Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

EnvGS: Modeling View-Dependent Appearance with Environment Gaussian

T0 review · 4 major / 5 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 EnvGS argues that explicit 3D Gaussian primitives, ray-traced along reflected rays, outperform environment maps for high-frequency and near-field reflections while staying real-time.

desk verdict EnvGS is a genuinely new representation for real-time reflections, but its headline quality margin over baselines is partly built on extra supervision those baselines never got. read the letter →

arxiv 2412.15215 v2 pith:77HC6Q6M submitted 2024-12-19 cs.CV

Tao Xie , Xi Chen , Zhen Xu , Yiman Xie , Yudong Jin , Yujun Shen , Sida Peng , Hujun Bao
show 1 more author
Xiaowei Zhou
This is my paper · ORCID
classification cs.CV
keywords 3DGaussiansplattingnovelviewsynthesisspecularreflectionraytracingnear-fieldenvironmentmapreal-timerenderingview-dependentappearance
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

EnvGS claims that reflections are better modeled as their own explicit set of 3D Gaussian primitives than as an environment map. The method separates a scene into base Gaussians that describe geometry and diffuse appearance and environment Gaussians that live in 3D space and are ray-traced along each reflected viewing direction. Because the reflection primitives occupy real positions, the representation can reproduce near-field reflections and high-frequency detail that distant-light environment maps miss. The paper reports that this runs in real time and matches the quality of a much slower state-of-the-art method.

What carries the argument

The environment Gaussian: a second set of 2D Gaussian primitives that carry reflection appearance and sit at explicit 3D positions, rendered by a custom differentiable ray tracer built on hardware-accelerated ray tracing. Each Gaussian is converted to two triangles and inserted into a bounding volume hierarchy; rays are traced against it in chunks, sorted by depth in a k-buffer, and integrated with the volume-rendering equation. The backward pass re-casts rays in the same order and includes gradients with respect to the ray origin and direction, which is what lets base geometry and environment appearance be optimized together. The blending weight on each base Gaussian decides how much reflected color replaces base color at each pixel.

What would settle it

Train the strongest environment-map baseline with exactly the same monocular-normal loss, perceptual loss, training schedule, and runtime budget as EnvGS; if its PSNR and LPIPS on the real-world reflective scenes then match or beat EnvGS, the claim that Gaussian primitives outperform environment maps for reflections is falsified. A second check: render a scene containing a known nearby reflector at a held-out view and compare near-field reflection sharpness directly between the environment-map variant and EnvGS.

Watch

Extended reading notes

Core claim

The core claim, stated in the authors' terms, is that explicit Gaussian primitives in 3D space—environment Gaussians—are a superior representation for view-dependent appearance than environment maps. Each pixel's reflected ray is traced through a bounding volume hierarchy built from these Gaussians, and the traced color is blended with a base-scene color produced by rasterized base Gaussians. The normal used to form the reflection direction, the surface position, and the blending weight all come from the base Gaussian render, so geometry and reflection appearance can be optimized jointly through gradients that flow back through the ray origin and direction. On real and synthetic benchmarks the paper reports the best quality among real-time methods and results competitive with a non-real-time baseline while roughly 100 times faster, and the ablation replacing environment Gaussians with an environment map confirms the representation is the source of the near-field and high-frequency gains.

Load-bearing premise

The paper credits its quality gains to the new environment-Gaussian representation, but EnvGS is also trained with extra monocular normal and perceptual-loss supervision that the compared real-time baselines do not receive; if that extra supervision does the heavy lifting, the representation-specific claim is not established.

Editorial extensions

If this is right

  • Real-time novel-view synthesis of reflective real-world scenes becomes feasible without sacrificing high-frequency specular detail.
  • Near-field reflections—objects reflecting nearby occluders—are representable without an explicit distant-lighting assumption.
  • Environment maps can be replaced by a learnable explicit 3D field of primitives in any Gaussian-splatting pipeline that needs view-dependent appearance.
  • Joint optimization of reflection primitives and base geometry, enabled by ray-origin and ray-direction gradients, should improve both geometry and reflection fidelity.

