joying fast optimization (training time) and real-time rendering (FPS), and supporting various down￾stream applications such as relighting and editing. Please refer to the video in our supplementary material for a more comprehensive and intuitive comparison. ABSTRACT Novel view synthesis has experienced significant advancements owi… view at source ↗
Figure 2
Figure 2. Overview of the Ref-Gaussian framework: First, we apply the splatting process to produce feature maps and perform ray-tracing on the extracted mesh to compute visibility for the specular term in the rendering equation. Next, we use the pixel-level feature maps to apply the rendering equation with split-sum approximation, yielding the final physically based rendering result. Here, x represents the homogeneous ray pas… view at source ↗
Figure 3
Figure 3. Qualitative comparisons on reflective scenes, including Bell, Luyu and Helmet. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons on Ref-Real dataset ( [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Inverse rendering with extracted mesh and indirect light. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparisons of the estimated environment maps on Glossy Synthetic [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Ablation study on PBR. GT 2DGS PBR Normal PBR Normal 3DGS [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 10
Figure 10. Figure 10: Effect of material-aware normal propa￾gation on normal maps at various steps. superior geometry reconstruction, as demonstrated by the high quality of the normal maps. This im￾provement is largely attributed to the carefully designed geometry optimization strategies d…
Figure 11
Figure 11. Figure 11: Relighting and editing on Luyu and Teapot with ground truth on the left. [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: Qualitative results of indirect lighting on Ref-Real dataset ( [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Per-scene qualitative comparisons of normals. [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: Qualitative decomposition results. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Beyond a Single Light: A Large-Scale Aerial Dataset for Urban Scene Reconstruction Under Varying Illumination

    cs.CV 2025-12 conditional novelty 7.0 of 10

    SkyLume contributes 10 real-world UAV urban regions captured at morning, noon, and afternoon with LiDAR-based ground truth, plus the Temporal Consistency Coefficient metric for cross-time albedo stability.

  2. AG$^2$aussian: Anchor-Graph Structured Gaussian Splatting for Instance-Level 3D Scene Understanding and Editing

    cs.CV 2025-08 conditional novelty 5.0 of 10

    An anchor-graph structured 3D Gaussians representation, with graph-based feature propagation and region growing, achieves cleaner instance-level object selection and better editing/simulation results than free-Gaussia...

  3. CTRL-GS: Cascaded Temporal Residue Learning for 4D Gaussian Splatting

    cs.CV 2025-05 conditional novelty 5.0 of 10

    CTRL-GS represents dynamic Gaussian scenes as cascaded video-segment-frame residuals, improving reconstruction quality over 4D-GS on several dynamic-view benchmarks.

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

    cs.CV 2025-08 conditional novelty 4.0 of 10

    GOGS rebuilds geometry of shiny objects using split-sum shading and monocular depth/normal priors, then decomposes materials with Monte Carlo sampling of the full rendering equation and a learnable specular compensation term.

Reference graph

Works this paper leans on

38 extracted references · 31 canonical work pages · cited by 4 Pith papers

  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]

    J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, and P. P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021

    work page 2021

  3. [3]

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022

    work page 2022

  4. [4]

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023

    work page 2023

  5. [5]

    M. Boss, R. Braun, V. Jampani, J. T. Barron, C. Liu, and H. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021 a

    work page 2021

  6. [6]

    M. Boss, V. Jampani, R. Braun, C. Liu, J. Barron, and H. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. NeurIPS, 2021 b

    work page 2021

  7. [7]

    M. Boss, A. Engelhardt, A. Kar, Y. Li, D. Sun, J. Barron, H. Lensch, and V. Jampani. Samurai: Shape and material from unconstrained real-world arbitrary image collections. NeurIPS, 2022

    work page 2022

  8. [8]

    Burley and W

    B. Burley and W. D. A. Studios. Physically-based shading at disney. In SIGGRAPH, 2012

    work page 2012

Show all 38 references
  1. [9]

