Pith. sign in

REVIEW 4 major objections 4 minor 47 references

A training-free post-processing step can compress 3D Gaussian Splat scenes to a thousandth of their primitives while keeping the best reported rendering quality at every tested budget.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 18:03 UTC pith:YXKQ54GN

load-bearing objection A promising training-free 3DGS simplifier whose reported gains depend on an unstated cost weight and an undescribed covariance-to-parameter conversion; worth reviewing, but not reproducible as written. the 4 major comments →

arxiv 2603.16103 v2 pith:YXKQ54GN submitted 2026-03-17 cs.CV cs.GR

NanoGS: Training-Free Gaussian Splat Simplification

Butian Xiong , Rong Liu , Tiantian Zhou , Meida Chen , Zhiwen Fan , Andrew Feng This is my paper
classification cs.CV cs.GR
keywords 3D Gaussian Splattingmodel simplificationprimitive reductionmoment matchingI-divergenceK-nearest-neighbor graphtraining-freenovel view synthesis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

NanoGS claims that a trained 3D Gaussian Splat scene can be drastically simplified—down to 0.1% of its original primitives—by a purely geometric merging procedure that never looks at rendered images and never retrains. The method treats each splat as a blob of mass, restricts candidate merges to a sparse nearest-neighbor graph, and only fuses pairs whose two-splat mixture is well approximated by a single Gaussian, judged by an information divergence plus an appearance-feature distance. The merged splat is built by mass-weighted moment matching, so larger, more opaque splats dominate, with opacity combined as a probabilistic union. The authors report the best PSNR on all four benchmarks (21 scenes) at all three compaction ratios, with the advantage growing as the budget shrinks. If correct, this would make Splat models far cheaper to store and stream, and would extend compaction to models generated, edited, or converted without access to training images.

Core claim

The paper claims Gaussian Splat simplification can be done as progressive local pairwise merging in 3D, with no image supervision. A KNN graph supplies candidate pairs; each edge is scored by a merge cost combining the I-divergence between the two-splat mixture and its single-Gaussian approximation with the L2 distance of appearance features. The lowest-cost disjoint edges are collapsed in parallel via Mass Preserved Moment Matching: merged mean and covariance are mass-weighted first and second moments, opacity is the probabilistic union. Repeating with periodic refresh yields a coarse-to-fine hierarchy. Across four benchmarks (21 scenes), it reports the best PSNR at every budget, beating co

What carries the argument

The merge cost C(i,j)=D_geo(i,j)+D_app(i,j). D_geo is the I-divergence between the normalized two-splat mixture and the single merged Gaussian, estimated by Monte Carlo sampling; D_app is the squared L2 distance between appearance feature vectors (SH coefficients). The merge operator MPMM computes the merged mean as a mass-weighted average, the merged covariance as the mass-weighted sum of within-splat covariance plus a between-mean dispersion term, opacity as 1-(1-α_i)(1-α_j), and appearance as a mass-weighted blend. The sparse KNN graph restricts candidate pairs to local neighborhoods, making cost evaluation O(kN) instead of O(N^2).

Load-bearing premise

The load-bearing premise is that the merged covariance matrix computed in Eq. (9)-(10) can be converted back into a scale-and-rotation (s,q) parameterization with negligible error; the paper never describes this decomposition, yet both the mass-preserving property and the reported PSNR depend on it.

