Pith. sign in

REVIEW 5 major objections 6 minor 43 references

LOD-GS: Level-of-Detail-Sensitive 3D Gaussian Splatting for Detail Conserved Anti-Aliasing

T0 review · 5 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 Making each 3D Gaussian's filter size depend on the camera sampling rate yields alias-free, detail-preserving rendering at every zoom level, with state-of-the-art results on synthetic and real-world benchmarks.

desk verdict LOD-GS is a credible incremental contribution to 3DGS anti-aliasing that replaces Mip-Splatting's fixed filter with a learnable sampling-rate-dependent GMM, but comparison confounds and an untested view-dependence axis keep the SOTA claim conditional. read the letter →

arxiv 2507.00554 v3 pith:6X5GXLM5 submitted 2025-07-01 cs.CV

Zhenya Yang , Bingchen Gong , Kai Chen This is my paper
classification cs.CV
keywords 3DGaussianSplattinganti-aliasinglevelofdetailsamplingratepre-filteredradiancefieldEWAfilteringnovelviewsynthesismulti-scalerendering
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

LOD-GS is a modification of 3D Gaussian Splatting designed to render the same scene correctly at every zoom level. Its central idea is that each Gaussian primitive should be filtered with a strength that depends on the sampling rate $\nu = f/d$, the ratio of focal length to camera distance, instead of the fixed filter used by earlier anti-aliasing methods. The paper argues that this sampling-rate sensitivity lets one model learn a pre-filtered radiance field from training images taken at mixed resolutions and distances, and reports state-of-the-art quality on multi-scale and multi-level benchmarks, including a new synthetic dataset rendered at three camera distances. This matters because zooming in or out of a trained 3DGS scene currently either leaves aliasing artifacts or blurs away fine texture.

What carries the argument

The shared Gaussian Mixture Model (GMM) filter is the load-bearing object: a mixture of $l = 20$ Gaussian basis functions with learnable weights, centers, and widths, evaluated at the scalar sampling rate $\nu = f/d$ and yielding a per-primitive blur added to the covariance and an opacity correction. Its work is to turn a cheap per-primitive scalar into a per-primitive filter strength, so that a primitive seen from near or far, at high or low resolution, is blurred to just below the Nyquist frequency of that view. The supporting mechanism is the EWA filter of Eq. 8, whose normalization term $\sqrt{|\Sigma^{2D}_k| / |\Sigma^{2D}_k + sI|}$ converts the vanilla 3DGS screen-space dilation into an energy-preserving low-pass filter. The paper's new dataset plays a supporting role by supplying training and test views whose sampling rates differ through camera distance rather than image downsampling.

What would settle it

Render a large oblique plane covered with fine texture while rotating it relative to a camera whose $f/d$ stays constant: if the scalar sampling-rate input is sufficient, rendering quality should stay uniform across the plane, whereas a view-dependent blur requirement would show uneven smearing. A matched-pair variant at $(f,d)$ and $(2f,2d)$, identical $f/d$ but different perspective foreshortening, tests whether the scalar conflates two physically different captures.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that aliasing in 3D Gaussian Splatting is a level-of-detail problem: the appearance of a 3D Gaussian should change with the camera's sampling rate, just as mipmap textures change with distance. To implement this, every Gaussian primitive is augmented with a shared set of learnable basis functions, a Gaussian mixture model $F(\nu) = \sum_{i=1}^{l} w_i \exp(-(\nu-\mu_i)^2 / 2\sigma_i^2)$ evaluated at the sampling rate $\nu = f/d$, which outputs a filter size $F_s(\nu)$ added to the primitive's covariance and an opacity residual $F_\alpha(\nu)$ added to its $\alpha$. The filtered primitive is then projected to screen space and passed through an EWA (Elliptical Weighted Average) filter whose normalization term suppresses the dilation artifact of the vanilla rasterizer. Because the sampling rate for all primitives from one camera costs $O(K)$, the filter can be recomputed at every iteration and optimized end-to-end with image supervision, which the paper says is what lets the model disentangle multi-scale training views and reconstruct fine detail while remaining alias-free.

