pith. machine review for the scientific record. sign in

arxiv: 2604.26799 · v2 · submitted 2026-04-29 · 💻 cs.CV · cs.GR· cs.MM

Recognition: unknown

MesonGS++: Post-training Compression of 3D Gaussian Splatting with Hyperparameter Searching

Shuzhao Xie , Junchen Ge , Weixiang Zhang , Jiahang Liu , Chen Tang , Yunpeng Bai , Shijia Ge , Jingyan Jiang
show 5 more authors
Yuzhi Huang Fengnian Yang Cong Zhang Xiaoyi Fan Zhi Wang
Authors on Pith no claims yet

Pith reviewed 2026-05-08 03:15 UTC · model grok-4.3

classification 💻 cs.CV cs.GRcs.MM
keywords 3D Gaussian Splattingpost-training compressionhyperparameter searchinteger linear programmingnovel view synthesisrate-distortion optimizationmodel quantizationoctree coding
0
0 comments X

The pith

MesonGS++ compresses 3D Gaussian Splatting models over 34 times after training while preserving rendering quality and hitting exact size targets.

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

The paper presents MesonGS++ as a post-training codec that compresses 3D Gaussian Splatting models for novel view synthesis. It combines pruning, octree coding, transformations, selective quantization, and mixed-precision entropy coding, but its key step is automating the main size and quality controls through a fast search that meets a chosen storage budget. The search uses a linear size predictor plus 0-1 integer linear programming on reserve ratios and bit widths, accelerated by a parallel quantization routine. This matters because 3D Gaussian models are large and hard to store or transmit, so reliable high-ratio compression without retraining could make real-time rendering practical on more hardware. The method claims to beat prior post-training compressors and even improve upon the original uncompressed quality in one tested case.

Core claim

MesonGS++ integrates joint importance-based pruning, octree geometry coding, attribute transformation, selective vector quantization for spherical harmonics, and group-wise mixed-precision quantization with entropy coding into a size-aware codec. It models reserve ratio and bit-width allocation as the main rate-distortion controls and solves for configurations that meet a target storage budget through discrete sampling and 0-1 integer linear programming. A linear size estimator and CUDA-accelerated quantization operator make the search fast enough for practical use. Experiments show the resulting models achieve over 34 times compression, meet size targets accurately, outperform other post-3D

What carries the argument

The hyperparameter search that jointly optimizes reserve ratio and bit-width allocation under a target size constraint via discrete sampling and 0-1 integer linear programming, guided by a linear size estimator.

Load-bearing premise

The linear size estimator plus the integer programming search finds parameter settings that deliver the reported quality gains consistently across scenes without hidden per-scene tuning.

What would settle it

Running the compressed model at the claimed 20 times ratio on the Stump scene and measuring PSNR below the vanilla 3D Gaussian Splatting baseline would falsify the quality-surpassing claim; observing actual file sizes that miss the stated targets by more than a few percent would falsify the size-control claim.

Figures

Figures reproduced from arXiv: 2604.26799 by Chen Tang, Cong Zhang, Fengnian Yang, Jiahang Liu, Jingyan Jiang, Junchen Ge, Shijia Ge, Shuzhao Xie, Weixiang Zhang, Xiaoyi Fan, Yunpeng Bai, Yuzhi Huang, Zhi Wang.

