Pith. sign in

REVIEW 4 major objections 6 minor 75 references

NeRF Is a Valuable Assistant for 3D Gaussian Splatting

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that jointly training a NeRF and a 3D Gaussian splatting model, with shared hash features, residual offsets, and cross-branch constraints, yields higher-quality 3D scene rendering than either representation alone while…

desk verdict A solid, well-ablated joint NeRF+3DGS training scheme whose 'state-of-the-art' claim is undercut by the missing direct comparison with RadSplat, the closest prior hybrid. read the letter →

arxiv 2507.23374 v1 pith:QEZMRBD4 submitted 2025-07-31 cs.CV

Shuangkang Fang , I-Chao Shen , Takeo Igarashi , Yufeng Wang , ZeSheng Wang , Yi Yang , Wenrui Ding , Shuchang Zhou This is my paper
classification cs.CV
keywords 3DGaussianSplattingNeuralRadianceFieldshybridscenerepresentationjointoptimizationhashfeaturegridnovelviewsynthesissparse-viewrenderingresidualvectors
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

NeRF-GS is a training framework that runs a NeRF and a 3D Gaussian splatting model together on the same scene, sharing one multi-resolution hash feature grid between them. The paper claims this collaboration fixes three known 3DGS weaknesses: sensitivity to how Gaussians are initialized, limited spatial awareness, and weak correlations between Gaussians. After pretraining a hash-based NeRF, the method initializes Gaussians along edge-weighted rays, gives each Gaussian its own residual feature and position offset, and aligns the two branches along rays through high-opacity Gaussians. On four benchmark suites, the GS branch of the resulting model reports higher PSNR, SSIM, and LPIPS than vanilla 3DGS and several recent hybrids, while remaining real-time and using fewer Gaussians. The authors read this as evidence that NeRF and 3DGS are complementary representations rather than competing ones.

What carries the argument

The load-bearing mechanism is a dual-branch training loop built around a shared multi-resolution hash feature grid. NeRF provides continuous spatial queries that support edge-based Gaussian initialization and NeRF-assisted adaptive growth of Gaussians in regions 3DGS would otherwise miss; the GS branch decodes color, opacity, scale, and rotation from the shared features plus per-Gaussian residual vectors $\Delta f$ and $\Delta p$; and GS-Rays, defined as rays from the camera through high-opacity Gaussian centers, focus NeRF rendering and enable opacity/RGB alignment losses that regularize both branches. After training, the GS branch can be detached and run without the NeRF branch, preserving real-time rendering.

What would settle it

Retrain the full NeRF-GS pipeline with the NeRF pretraining replaced by a randomly initialized hash grid, keeping all other components identical; if the GS branch's PSNR on Mip-NeRF360 stays within noise of the reported 28.32 dB, then the continuous prior is not load-bearing.

Watch

Extended reading notes

Core claim

NeRF-GS trains a hash-based NeRF and a 3D Gaussian splatting model jointly on the same scene, with one multi-resolution hash feature grid shared between them. The paper claims that this dual-branch collaboration directly addresses three limitations of vanilla 3DGS: sensitivity to Gaussian initialization, limited spatial awareness, and weak inter-Gaussian correlation. Concretely, a pretrained NeRF produces edge-weighted rays whose median depths become candidate Gaussian positions; each Gaussian then reads shared features and decodes its attributes through a small MLP, augmented by a per-Gaussian residual feature vector and a residual position offset that absorb the representational gap between the two branches. During joint optimization, NeRF renders only along GS-Rays, rays through high-opacity Gaussians, and the branches are aligned with opacity and color losses while residual regularization keeps them close. The paper reports state-of-the-art or on-par results across PSNR, SSIM, and LPIPS on Mip-NeRF360, Tanks&Temples, DeepBlending, and Blender, with the GS branch retaining real-time inference and using fewer Gaussians than vanilla 3DGS.

Load-bearing premise

The framework's gains rest on the pretrained NeRF providing an accurate enough density field that the edge-based Gaussian positions and shared hash features are a reliable starting point for the joint optimization.