Load-bearing premise

The load-bearing premise is that the correct blur for a Gaussian is fixed by the single scalar $f/d$, the focal length divided by the distance from the camera to the primitive's center, so all primitives at the same depth receive the same filter; if the needed blur also changes with viewing angle, screen-space elongation, or occlusion, the learned filter will mis-adapt for exactly those primitives.

Editorial extensions

If this is right

  • A single trained LOD-GS model renders the same scene at multiple resolutions and camera distances without retraining, from full resolution down to 1/8 scale and from near to far camera levels, because the per-primitive filter adapts to each view's sampling rate.
  • Mixed-scale training becomes consistent: images with different sampling rates can be optimized together without the ambiguity that blurs a vanilla 3DGS model, shown by the reconstruction of fine textures such as the silk ribbon that competing methods smooth away.
  • Single-scale-trained models generalize to other scales: in the STMT (single-scale training, multi-scale testing) setup, the LOD filter improves both zoomed-out and zoomed-in rendering even though training saw only one scale.
  • Anti-aliasing quality no longer depends on a hand-tuned filter-size hyperparameter as in Mip-Splatting; the GMM learns the correct strength from data, and ablations removing either the LOD filter or the EWA filter drop performance on both synthetic and real scenes.

Reading between the lines

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

  • The shared GMM is optimized only over sampling rates seen in training; whether it extrapolates to zooms far beyond the trained range is untested, and a stress test at extreme $f/d$ values would probe how well the learned pre-filter generalizes.
  • Because the filter input is the single scalar $\nu = f/d$, primitives at the same depth but different orientations receive identical blur; feeding view-dependent information such as projected anisotropy or view direction into the GMM is an obvious next step the current design cannot express.
  • The same $O(K)$ per-view machinery would support spatially varying sampling rates inside one image, for example foveated or non-perspective rendering, since only the scalar fed to the shared GMM would need to vary per region.
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

5 major / 6 minor

Summary. The paper proposes LOD-GS, a 3D Gaussian Splatting anti-aliasing method that introduces per-primitive level-of-detail filters. A Gaussian mixture module takes the sampling rate ν = f/d (focal length over camera-to-primitive distance) as input and predicts a filtering strength and an opacity residual for each 3D Gaussian; an EWA filter is then applied in 2D screen space. The authors also introduce a new synthetic 'Multi-level Blender Dataset' rendered at three camera distances to complement the existing multi-scale Blender dataset, which only varies focal length via downsampling. Experiments on multi-scale Blender, Mip-NeRF 360, and the new dataset report improved PSNR, SSIM, and LPIPS over prior 3DGS anti-aliasing baselines, alongside ablations and a single-scale-training/multi-scale-testing (STMT) generalization experiment.

Significance. If the reported results hold, LOD-GS offers a simple, efficient mechanism for making 3DGS filtering adaptive to sampling rate, and the new multi-level dataset addresses a genuine gap in anti-aliasing evaluation. The paper ships open-source code and data, and the STMT experiment is a real held-out-scale test that goes beyond the standard multi-scale training protocol. These are concrete strengths. However, the core technical claims are undermined by an under-specified filter formulation, a potentially confounded baseline comparison, and an untested limitation in the scalar sampling-rate input. The contribution is promising but requires substantial clarification and additional experiments before the state-of-the-art claim can be accepted.

