Multi-Sample Anti-Aliasing and Constrained Optimization for 3D Gaussian Splatting

Chun-Ming Xia; Jia-Chen Zhang; Yu-Jie Xiong; Zheng Zhou

REVIEW 5 major objections 5 minor 35 references

Multi-Sample Anti-Aliasing and Constrained Optimization for 3D Gaussian Splatting

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that adding 4x multisample anti-aliasing, an adaptive error-weighting loss, and a gradient-difference loss to 3D Gaussian Splatting preserves high-frequency textures and sharp edges while keeping real-time rendering.

desk verdict Real but modest gains on Mip-NeRF360; the SOTA claim is contradicted by its own table and Eq. (3) misdescribes the forward model. read the letter →

arxiv 2508.10507 v1 pith:C73PAKOV submitted 2025-08-14 cs.CV cs.AI

Zheng Zhou , Jia-Chen Zhang , Yu-Jie Xiong , Chun-Ming Xia This is my paper

classification cs.CVcs.AI

keywords 3DGaussiansplattingmulti-sampleanti-aliasingnovelviewsynthesisadaptiveweightinglossgradientdifferencehigh-frequencydetailpreservationreal-timerendering

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

The paper tries to show that the blur and jagged edges seen in 3D Gaussian Splatting come from evaluating each pixel at a single point and from losses that treat every pixel equally. Its remedy is to render each pixel as the average of four subpixel samples, and to add two loss terms: one that up-weights pixels with large reconstruction error, and one that matches the predicted image's horizontal and vertical gradients to the ground truth. On the Mip-NeRF360, Tanks and Temples, and Deep Blending benchmarks, the combined method reports higher SSIM and LPIPS than 3DGS and than two recent variants, while keeping the real-time rendering property. If true, the result means a modest change to the rasterizer and the loss can recover fine textures and sharp discontinuities without replacing the underlying Gaussian representation.

What carries the argument

The load-bearing mechanism is 4x MSAA rasterization: for each pixel, four subpixel positions are $\alpha$-blended and averaged, making the color computation differentiable across all four sampling paths. Around it, the adaptive weight map $w_{i,j}$ in Eq. (7) re-weights the $\ell^1$ loss toward under-reconstructed regions, and the gradient-difference loss in Eq. (14) enforces matching of horizontal and vertical forward differences between prediction and ground truth. The three components work together to allocate optimization effort to the regions where single-sample splatting fails.

What would settle it

Re-run the three benchmark suites with several random seeds per scene, reporting means and variances, with the adaptive-weight floor $\alpha$ and subpixel offsets $\delta_k$ chosen by cross-validation instead of fixed once. If the PSNR, SSIM, and LPIPS advantages over 3DGS do not persist outside the paper's single configuration, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's claim is that 3DGS's characteristic blur and aliasing are not intrinsic to the Gaussian representation but artifacts of single-point pixel sampling and unweighted reconstruction loss. The method computes each pixel color as the average of four subpixel renderings, each obtained by $\alpha$-compositing the Gaussian splats at a fractional offset, and trains with a composite loss $L = \lambda_1 L_w + \lambda_2 L_{\mathrm{DSSIM}} + \lambda_3 L_{\mathrm{grad}}$ in which the $\ell^1$ term is multiplied by a per-pixel weight proportional to local error and a gradient-difference term penalizes mismatches in horizontal and vertical forward differences. In the paper's experiments this configu

Load-bearing premise

The whole comparison rests on the assumption that the fixed loss weights, the baseline weight floor, and the subpixel offsets were not tuned to the test scenes and that the single-run metric improvements are stable enough to be called statistically significant.

Editorial extensions

If this is right

Vanilla 3DGS can be upgraded by changing the rasterizer and loss only, without altering the Gaussian representation or density-control logic.
The same modifications transfer to newer baselines: wrapping them around Pixel-GS and AbsGS yields better SSIM/LPIPS than either baseline alone in the reported tables.
Real-time rendering is preserved because MSAA is four subpixel evaluations per pixel rather than a different scene representation.
Gains should concentrate in scenes with high-frequency textures and sharp discontinuities, since the gradient loss directly targets edges and the adaptive weight targets under-reconstructed regions.
Perceptual quality improves even when PSNR gains are small, consistent with the loss focusing on gradient and structural fidelity rather than raw pixel error.