Editorial extensions

If this is right

  • The GS branch of NeRF-GS can be separated after training and still deliver the reported rendering quality, meaning hybrid training does not sacrifice the real-time property of 3DGS.
  • On sparse-view benchmarks (12 and 8 views), NeRF-GS stays competitive with SplatFields, a method purpose-built for sparse-view reconstruction, suggesting the joint regularization acts as a strong prior.
  • NeRF-assisted Gaussian growth reconstructs regions missing from the initial point cloud, reducing 3DGS's sensitivity to poor initialization.
  • Joint training also improves the NeRF branch relative to an equally trained standalone NeRF, indicating the gains are mutual rather than one-sided.
  • The reported results beat several recent NeRF-3DGS hybrids (VDGS, Hash-GS), implying that systematically sharing features and modeling inter-branch differences matters more than adding NeRF attributes to 3DGS.

Reading between the lines

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

  • A natural extension would replace the NeRF branch with another continuous implicit field (e.g., a signed distance function or triplane) to test whether the spatial regularizer, rather than NeRF specifically, is what drives the improvement.
  • Because the two branches share a hash grid, an edit applied in the NeRF branch's feature space could propagate to the rendered Gaussians, opening a path to editable hybrid scene representations.
  • The per-scene breakdown shows larger gains on some scenes than others, so a testable hypothesis is that NeRF assistance matters most for texture-rich or geometrically discontinuous content.
  • The sparse-view results suggest the method could be stressed further at very few views (e.g., 3–4), where the residual vectors may either prevent overfitting or become a new weak point.
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 NeRF-GS, a hybrid framework that jointly trains a hash-based NeRF branch and a 3D Gaussian Splatting (3DGS) branch. The authors identify three design components: a sharing mechanism that initializes Gaussian positions from NeRF depth and shares hash-grid features, residual feature/position vectors that adapt NeRF-shared information to the Gaussian branch, and a joint optimization scheme with GS-Rays, opacity/RGB alignment losses, and NeRF-assisted Gaussian growth. Experiments on Mip-NeRF360, Tanks&Temples, DeepBlending, and Blender report improvements over vanilla 3DGS and several variants in PSNR, SSIM, and LPIPS, while preserving real-time rendering from the GS branch at inference time. The central claim is that NeRF and 3DGS are complementary, and that NeRF-GS achieves state-of-the-art performance.

Significance. If the reported gains are reproducible, the paper makes a useful empirical contribution: it demonstrates a concrete training scheme in which a NeRF prior improves 3DGS quality, especially in sparse-view settings, and it provides ablations isolating the contributions of feature sharing, residual vectors, joint losses, and NeRF-assisted growth. The supplementary analysis of NeRF-assisted growth on deliberately incomplete initializations and the mutual-promotion experiment for the NeRF branch are valuable and go beyond a simple combination of two off-the-shelf models. However, the strength of the claims is currently limited by the absence of a direct comparison with the closest hybrid baseline (RadSplat), an arithmetic inconsistency in Table 4, and an overstatement of the sparse-view results in the abstract.