What would settle it

Run NanoGS on the 'chair' scene at ρ=0.01 and, for every merged splat, decompose the merged covariance Σ_m from Eq. (9) into its eigen-system and compare the reconstructed covariance from the stored (scale, rotation) parameters to the original Σ_m. If the round-trip error is non-negligible (for example, relative Frobenius error above 1%), or if the PSNR in Table 1 for that scene (22.28 dB at ρ=0.01) cannot be reproduced from the released code and model without an additional conversion step, the central claim is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Compaction becomes a drop-in post-processing step: any existing 3DGS model can be simplified without training images, camera poses, or differentiable rendering.
  • Because the output keeps the standard 3DGS parameterization, it can be fed directly into existing renderers and combined with bit-level compression (quantization, entropy coding) for further storage reduction.
  • At ρ=0.001 (a 1000x primitive reduction), the method still maintains usable rendering quality, suggesting LOD-style multi-resolution rendering from a single model.
  • The CPU-only, model-only design means it can be applied to Splat assets produced by generative models, editors, or mesh/point-cloud converters, where training supervision is unavailable.
  • The progressive hierarchy of representations supports choosing any target compaction ratio after the fact, without re-running the pipeline.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves unspecified how the merged covariance Σ_m is converted back into the scale/rotation parameters (s_m, q_m) of standard 3DGS; if that decomposition is lossy or numerically unstable, the mass-preservation argument is weaker than stated. Testing the round-trip error on a single scene would quantify this.
  • The Monte Carlo estimator for the I-divergence uses a small fixed sample count per pair; over tens of thousands of merges, estimator variance could bias edge selection. A variance-reduction or deterministic approximation could change which pairs are merged.
  • Because the merge cost is purely local and appearance is compared only via SH-coefficient L2 distance, scenes dominated by view-dependent specular effects may require an appearance-aware cost that the paper leaves as future work; a natural test is to compare NanoGS against image-supervised pruning on a specular-heavy scene.
  • NanoGS is orthogonal to bit-level compression methods, so a combined pipeline (NanoGS for primitive reduction, then quantization/entropy coding on the survivors) is an obvious next step; the paper does not measure the combined savings, and the effective end-to-end footprint could be much smaller than either alone.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper presents NanoGS, a training-free post-hoc simplification method for 3D Gaussian Splatting. Given a trained splat set, it prunes low-opacity splats, builds a k-NN graph, scores candidate edges by a merge cost combining a geometric I-divergence between the two-splat mixture and its moment-matched Gaussian approximation with an appearance-feature discrepancy, greedily collapses disjoint low-cost pairs via mass-preserving moment matching, and iterates with periodic neighborhood refresh. The authors report consistently higher PSNR than LightGS, PUP3DGS, and GHAP at compaction ratios 0.1, 0.01, and 0.001 across NeRF-Synthetic, Mip-NeRF 360, Tanks & Temples, and Deep Blending, with average gains up to +5.46 dB.

Significance. If the claims hold, NanoGS would be a practically valuable contribution: it works without training images or GPU optimization, preserves the standard 3DGS representation, and scales via local graph operations. The use of mass-preserving moment matching and an I-divergence cost is principled, and the paper provides per-scene tables and ablations supporting the design choices. The main limitation at this stage is reproducibility: the cost definition and several implementation parameters are underspecified, and the covariance-to-scale/quaternion conversion is missing. These gaps must be closed before the empirical claims can be independently verified.

major comments (4)
  1. [§3.2, Eq. (1); Fig. 2] The merge cost is stated inconsistently. Eq. (1) defines C(i,j)=Dgeo(i,j)+Dapp(i,j), but Fig. 2's caption and diagram give C(i,j)=D_app(i,j)+λ_a D_geo(i,j). λ_a is never introduced or assigned a value anywhere in the text or appendices. This is load-bearing because Dgeo is an I-divergence in nats, while Dapp is a squared L2 distance on SH coefficients; these quantities have different units and scales, so without a weight or normalization the greedy edge selection in Algorithm 1 is not actually defined by the paper. The reported PSNR improvements cannot be reproduced or attributed to the proposed cost without specifying the exact combination and the value of λ_a.
  2. [§3.3, Eqs. (9)–(10); Algorithm 1] The merge operator outputs a 3×3 covariance Σ_m, but 3DGS stores each splat by scale s and quaternion q. The conversion from Σ_m back to (s_m,q_m) is never described. This step is not cosmetic: subsequent merge-cost evaluations and rendering use the parameterized Gaussian, and an incorrect decomposition (e.g., unconstrained eigen-decomposition with sign/ordering ambiguities, or negative eigenvalues from the moment update) changes all later merges. The paper's claim to 'preserve the standard 3DGS parameterization' therefore lacks a needed technical detail. Please specify the decomposition method and any regularization used in MergePairs.
  3. [§3.2, Appendix A, Algorithm 1] Several implementation parameters that directly determine the cost and the schedule are left unspecified: the Monte Carlo sample count S is described only as a 'small fixed constant'; Algorithm 1 uses an edge-cost block size B and 'cost params c_p' that are never defined in §3.2/3.4; and the opacity threshold τ is given only in the pseudocode. Since the greedy merging order and the I-divergence estimates depend on these choices, the paper should state concrete values (or a sensitivity study) for k, S, τ, B, c_p, and λ_a.
  4. [§4, Table 1] The comparison protocol evaluates LightGS and PUP3DGS only in their pruning/selection stage and not with the fine-tuning stage those methods are designed to use. This is stated, but the paper's summary claims improvement over 'state-of-the-art compression methods' at every budget. The claim is valid only for the pruning-only comparison, not for the complete published pipelines. Either include full-pipeline numbers for the baselines or adjust the wording of the central claim.
minor comments (4)
  1. [Table 1] The dataset column label 'Mips-360 [4]' contradicts the text, which refers to Mip-NeRF 360 [5]; please correct the reference and label.
  2. [Figs. 3 and 4 captions] The figure captions appear truncated ('Ground Trut', 'ratio = 0.'). Please ensure the final captions are complete.
  3. [§3.3, Eq. (11)] The term 'mass-preserving' is potentially misleading: Eq. (11) does not conserve the total unnormalized mass w_i+w_j, since α_m = α_i+α_j−α_i α_j and the merged scale changes the mass factor. A brief clarification of why the name is used would help.
  4. [Appendix C] The text promises 'additional qualitative comparison will be provided in appendix,' but Appendix C mostly contains per-scene quantitative tables and mentions a video. Please make the appendix content match the promise.