Reading between the lines

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

  • Because the compared real-time baselines were not given the same monocular-normal and perceptual-loss supervision, a strict reading of the experiments leaves open how much of the gain is representation-specific; feeding baseline environment-map variants the same extra losses would isolate the effect.
  • The environment Gaussians could double as an editable reflection field: moving, deleting, or re-coloring a primitive would change reflections in a spatially localized way, which environment maps cannot do.
  • The same ray-tracing tracer could be extended to secondary and higher-order bounces, or to refraction by tracing transmitted directions instead of reflected ones—the paper itself lists transparency and refraction as a limitation.
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 / 5 minor

Summary. EnvGS introduces an explicit set of 3D Gaussian primitives, called environment Gaussians, to represent reflections separately from the base 2D Gaussian scene geometry. The method rasterizes the base Gaussians to obtain surface position, normal, base color, and a per-Gaussian blending weight, then casts reflected rays through a custom differentiable OptiX-based ray tracer to render the environment Gaussians, and finally blends the base and reflection colors. Training jointly optimizes base and environment Gaussians with photometric, normal-consistency, monocular-normal, and VGG perceptual losses. The paper reports state-of-the-art or competitive rendering quality among real-time methods on Ref-Real, NeRF-Casting Shiny Scenes, self-captured scenes, Shiny Blender, and Mip-NeRF 360, at roughly 26 FPS on an RTX 4090, and claims to be the first method to achieve real-time photorealistic specular reflections in real-world scenes.

Significance. If the central claim holds, EnvGS is a practically valuable contribution: it offers a real-time, differentiable ray-tracing pipeline for 2DGS with explicit 3D reflection primitives, and the paper includes a thorough set of ablations, per-scene breakdowns, runtime analyses, and a public code release. The design choice of representing reflections as Gaussian primitives rather than environment maps is well motivated and the internal ablations (Table 4, Fig. 8) support the value of that representation within the full pipeline. However, the headline comparison against external baselines is confounded by auxiliary supervision, and the region-based evaluation uses self-derived masks. These issues are fixable but currently prevent a clean attribution of the reported margins to the environment Gaussian representation alone.