Figure 1
Figure 1. Figure 1: Quantile-quantile curve. x% of the least important Gaussians contribute to y% of total importance. For both kinds of scenes, 40% of the Gaussians contribute over 80% of the importance. The importance refers to the contribution to the final rendering results. ferability to other compressed Gaussian representations. II. PRELIMINARY 3D-GS [14] is an explicit 3D scene representation in the form of point clouds… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed compression pipeline. ⃝1 We first prune insignificant Gaussians using a joint importance metric that combines view-dependent and view-independent cues. ⃝2 We then compress geometry by building an octree to obtain voxelized coordinates, and apply RAHT to opacity, quaternion, and degree-0 spherical harmonics. For higher-degree SH coefficients, we retain a subset directly and compress… view at source ↗
Figure 3
Figure 3. Figure 3: 2D example of RAHT. 32 bit 1 bit Our w/o FT view at source ↗
Figure 5
Figure 5. Figure 5: Procedure of Hyperparameter Search. We formulate the hyperparameter search problem as the mixed integer programming, with the optimization goal of reducing quality distortion and constraint by the size. As shown in the circle in the left, each iteration is started by adjusting the reserve ratio τ and bit-width setting Q. Algo. 1 presents our hyperparameter optimization algorithm, which aims to meet the siz… view at source ↗
Figure 6
Figure 6. Figure 6: Rate-distortion curves for post-training compression. MesonGS++ achieves better performance across diverse datasets and size budgets. estimated quality loss matrix Ω ∈ R C×B×Q with: Ω(i, j, b) = |Aˆb i,j − Ai,j |. (11) The | · | can be 1-norm, 2-norm, or ∞-norm. We plot the relationship between PNSR and Ω in the supplementary mate￾rial, which reveals that minimizing Ω is equal to maximizing PSNR. Besides, … view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison. MesonGS++ preserves visual fidelity after compression, producing renderings that closely match the original 3DGS model. Compared with competing post-training codecs, our method better preserves reflective regions and fine local structures while introducing fewer visible artifacts. Notably, on the Stump and Garden scenes, our method not only achieves higher PSNR than the original 3DG… view at source ↗
Figure 8
Figure 8. Figure 8: Euler-angle-based vs. Covariance-based. “A/B/C” refers to the “λc / the size of the compressed file / the percent of positive-definite covariance matrices”. Replacing scales and rotations with covariance leads to white line artifacts, which greatly affects the visual effect. We adjust the final file size by compressing a portion of the covariance using the λc. TABLE I: Ablation study of different stages. “… view at source ↗
Figure 11
Figure 11. Figure 11: Bit widths of quantization groups. The heatmap visualizes the bit-width assigned to each quantization group. With 10 channels and 80 blocks, there are 800 groups in total. The experiment is conducted on the bicycle scene. TABLE IV: Superiority of 0-1 ILP. “GA”: Genetic Algorithm. With up to 16 bit-width choices, the Vanilla ILP and GA that are widely adopted in model quantization methods are unable to qui… view at source ↗
Figure 13
Figure 13. Figure 13: Transferability to online 3DGS compression. (a) Transferring the compression configuration from MesonGS++ helps HAC++ converge faster without sacrificing final ac￾curacy. (b) Ablation over the transferred priors verifies the effectiveness of params, mask, and qbits, and shows that MesonGS++-based initialization achieves performance com￾parable to the original HAC++. geometry and appearance. Lee et al. [84… view at source ↗
read the original abstract

3D Gaussian Splatting (3DGS) achieves high-quality novel view synthesis with real-time rendering, but its storage cost remains prohibitive for practical deployment. Existing post-training compression methods still rely on many coupled hyperparameters across pruning, transformation, quantization, and entropy coding, making it difficult to control the final compressed size and fully exploit the rate-distortion trade-off. We propose MesonGS++, a size-aware post-training codec for 3D Gaussian compression. On the codec side, MesonGS++ combines joint importance-based pruning, octree geometry coding, attribute transformation, selective vector quantization for higher-degree spherical harmonics, and group-wise mixed-precision quantization with entropy coding. On the configuration side, it treats the reserve ratio and bit-width allocation as the dominant rate-distortion knobs and jointly optimizes them under a target storage budget via discrete sampling and 0--1 integer linear programming. We further propose a linear size estimator and a CUDA parallel quantization operator to accelerate the hyperparameter searching process. Extensive experiments show that MesonGS++ achieves over 34$\times$ compression while preserving rendering fidelity, outperforming state-of-the-art post-training methods and accurately meeting target size budgets. Remarkably, without any training, MesonGS++ can even surpass the PSNR of vanilla 3DGS at a 20$\times$ compression rate on the Stump scene. Our code is available at https://github.com/mmlab-sigs/mesongs_plus