major comments (4)
  1. [Sec. 4.1, Table 1, Abstract] RadSplat [46] is discussed at length in Sec. 4.1 as the closest prior work on NeRF-initialized 3DGS, and the text explicitly differentiates NeRF-GS from it ('Similar to RadSplat... unlike RadSplat... completely different from RadSplat and NeRF-init'), yet Table 1 and the rest of the paper contain no quantitative comparison with RadSplat. Since the abstract claims that NeRF-GS 'surpasses existing methods and achieves state-of-the-art performance,' and RadSplat is the most natural competitive baseline for this exact claim, omitting it leaves the central empirical assertion unsupported. I request a direct comparison on Mip-NeRF360, Tanks&Temples, and DeepBlending under the same evaluation protocol, reported in Table 1 and discussed in Sec. 5.2.
  2. [Table 4] The DeepBlending ablation row 'w/o Edge-based Init' reports per-scene PSNRs of 28.65 (Drjohnson) and 29.8 (Playroom) with an average of 29.8. The average of these two numbers is 29.23, not 29.8, so the table contains an arithmetic inconsistency. This is not a rounding artifact, and it raises concerns about the reliability of the other ablation entries in the same table. Please correct the value and audit all averaged rows for internal consistency; the corrected average still supports a degradation from the full model, but the current presentation is not acceptable as is.
  3. [Abstract, Table 2] The abstract's unconditional claim that NeRF-GS 'surpasses existing methods' is contradicted by the sparse-view results in Table 2: on Blender with 8 views, SplatFields achieves PSNR 23.98 and SSIM 0.889, while NeRF-GS achieves 23.92 and 0.881. The paper acknowledges 'comparable to or even surpassing' SplatFields, but the abstract and Sec. 5.2 ('NeRF-GS consistently surpasses corresponding baselines') are stronger than the evidence. Please either soften the claims to match the data or provide additional experiments that justify the state-of-the-art statement across all reported settings.
  4. [Tables 1-2, Sec. 5.2] The phrase 'significantly outperforms' is used repeatedly, but no error bars, multiple-seed standard deviations, or significance tests are reported. Some of the margins are small (e.g., Mip-NeRF360 SSIM 0.817 vs. 0.813 for 3DGS), so without run-to-run variance it is not possible to assess whether the differences are statistically meaningful. Please add repeated-run statistics or, at minimum, state the number of runs and the observed spread for the main tables.
minor comments (6)
  1. [Sec. 5.4, Table 4] The text says that 'w/o Edge-based Init' refers to 'the alternative initialization from SFM,' but the terminology in Sec. 4.1 and in the supplementary (D_edge_3dgs and D_random_3dgs) is different. Please clarify the exact initialization used in each ablation row to avoid confusion.
  2. [Eq. (12), Table 4] The loss terms in Eq. (12) are called L_fea_reg and L_pos_reg, but Table 4 uses the headers 'w/o L_fea_joint' and 'w/o L_pos_joint'. Please use consistent notation throughout.
  3. [Supplementary Sec. 10, Table 5] Table 5 shows that the NeRF branch of NeRF-GS underperforms Instant-NGP on all three datasets, while the text emphasizes mutual promotion. The mutual-promotion claim is only supported by qualitative Fig. 9 and by comparing the branch against itself, not against Instant-NGP; please make this distinction explicit.
  4. [Fig. 5] The figure caption 'GT w/o feature share Full w/o joint optimization' is ambiguous; it should be rephrased to indicate which panel corresponds to which ablation condition.
  5. [Sec. 3] There is a typo in the sentence 'inheriting the EW A volume splatting method' -- it should read 'EWA volume splatting.'
  6. [Sec. 5.1] The implementation details state that all experiments are run on an NVIDIA A100, but do not specify how the baseline timings in Table 3 were measured (same hardware, same protocol). Please state whether the 3DGS timings were re-run locally or taken from prior papers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: NeRF-GS is an empirically evaluated training scheme; the SOTA claim is under-supported by the missing RadSplat comparison, but that is an evidence gap, not circularity.

full rationale

The claimed derivation is not circular. NeRF-GS pretrains a hash-based NeRF, initializes Gaussians from NeRF ray depths biased toward edges (Eq. 5), decodes Gaussian attributes from shared features plus learned residuals (Eqs. 6-7), and jointly optimizes branch-specific and cross-branch losses (Eq. 12). Every reported quantity (PSNR/SSIM/LPIPS on Mip-NeRF360, Tanks&Temples, DeepBlending, Blender) is measured on held-out test views and compared against external baselines in Tables 1-3; no fitted parameter is relabeled as a prediction, and no performance number follows from a loss definition by construction. Self-citations (PVD, shared feature field, ChatEdit-3D) appear only as related-work context and are not used to justify the central claims; no uniqueness theorem or ansatz is imported from the authors' prior work. The abstract's 'surpasses existing methods' is weakened by the paper's own differentiation from RadSplat in Sec. 4.1 without a RadSplat row in Table 1, and Table 4 lists an internally inconsistent DeepBlending average (28.65 and 29.8 listed as 29.8); however, these are benchmark-coverage and reporting gaps, not reductions of the result to its inputs. The conclusion that NeRF and 3DGS are complementary is a post-hoc interpretation of the experiments, not an assumption baked into the loss. Thus no significant circularity.