major comments (4)
  1. [Sec. 5.3, Eq. (11), Table 1] The claimed superiority of EnvGS over real-time baselines (3DGS, 2DGS, GaussianShader, 3DGS-DR) is confounded by supervision: EnvGS is trained with the monocular-normal loss L_mono of Eq. (9) and the VGG perceptual loss L_perc of Eq. (10), while the baselines are not given either signal. The reported gaps, e.g., 24.617 PSNR versus 23.522 for 3DGS-DR on Ref-Real/NeRF-Casting, therefore conflate the effect of the new representation with the effect of stronger external supervision. Please re-run the baselines with the same auxiliary losses (or with equivalent supervision), or at least provide a controlled comparison in which only the environment representation is varied and all other losses are fixed.
  2. [Sec. 5.4, Table 4] The 'w/ environment map' ablation is informative but does not fully resolve the attribution question. This variant replaces environment Gaussians with an environment map while retaining L_mono and L_perc, so it isolates the representation within EnvGS's own pipeline. However, it does not tell the reader how much of the margin over the published baselines in Table 1 comes from the representation versus from the auxiliary losses. Please add a variant that uses an environment map together with monocular-normal and perceptual losses and compare it directly against both EnvGS and the baselines; this would make the representation's contribution to the state-of-the-art claim unambiguous.
  3. [Supplementary A.1, Table 2] The reflective-region metrics are evaluated under masks that are generated from EnvGS's own trained model: the paper trains EnvGS, exports its Gaussians, removes non-foreground points, renders an accumulated alpha map, and binarizes it. Applying this self-derived mask to all methods creates a systematic bias in favor of EnvGS, because the mask follows EnvGS's geometry and coverage. Please use fixed, method-independent annotations (for example, manually labeled or segment-based masks defined in a common coordinate frame) and report the masked metrics under identical masks for every compared method. Without this, the 'over 1.0 PSNR improvement on reflective regions' claim in Supplementary A.1 is not fully trustworthy.
  4. [Table 1 and Table 7] The headline aggregate metrics are reported without error bars or statistical significance, and the per-scene breakdown in Table 7 shows that the advantage over baselines is not uniform across scenes (for example, on 'spheres' EnvGS is 22.949 PSNR versus 23.040 for NeRF-Casting, and on 'sedan' the margin over 3DGS is about 0.9 PSNR). Please report the number of seeds/runs and standard deviations, or perform a per-scene paired significance test, so that claims such as 'considerably surpasses existing real-time methods' are supported by the actual variability of the measurements.
minor comments (5)
  1. [Sec. 4.1, Eq. (4)] There is a typo in the text: 'we compute the reflection direction d_ref based on the camera ray direction d_ref' should refer to the camera ray direction d_cam, matching Eq. (4).
  2. [Sec. 5.3] The baseline citation 'Zip-NeRF [13]' appears to be a numbering error; Zip-NeRF is reference [4], not [13].
  3. [Sec. 5.3 and Table 5] The concurrent work Ref-GS [54] is mentioned in Section 2 but is not included in any quantitative comparison. If the code is available, adding it to Table 1 or Table 8 would strengthen the state-of-the-art claim; if not, a brief note in Section 5.3 explaining the omission would help.
  4. [Sec. 4.2] The statement that 'r is set to 3 representing three times the sigma range' is slightly unclear when combined with the sign function in V_local; please clarify that this defines a 3-sigma bounding square and that the two triangles share the four corners.
  5. [Sec. 6] The limitation on transparent and refractive materials is stated only briefly; since the ray tracer already supports arbitrary secondary rays, a sentence on whether refraction could be handled by the same framework would help readers understand the scope of the limitation.

Circularity Check

1 steps flagged · score 4.0 of 10

Central representation claim is non-circular (held-out benchmarks plus a clean env-map ablation), but the supplementary reflective-region metric is self-referential because the mask is generated from EnvGS's own Gaussians.

  1. other [Supplementary A.1 (mask generation and Table 2)]
    "The reflective masks mentioned above are obtained through the following steps. First, we train our EnvGS on each scene, then export the trained Gaussian and remove the Gaussian points in 3D space except for those in the foreground reflective region. We render the remaining Gaussian to generate an accumulated alpha map. Finally, we binarize this alpha map to obtain the foreground reflective masks."

    The reflective-region comparison in Table 2 is evaluated only inside masks that are outputs of the method being claimed superior. Because the mask is the binarized alpha map of EnvGS's own exported Gaussians, the statement in A.1 that EnvGS gives 'a significant improvement of over 1.0 PSNR on the reflective region' is measured in a region defined by EnvGS itself rather than by an independent annotation or by a baseline-derived mask. Applying the same mask to all methods prevents the scores from being algebraically forced, but it makes the region-specific subclaim self-referential and weakens the attribution of the reflective-region gain to the environment-Gaussian representation.

full rationale

The paper's central claim is not derivationally circular. Environment Gaussians are optimized against training views and evaluated on held-out novel views of external datasets (Table 1, Tables 7-9), and the key representation question is tested by the 'w/ environment map' ablation, which swaps the environment representation while keeping the base Gaussians, losses, and tracer unchanged; that is the correct non-circular control. There is no load-bearing self-citation: the authors' own prior work (SfM initializer [11], EasyVolcap [47]) is incidental infrastructure, and the cited ray-tracing techniques [29] are external. The reported margins over real-time baselines are attributionally confounded because EnvGS alone trains with monocular normal estimates (Eq. 9) and VGG perceptual loss (Eq. 10), which the baselines do not receive; this is a fairness/correctness gap rather than a circularity and should be resolved by rerunning baselines with the same auxiliary supervision. The one self-referential element is the supplementary reflective-region mask described in A.1, which injects a localized circularity into the region-specific metrics (Table 2) without undermining the global held-out evaluation. Score 4 reflects that single non-central, self-referential evaluation step while acknowledging the independent content of the central claim.