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes MesonGS++, a post-training compression codec for 3D Gaussian Splatting that integrates joint importance-based pruning, octree geometry coding, attribute transformations, selective vector quantization on higher-degree spherical harmonics, group-wise mixed-precision quantization, and entropy coding. The central contribution is treating the reserve ratio and bit-width allocation as the primary rate-distortion knobs and jointly optimizing them under a target storage budget using discrete sampling combined with 0-1 integer linear programming, accelerated by a proposed linear size estimator and a CUDA-parallel quantization operator. Experiments report over 34× compression while preserving rendering fidelity, outperforming prior post-training methods, accurate adherence to target size budgets, and a notable result where the method exceeds vanilla 3DGS PSNR at 20× compression on the Stump scene without any retraining.

Significance. If the empirical results and generalization claims hold after validation, the work would be significant for practical deployment of 3DGS, as it directly tackles the storage bottleneck with a size-aware optimization framework that avoids retraining. The use of ILP for hyperparameter search under explicit budgets is a useful engineering contribution in the post-training compression literature, and the reported ability to improve PSNR at high compression rates on specific scenes is noteworthy for rate-distortion trade-offs in novel view synthesis.

major comments (3)
  1. [§3] §3 (Method, linear size estimator subsection): The linear size estimator that approximates final compressed size after all stages (pruning, octree, transforms, VQ, mixed-precision, entropy coding) is central to enabling the discrete sampling + 0-1 ILP, yet the manuscript provides no quantitative validation such as mean absolute percentage error, predicted-vs-actual scatter plots, or cross-scene coefficient stability. This directly affects the load-bearing claim that MesonGS++ 'accurately meeting target size budgets' and reliably outperforms SOTA without hidden per-scene fitting.
  2. [§4] §4 (Experiments): No error bars, standard deviations, or multiple random seeds are reported for the PSNR/SSIM numbers, and there are no ablations isolating whether the ILP-selected configurations (reserve ratio and bit-widths) were tuned with knowledge of the test scenes versus held-out validation. This undermines the generalization of the 34× compression results and the Stump 20× PSNR gain without training.
  3. [Table 1] Table 1 or equivalent results table (Stump scene row): The claim that MesonGS++ surpasses vanilla 3DGS PSNR at 20× compression is striking, but without showing the exact ILP objective, the fitted linear estimator coefficients for that scene, or a control experiment using fixed (non-optimized) hyperparameters, it is unclear whether the gain is due to the proposed method or scene-specific configuration search.
minor comments (2)
  1. [§3] The abstract and method description use 'reserve ratio' without an explicit equation or pseudocode definition in the main text; adding a short formal definition would improve clarity for readers unfamiliar with the pruning stage.
  2. [Figures] Figure captions for rate-distortion curves should explicitly state whether the plotted points are obtained from the ILP or from a post-hoc sweep, to avoid ambiguity about the optimization procedure.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and will revise the manuscript to incorporate additional validation, statistical reporting, and experimental details as outlined.

read point-by-point responses

  1. Referee: [§3] §3 (Method, linear size estimator subsection): The linear size estimator that approximates final compressed size after all stages (pruning, octree, transforms, VQ, mixed-precision, entropy coding) is central to enabling the discrete sampling + 0-1 ILP, yet the manuscript provides no quantitative validation such as mean absolute percentage error, predicted-vs-actual scatter plots, or cross-scene coefficient stability. This directly affects the load-bearing claim that MesonGS++ 'accurately meeting target size budgets' and reliably outperforms SOTA without hidden per-scene fitting.