Circularity Check

0 steps flagged

No significant circularity; NanoGS's derivation and benchmark results are self-contained and not forced by construction.

full rationale

I examined the paper's derivation chain: the merge cost in Eq. (1)-(6) combines a geometric I-divergence term with an appearance distance, the MPMM operator in Eq. (7)-(11) is defined by explicit mass-weighted moment formulas, and the greedy edge-collapse algorithm is a standard local optimization procedure. The cost function does use the merged Gaussian q_m produced by the same merge operator (Eq. 5), but this is the normal self-referential structure of edge-collapse methods; it does not encode any target PSNR or benchmark value, and the method is not trained or fitted to the evaluation data. The paper reports external benchmark comparisons, and the claimed improvements are empirical outcomes rather than consequences of the definitions. The self-citations to the authors' prior work ([25], [26], [40]) appear only in related-work/application contexts and are not load-bearing for the central derivation. The discrepancies flagged by the reader—the un-introduced λ_a in Fig. 2 versus Eq. (1), and the unspecified conversion from Σ_m back to scale/quaternion parameters—are reproducibility and correctness concerns, not circularity. No step in the paper reduces a 'prediction' to a fitted input or imports a load-bearing conclusion from a self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities; it reuses the 3DGS representation. The main burden is on several unspecified hyperparameters and unstated assumptions about the covariance decomposition and Monte Carlo accuracy, which are needed to make the method reproducible.

free parameters (5)
  • k (neighborhood size)
    Controls the KNN graph density and the candidate merge set; value not stated.
  • S (Monte Carlo samples)
    Number of samples per pair for estimating the I-divergence; described as 'small fixed constant' but not quantified.
  • τ (opacity prune threshold)
    Used for initial filtering; adapted as τ' = min(τ, median(α)) but base τ is not given.
  • λ_a (appearance-geometry trade-off)
    Appears in Fig. 2 caption as C = D_app + λ_a D_geo but is absent from Eq. (1); no value or tuning described.
  • Edge-cost block size B and cost params c_p
    Algorithm paramaters for blockwise edge scoring; values not specified.
axioms (5)
  • standard math Splat mass is proportional to (2π)^(3/2) α ∏ s_k, i.e., the integral of an unnormalized Gaussian.
    Used to define mass weights in Eq. (7) and derived from the Gaussian integral.
  • standard math Moment matching yields the Gaussian that minimizes KL(˜pij || q_m) within the Gaussian family.
    Standard exponential-family result invoked implicitly in Section 3.3.
  • domain assumption The Monte Carlo estimator of the I-divergence is unbiased and sufficiently accurate with the (unspecified) number of samples S.
    Appendix A states S is small and shared across pairs; no error analysis is provided.
  • domain assumption The Porter-Duff 'over' operator correctly models opacity aggregation for merged splats.
    Used in Eq. (11) to combine opacities as probabilistic union; assumes independence and no spatial correlation.
  • ad hoc to paper Greedy disjoint edge selection approximates the global optimal simplification.
    The greedy matching is a heuristic; no bound or proof of near-optimality is given relative to a global merge objective.

pith-pipeline@v1.3.0-alltime-deepseek · 121 in / 7930 out tokens · 160493 ms · 2026-08-02T18:03:58.785850+00:00 · methodology