Assumptions & free parameters 7 free parameters · 6 assumptions · 2 invented entities

The central claim rests on the standard 2DGS machinery, a ray-traced volume-integration approximation, and two external supervision priors (StableNormal normals, VGG perceptual loss). The hand-chosen weights and the two-triangle Gaussian approximation are the most fragile design choices, though none appears to invalidate the method. The environment Gaussian is an invented representation with no independent physical evidence, but that is expected for an appearance model.

free parameters (7)
  • lambda_1 (normal consistency weight) = 0.04
    Hand-chosen loss weight in Eq. (11). No sensitivity analysis is reported in the main text.
  • lambda_2 (monocular normal weight) = 0.01
    Hand-chosen weight for the StableNormal supervision in Eq. (11). This external signal is not used by the compared baselines.
  • lambda_3 (perceptual loss weight) = 0.01
    Hand-chosen weight for the VGG perceptual loss in Eq. (11).
  • k (ray tracing k-buffer size) = 16
    Empirically chosen trade-off between traversal counts and sorted Gaussians per ray, stated in Sec. 4.2.
  • N (sub-grid resolution for environment Gaussian initialization) = 32
    Set in Sec. 4.3 to partition the scene bounding box; no ablation is shown for this value.
  • K (environment Gaussians per sub-grid) = 5
    Random initialization count per grid cell in Sec. 4.3.
  • environment Gaussian pruning threshold = top 630k by rendering weight
    Reported in Supplementary C; the pruning policy is inherited from 2DGS with a hard cap.
assumptions (6)
  • domain assumption The 2D Gaussian Splatting representation and its rasterization/volume rendering equations are valid and accurate enough for the base scene.
    The method builds entirely on 2DGS [13]; the base Gaussian rendering and normal extraction rely on its correctness (Sec. 3 and Eq. 2).
  • standard math The volume rendering integral in Eq. (2) is the correct model for integrating Gaussian contributions along a ray.
    This is the standard alpha-compositing approximation used throughout NeRF/Gaussian splatting literature, cited to Kajiya and Herzen [16].
  • ad hoc to paper Each 2D Gaussian can be faithfully approximated for ray tracing by two triangles spanning a 3-sigma bounding square.
    Stated in Sec. 4.2: the four bounding vertices map to Vworld and form two triangles. This truncation is a modeling choice that could affect accuracy for large or thin Gaussians.
  • domain assumption Monocular normal estimates from StableNormal [49] provide a reliable geometric prior for reflective scenes.
    The Lmono loss in Eq. (9) treats these externally predicted normals as pseudo-ground truth. If the prior is biased, the reconstructed geometry and reflections could be skewed.
  • domain assumption The VGG-16 perceptual features used in Eq. (10) are a valid proxy for perceived image quality.
    Standard practice in image synthesis; the authors rely on this pretrained network without re-validating it for reflection rendering.
  • standard math The reflection direction formula d_ref = d_cam - 2(d_cam dot n)n is the correct specular reflection of the viewing ray around the surface normal.
    This is a basic vector reflection identity, used in Sec. 4.1, Eq. (4).
invented entities (2)
  • Environment Gaussian primitives
    purpose: An explicit 3D set of 2D Gaussian blobs placed in the scene bounding box to represent incoming radiance along reflection rays; traced instead of sampling an environment map.
    These are a learned internal representation of the scene's reflection field, not a physically measured entity. They are optimized per scene and have no observable handle outside the method's own training and rendering.
  • Per-base-Gaussian blending weight beta
    purpose: A scalar per base Gaussian that controls how much the reflection color contributes to the final pixel color in Eq. (5).
    An internal parameter learned during optimization; has no independent physical meaning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EnvGS: Modeling View-Dependent Appearance with Environment Gaussian." pith.science (2026). https://pith.science/paper/77HC6Q6M