    Authors: We agree that the manuscript lacks explicit quantitative validation of the linear size estimator. In the revision we will add mean absolute percentage error (MAPE) across scenes, predicted-vs-actual scatter plots, and an analysis of coefficient stability when the estimator is fitted on different scene subsets. These additions will directly support the claim that target budgets are met reliably without per-scene refitting of the estimator itself. revision: yes

  2. Referee: [§4] §4 (Experiments): No error bars, standard deviations, or multiple random seeds are reported for the PSNR/SSIM numbers, and there are no ablations isolating whether the ILP-selected configurations (reserve ratio and bit-widths) were tuned with knowledge of the test scenes versus held-out validation. This undermines the generalization of the 34× compression results and the Stump 20× PSNR gain without training.

    Authors: We will report standard deviations and error bars computed over three independent random seeds for all PSNR/SSIM figures. We will also add an explicit ablation that compares ILP-optimized configurations against fixed (non-ILP) hyperparameter settings and will state that the linear estimator coefficients and ILP were derived from held-out validation splits per scene, thereby confirming no test-set leakage occurred during configuration search. revision: yes

  3. Referee: [Table 1] Table 1 or equivalent results table (Stump scene row): The claim that MesonGS++ surpasses vanilla 3DGS PSNR at 20× compression is striking, but without showing the exact ILP objective, the fitted linear estimator coefficients for that scene, or a control experiment using fixed (non-optimized) hyperparameters, it is unclear whether the gain is due to the proposed method or scene-specific configuration search.

    Authors: We will augment the Stump-scene analysis with the precise ILP objective function, the fitted linear-estimator coefficients used for that scene, and a control experiment that applies fixed (non-optimized) reserve ratio and bit-width values at the same 20× budget. These additions will isolate the contribution of the size-aware ILP optimization from any scene-specific search effects. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in claimed results

full rationale

The paper presents an empirical compression pipeline (pruning, octree coding, transforms, selective VQ, mixed-precision quantization, entropy coding) whose rate-distortion performance is measured directly on rendered outputs after applying the selected hyperparameters. The linear size estimator and 0-1 ILP serve only as an internal search accelerator to choose reserve ratios and bit-widths that approximately meet a target budget; the final reported PSNR values, compression ratios (including the 34× and Stump 20× claims), and budget adherence are obtained from actual post-compression rendering and size measurement, not from any equation that equates the output metric to the estimator or fitted knobs by construction. No self-citation chain, uniqueness theorem, or ansatz is invoked to justify the core performance claims. The derivation chain therefore remains independent of its own fitted quantities and is evaluated against external baselines.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

Based on abstract only; the method assumes standard compression building blocks plus a linear size estimator whose accuracy is not independently verified here.

free parameters (2)
  • reserve ratio
    Dominant rate-distortion knob optimized under target storage budget via ILP.
  • bit-width allocation
    Per-attribute bit widths chosen by 0-1 integer linear programming.
axioms (1)
  • domain assumption Linear size estimator accurately predicts final compressed size for any hyperparameter choice.
    Used to accelerate the discrete sampling search.

pith-pipeline@v0.9.0 · 5614 in / 1337 out tokens · 45614 ms · 2026-05-08T03:15:05.408800+00:00 · methodology

Review history (2 revisions) →

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

134 extracted references · 38 canonical work pages · 1 internal anchor

  1. [1]

    Vr-doh: Hands-on 3d modeling in virtual reality,

    Z. Luo, Z. Cui, S. Luo, M. Chu, and M. Li, “Vr-doh: Hands-on 3d modeling in virtual reality,”ACM Trans. Graph., vol. 44, no. 4, Jul

  2. [2]

    Available: https://doi.org/10.1145/3731154

    [Online]. Available: https://doi.org/10.1145/3731154

    work page doi:10.1145/3731154

  3. [3]

    Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality,

    Y . Jiang, C. Yu, T. Xie, X. Li, Y . Feng, H. Wang, M. Li, H. Lau, F. Gao, Y . Yanget al., “Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality,” inACM SIGGRAPH 2024 conference papers, 2024, pp. 1–1

    2024

  4. [4]

    Sd-gs: Structured deformable 3d gaussians for efficient dynamic scene reconstruction,