0 comments
read the original abstract

3D Gaussian Splat (3DGS) enables high-fidelity, real-time novel view synthesis by representing scenes with large sets of anisotropic primitives, but often requires millions of Splats, incurring significant storage and transmission costs. Most existing compression methods rely on GPU-intensive post-training optimization with calibrated images, limiting practical deployment. We introduce \textbf{NanoGS}, a training-free and lightweight framework for Gaussian Splat simplification. Instead of relying on image-based rendering supervision, NanoGS formulates simplification as local pairwise merging over a sparse spatial graph. The method approximates a pair of Gaussians with a single primitive using mass preserved moment matching and evaluates merge quality through a principled merge cost between the original mixture and its approximation. By restricting merge candidates to local neighborhoods and selecting compatible pairs efficiently, NanoGS produces compact Gaussian representations while preserving scene structure and appearance. NanoGS operates directly on existing Gaussian Splat models, runs efficiently on CPU, and preserves the standard 3DGS parameterization, enabling seamless integration with existing rendering pipelines. Experiments demonstrate that NanoGS substantially reduces primitive count while maintaining high rendering fidelity, providing an efficient and practical solution for Gaussian Splat simplification. Our project website is available at \href{https://saliteta.github.io/NanoGS/}{https://saliteta.github.io/NanoGS/}.

Figures

Figures reproduced from arXiv: 2603.16103 by Andrew Feng, Butian Xiong, Meida Chen, Rong Liu, Tiantian Zhou, Zhiwen Fan.

Figure 1
Figure 1. Figure 1: NanoGS reduces Gaussian Splat primitive count from the full model to in￾creasingly compact representations. NanoGS achieves substantial compaction ratio while preserving visual fidelity and geometric structure without GPU-intensive opti￾mization or calibrated images. Abstract. 3D Gaussian Splat (3DGS) enables high-fidelity, real-time novel view synthesis by representing scenes with large sets of anisotropi… view at source ↗
Figure 2
Figure 2. Figure 2: NanoGS pipeline overview. (Top) Starting from an initial set of Gaus￾sian splats, NanoGS constructs a sparse k-nearest-neighbor graph, evaluates a merge cost for each candidate edge, and collapses the lowest-cost disjoint pairs progressively. (Bottom-left) The merge cost combines an appearance term and a geometry term. The geometry cost measures the I-divergence between the original two-splat mixture and i… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on chair from NeRF Synthetic dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on Mip-NeRF 360 dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: 3 OpenAI, “ChatGPT.” https://chatgpt.com/ (accessed March 12, 2026). 4 Google, “Gemini.” https://gemini.google.com/ (accessed March 12, 2026). 5 Jianfeng Xiang et al., “Structured 3D Latents for Scalable and Versatile 3D Genera￾tion,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025 [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Generated results and corresponding compression results [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Generated results and corresponding compression results with multimodal in￾puts [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 5 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2502.19457 (2025)

    Ali, M.S., Zhang, C., Cagnazzo, M., Valenzise, G., Tartaglione, E., Bae, S.H.: Compression in 3d gaussian splatting: A survey of methods, trends, and future directions. arXiv preprint arXiv:2502.19457 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    Computer Graphics Forum (2025)

    Bagdasarian, M.T., Knoll, P., Li, Y., Barthel, F., Hilsmann, A., Eisert, P., Mor- genstern, W.: 3dgs.zip: A survey on 3d gaussian splatting compression methods. Computer Graphics Forum (2025)

    2025

  3. [3]

    Journal of machine learning research6(Oct), 1705–1749 (2005)

    Banerjee, A., Merugu, S., Dhillon, I.S., Ghosh, J.: Clustering with bregman diver- gences. Journal of machine learning research6(Oct), 1705–1749 (2005)

    2005

  4. [4]

    In: CVPR (2022)

    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: CVPR (2022)

    2022

  5. [5]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip- nerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5470–5479 (2022)

    2022

  6. [6]

    In: ICCV (2023)

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

    2023

  7. [7]

    USSR computational mathematics and mathematical physics7(3), 200–217 (1967)

    Bregman, L.M.: The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming. USSR computational mathematics and mathematical physics7(3), 200–217 (1967)

    1967

  8. [8]

    In: European Conference on Computer Vision

    Chen,Y.,Wu,Q.,Lin,W.,Harandi,M.,Cai,J.:Hac:Hash-gridassistedcontextfor 3d gaussian splatting compression. In: European Conference on Computer Vision. Springer (2024)

    2024

  9. [9]

    arXiv preprint arXiv:2501.12255 (2025)