    A. Chen, Z. Xu, A. Geiger, J. Yu, and H. Su. Tensorf: Tensorial radiance fields. In ECCV, 2022

    2022

  2. [10]

    Fridovich-Keil, A

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

    2022

  3. [11]

    J. Gao, C. Gu, Y. Lin, H. Zhu, X. Cao, L. Zhang, and Y. Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint, 2023

    2023

  4. [12]

    Huang, Z

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH, 2024

    2024

  5. [13]

    Jiang, J

    Y. Jiang, J. Tu, Y. Liu, X. Gao, X. Long, W. Wang, and Y. Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In CVPR, 2024

    2024

  6. [14]

    H. Jin, I. Liu, P. Xu, X. Zhang, S. Han, S. Bi, X. Zhou, Z. Xu, and H. Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023

    2023

  7. [15]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk \"u hler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023

    2023

  8. [16]

    Liang, H

    R. Liang, H. Chen, C. Li, F. Chen, S. Panneer, and N. Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In ICCV, 2023

    2023

  9. [17]

    Liang, Q

    Z. Liang, Q. Zhang, Y. Feng, Y. Shan, and K. Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In CVPR, 2024

    2024

  10. [18]

    Y. Liu, P. Wang, C. Lin, X. Long, J. Wang, L. Liu, T. Komura, and W. Wang. Nero: Neural geometry and brdf reconstruction of reflective objects from multiview images. ACM Trans. Graph., 2023

    2023

  11. [19]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020

    2020

  12. [20]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Commun. ACM, 2021

    2021

  13. [21]

    M \"u ller, A

    T. M \"u ller, A. Evans, C. Schied, and A. Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 2022

    2022

  14. [22]

    Munkberg, J

    J. Munkberg, J. Hasselgren, T. Shen, J. Gao, W. Chen, A. Evans, T. M \"u ller, and S. Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2022

    2022

  15. [23]

    Y. Shi, Y. Wu, C. Wu, X. Liu, C. Zhao, H. Feng, J. Liu, L. Zhang, J. Zhang, B. Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint, 2023

    2023

  16. [24]

    P. P. Srinivasan, B. Deng, X. Zhang, M. Tancik, B. Mildenhall, and J. T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021

    2021

  17. [25]

    C. Sun, M. Sun, and H. Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022

    2022

  18. [26]

    Verbin, P

    D. Verbin, P. Hedman, B. Mildenhall, T. Zickler, J. T. Barron, and P. P. Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022

    2022

  19. [27]

    P. Wang, L. Liu, Y. Liu, C. Theobalt, T. Komura, and W. Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint, 2021

    2021

  20. [28]

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 2004

    2004

  21. [29]

    L. Wu, S. Bi, Z. Xu, F. Luan, K. Zhang, I. Georgiev, K. Sunkavalli, and R. Ramamoorthi. Neural directional encoding for efficient and accurate view-dependent appearance modeling. In CVPR, 2024

    2024

  22. [30]

    Y. Yao, J. Zhang, J. Liu, Y. Qu, T. Fang, D. McKinnon, Y. Tsin, and L. Quan. Neilf: Neural incident light field for physically-based material estimation. In ECCV, 2022

    2022

  23. [31]

    K. Ye, Q. Hou, and K. Zhou. 3d gaussian splatting with deferred reflection. In SIGGRAPH, 2024

    2024

  24. [32]

    Z. Yu, T. Sattler, and A. Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint, 2024

    2024

  25. [33]

    Zhang, Y

    J. Zhang, Y. Yao, S. Li, J. Liu, T. Fang, D. McKinnon, Y. Tsin, and L. Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023

    2023

  26. [34]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018

    2018

  27. [35]

    Zhang, P

    X. Zhang, P. P. Srinivasan, B. Deng, P. Debevec, W. T. Freeman, and J. T. Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021

    2021

  28. [36]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  29. [37]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  30. [38]

    However, reflective object reconstruction remains challenging, lacking a proper solution to achieve real-time, high-quality rendering while accommodating inter-reflection

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

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