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

The paper relies on standard neural rendering equations and a small number of domain assumptions about where Gaussians should be placed and how NeRF opacity correlates with surface presence. The hand-tuned hyperparameters listed above are part of the method and are not independently justified, but they do not constitute new physical or conceptual entities.

free parameters (4)
  • Edge-to-random ray sampling ratio = 8:2
    Set by hand for initializing Gaussian positions; higher weight on edge rays, but no sensitivity study is provided to justify this ratio.
  • Hash grid levels and feature dimension = 16 levels, 2-dim per level (total 32-dim)
    Architecture choice borrowed from Instant-NGP; not tuned in this paper, yet it defines the shared feature capacity.
  • Max Gaussian additions per growth step = 200 every 100 iterations
    Heuristic controlling NeRF-assisted growth; no analysis of sensitivity to this number is given.
  • Joint loss weights (lambda_nerf, lambda_rgb, lambda_op, lambda_fea, lambda_pos) = 0.1, 0.05, 1e-3, 1e-4, 1e-4
    Hand-tuned to balance the NeRF and GS branches; no principled criterion or sensitivity analysis is reported.
assumptions (4)
  • standard math Volume rendering integral (Eq. 1) correctly models light transport for opaque and semi-transparent media.
    Standard NeRF assumption used throughout Section 3.
  • standard math Gaussian splatting compositing (Eq. 3) approximates the true radiance field well.
    Standard 3DGS rendering model from Kerbl et al., used throughout Section 3.
  • domain assumption High NeRF opacity at a sampled point indicates the presence of a surface that warrants a new Gaussian.
    Used for NeRF-assisted growth in Section 4.3; could add floaters if NeRF is overconfident in empty space.
  • domain assumption Edge-detected high-frequency image regions are the most informative for Gaussian placement.
    Motivates edge-based initialization in Section 4.1; may not hold for low-texture scenes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NeRF Is a Valuable Assistant for 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/QEZMRBD4

@misc{pith2026250723374,
  author       = {Pith},
  title        = {Pith review of: NeRF Is a Valuable Assistant for 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QEZMRBD4}},
  note         = {Machine review of arXiv:2507.23374}
}
read the original abstract

We introduce NeRF-GS, a novel framework that jointly optimizes Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS). This framework leverages the inherent continuous spatial representation of NeRF to mitigate several limitations of 3DGS, including sensitivity to Gaussian initialization, limited spatial awareness, and weak inter-Gaussian correlations, thereby enhancing its performance. In NeRF-GS, we revisit the design of 3DGS and progressively align its spatial features with NeRF, enabling both representations to be optimized within the same scene through shared 3D spatial information. We further address the formal distinctions between the two approaches by optimizing residual vectors for both implicit features and Gaussian positions to enhance the personalized capabilities of 3DGS. Experimental results on benchmark datasets show that NeRF-GS surpasses existing methods and achieves state-of-the-art performance. This outcome confirms that NeRF and 3DGS are complementary rather than competing, offering new insights into hybrid approaches that combine 3DGS and NeRF for efficient 3D scene representation.

Figures

Figures reproduced from arXiv: 2507.23374 by the authors.

Figure 1
Figure 1. NeRF-GS establishes a bridge of communication be [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of NeRF-GS. (a) We first pretrain a Hash-based NeRF network to acquire continuous spatial encoding capabilities and implicit scene representation. (b) Utilizing the preliminary scene carved by NeRF, we resample rays corresponding to image edges to obtain potential Gaussian positions, facilitating Gaussian initialization. (c) During joint optimization, the GS branch queries corresponding features f from the … view at source ↗
Figure 3
Figure 3. Qualitative comparison on real-world datasets. The numbers indicate the PSNR. Our method demonstrates a significant advantage over 3DGS and its variants, achieving a more faithful representation of scene details. from limited 3D scene information. Additionally, the col￾laborative optimization between NeRF and GS branches, facilitated by this shared information, creates mutual con￾straints and regularization effects,… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison under 12 input views on the Blender dataset. The numbers indicate the PSNR [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Visualization of position residuals. The points repre￾sent the initial Gaussian positions, with the top 20% of points hav￾ing the largest optimized residuals highlighted in red. We compare this with the results obtained by fixed Gaussian positions during training, demo…
Figure 5
Figure 5. Figure 5: Impact of feature share and joint optimization on sparse view scenes. These two key designs enable mutual regular￾ization constraints between NeRF and GS branches, significantly improving the visual quality of NeRF-GS in sparse views [PITH_FULL_IMAGE:figures/full_fig_…
Figure 7
Figure 7. Figure 7: Impact of NeRF-assisted Gaussian growth. We initialize 3DGS using point clouds with missing regions to evaluate its scene perception range and sensitivity to initialization. Without NeRF-assisted Gaussian growth, 3DGS exhibits insufficient reconstruction (a) or incompl…
Figure 8
Figure 8. Figure 8: Comparison of initialization with RadSplat. NeRF-GS focuses more on the contours of the scene during ray sampling, alleviating the burden of position optimization in the GS branch while achieving superior visual results in regions with complex textures. Supp [PITH_FUL…
Figure 9
Figure 9. Figure 9: Impact of joint optimization on the NeRF branch. The dashed line indicates the mean PSNR. Given equivalent training iter￾ations, the NeRF obtained through NeRF-GS outperforms training this NeRF independently. This demonstrates that dual-branch training not only benefit…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