    Chen, Y., Wu, Q., Lin, W., Harandi, M., Cai, J.: Hac++: Towards 100x compres- sion of 3d gaussian splatting. arXiv preprint arXiv:2501.12255 (2025)

    Pith/arXiv arXiv 2025

  10. [10]

    arXiv preprint arXiv:2403.11247 (2024)

    Deng, T., Chen, Y., Zhang, L., Yang, J., Yuan, S., Liu, J., Wang, D., Wang, H., Chen, W.: Compact 3d gaussian splatting for dense visual slam. arXiv preprint arXiv:2403.11247 (2024)

    Pith/arXiv arXiv 2024

  11. [11]

    In: Pro- ceedings of the ninth ACM SIGKDD international conference on Knowledge dis- covery and data mining

    Dhillon, I.S., Mallela, S., Modha, D.S.: Information-theoretic co-clustering. In: Pro- ceedings of the ninth ACM SIGKDD international conference on Knowledge dis- covery and data mining. pp. 89–98 (2003)

    2003

  12. [12]

    electronicarts: mesh2splat: Fast mesh to 3d gaussian splat conversion (2025), https://github.com/electronicarts/mesh2splat

    2025

  13. [13]

    Advances in neural information processing systems37, 140138–140158 (2024)

    Fan, Z., Wang, K., Wen, K., Zhu, Z., Xu, D., Wang, Z., et al.: Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. Advances in neural information processing systems37, 140138–140158 (2024)

    2024

  14. [14]

    Franke, L., Fink, L., Stamminger, M.: Vr-splatting: Foveated radiance field ren- dering via 3d gaussian splatting and neural points. Proc. ACM Comput. Graph. Interact. Tech. (2025)

    2025

  15. [15]

    In: European Conference on Computer Vision

    Girish, S., Gupta, K., Shrivastava, A.: Eagles: Efficient accelerated 3d gaussians with lightweight encodings. In: European Conference on Computer Vision. pp. 54–71. Springer (2024) 16 Xiong et al

    2024

  16. [16]

    ECCV (2024)

    Guédon, A., Lepetit, V.: Gaussian frosting: Editable complex radiance fields with real-time rendering. ECCV (2024)

    2024

  17. [17]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Hanson, A., Tu, A., Singla, V., Jayawardhana, M., Zwicker, M., Goldstein, T.: Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splatting. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5949–5958 (2025)

    2025

  18. [18]

    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

  19. [19]

    In: SIGGRAPH (2024)

    Jiang, Y., Yu, C., Xie, T., Li, X., Feng, Y., Wang, H., Li, M., Lau, H., Gao, F., Yang, Y., Jiang, C.: Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality. In: SIGGRAPH (2024)

    2024

  20. [20]

    In: SIGGRAPH (2023)

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. In: SIGGRAPH (2023)

    2023

  21. [21]

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

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

    2017

  22. [22]

    Lee, J.C., Ko, J.H., Park, E.: Optimized minimal 3d gaussian splatting (2025), https://arxiv.org/abs/2503.16924

    arXiv 2025

  23. [23]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lee, J.C., Rho, D., Sun, X., Ko, J.H., Park, E.: Compact 3d gaussian representation for radiance field. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21719–21728 (2024)

    2024

  24. [24]

    arXiv preprint arXiv:2411.18473 (2024)

    Liu, L., Chen, Z., Jiang, W., Wang, W., Xu, D.: Hemgs: A hybrid entropy model for 3d gaussian splatting data compression. arXiv preprint arXiv:2411.18473 (2024)

    Pith/arXiv arXiv 2024

  25. [25]

    arXiv preprint arXiv:2510.03312 (2025)

    Liu, R., Gao, Z., Planche, B., Chen, M., Nguyen, V.N., Zheng, M., Choudhuri, A., Chen, T., Wang, Y., Feng, A., et al.: Universal beta splatting. arXiv preprint arXiv:2510.03312 (2025)

    arXiv 2025

  26. [26]

    In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers

    Liu, R., Sun, D., Chen, M., Wang, Y., Feng, A.: Deformable beta splatting. In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers. pp. 1–11 (2025)

    2025

  27. [27]

    In: SIGGRAPH Asia 2024 Conference Papers

    Mallick, S.S., Goel, R., Kerbl, B., Steinberger, M., Carrasco, F.V., De La Torre, F.: Taming 3dgs: High-quality radiance fields with limited resources. In: SIGGRAPH Asia 2024 Conference Papers. pp. 1–11 (2024)