@misc{pith2026241215215,
  author       = {Pith},
  title        = {Pith review of: EnvGS: Modeling View-Dependent Appearance with Environment Gaussian},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/77HC6Q6M}},
  note         = {Machine review of arXiv:2412.15215}
}
read the original abstract

Reconstructing complex reflections in real-world scenes from 2D images is essential for achieving photorealistic novel view synthesis. Existing methods that utilize environment maps to model reflections from distant lighting often struggle with high-frequency reflection details and fail to account for near-field reflections. In this work, we introduce EnvGS, a novel approach that employs a set of Gaussian primitives as an explicit 3D representation for capturing reflections of environments. These environment Gaussian primitives are incorporated with base Gaussian primitives to model the appearance of the whole scene. To efficiently render these environment Gaussian primitives, we developed a ray-tracing-based renderer that leverages the GPU's RT core for fast rendering. This allows us to jointly optimize our model for high-quality reconstruction while maintaining real-time rendering speeds. Results from multiple real-world and synthetic datasets demonstrate that our method produces significantly more detailed reflections, achieving the best rendering quality in real-time novel view synthesis. The code is available at https://zju3dv.github.io/envgs.

Figures

Figures reproduced from arXiv: 2412.15215 by the authors.

Figure 1
Figure 1. Photorealistic, real-time rendering of real-world scenes with complex reflections. Our proposed EnvGS outperforms prior works in capturing complex reflection effects, especially near-field reflections and high-frequency details while maintaining real-time rendering speed. Please see our supplementary video for better visualizations. Abstract Reconstructing complex reflections in real-world scenes from 2D images is e… view at source ↗
Figure 2
Figure 2. Overview of EnvGS. The rendering process begins by rasterizing the base Gaussian to obtain per-pixel normals, base colors, and blending weights. Next, we render the environment Gaussian in the reflection direction using our ray-tracing-based Gaussian renderer to capture the reflection colors. Finally, we combine the reflection and base colors for the final output. We jointly optimize the environment Gaussian and bas… view at source ↗
Figure 3
Figure 3. Visualization of reflection and base color. Our method successfully reconstructs near-field and distant reflections using the environment Gaussian instead of baking into the base color. ray-primitive intersections, we need to convert each 2D Gaussian into a geometric primitive compatible with GPU processing and insert it into a bounding volume hier￾archy (BVH). In light of this, we propose to represent each 2D Gauss… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on real scenes. Our method significantly improves rendering quality over previous approaches, partic￾ularly in producing more detailed reflections. Zoom in for more details. In addition to the L1 image loss and D-SSIM loss Lssim employed by 3DGS …
Figure 5
Figure 5. Figure 5: Ablation study of proposed components on the Ref-Real dataset [37]. Removing either the monocular normal constraint or the joint optimization of base and environment Gaussians results in noisy geometry and inaccurate reflection reconstruction. The “w/ environment map” …
Figure 6
Figure 6. Figure 6: Qualitative comparison on reflective foreground and near-field reflection regions. We also provide visualizations of the foreground and near-field region mask we annotated. The reflective masks mentioned above are obtained through the following steps. First, we train o…
Figure 7
Figure 7. Figure 7: Qualitative comparison with object-level methods. A.5. Additional Baselines We also compare our method with object-baselines includ￾ing ENVIDR [22] and NDE [45]. While object-level meth￾ods perform well on synthetic data, they often struggle with real-world scenes and …
Figure 9
Figure 9. Figure 9: Qualitative comparison on real scenes. Our method significantly improves rendering quality over previous approaches, partic￾ularly in producing more detailed reflections. Zoom in for more details. vironment Gaussian, and the final color is the blending of the two. Base…
Figure 10
Figure 10. Figure 10: Qualitative comparison on synthetic scenes. Despite being designed for robustness on real-world data, our method effectively reconstructs accurate distant specular reflections and effectively captures near-field reflections caused by self-occlusions. sian primitives u…
Figure 11
Figure 11. Figure 11: Qualitative comparisons of normal produced by different methods. the derivatives w.r.t. the ray origin and direction: \begin {aligned} \frac {\mathrm {d} \mathcal {L}}{\mathrm {d} \mathbf {o}} & = \frac {\mathrm {d} \mathcal {L}}{\mathrm {d} \mathbf {x}_i} \frac {\mat…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Decomposing Densification in Gaussian Splatting for Faster 3D Scene Reconstruction