75 extracted references · 47 canonical work pages

  1. [46]

    Radsplat: Radiance field-informed gaussian splat- ting for robust real-time rendering with 900+ fps

    Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto- saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splat- ting for robust real-time rendering with 900+ fps. arXiv preprint arXiv:2403.13806, 2024. 1, 2, 4

    arXiv 2024

  2. [1]

    Trimming the fat: Efficient compres- sion of 3d gaussian splats through pruning

    Muhammad Salman Ali, Maryam Qamar, Sung-Ho Bae, and Enzo Tartaglione. Trimming the fat: Efficient compres- sion of 3d gaussian splats through pruning. arXiv preprint arXiv:2406.18214, 2024. 2

    arXiv 2024

  3. [2]

    Controlling neural level sets

    Matan Atzmon, Niv Haim, Lior Yariv, Ofer Israelov, Haggai Maron, and Yaron Lipman. Controlling neural level sets. Advances in Neural Information Processing Systems(NIPS),

    work page

  4. [3]

    Milena T Bagdasarian, Paul Knoll, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgenstern. 3dgs. zip: A survey on 3d gaussian splatting compression meth- ods. arXiv preprint arXiv:2407.09510, 2024. 2

    arXiv 2024

  5. [4]

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

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 5855–5864,

  6. [5]

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

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 5

    work page 2022

  7. [6]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InEuropean con- ference on computer vision, pages 333–350. Springer, 2022. 1

    work page 2022

  8. [7]

    A survey on 3d gaussian splatting

    Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 1

    arXiv 2024

Show all 75 references
  1. [8]

    Hac: Hash-grid assisted context for 3d gaussian splatting compression

    Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. Hac: Hash-grid assisted context for 3d gaussian splatting compression. In European Conference on Computer Vision, pages 422–438. Springer, 2025. 1, 2, 5, 6, 7

    2025

  2. [9]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370–386. Springer, 2025. 1

    2025

  3. [10]

    Neural parametric gaussians for monocular non-rigid object reconstruction

    Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 10715–10725, 2024. 2

    2024

  4. [11]

    Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, De- jia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. arXiv preprint arXiv:2311.17245, 2023. 2

    2023 arXiv

  5. [12]

    Instantsplat: Un- bounded sparse-view pose-free gaussian splatting in 40 sec- onds

    Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Un- bounded sparse-view pose-free gaussian splatting in 40 sec- onds. arXiv preprint arXiv:2403.20309, 2, 2024. 1

    2024 arXiv

  6. [13]

    Edit- ing 3d scenes via text prompts without retraining

    Shuangkang Fang, Yufeng Wang, Yi Yang, Yi-Hsuan Tsai, Wenrui Ding, Shuchang Zhou, and Ming-Hsuan Yang. Edit- ing 3d scenes via text prompts without retraining. arXiv e- prints, pages arXiv–2309, 2023. 1

    2023

  7. [14]

    Pvd-al: Progres- sive volume distillation with active learning for efficient con- version between different nerf architectures

    Shuangkang Fang, Yufeng Wang, Yi Yang, Weixin Xu, Heng Wang, Wenrui Ding, and Shuchang Zhou. Pvd-al: Progres- sive volume distillation with active learning for efficient con- version between different nerf architectures. arXiv preprint arXiv:2304.04012, 2023. 1, 2

    2023 arXiv

  8. [15]

    One is all: Bridging the gap between neural radiance fields architectures with progressive volume distillation

    Shuangkang Fang, Weixin Xu, Heng Wang, Yi Yang, Yufeng Wang, and Shuchang Zhou. One is all: Bridging the gap between neural radiance fields architectures with progressive volume distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 597–605, 2023. 2