major comments (5)
  1. [§5.1] The comparison with Mip-Splatting is confounded by densification scheme. The paper states that Mip-Splatting uses the densification scheme from GOF [37] while all other methods, including LOD-GS, use the original 3DGS densification. Since the densification strategy changes the number and distribution of Gaussian primitives, it directly affects rendering quality and cannot be separated from the filtering mechanism. This confound affects the headline results in Tables 1, 2, and 3. Please rerun Mip-Splatting with the same densification scheme as the other baselines (or run all methods under both schemes) and report the comparison under matched conditions.
  2. [§4.1, Eqs. (6)–(7)] The LOD filter is mathematically under-specified. Eq. (6) defines F(x) as a scalar Gaussian mixture, but Eq. (7) writes G_k(x) = exp(-1/2 (x-p_k)^T (Σ_k + F_s(ν))(x-p_k)) without defining F_s or explaining how a scalar function is added to a covariance matrix. If F_s(ν) is intended as a scalar times the identity, the equation should state F_s(ν) I; if it is a matrix, its construction from the GMM is not described. The paper also does not state any constraints ensuring that Σ_k + F_s(ν) remains positive definite or that the opacity residual in Eq. (10), alpha_k := alpha_k + F_alpha(ν), stays in [0,1]. Please specify the exact functional form, including the dimensions and constraints on all outputs, so the method can be reproduced from the text alone.
  3. [§4.1, Eq. (2)] The filter depends only on the scalar ν = f/d, which cannot capture the view-dependent projected footprint. The screen-space covariance of a Gaussian is Σ' = JWΣW^T J^T (Eq. 2), which depends on the view rotation, the primitive's own anisotropy, and its off-axis position. Two views with identical f/d can project the same primitive with very different footprint shapes, e.g., a thin ribbon viewed end-on versus obliquely, or a primitive near the image edge. The GMM input ν alone cannot express this distinction, and the EWA filter in Eq. (8) uses a fixed scalar s times the identity, so it cannot repair the shape mismatch. The current evaluation does not isolate the view-direction axis at fixed ν; the training views in the Blender and 360 datasets are all captured from constrained trajectories around the scene. Please add an experiment that varies view direction and off-axis position while keeping f/d fixed, for example by rendering a scene with a wide-FOV camera or a fixed-distance orbit over different elevations, to test whether the learned filter adapts correctly, or explicitly state this as a limitation of the method.
  4. [§5.2, Table 1] The state-of-the-art claim is not fully supported at the original resolution. In Table 1, LOD-GS achieves full-resolution PSNR 32.90, which is below Analytic-Splatting's 33.22 and only slightly above Mip-Splatting's 32.81. The averaged gains are modest (0.34 dB over Analytic-Splatting), and given the densification confound with Mip-Splatting noted above, the margin may not be robust. Please report results from multiple random seeds with standard deviations, and discuss the per-scene consistency of the improvement (the supplement's Table 8 shows LOD-GS is not the best on all scenes).
  5. [§5.3] The new Multi-level Blender Dataset is a central contribution, but its construction is under-described. The paper does not specify the three camera distances, the number of training and test views per level, whether the camera trajectories are the same across levels, or how the focal length is adjusted when re-rendering. Without these details the dataset cannot be reproduced or used as a benchmark by other researchers. Please provide a complete dataset specification, including rendering settings and a description of the removal of the drum scene noted in the supplement.
minor comments (6)
  1. [§4.1] Eq. (6) uses x as the variable of the Gaussian mixture, while Eq. (7) uses x as the 3D position in the exponential; these are different quantities but share a symbol. Please use distinct notation, for example ν for the GMM input and p for position.
  2. [§4.1] The paper never defines the subscripts in F_s and F_α, nor the quantities Δs and Δα indicated in Figure 3. Please introduce all notation explicitly.
  3. [§4.1] The claim that the added parameters are 'minor' is not quantified. Since a GMM with 20 basis functions is introduced per Gaussian primitive, each with weights, means, and standard deviations, the total parameter increase could be substantial for scenes with hundreds of thousands of primitives. Please report the additional memory and parameter counts.
  4. [§5.7] The analysis of the number of basis functions in Figure 7 reports a single run per setting; please add error bars or multiple seeds to support the conclusion that more basis functions consistently improve results.
  5. [References] The reference list contains formatting irregularities, including duplicated citations (e.g., [2] and [3] appear twice) and malformed page ranges in several entries. Please clean up the bibliography.
  6. [§5.4] In the Mip-NeRF 360 experiments, the paper reports LPIPS scores that are not consistently better than Analytic-Splatting (e.g., Table 3, 1/8 resolution, LOD-GS LPIPS 0.133 vs. Analytic-Splatting 0.128). Please address this in the discussion or soften the claim that the method achieves SOTA on all metrics.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the LOD filter is a learned mapping tested on held-out views and scales, not a fitted parameter renamed as a prediction.