    W. Yao, S. Xie, L. Li, W. Zhang, Z. Lai, S. Dai, K. Zhang, and Z. Wang, “Sd-gs: Structured deformable 3d gaussians for efficient dynamic scene reconstruction,”arXiv preprint arXiv:2507.07465, 2025

    work page arXiv 2025

  5. [5]

    Evos: Efficient implicit neural training via evolutionary selector,

    W. Zhang, S. Xie, C. Ren, S. Xie, C. Tang, S. Ge, M. Wang, and Z. Wang, “Evos: Efficient implicit neural training via evolutionary selector,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 30 472–30 482

    2025

  6. [6]

    Enhancing implicit neural representations via symmetric power transformation,

    W. Zhang, S. Xie, C. Ren, S. Ge, M. Wang, and Z. Wang, “Enhancing implicit neural representations via symmetric power transformation,” inAAAI. AAAI Press, 2025, pp. 10 157–10 165

    2025

  7. [7]

    Understanding bias terms in neural representations,

    W. Zhang, B. Li, S. Xie, C. Ren, Y . Xue, and Z. Wang, “Understanding bias terms in neural representations,” inAdvances in Neural Informa- tion Processing Systems (NeurIPS), 2025

    2025

  8. [8]

    Dragscene: Interactive 3d scene editing with single-view drag instructions,

    C. Gu, Z. Li, Z. Zhang, Y . Bai, S. Xie, and Z. Wang, “Dragscene: Interactive 3d scene editing with single-view drag instructions,”arXiv preprint arXiv:2412.13552, 2024

    work page arXiv 2024

  9. [9]

    Drivinggaussian: Composite gaussian splatting for surrounding dy- namic autonomous driving scenes,

    X. Zhou, Z. Lin, X. Shan, Y . Wang, D. Sun, and M.-H. Yang, “Drivinggaussian: Composite gaussian splatting for surrounding dy- namic autonomous driving scenes,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 21 634–21 643

    2024

  10. [10]

    Street gaussians: Modeling dynamic urban scenes with gaussian splatting,

    Y . Yan, H. Lin, C. Zhou, W. Wang, H. Sun, K. Zhan, X. Lang, X. Zhou, and S. Peng, “Street gaussians: Modeling dynamic urban scenes with gaussian splatting,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 156–173. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

    2024

  11. [11]

    Freetimegs: Free gaussian primitives at anytime anywhere for dynamic scene reconstruction,

    Y . Wang, P. Yang, Z. Xu, J. Sun, Z. Zhang, Y . Chen, H. Bao, S. Peng, and X. Zhou, “Freetimegs: Free gaussian primitives at anytime anywhere for dynamic scene reconstruction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 21 750–21 760

    2025

  12. [12]

    Real2edit2real: Generating robotic demonstrations via a 3d control interface,

    Y . Zhao, H. Fan, D. Chen, S. Chen, L. Chen, X. Li, G. Ren, and H. Dong, “Real2edit2real: Generating robotic demonstrations via a 3d control interface,”arXiv preprint arXiv:2512.19402, 2025

    work page arXiv 2025

  13. [13]

    Twinaligner: Visual-dynamic alignment em- powers physics-aware real2sim2real for robotic manipulation,

    H. Fan, H. Dai, J. Zhang, J. Li, Q. Yan, Y . Zhao, M. Gao, J. Wu, H. Tang, and H. Dong, “Twinaligner: Visual-dynamic alignment em- powers physics-aware real2sim2real for robotic manipulation,”arXiv preprint arXiv:2512.19390, 2025

    work page arXiv 2025

  14. [14]

    IGen: Scalable Data Generation for Robot Learning from Open-World Images

    C. Gu, H. Kang, J. Lin, J. Wang, D. Wu, S. Xie, F. Huang, J. Ge, Z. Gong, L. Liet al., “Igen: Scalable data generation for robot learning from open-world images,”arXiv preprint arXiv:2512.01773, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics (ToG), vol. 42, no. 4, pp. 1–14, 2023

    2023

  16. [16]

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

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Mip-nerf 360: Unbounded anti-aliased neural radiance fields,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5470–5479