    2023

  9. [16]

    Efficient implicit sdf and color recon- struction via shared feature field

    Shuangkang Fang, Dacheng Qi, Weixin Xu, Yufeng Wang, Zehao Zhang, Xiaorong Zhang, Huayu Zhang, Zeqi Shao, and Wenrui Ding. Efficient implicit sdf and color recon- struction via shared feature field. InProceedings of the Asian Conference on Computer Vision, pages 3499–3516, 2024. 1

    2024

  10. [17]

    Chat- edit-3d: Interactive 3d scene editing via text prompts

    Shuangkang Fang, Yufeng Wang, Yi-Hsuan Tsai, Yi Yang, Wenrui Ding, Shuchang Zhou, and Ming-Hsuan Yang. Chat- edit-3d: Interactive 3d scene editing via text prompts. arXiv preprint arXiv:2407.06842, 2024. 1

    2024 arXiv

  11. [18]

    Does gaussian splatting need sfm initializa- tion? arXiv preprint arXiv:2404.12547, 2024

    Yalda Foroutan, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasacchi. Does gaussian splatting need sfm initializa- tion? arXiv preprint arXiv:2404.12547, 2024. 1, 2, 4

    2024 arXiv

  12. [19]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 1, 2

    2022

  13. [20]

    Efros, and Xiaolong Wang

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splat- ting. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 20796– 20805, 2024. 1

    2024

  14. [21]

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

    Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray trac- ing. arXiv preprint arXiv:2311.16043, 2023. 2

    2023 arXiv

  15. [22]

    Fastnerf: High-fidelity neu- ral rendering at 200fps

    Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neu- ral rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision , pages 14346– 14355, 2021. 1

    2021

  16. [23]

    Deep blending for free-viewpoint image-based rendering

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1–15, 2018. 5

    2018

  17. [24]

    2d gaussian splatting for geometrically accu- rate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accu- rate radiance fields. arXiv preprint arXiv:2403.17888, 2024. 2, 5, 7

    2024 arXiv

  18. [25]

    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

  19. [26]

    Relaxing accu- rate initialization constraint for 3d gaussian splatting

    Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Seungryong Kim. Relaxing accu- rate initialization constraint for 3d gaussian splatting. arXiv preprint arXiv:2403.09413, 2024. 2

    2024 arXiv

  20. [27]

    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):1–14, 2023. 1, 2, 4, 5, 6, 7

    2023

  21. [28]

    Infonerf: Ray entropy minimization for few-shot neural volume ren- dering

    Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume ren- dering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 12912– 12921, 2022. 5

    2022

  22. [29]

    Tanks and temples: Benchmarking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) , 36 (4):1–13, 2017. 5

    2017

  23. [30]

    Decomposing nerf for editing via feature field distil- lation

    Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitz- mann. Decomposing nerf for editing via feature field distil- lation. Advances in Neural Information Processing Systems, 35:23311–23330, 2022. 2

    2022

  24. [31]

    Compact 3d gaussian representation for radiance field

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719– 21728, 2024. 2

    2024

  25. [32]

    Analytic-splatting: Anti-aliased 3d gaussian splatting via analytic integration

    Zhihao Liang, Qi Zhang, Wenbo Hu, Ying Feng, Lei Zhu, and Kui Jia. Analytic-splatting: Anti-aliased 3d gaussian splatting via analytic integration. arXiv preprint arXiv:2403.11056, 2024. 2

    2024 arXiv

  26. [33]

    Autoint: Automatic integration for fast neural volume ren- dering

    David B Lindell, Julien NP Martel, and Gordon Wetzstein. Autoint: Automatic integration for fast neural volume ren- dering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14556– 14565, 2021. 2

    2021

  27. [34]

    Compgs: Efficient 3d scene repre- sentation via compressed gaussian splatting

    Xiangrui Liu, Xinju Wu, Pingping Zhang, Shiqi Wang, Zhu Li, and Sam Kwong. Compgs: Efficient 3d scene repre- sentation via compressed gaussian splatting. arXiv preprint arXiv:2404.09458, 2024. 2

    2024 arXiv

  28. [35]

    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. 1, 2, 5, 6, 7, 3

    2024

  29. [36]