full rationale

The paper's central mechanism (Eqs. 6-7 and 9-10) is a learnable GMM whose basis-function parameters are explicitly stated to be 'jointly optimized with the 3D Gaussian in an end-to-end manner' under photometric image loss. The filter output is a function of the sampling rate ν = f/d, and its generalization is evaluated on held-out test views (Tables 1-3) and, in the STMT experiment (Table 5), on zoom levels that were not used during training. No fitted constant is repackaged as a prediction: the GMM parameters are trained, not analytically derived from the evaluation targets. The conditioning on scalar f/d is an architectural assumption about what information controls filter strength, and it may limit generalization to oblique or off-axis views, but an assumption is not circularity. Citations in the paper are to external prior work such as Mip-Splatting, Analytic-Splatting, and EWA volume splatting; there are no load-bearing self-citations, no imported uniqueness theorems, and no renaming of a known result as a new derivation. The ablation study and the new dataset are empirical contributions rather than steps that reduce to their own inputs. Overall, the derivation chain is self-contained: the learned filter is supervised by images, and the reported gains are evaluated out-of-distribution with respect to view and scale, so no significant circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method's success depends on the learned GMM filter, whose parameters are free, and on the modeling assumption that a scalar sampling rate fully determines the filter. No new physical entities are postulated; the GMM is a learned component rather than an invented entity. The novelty claim is therefore empirical rather than derived.

free parameters (3)
  • GMM filter parameters = not reported
    Weights, centers, and standard deviations of the l=20 Gaussian basis functions in Eq. 6, plus the opacity-residual output F_alpha(nu), are optimized on the training set. The learned mapping from sampling rate to filter size is the core of the method, so the result depends on these fitted values.
  • Number of basis functions l = 20
    Chosen by hand; Section 5.7 ablates 5, 10, 15, and 20 and shows performance improves with more functions, so this is a free design choice affecting the result.
  • EWA filter scale s = not reported
    Eq. 8 uses an unspecified scale parameter s in the EWA normalization; the paper does not state whether s is fixed, learned, or how it is set, yet it affects the 2D filtering strength.
assumptions (4)
  • domain assumption The correct per-primitive filter strength is a function of the scalar sampling rate ν = f/d
    Invoked by Eq. 6-7: the GMM takes only ν as input; no dependence on view direction or projected anisotropy.
  • standard math Nyquist-Shannon sampling theorem justifies pre-filtering the Gaussians to the Nyquist frequency
    Section 3.2 uses this theorem as the theoretical basis for filtering.
  • standard math EWA filtering with the normalization factor sqrt(|Σ2D|/|Σ2D+sI|) removes dilation artifacts
    Borrowed from Zwicker et al. 2001, Eq. 8; the paper does not re-derive it.
  • ad hoc to paper A Gaussian mixture with 20 basis functions can represent the needed filter mapping
    The choice l=20 is justified only empirically by the ablation in Section 5.7; there is no theoretical guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LOD-GS: Level-of-Detail-Sensitive 3D Gaussian Splatting for Detail Conserved Anti-Aliasing." pith.science (2026). https://pith.science/paper/6X5GXLM5

@misc{pith2026250700554,
  author       = {Pith},
  title        = {Pith review of: LOD-GS: Level-of-Detail-Sensitive 3D Gaussian Splatting for Detail Conserved Anti-Aliasing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6X5GXLM5}},
  note         = {Machine review of arXiv:2507.00554}
}
read the original abstract