Reading between the lines

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

A natural test the paper leaves implicit is whether the 4x MSAA gain grows with output resolution, since aliasing is a sampling phenomenon and the single-sample baseline should degrade faster at higher resolutions.
The error weight map $w_{i,j}$ could be reused as a densification signal, guiding where new Gaussians are split or cloned; the paper does not explore this connection.
The gradient-difference term is defined on the image grid with forward differences, so a multi-scale or wavelet-domain variant might extend the reported boundary improvements to texture at different frequencies.
The paper fixes $\lambda_1=0.8$, $\lambda_2=0.2$, $\lambda_3=0.1$ but leaves the floor $\alpha$ in Eq. (7) and the offsets $\delta_k$ in Eq. (4) unspecified; measuring sensitivity to those choices would show how much of the gain is from the recipe rather than the components.

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.

Desk Editor's Note

read the letter

Short version: the empirical core is real but narrower than the abstract claims. The method reliably improves over vanilla 3DGS on Mip-NeRF360 (every scene in Table 2, small gains), and the ablations suggest all three components contribute. That's worth a serious look. But the paper overstates its own results: Table 1 shows Ours(3DGS) below Pixel-GS and AbsGS on SSIM and LPIPS, the 'statistically significant' in the abstract has no error bars or tests, and Deep Blending numbers in Table 3 have the full method slightly worse than vanilla 3DGS. The SOTA claim has to go.

More technically, Eq. (3) is not 3DGS alpha compositing. It's a normalized weighted sum with no per-Gaussian transmittance. The surrounding text calls it alpha compositing, but the equation doesn't describe it. If the implementation actually uses standard front-to-back compositing, the equation is a misdescription; if it doesn't, the method changes the forward model and the comparison against 3DGS is apples-to-oranges. No code or clarification is given, so this is a load-bearing open question. It's the first thing a referee should ask.

Other soft spots are smaller but add up: the adaptive floor alpha, subpixel offsets, and epsilon are unspecified, and we don't know if the loss weights were tuned per scene. Single-run point metrics without variance don't support the phrase 'statistically significant.' The paper also only compares against 3DGS, Pixel-GS, and AbsGS in the main table, even though it cites Mip-splatting and FreGS as related anti-aliasing work—those comparisons should be there.

To be fair, the per-scene improvement pattern on Mip-NeRF360 is consistent, which is stronger than cherry-picking. The ablation design is sensible, and the wavelet visualization is a nice qualitative touch. The combination is a legitimate extension, even if each ingredient is known.

This deserves peer review, not desk rejection. A referee can verify the Eq. (3) issue and ask for code, error bars, and corrected claims. The core idea is plausible and the data is mostly there; it's just not ready as submitted. If the authors fix the rendering equation description, release code or at least report alpha/offsets/sensitivity, and remove the SOTA language, it would be a reasonable modest contribution.

Referee Report

5 major / 5 minor

Summary. The paper proposes three modifications to 3D Gaussian Splatting: a 4× multi-sample anti-aliasing scheme that averages pixel colors from four subpixel offsets, an adaptive weighting loss that upweights high-error pixels based on the ℓ1 error to ground truth, and a gradient-difference loss that matches horizontal and vertical image gradients between prediction and ground truth. Experiments are reported on Mip-NeRF360, Tanks&Temples, and Deep Blending, with per-scene results on Mip-NeRF360 and ablations. The abstract and conclusion claim state-of-the-art detail preservation and statistically significant improvements over baseline approaches.