    2024

  28. [28]

    In: ECCV (2020)

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

    2020

  29. [29]

    In: SIGGRAPH (2022)

    Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: SIGGRAPH (2022)

    2022

  30. [30]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Niedermayr, S., Stumpfegger, J., Westermann, R.: Compressed 3d gaussian splat- ting for accelerated novel view synthesis. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 10349–10358 (2024)

    2024

  31. [31]

    Proceedings of the ACM on Computer Graphics and Interactive Techniques7(1) (May 2024)

    Papantonakis, P., Kopanas, G., Kerbl, B., Lanvin, A., Drettakis, G.: Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques7(1) (May 2024)

    2024

  32. [32]

    PlayCanvas: Supersplat: 3d gaussian splat editor (2024),https://github.com/ playcanvas/supersplat

    2024

  33. [33]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Qin, M., Li, W., Zhou, J., Wang, H., Pfister, H.: Langsplat: 3d language gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20051–20060 (2024)

    2024

  34. [34]

    IEEE Transactions on Aerospace and Electronic Systems43(3), 989–999 (2007) NanoGS 17

    Runnalls, A.R.: Kullback-leibler approach to gaussian mixture reduction. IEEE Transactions on Aerospace and Electronic Systems43(3), 989–999 (2007) NanoGS 17

    2007

  35. [35]

    In: ICLR (2024)

    Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In: ICLR (2024)

    2024

  36. [36]

    In: Computer Graphics Forum

    Wang, C., Wolski, K., Kerbl, B., Serrano, A., Bemana, M., Seidel, H.P., Myszkowski,K.,Leimkühler,T.:Cinematicgaussians:Real-timehdrradiancefields with depth of field. In: Computer Graphics Forum. p. e15214. Wiley Online Library (2024)

    2024

  37. [37]

    In: Advances in Neural Information Processing Systems (NeurIPS 2025) (2025)

    Wang, T., Li, M., Zeng, G., Meng, C., Zhang, Q.: Gaussian herding across pens: An optimal transport perspective on global gaussian reduction for 3dgs. In: Advances in Neural Information Processing Systems (NeurIPS 2025) (2025)

    2025

  38. [38]

    Advances in neural information processing systems37, 51532–51551 (2024)

    Wang, Y., Li, Z., Guo, L., Yang, W., Kot, A., Wen, B.: Contextgs: Compact 3d gaussian splatting with anchor level context model. Advances in neural information processing systems37, 51532–51551 (2024)

    2024

  39. [39]

    In: CVPR (2024)

    Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. In: CVPR (2024)

    2024

  40. [40]

    In: ICLR (2026)

    Xiong, B., Liu, R., Xu, K., Chen, M., Feng, A.: Splat feature solver. In: ICLR (2026)

    2026

  41. [41]

    In: ICLR (2024)

    Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In: ICLR (2024)

    2024

  42. [42]

    In: CVPR (2024)

    Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In: CVPR (2024)

    2024

  43. [43]

    In: CVPR (2024)

    Yi, T., Fang, J., Wang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In: CVPR (2024)

    2024

  44. [44]

    optimizing-sparsifying

    Zhang, Y., Jia, W., Niu, W., Yin, M.: Gaussianspa: An" optimizing-sparsifying" simplification framework for compact and high-quality 3d gaussian splatting. In: CVPR (2025)

    2025

  45. [45]

    Advances in Neural Information Processing Systems37, 122434–122457 (2024)

    Zhang, Z., Song, T., Lee, Y., Yang, L., Peng, C., Chellappa, R., Fan, D.: Lp- 3dgs: Learning to prune 3d gaussian splatting. Advances in Neural Information Processing Systems37, 122434–122457 (2024)

    2024

  46. [46]

    Zhou, J., Fan, L., Chen, X., Huang, L., Liu, S., Li, H.: Gaussianpainter: Painting point cloud into 3d gaussians with normal guidance (2024),https://arxiv.org/ abs/2412.17715

    Pith/arXiv arXiv 2024

  47. [47]

    ChatGPT

    Zhou,S.,Chang,H.,Jiang,S.,Fan,Z.,Zhu,Z.,Xu,D.,Chari,P.,You,S.,Wang,Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In: CVPR (2024) 18 Xiong et al. NanoGS: Training-Free Gaussian Splat Simplification Supplementary Material A Special Case for I Divergence Analytical cases and practical estimator.For two...

    2024