Despite the advancements in quality and efficiency achieved by 3D Gaussian Splatting (3DGS) in 3D scene rendering, aliasing artifacts remain a persistent challenge. Existing approaches primarily rely on low-pass filtering to mitigate aliasing. However, these methods are not sensitive to the sampling rate, often resulting in under-filtering and over-smoothing renderings. To address this limitation, we propose LOD-GS, a Level-of-Detail-sensitive filtering framework for Gaussian Splatting, which dynamically predicts the optimal filtering strength for each 3D Gaussian primitive. Specifically, we introduce a set of basis functions to each Gaussian, which take the sampling rate as input to model appearance variations, enabling sampling-rate-sensitive filtering. These basis function parameters are jointly optimized with the 3D Gaussian in an end-to-end manner. The sampling rate is influenced by both focal length and camera distance. However, existing methods and datasets rely solely on down-sampling to simulate focal length changes for anti-aliasing evaluation, overlooking the impact of camera distance. To enable a more comprehensive assessment, we introduce a new synthetic dataset featuring objects rendered at varying camera distances. Extensive experiments on both public datasets and our newly collected dataset demonstrate that our method achieves SOTA rendering quality while effectively eliminating aliasing. The code and dataset have been open-sourced.

Figures

Figures reproduced from arXiv: 2507.00554 by the authors.

Figure 1
Figure 1. (a) 3DGS treats a 3D Gaussian primitive (in blue) uni [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of Detail Reconstruction. All methods are trained on inputs with varying sampling rates. In comparing the reconstruction results of different methods, only our LOD-GS successfully reconstructs the intricate texture of the silk ribbon. Other methods share the smoothing problem to different extents. 2. Related Work Novel View Synthesis Novel View Synthesis (NVS) is a fundamental vision task aimed at generat… view at source ↗
Figure 3
Figure 3. Overview of our pipeline. During rendering, the sampling rate for each Gaussian primitive is sent to a learnable module that predicts the appropriate 3D filters for them. The filtered 3D Gaussians are then projected into 2D space and further processed using EWA filter before rasterization. The learnable 3D filters and the 3D Gaussians are jointly optimized end-to-end with image supervision. Anti-Aliasing of Neural R… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on our extended Blender dataset. All methods are trained across three levels: L1 (near), L2 (middle), and L3 (far). Our method more effectively captures the details of microstructures and textures during the multi-level training. PSNR ↑ SSIM ↑ LP…
Figure 6
Figure 6. Figure 6: Qualitative comparison on the Mip-NeRF 360 dataset [3]. All methods are trained and tested at three downsam￾pled resolutions (1/8, 1/16, and 1/32). We recommend that readers scale up this image and compare the region marked with red boxes across different methods. cont…
Figure 7
Figure 7. Figure 7: Performance of Models Using Different Numbers of Basis Functions: We report our model’s changes in PSNR, SSIM, and LPIPS by using 5, 10, 15, and 20 basis functions. Metrics for Mip-Splatting and Analytic-Splatting are visualized with dashed lines [PITH_FULL_IMAGE:figu…
Figure 8
Figure 8. Figure 8: Qualitative Results of the Ablation Study on the Blender Dataset [2, 20]. We render the objects from both far and near distances to test the anti-aliasing and detail conservation abilities [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Qualitative Results of the Ablation Study on the Mip-NeRF 360 Dataset [3]. All methods are trained and tested on multiscale inputs (1/8, 1/16, 1/32). We present rendering results from different methods at 1/8 and 1/32 scales to evaluate detail conservation and anti￾ali…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 37 canonical work pages

  1. [37]

    Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Trans. Graph., 43(6), 2024. 5

    work page 2024

  2. [1]