Significance. If the method were implemented as described and the empirical claims held, the work would offer a simple, practical recipe for improving detail reconstruction in 3DGS while preserving real-time rendering. The per-scene Mip-NeRF360 results in Table 2 are consistently positive, which is encouraging. However, the paper currently provides no code, no variance estimates, no significance tests, and no sensitivity analysis, so the broader significance as a state-of-the-art method is not established. The forward-model ambiguity in Eq. (3) is a more fundamental issue that must be resolved before the experimental comparison has its claimed meaning.

major comments (5)

[Section 3.1.2, Eq. (3)] The stated forward model is not 3D Gaussian Splatting. Standard 3DGS uses front-to-back alpha compositing with per-Gaussian transmittance T_n = ∏_{m<n}(1 − α_m w_m(u)), whereas Eq. (3) is a normalized sum over all Gaussians with no ordering or transmittance. If Eq. (3) is what is implemented, the method changes the rendering model itself, so the reported gains over the 3DGS baseline cannot be attributed to MSAA/losses. If standard 3DGS is used, the paper misdescribes its own forward pass and the gradient discussion in §3.2.2 is not tied to the implemented function. No code or clarification is provided. Please specify the exact rasterization function and, if Eq. (3) is a typo, correct it.
[Abstract and Table 1] The claim of state-of-the-art performance is contradicted by the paper's own numbers. On Mip-NeRF360, Ours(3DGS) has SSIM 0.819 vs. Pixel-GS 0.823 and LPIPS 0.207 vs. 0.193; only Ours(Pixel-GS) and Ours(AbsGS) exceed the corresponding baselines on some metrics, and not uniformly. The abstract's 'state-of-the-art' claim is not supported by Table 1. Please either revise the SOTA claim to reflect the actual ranking or provide a stronger comparison.
[Abstract and Section 4.2] 'Statistically significant improvements' are asserted but no variance, error bars, or significance tests are reported. All tables give single-run point estimates. Without multiple seeds or confidence intervals, the significance claim is unsupported. Please provide at least standard deviations over multiple runs or remove the statistical claim.
[Section 4.3, Table 3] The ablation does not demonstrate synergy. On Deep Blending, Ours has LPIPS 0.246, identical to 3DGS+MSAA and worse than 3DGS+AWS+GDC (0.244), and SSIM 0.900 vs. 3DGS+AWS+GDC 0.901. The claim that the full model 'achieves superior' quality to both ablations is not supported by the table. Please discuss these cases or adjust the claim.
[Section 3.3.1 and Section 3.2.1] Reproducibility-critical hyperparameters are missing. α in Eq. (7) is not specified, and the subpixel offsets δ_k in Eq. (4) are not defined. The fixed λ1=0.8, λ2=0.2, λ3=0.1 are reported, but no sensitivity analysis is given. Since the reported improvements are small (e.g., +0.01 PSNR on some scenes), these choices are load-bearing; please specify all values and provide a sensitivity study.

minor comments (5)

[Section 4.1] '13 real-world images' should be '13 scenes' (9 Mip-NeRF360 + 2 Tanks & Temples + 2 Deep Blending), and Section 4.2 says '14 scenarios' inconsistently.
[Section 4.2] The method list includes both 'MipNerf360 [4]' and 'Mip-NeRF360 [5]'. Reference [4] is Mip-NeRF, not Mip-NeRF360; please correct the citation and remove the duplicate.
[Figure 4 caption] The caption contains unreadable placeholder characters. Replace with a proper description of the visualization.
[Table 1] The legend says best, second best, and third best scores are colored red, orange, and yellow, but most table entries are not colored; formatting is inconsistent.
[Section 3.5.1] The text says λ3 'can be gradually increased during training', but Section 4.1.2 reports fixed values. State whether a schedule was used and, if so, specify it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's improvements are empirical, benchmark-measured additions to 3DGS, and no load-bearing step reduces to its own inputs.

full rationale

I walked the paper's derivation chain: Eq. (2)-(3) define the forward splatting color, Eq. (4)-(5) define MSAA as an average of subpixel composites, Eq. (6)-(9) define the adaptive weighting loss, Eq. (10)-(14) define the gradient-difference constraint, and Eq. (15)-(16) combine them. None of these quantities is defined in terms of the PSNR/SSIM/LPIPS results it is claimed to improve. The adaptive weight w_{i,j} depends on the per-pixel training error e_{i,j}, which is a standard supervised loss-weighting mechanism, not a fitted parameter renamed as a prediction. The gradient constraint compares predicted gradients to ground-truth gradients, and the MSAA output is a fixed average of four subpixel renderings; neither is constructed to equal the evaluation metrics. The paper contains no self-citations and no uniqueness theorem imported from the authors' prior work. The reported gains are measured on held-out test views against external baselines. Concerns about Eq. (3) not matching the canonical 3DGS alpha-compositing formula, and about unspecified hyperparameters (alpha, delta_k), are reproducibility/correctness issues, not circularity: they do not make the output equal to the input by construction.