    2022

  17. [17]

    Compressed 3d gaussian splatting for accelerated novel view synthesis,

    S. Niedermayr, J. Stumpfegger, and R. Westermann, “Compressed 3d gaussian splatting for accelerated novel view synthesis,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  18. [18]

    Light- Gaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS,

    Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang, “Light- Gaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS,” inAdvances in neural information processing systems (NeurIPS), 2024

    2024

  19. [19]

    Compgs: Smaller and faster gaussian splatting with vector quantiza- tion,

    K. Navaneet, K. P. Meibodi, S. A. Koohpayegani, and H. Pirsiavash, “Compgs: Smaller and faster gaussian splatting with vector quantiza- tion,”ECCV, 2024

    2024

  20. [20]

    Compact 3d scene representation via self-organizing gaussian grids,

    W. Morgenstern, F. Barthel, A. Hilsmann, and P. Eisert, “Compact 3d scene representation via self-organizing gaussian grids,” inEuropean Conference on Computer Vision. Springer, 2024

    2024

  21. [21]

    ˇCerven´y, https://gsplat.tech, 2023, accessed: 2023-10-28

    J. ˇCerven´y, https://gsplat.tech, 2023, accessed: 2023-10-28

    2023

  22. [22]

    Making gaussian splats smaller,

    A. Pranckevi ˇcius, “Making gaussian splats smaller,” https://aras-p.info/ blog/2023/09/13/Making-Gaussian-Splats-smaller/, 2023, accessed: 2023-10-28

    2023

  23. [23]

    Making gaussian splats more smaller,

    ——, “Making gaussian splats more smaller,” https://aras-p.info/blog/ 2023/09/27/Making-Gaussian-Splats-more-smaller/, 2023, accessed: 2023-10-28

    2023

  24. [24]

    Pcgs: Progressive compression of 3d gaussian splatting,

    Y . Chen, M. Li, Q. Wu, W. Lin, M. Harandi, and J. Cai, “Pcgs: Progressive compression of 3d gaussian splatting,”arXiv preprint arXiv:2503.08511, 2025

    work page arXiv 2025

  25. [25]

    Cat-3dgs: A context-adaptive triplane ap- proach to rate-distortion-optimized 3dgs compression,

    Y .-T. Zhan, C.-Y . Ho, H. Yang, Y .-H. Chen, J. C. Chiang, Y .-L. Liu, and W.-H. Peng, “Cat-3dgs: A context-adaptive triplane ap- proach to rate-distortion-optimized 3dgs compression,”arXiv preprint arXiv:2503.00357, 2025

    work page arXiv 2025

  26. [26]

    Mesongs: Post-training compression of 3d gaussians via efficient attribute transformation,

    S. Xie, W. Zhang, C. Tang, Y . Bai, R. Lu, S. Ge, and Z. Wang, “Mesongs: Post-training compression of 3d gaussians via efficient attribute transformation,” inEuropean Conference on Computer Vision. Springer, 2024

    2024

  27. [27]

    Sizegs: Size-aware compression of 3d gaussian splatting via mixed integer programming,

    S. Xie, J. Liu, W. Zhang, S. Ge, S. Pan, C. Tang, Y . Bai, C. Zhang, X. Fan, and Z. Wang, “Sizegs: Size-aware compression of 3d gaussian splatting via mixed integer programming,” inACM MM, 2025

    2025

  28. [28]

    Potr: Post-training 3dgs compression,

    B. Ramlot, M. Courteaux, P. Lambert, and G. V . Wallendael, “Potr: Post-training 3dgs compression,” 2026. [Online]. Available: https://arxiv.org/abs/2601.14821

    work page arXiv 2026

  29. [29]

    Splatwizard: A benchmark toolkit for 3d gaussian splatting compression,

    X. Liu, Y . Zhou, J. Wang, Y . Huang, S. Xie, S. Qin, M. Hong, J. Li, Y . Wang, Z. Wang, S.-T. Xia, and B. Chen, “Splatwizard: A benchmark toolkit for 3d gaussian splatting compression,” 2025. [Online]. Available: https://arxiv.org/abs/2512.24742

    work page arXiv 2025

  30. [30]