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

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields. In ICCV, 2021. 5

    work page 2021

  3. [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. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), pages 5855– 5864, 2021. 2, 3, 5, 1, 4, 6

    work page 2021

  4. [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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470–5479, 2022. 2, 3, 5, 7, 8, 1

    work page 2022

  5. [4]

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

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 19697–19705, 2023. 3

    work page 2023

  6. [5]

    V ol- ume rendering

    Robert A Brebin, Loren Carpenter, and Pat Hanrahan. V ol- ume rendering. In Seminal graphics: pioneering efforts that shaped the field, pages 363–372. 1998. 1

    work page 1998

  7. [6]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2, 6

    work page 2022

  8. [7]

    The aliasing problem in computer- generated shaded images

    Franklin C Crow. The aliasing problem in computer- generated shaded images. Communications of the ACM, 20 (11):799–805, 1977. 1

    work page 1977

Show all 43 references
  1. [8]

    Hierarchical triangula- tion for multiresolution surface description

    Leila De Floriani and Enrico Puppo. Hierarchical triangula- tion for multiresolution surface description. ACM Transac- tions On Graphics (TOG), 14(4):363–411, 1995. 2, 4

    1995

  2. [9]

    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, 6

    2022

  3. [10]

    Tri-miprf: Tri-mip represen- tation for efficient anti-aliasing neural radiance fields

    Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip represen- tation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19774–19783, 2023. 3, 5, 6

    2023

  4. [11]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5, 6, 8, 7

    2023

  5. [12]

    Multireso- lution techniques for interactive texture-based volume visu- alization

    Eric LaMar, Bernd Hamann, and Kenneth I Joy. Multireso- lution techniques for interactive texture-based volume visu- alization. IEEE, 1999. 2, 4

    1999

  6. [13]

    Pulsar: Effi- cient sphere-based neural rendering

    Christoph Lassner and Michael Zollhofer. Pulsar: Effi- cient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1440–1449, 2021. 2

    2021

  7. [14]

    Efficient ray tracing of volume data

    Marc Levoy. Efficient ray tracing of volume data. ACM Transactions on Graphics (ToG), 9(3):245–261, 1990. 1

    1990

  8. [15]

    Mipmap-gs: Let gaus- sians deform with scale-specific mipmap for anti-aliasing rendering

    Jiameng Li, Yue Shi, Jiezhang Cao, Bingbing Ni, Wenjun Zhang, Kai Zhang, and Luc Van Gool. Mipmap-gs: Let gaus- sians deform with scale-specific mipmap for anti-aliasing rendering. In International Conference on 3D Vision 2025 . 3, 5, 6, 8, 1, 7

    2025

  9. [16]

    Analytic-splatting: Anti-aliased 3d gaus- sian splatting via analytic integration

    Zhihao Liang, Qi Zhang, Wenbo Hu, Lei Zhu, Ying Feng, and Kui Jia. Analytic-splatting: Anti-aliased 3d gaus- sian splatting via analytic integration. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part XVII , ...

    2024

  10. [17]

    Neural sparse voxel fields

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. 2020. 2

    2020

  11. [18]

    Cohen, Mar- tin Reddy, and Amitabh Varshney

    David Luebke, Benjamin Watson, Jonathan D. Cohen, Mar- tin Reddy, and Amitabh Varshney. Level of Detail for 3D Graphics. Elsevier Science Inc., USA, 2002. 2, 4

    2002

  12. [19]

    Optical models for direct volume rendering

    Nelson Max. Optical models for direct volume rendering. TVCG, 1(2):99–108, 1995. 1

    1995

  13. [20]

    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, 5, 6, 4, 7

    2020

  14. [21]

    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. TOG, 41(4):1–15, 2022. 2, 6

    2022

  15. [22]

    Certain topics in telegraph transmission the- ory

    Harry Nyquist. Certain topics in telegraph transmission the- ory. Transactions of the American Institute of Electrical En- gineers, 47(2):617–644, 2009. 2

    2009

  16. [23]

    Dynamic point fields

    Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic point fields. In Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 7964–7976, 2023. 2