    cs.CV 2025-07 conditional novelty 5.0 of 10

    A split-then-clone densification schedule with energy-guided multi-resolution training roughly halves 3D Gaussian Splatting training time while keeping reconstruction quality.

  2. 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

54 extracted references · 33 canonical work pages · cited by 2 Pith papers

  1. [1]

    Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) ,

    work page

  2. [2]

    Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P

    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, 2021. 2

    2021

  3. [3]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 7, 1, 8

    work page 2022

  4. [4]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid- based neural radiance fields. ICCV, 2023. 2, 7, 1, 6, 8

    work page 2023

  5. [5]

    Gs3: Efficient relighting with triple gaussian splatting

    Zoubin Bi, Yixin Zeng, Chong Zeng, Fan Pei, Xiang Feng, Kun Zhou, and Hongzhi Wu. Gs3: Efficient relighting with triple gaussian splatting. In SIGGRAPH Asia 2024 Confer- ence Papers, pages 1–12, 2024. 3

    work page 2024

  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 (ECCV), 2022. 2

    2022

  7. [7]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2

    2022

  8. [8]

    Pisr: Polarimetric neural implicit surface reconstruction for textureless and specular objects

    Chen Guangcheng, He Yicheng, He Li, and Zhang Hong. Pisr: Polarimetric neural implicit surface reconstruction for textureless and specular objects. In Proceedings of the Euro- pean Conference on Computer Vision (ECCV), 2024. 2

    work page 2024

Show all 54 references
  1. [9]

    The growth of image-related three dimensional re- construction techniques in deep learning-driven era: a criti- cal summary, 2023

    Yang Hang, Chen Rui, An Shipeng, Wei Hao, and Zhang Heng. The growth of image-related three dimensional re- construction techniques in deep learning-driven era: a criti- cal summary, 2023. 2

    2023

  2. [10]

    Lotus: Diffusion-based visual foundation model for high-quality dense prediction

    Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and Ying- Cong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 5

    2024 arXiv

  3. [11]

    Detector-free struc- ture from motion

    Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free struc- ture from motion. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 21594–21603, 2024. 5

    2024

  4. [12]

    A survey on point cloud completion algorithms for deep learning, 2025

    Fuyuan Hu, Chenlu Li, Tao Zhou, Hongfu Cheng, and Min- ming Gu. A survey on point cloud completion algorithms for deep learning, 2025. 2

    2025

  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 ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3, 4, 5, 6, 7, 8

    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 arXiv:2311.17977, 2023. 3, 7, 6, 8

    2023 arXiv

  7. [15]

    Gaussian- shader: 3d gaussian splatting with shading functions for re- flective surfaces

    Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xi- aoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian- shader: 3d gaussian splatting with shading functions for re- flective surfaces. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , p...