    Gaussian splatting with nerf-based color and opacity

    Dawid Malarz, Weronika Smolak, Jacek Tabor, Sławomir Tadeja, and Przemysław Spurek. Gaussian splatting with nerf-based color and opacity. arXiv preprint arXiv:2312.13729, 2023. 1, 2, 5, 6, 7

    2023 arXiv

  30. [37]

    Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duck- worth. NeRF in the Wild: Neural Radiance Fields for Un- constrained Photo Collections. In CVPR, 2021. 1

    2021

  31. [38]

    Occupancy networks: Learning 3d reconstruction in function space

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Se- bastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition(CVPR), 2019. 2

    2019

  32. [39]

    Implicit sur- face representations as layers in neural networks

    Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit sur- face representations as layers in neural networks. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision(ICCV), 2019. 2

    2019

  33. [40]

    Splatfields: Neural gaussian splats for sparse 3d and 4d re- construction

    Marko Mihajlovic, Sergey Prokudin, Siyu Tang, Robert Maier, Federica Bogo, Tony Tung, and Edmond Boyer. Splatfields: Neural gaussian splats for sparse 3d and 4d re- construction. In European Conference on Computer Vision, pages 313–332. Springer, 2025. 1, 2, 5, 7

    2025

  34. [41]

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

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. In Proceedings of the European Conference on Com- puter Vision(ECCV), 2020. 1, 2, 5

    2020

  35. [42]

    Instant neural graphics primitives with a multires- olution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a multires- olution hash encoding. ACM Trans. Graph. , 41(4):102:1– 102:15, 2022. 1, 2, 4, 6, 7, 3

    2022

  36. [43]

    Compact3d: Com- pressing gaussian splat radiance field models with vector quantization

    KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compact3d: Com- pressing gaussian splat radiance field models with vector quantization. arXiv preprint arXiv:2311.18159, 2023. 2

    2023 arXiv

  37. [44]

    Compressed 3d gaussian splatting for accelerated novel view synthesis

    Simon Niedermayr, Josef Stumpfegger, and R ¨udiger West- ermann. Compressed 3d gaussian splatting for accelerated novel view synthesis. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10349–10358, 2024. 2, 5, 6

    2024

  38. [45]

    Occupancy flow: 4d reconstruction by learning particle dynamics

    Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proceedings of the IEEE/CVF International Conference on Computer Vision(ICCV), 2019. 2

    2019

  39. [47]

    Texture fields: Learning tex- ture representations in function space

    Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning tex- ture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion(ICCV), 2019. 2

    2019

  40. [48]

    Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction

    Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion(ICCV), 2021. 2

    2021

  41. [49]

    Deepsdf: Learning con- tinuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning con- tinuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR), 2019. 2

    2019

  42. [50]

    Convolutional occupancy networks

    Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Proceedings of the European Conference on Computer Vision(ECCV), 2020. 2

    2020

  43. [51]

    A survey on real-time 3d scene recon- struction with slam methods in embedded systems

    Quentin Picard, Stephane Chevobbe, Mehdi Darouich, and Jean-Yves Didier. A survey on real-time 3d scene recon- struction with slam methods in embedded systems. arXiv preprint arXiv:2309.05349, 2023. 1

    2023 arXiv

  44. [52]

    Derf: Decom- posed radiance fields

    Daniel Rebain, Wei Jiang, Soroosh Yazdani, Ke Li, Kwang Moo Yi, and Andrea Tagliasacchi. Derf: Decom- posed radiance fields. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 14153–14161, 2021. 2

    2021

  45. [53]

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

    Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision , pages 14335– 14345, 2021. 1, 2

    2021

  46. [54]

    Deep learning-based 3d reconstruction: a survey

    Taha Samavati and Mohsen Soryani. Deep learning-based 3d reconstruction: a survey. Artificial Intelligence Review , 56(9):9175–9219, 2023. 1

    2023

  47. [55]

    Sa- gs: Scale-adaptive gaussian splatting for training-free anti- aliasing

    Xiaowei Song, Jv Zheng, Shiran Yuan, Huan-ang Gao, Jing- wei Zhao, Xiang He, Weihao Gu, and Hao Zhao. Sa- gs: Scale-adaptive gaussian splatting for training-free anti- aliasing. arXiv preprint arXiv:2403.19615, 2024. 2

    2024 arXiv

  48. [56]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 5459– 5469, 2022. 2