    2023

  17. [24]

    Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians

    Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. IEEE Transactions on Pattern Analysis and Machine Intelligence,

  18. [25]

    Communication in the presence of noise

    Claude E Shannon. Communication in the presence of noise. Proceedings of the IRE, 37(1):10–21, 1949. 2

    1949

  19. [26]

    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. 3, 5

    2024 arXiv

  20. [27]

    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

  21. [28]

    Pointnerf++: a multi-scale, point-based neural radiance field

    Weiwei Sun, Eduard Trulls, Yang-Che Tseng, Sneha Sambandam, Gopal Sharma, Andrea Tagliasacchi, and Kwang Moo Yi. Pointnerf++: a multi-scale, point-based neural radiance field. In European Conference on Computer Vision, pages 221–238. Springer, 2024. 2

    2024

  22. [29]

    Nerf-sr: High-quality neural radiance fields using supersampling

    Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High-quality neural radiance fields using supersampling. In Proceedings of the 30th ACM International Conference on Multimedia , pages 6445–6454, 2022. 4

    2022

  23. [30]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 13(4):600–612, 2004. 6

    2004

  24. [31]

    Level-of-detail volume ren- dering via 3d textures

    Manfred Weiler, R ¨udiger Westermann, Chuck Hansen, Kurt Zimmermann, and Thomas Ertl. Level-of-detail volume ren- dering via 3d textures. In Proceedings of the 2000 IEEE symposium on Volume visualization, pages 7–13, 2000. 2, 4

    2000

  25. [32]

    Pyramidal parametrics

    Lance Williams. Pyramidal parametrics. In Proceedings of the 10th Annual Conference on Computer Graphics and In- teractive Techniques, page 1–11, New York, NY , USA, 1983. Association for Computing Machinery. 2

    1983

  26. [33]

    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. 2

    2022

  27. [34]

    Multi-scale 3d gaussian splatting for anti-aliased render- ing

    Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased render- ing. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 20923– 20931, 2024. 3, 5, 6, 8, 1, 7

    2024

  28. [35]

    Differentiable surface splatting for point-based geometry processing

    Wang Yifan, Felice Serena, Shihao Wu, Cengiz ¨Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions On Graphics (TOG), 38(6):1–14, 2019. 2

    2019

  29. [36]

    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 (CVPR), pages 19447– 19456, 2024. 3, 4, 5, 6, 8, 1, 7

    2024

  30. [38]

    The unreasonable effectiveness of deep features as a perceptual metric

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

    2018

  31. [39]

    Pointavatar: Deformable point-based head avatars from videos

    Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21057–21067, 2023. 2

    2023

  32. [40]

    Anti-aliased neural implicit surfaces with encoding level of detail

    Yiyu Zhuang, Qi Zhang, Ying Feng, Hao Zhu, Yao Yao, Xi- aoyu Li, Yan-Pei Cao, Ying Shan, and Xun Cao. Anti-aliased neural implicit surfaces with encoding level of detail. New York, NY , USA, 2023. Association for Computing Machin- ery. 3

    2023

  33. [41]

    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. 2, 3, 5 LOD-GS: Level of Detail-Sensitive 3D Gaussian Splatting for Detail Conserved Anti-Aliasing Supplement...

    2001

  34. [42]

    Ablation The capability of our method to achieve aliasing-free and detail-conserved rendering relies on two components: the level-of-detail sensitive filter (denoted as LOD filter here- after) and the EW A filter. To assess the contribution of these modules to overall performa...

  35. [43]

    We analyze the experimental results on the Blender Dataset in Section 8.1 and the Mip-NeRF 360 Dataset in Section 8.2

    Additional Results In this section, we report the detail of our comparison ex- periment on each scene across different sampling rates. We analyze the experimental results on the Blender Dataset in Section 8.1 and the Mip-NeRF 360 Dataset in Section 8.2. 8.1. Blender Dataset We...

Pith tools

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