    2024

  8. [16]

    Ray tracing volume densities

    James T Kajiya and Brian P V on Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165– 174, 1984. 2

    1984

  9. [17]

    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 Transactions on Graphics, 42 (4), 2023. 1, 3, 5, 6, 7, 2, 8

    2023

  10. [18]

    3d gaussian splat- ting with deferred reflection

    Ye Keyang, Hou Qiming, and Zhou Kun. 3d gaussian splat- ting with deferred reflection. 2024. 3, 6, 7, 1

    2024

  11. [19]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 6

    2014 arXiv

  12. [20]

    Deep learning-based foveated rendering in 3d space: a re- view, 2024

    Yingqun Li, Xiao Hu, Xiang Xu, Yanning Xu, and Lu Wang. Deep learning-based foveated rendering in 3d space: a re- view, 2024. 2

    2024

  13. [21]

    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 IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2023. 2

    2023

  14. [22]

    Envidr: Im- plicit differentiable renderer with neural environment light- ing

    Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Sel- vakumar Panneer, and Nandita Vijaykumar. Envidr: Im- plicit differentiable renderer with neural environment light- ing. arXiv preprint arXiv:2303.13022, 2023. 2, 7

    2023 arXiv

  15. [23]

    Prompting depth anything for 4k resolution accurate metric depth estimation

    Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji- aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth anything for 4k resolution accurate metric depth estimation. 2025. 2

    2025

  16. [24]

    Neural sparse voxel fields

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS,

  17. [25]

    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. 2

    2023

  18. [26]

    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 Computer Vision and Pattern Recognition, pages 20654–20664, 2024. 3

    2024

  19. [27]

    Specnerf: Gaussian directional encoding for spec- ular reflections, 2023

    Li Ma, Vasu Agrawal, Haithem Turki, Changil Kim, Chen Gao, Pedro Sander, Michael Zollh ¨ofer, and Christian Richardt. Specnerf: Gaussian directional encoding for spec- ular reflections, 2023. 2

    2023

  20. [28]

    Srinivasan, Matthew Tancik, Jonathan T

    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

  21. [29]

    3D Gaussian Ray Tracing: Fast tracing of particle scenes

    Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ric- cardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian Ray Tracing: Fast tracing of particle scenes. ACM Transactions on Graphics and SIGGRAPH Asia, 2024. 4, 5, 6

    2024

  22. [30]

    Instant neural graphics primitives with a multires- 9 olution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a multires- 9 olution hash encoding. ACM Trans. Graph. , 41(4):102:1– 102:15, 2022. 2

    2022

  23. [31]

    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. Acm transac- tions on graphics (tog), 29(4):1–13, 2010. 2, 4

    2010

  24. [32]

    Pytorch: An im- perative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im- perative style, high-performance deep learning library. Ad- vances in neural information processing systems ...

    2019

  25. [33]

    Very deep convo- lutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6

    2014 arXiv

  26. [34]

    Photo tourism: exploring photo collections in 3d

    Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835–846. 2006. 5

    2006

  27. [35]

    Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction

    Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2

    2022

  28. [36]

    3igs: Factorised tenso- rial illumination for 3d gaussian splatting

    Zhe Jun Tang and Tat-Jen Cham. 3igs: Factorised tenso- rial illumination for 3d gaussian splatting. arXiv preprint arXiv:2408.03753, 2024. 3, 7

    2024 arXiv

  29. [37]

    Barron, and Pratul P

    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. CVPR, 2022. 2, 3, 7, 8, 1, 6

    2022

  30. [38]

    Nerf-casting: Improved view-dependent ap- pearance with consistent reflections

    Dor Verbin, Pratul P Srinivasan, Peter Hedman, Ben Milden- hall, Benjamin Attal, Richard Szeliski, and Jonathan T Barron. Nerf-casting: Improved view-dependent ap- pearance with consistent reflections. arXiv preprint arXiv:2405.14871, 2024. 2, 7, 1, 6, 8

    2024 arXiv

  31. [39]

    Unisdf: Unifying neural representations for high- fidelity 3d reconstruction of complex scenes with reflections