    2022

  49. [57]

    Sparsenerf: Distilling depth ranking for few-shot novel view synthesis

    Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zi- wei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 9065–9076,

  50. [58]

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

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1

    2021 arXiv

  51. [59]

    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. 5

    2004

  52. [60]

    Agg: Amor- tized generative 3d gaussians for single image to 3d

    Dejia Xu, Ye Yuan, Morteza Mardani, Sifei Liu, Jiaming Song, Zhangyang Wang, and Arash Vahdat. Agg: Amor- tized generative 3d gaussians for single image to 3d. arXiv preprint arXiv:2401.04099, 2024. 2

    2024 arXiv

  53. [61]

    Point- nerf: Point-based neural radiance fields

    Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point- nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438–5448, 2022. 4

    2022

  54. [62]

    Multi-scale 3d gaussian splatting for anti-aliased rendering

    Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 2

    2024

  55. [63]

    Gaussianob- ject: Just taking four images to get a high-quality 3d object with gaussian splatting

    Chen Yang, Sikuang Li, Jiemin Fang, Ruofan Liang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Gaussianob- ject: Just taking four images to get a high-quality 3d object with gaussian splatting. arXiv preprint arXiv:2402.10259 ,

    arXiv

  56. [64]

    Plenoctrees for real-time rendering of neural radiance fields

    Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752– 5761, 2021. 1, 2

    2021

  57. [65]

    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. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 19447–19456,

  58. [66]

    Soundstream: An end- to-end neural audio codec

    Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end- to-end neural audio codec. IEEE/ACM Transactions on Au- dio, Speech, and Language Processing , 30:495–507, 2021. 2

    2021

  59. [67]

    Ray priors through reprojection: Improving neu- ral radiance fields for novel view extrapolation

    Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray priors through reprojection: Improving neu- ral radiance fields for novel view extrapolation. In Proceed- ings of the IEEE/CVF Conference on Computer Visi...

    2022

  60. [68]

    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. 5

    2018

  61. [69]

    Lp-3dgs: Learning to prune 3d gaussian splatting

    Zhaoliang Zhang, Tianchen Song, Yongjae Lee, Li Yang, Cheng Peng, Rama Chellappa, and Deliang Fan. Lp-3dgs: Learning to prune 3d gaussian splatting. arXiv preprint arXiv:2405.18784, 2024. 2

    2024 arXiv

  62. [70]

    Ewa volume splatting

    Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visu- alization, 2001. VIS’01., pages 29–538. IEEE, 2001. 3 NeRF Is a Valuable Assistant for 3D Gaussian Splatting Supplementary Material w/o NeRF-assisted Gaussian gro...

    2001

  63. [71]

    In contrast, individual Gaussian sphere in 3DGS has a limited perceptual range, making 3DGS sen- sitive to initialization and less effective in adaptive control

    Analysis of Gaussian Adaptive Control from NeRF Branch The continuous spatial representation of NeRF enables queries at any spatial location, allowing it to perceive the entire 3D scene. In contrast, individual Gaussian sphere in 3DGS has a limited perceptual range, making 3DG...

  64. [72]

    To enhance initialization efficiency, we incorporate an edge detection step that pre-identifies critical rays and increases their sampling probability during initial- ization

    Analysis of Edge-based Initialization NeRF-GS utilizes pre-trained NeRF to obtain candidate Gaussian positions. To enhance initialization efficiency, we incorporate an edge detection step that pre-identifies critical rays and increases their sampling probability during initial...

  65. [73]

    Analysis of NeRF Branch Performance Mutual Promotion between NeRF and GS Branches . While the primary aim of this work is to leverage NeRF characteristics to address 3DGS limitations, we have found that the GS branch also positively impacts the NeRF branch during joint trainin...

  66. [74]

    Additional Ablation Studies We further conduct ablation studies on additional loss terms, including the introduced volume regularization [35] and the overall loss term of the NeRF branch, Lnerf. Addi- tionally, we evaluate the performance of directly optimiz- ing 3DGS after in...

  67. [75]

    Per-scene Breakdown Results of NeRF-GS We provide a detailed quantitative assessment of NeRF-GS across various scenes in Tables 7, 8 and 9, including metrics such as PSNR, SSIM, and LPIPS. Table 7. Per-scene results of Blender dataset of our method. Full views chair drums ficu...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.