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

No new physical or representational entities are introduced. The MSAA subsamples are sampling positions within existing pixels, and the adaptive weights are scalars in the loss. The main uncharged entries are hyperparameters, not entities.

free parameters (4)

lambda1, lambda2, lambda3 = 0.8, 0.2, 0.1
Composite loss coefficients in Eq. (15). Reported as fixed constants with no sensitivity analysis or search procedure; the text also says lambda3 'can be gradually increased during training.'
alpha = not reported
Baseline weight floor in the adaptive weight map, Eq. (7). Controls how strongly high-error pixels are amplified, but its value is never given.
subpixel offsets delta_k = not reported
The four MSAA sample positions in Eq. (4) are described only as 'strategically offset'; the actual offsets are not provided, making the rendering path underdetermined.
epsilon = not reported
Numerical stability constant in Eq. (7). Unspecified, though it affects the normalized weight map only weakly.

assumptions (3)

domain assumption Eq. (2)-(3) describe 3DGS rendering correctly.
Eq. (3) is a normalized weighted average without the transmittance-based ordering used by actual 3DGS. If this is only a schematic, the text should say so; if it is the implemented model, the baseline is misstated.
domain assumption Four subpixel samples per pixel with uniform averaging provide an effective differentiable anti-aliasing approximation.
This is the core design premise of Section 3.2. No analysis of sample count, offset placement, or interaction with the Gaussian splatting optimization is provided.
ad hoc to paper The loss weights and adaptive floor generalize across all evaluation scenes.
The fixed values in Section 4.1.2 and Eq. (7) are asserted without cross-validation or sensitivity analysis, yet they underpin the reported gains.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-Sample Anti-Aliasing and Constrained Optimization for 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/C73PAKOV

@misc{pith2026250810507,
  author       = {Pith},
  title        = {Pith review of: Multi-Sample Anti-Aliasing and Constrained Optimization for 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C73PAKOV}},
  note         = {Machine review of arXiv:2508.10507}
}

read the original abstract

Recent advances in 3D Gaussian splatting have significantly improved real-time novel view synthesis, yet insufficient geometric constraints during scene optimization often result in blurred reconstructions of fine-grained details, particularly in regions with high-frequency textures and sharp discontinuities. To address this, we propose a comprehensive optimization framework integrating multisample anti-aliasing (MSAA) with dual geometric constraints. Our system computes pixel colors through adaptive blending of quadruple subsamples, effectively reducing aliasing artifacts in high-frequency components. The framework introduces two constraints: (a) an adaptive weighting strategy that prioritizes under-reconstructed regions through dynamic gradient analysis, and (b) gradient differential constraints enforcing geometric regularization at object boundaries. This targeted optimization enables the model to allocate computational resources preferentially to critical regions requiring refinement while maintaining global consistency. Extensive experimental evaluations across multiple benchmarks demonstrate that our method achieves state-of-the-art performance in detail preservation, particularly in preserving high-frequency textures and sharp discontinuities, while maintaining real-time rendering efficiency. Quantitative metrics and perceptual studies confirm statistically significant improvements over baseline approaches in both structural similarity (SSIM) and perceptual quality (LPIPS).

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 23 canonical work pages

[1]

Visual Communications and Image Processing 20004067, 2–13 (2000) 18

Shum, H., Kang, S.B.: Review of image-based rendering techniques. Visual Communications and Image Processing 20004067, 2–13 (2000) 18

work page 2000
[2]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), vol

Seitz, S.M., Curless, B., Diebel, J., Scharstein, D., Szeliski, R.: A comparison and evaluation of multi-view stereo reconstruction algorithms. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), vol. 1, pp. 519–528 (2006).https://doi.org/10.1109/CVPR.2006.19

doi:10.1109/cvpr.2006.19 2006
[3]

Communications of the ACM65(1), 99–106 (2021)

Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM65(1), 99–106 (2021)

2021
[4]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp

Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srini- vasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 5855–5864 (2021)

work page 2021
[6]

ACM transactions on graphics (TOG) 41(4), 1–15 (2022)

Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG) 41(4), 1–15 (2022)

2022
[8]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp

Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High- fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 14346–14355 (2021)

work page 2021
[9]

ACM Trans

Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023)

2023
[10]

SIGGRAPH Courses2(3), 4 (2011)

Jimenez, J., Gutierrez, D., Yang, J., Reshetov, A., Demoreuille, P., Berghoff, T., Perthuis, C., Yu, H., McGuire, M., Lottes, T.,et al.: Filtering approaches for real-time anti-aliasing. SIGGRAPH Courses2(3), 4 (2011)

work page 2011

Show all 35 references

[11]

In: Proceedings of the 20th Annual Con- ference on Computer Graphics and Interactive Techniques, pp

Akeley, K.: Reality engine graphics. In: Proceedings of the 20th Annual Con- ference on Computer Graphics and Interactive Techniques, pp. 109–116 (1993). https://doi.org/10.1145/166117.166131

1993
[12]

In: International Conference on Machine Learning, pp

Xu, L., Ren, J., Yan, Q., Liao, R., Jia, J.: Deep edge-aware filters. In: International Conference on Machine Learning, pp. 1669–1678 (2015)

2015
[13]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

Hua, M., Bie, X., Zhang, M., Wang, W.: Edge-aware gradient domain optimization 19 framework for image filtering by local propagation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2838–2845 (2014). https://doi.org/10.1109/CVPR.2014.363

2014 doi
[14]

arXiv preprint arXiv:2410.01804 (2024)

Mai, A., Hedman, P., Kopanas, G., Verbin, D., Futschik, D., Xu, Q., Kuester, F., Barron, J.T., Zhang, Y.: Ever: Exact volumetric ellipsoid rendering for real-time view synthesis. arXiv preprint arXiv:2410.01804 (2024)

2024 arXiv
[15]

In: Proceedings of the IEEE European Conference on Computer Vision (ECCV), pp

Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Proceedings of the IEEE European Conference on Computer Vision (ECCV), pp. 694–711 (2016)

2016
[16]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

Zhang, K., Zuo, W., Zhang, L.: Learning a single convolutional super-resolution network for multiple degradations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3262–3271 (2018). https://doi.org/10.1109/CVPR.2018.00344

2018
[17]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual net- works for single image super-resolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 136–144 (2017). https://doi.org/10.1109/CVPRW.2017.151

2017 doi
[18]

In: IEEE Transactions on Computational Imaging, vol

Zhao, H., Gallo, O., Frosio, I., Kautz, J.: Loss functions for image restoration with neural networks. In: IEEE Transactions on Computational Imaging, vol. 3, pp. 47–57 (2017).https://doi.org/10.1109/TCI.2016.2644865

2017
[19]

IEEE transactions on image processing 13(4), 600–612 (2004)

Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600–612 (2004)

2004
[20]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

Flynn, J., Neulander, I., Philbin, J., Snavely, N.: Deepstereo: Learning to predict new views from the world’s imagery. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5515–5524 (2016). https://doi.org/10.1109/CVPR.2016.595

2016 doi
[21]

In: ACM Transactions on Graphics (TOG), vol

Hedman, P., Kopf, J., Langguth, F., Goesele, M.: Deep blending for free-viewpoint image-based rendering. In: ACM Transactions on Graphics (TOG), vol. 37, pp. 1–15 (2018). https://doi.org/10.1145/3272127.3275084

2018
[22]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhöfer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2437–2446 (2019). https://doi.org/10.1109/CVPR.2019.00254

2019
[23]

Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition 20 (CVPR), 3504–3515 (2020)

Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition 20 (CVPR), 3504–3515 (2020)

2020
[24]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for uncon- strained photo collections. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7206–7215 ...