    Fangjinhua Wang, Marie-Julie Rakotosaona, Michael Niemeyer, Richard Szeliski, Marc Pollefeys, and Federico Tombari. Unisdf: Unifying neural representations for high- fidelity 3d reconstruction of complex scenes with reflections. In NeurIPS, 2024. 2, 7, 6, 8

    2024

  32. [40]

    Review of multi-source fusion slam: current status and challenges, 2022

    Jinke Wang, Xingxing Zuo, Xiangrui Zhao, Jiajun Lyu, and Yong Liu. Review of multi-source fusion slam: current status and challenges, 2022. 2

    2022

  33. [41]

    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. NeurIPS, 2021. 2

    2021

  34. [42]

    Neus2: Fast learning of neural implicit surfaces for multi-view recon- struction

    Yiming Wang, Qin Han, Marc Habermann, Kostas Dani- ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view recon- struction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2

    2023

  35. [43]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 6, 7

    2004

  36. [44]

    A hardware architec- ture for surface splatting

    Tim Weyrich, Simon Heinzle, Timo Aila, Daniel B Fasnacht, Stephan Oetiker, Mario Botsch, Cyril Flaig, Simon Mall, Kaspar Rohrer, Norbert Felber, et al. A hardware architec- ture for surface splatting. ACM Transactions on Graphics (TOG), 26(3):90–es, 2007. 3

    2007

  37. [45]

    Neural directional encoding for efficient and accurate view- dependent appearance modeling

    Liwen Wu, Sai Bi, Zexiang Xu, Fujun Luan, Kai Zhang, Iliyan Georgiev, Kalyan Sunkavalli, and Ravi Ramamoorthi. Neural directional encoding for efficient and accurate view- dependent appearance modeling. In CVPR, 2024. 7, 2

    2024

  38. [46]

    Physically based differentiable rendering: a survey, 2024

    Jiankai Xing and Kun Xu. Physically based differentiable rendering: a survey, 2024. 2

    2024

  39. [47]

    Easyvolcap: Accelerating neural volumetric video research

    Zhen Xu, Tao Xie, Sida Peng, Haotong Lin, Qing Shuai, Zhiyuan Yu, Guangzhao He, Jiaming Sun, Hujun Bao, and Xiaowei Zhou. Easyvolcap: Accelerating neural volumetric video research. In SIGGRAPH Asia 2023 Technical Commu- nications, pages 1–4. 2023. 6

    2023

  40. [48]

    Multiview neu- ral surface reconstruction by disentangling geometry and ap- pearance

    Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neu- ral surface reconstruction by disentangling geometry and ap- pearance. Advances in Neural Information Processing Sys- tems, 33, 2020. 2

    2020

  41. [49]

    Stablenormal: Reducing diffusion variance for stable and sharp normal.ACM Transactions on Graphics,

    Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal.ACM Transactions on Graphics,

  42. [50]

    Lightweight visual-based localization technology, 2024

    Hanqiao Ye, Yangdong Liu, and Shuhan Shen. Lightweight visual-based localization technology, 2024. 2

    2024

  43. [51]

    3d gaussian splat- ting with deferred reflection

    Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splat- ting with deferred reflection. InACM SIGGRAPH 2024 Con- ference Papers, pages 1–10, 2024. 2, 4, 7, 1, 6, 8

    2024

  44. [52]

    Mip-splatting: Alias-free 3d gaussian splat- ting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splat- ting. Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2024. 3

    2024

  45. [53]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 6, 7

    2018

  46. [54]

    w/ env. map 128

    Youjia Zhang, Anpei Chen, Yumin Wan, Zikai Song, Jun- qing Yu, Yawei Luo, and Wei Yang. Ref-gs: Direc- tional factorization for 2d gaussian splatting. arXiv preprint arXiv:2412.00905, 2024. 3 10 EnvGS: Modeling View-Dependent Appearance with Environment Gaussian Supplementary ...

    2024 arXiv

Pith tools

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