2021
[25]

In: Proceedings of the IEEE International Conference on Computer Vision (ICCV), pp

Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV), pp. 14335–14345 (2021)

2021
[26]

In: Proceedings of the IEEE European Conference on Computer Vision (ECCV), pp

Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields for compact neural scene representation. In: Proceedings of the IEEE European Conference on Computer Vision (ECCV), pp. 1–17 (2022)

2022
[27]

arXiv preprint arXiv:2010.08422 (2020)

Zhang, K., Barron, J.T., Tancik, M., Hedman, P., Srinivasan, P.P.: Nerf- gan: Learning implicit generative representations of 3d scenes. arXiv preprint arXiv:2010.08422 (2020)

2010 arXiv
[28]

In: IEEE Visualization Conference (VIS), pp

Linsen, L.: Point cloud representation through 3d gaussian models. In: IEEE Visualization Conference (VIS), pp. 27–34 (2001)

2001
[32]

In: Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pp

Deng, K., Liu, A., Zhu, J.-Y.: Depth-supervised nerf: Fewer views and faster training for free. In: Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pp. 12735–12744 (2022). https://doi.org/10.1109/CVPR52688.2022.01254

2022
[33]

IEEE computer graphics and applications14(4), 23–32 (1994)

Molnar, S., Cox, M., Ellsworth, D., Fuchs, H.: A sorting classification of parallel rendering. IEEE computer graphics and applications14(4), 23–32 (1994)

1994
[34]

Akenine-Moller, T., Haines, E., Hoffman, N.: Real-time Rendering, (2019) 21

2019
[35]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

Zhang, J., Zhan, F., Xu, M., Lu, S., Xing, E.: Fregs: 3d gaussian splatting with progressive frequency regularization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 21424–21433 (2024). https://doi.org/10.1109/CVPR52733.2024.02024

2024
[36]

ACM Transactions on Graphics (ToG)36(4), 1–13 (2017)

Knapitsch, A., Park, J., Zhou, Q.-Y., Koltun, V.: Tanks and temples: Benchmark- ing large-scale scene reconstruction. ACM Transactions on Graphics (ToG)36(4), 1–13 (2017)

2017
[37]

ACM Transactions on Graphics (ToG)37(6), 1–15 (2018)

Hedman, P., Philip, J., Price, T., Frahm, J.-M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG)37(6), 1–15 (2018)

2018
[38]

Advances in neural information processing systems 32 (2019)

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. Advances in neural information processing systems 32 (2019)

2019
[39]

In: Proceedings of the 32nd ACM International Conference on Multimedia (ACM MM), pp

Ye, Z., Li, W., Liu, S., Qiao, P., Dou, Y.: Absgs: Recovering fine details in 3d gaussian splatting. In: Proceedings of the 32nd ACM International Conference on Multimedia (ACM MM), pp. 1053–1061 (2024)

2024
[40]

In: European Conference on Computer Vision (ECCV), pp

Zhang, Z., Hu, W., Lao, Y., He, T., Zhao, H.: Pixel-gs: Density control with pixel- aware gradient for 3d gaussian splatting. In: European Conference on Computer Vision (ECCV), pp. 326–342 (2024). Springer 22

2024

Pith tools

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

[1] [1]

Visual Communications and Image Processing 20004067, 2–13 (2000) 18

Shum, H., Kang, S.B.: Review of image-based rendering techniques. Visual Communications and Image Processing 20004067, 2–13 (2000) 18

work page 2000

[2] [2]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), vol

Seitz, S.M., Curless, B., Diebel, J., Scharstein, D., Szeliski, R.: A comparison and evaluation of multi-view stereo reconstruction algorithms. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), vol. 1, pp. 519–528 (2006).https://doi.org/10.1109/CVPR.2006.19

doi:10.1109/cvpr.2006.19 2006

[3] [3]

Communications of the ACM65(1), 99–106 (2021)

Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM65(1), 99–106 (2021)

2021

[4] [4]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp

Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srini- vasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 5855–5864 (2021)

work page 2021

[5] [6]

ACM transactions on graphics (TOG) 41(4), 1–15 (2022)

Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG) 41(4), 1–15 (2022)

2022

[6] [8]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp

Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High- fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 14346–14355 (2021)

work page 2021

[7] [9]

ACM Trans

Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023)

2023

[8] [10]

SIGGRAPH Courses2(3), 4 (2011)

Jimenez, J., Gutierrez, D., Yang, J., Reshetov, A., Demoreuille, P., Berghoff, T., Perthuis, C., Yu, H., McGuire, M., Lottes, T.,et al.: Filtering approaches for real-time anti-aliasing. SIGGRAPH Courses2(3), 4 (2011)

work page 2011