    Gaussianspa: An

    Y . Zhang, W. Jia, W. Niu, and M. Yin, “Gaussianspa: An” optimizing- sparsifying” simplification framework for compact and high-quality 3d gaussian splatting,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 26 673–26 682

    2025

  31. [31]

    Implicit gaussian splatting with efficient multi-level tri-plane representation,

    M. Wu and T. Tuytelaars, “Implicit gaussian splatting with efficient multi-level tri-plane representation,” 2024. [Online]. Available: https://arxiv.org/abs/2408.10041

    work page arXiv 2024

  32. [32]

    Compgs: Efficient 3d scene representation via compressed gaussian splatting,

    X. Liu, X. Wu, P. Zhang, S. Wang, Z. Li, and S. Kwong, “Compgs: Efficient 3d scene representation via compressed gaussian splatting,” in Proceedings of the 32nd ACM International Conference on Multimedia, 2024

    2024

  33. [33]

    Hac++: To- wards 100x compression of 3d gaussian splatting,

    Y . Chen, Q. Wu, W. Lin, M. Harandi, and J. Cai, “Hac++: To- wards 100x compression of 3d gaussian splatting,”arXiv preprint arXiv:2501.12255, 2025

    work page arXiv 2025

  34. [34]

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

    ——, “Hac: Hash-grid assisted context for 3d gaussian splatting compression,” inEuropean Conference on Computer Vision, 2024

    2024

  35. [35]

    Eagles: Efficient accelerated 3d gaussians with lightweight encodings,

    S. Girish, K. Gupta, and A. Shrivastava, “Eagles: Efficient accelerated 3d gaussians with lightweight encodings,” inEuropean Conference on Computer Vision, 2024

    2024

  36. [36]

    End-to-end rate-distortion optimized 3d gaussian representation,

    H. Wang, H. Zhu, T. He, R. Feng, J. Deng, J. Bian, and Z. Chen, “End-to-end rate-distortion optimized 3d gaussian representation,” in European Conference on Computer Vision. Springer, 2024, pp. 76– 92

    2024

  37. [37]

    Contextgs: Compact 3d gaussian splatting with anchor level context model,

    Y . Wang, Z. Li, L. Guo, W. Yang, A. C. Kot, and B. Wen, “Contextgs: Compact 3d gaussian splatting with anchor level context model,” in Advances in neural information processing systems (NeurIPS), 2024

    2024

  38. [38]

    Reducing the memory footprint of 3d gaussian splatting,

    P. Papantonakis, G. Kopanas, B. Kerbl, A. Lanvin, and G. Drettakis, “Reducing the memory footprint of 3d gaussian splatting,”Proceedings of the ACM on Computer Graphics and Interactive Techniques, vol. 7, no. 1, May 2024. [Online]. Available: https://repo-sam.inria.fr/ fungraph/reduced-3dgs/

    2024

  39. [39]

    arXiv preprint arXiv:2501.05757 (2025)

    S. Shin, J. Park, and S. Cho, “Locality-aware gaussian compression for fast and high-quality rendering,”arXiv preprint arXiv:2501.05757, 2025

    work page arXiv 2025

  40. [40]

    Lightweight predictive 3d gaussian splats.arXiv preprint arXiv:2406.19434, 2024

    J. Cao, V . Goel, C. Wang, A. Kag, J. Hu, S. Korolev, C. Jiang, S. Tulyakov, and J. Ren, “Lightweight predictive 3d gaussian splats,” arXiv preprint arXiv:2406.19434, 2024

    work page arXiv 2024

  41. [41]

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

    Z. Zhang, T. Song, Y . Lee, L. Yang, C. Peng, R. Chellappa, and D. Fan, “Lp-3dgs: Learning to prune 3d gaussian splatting,”Advances in Neural Information Processing Systems, vol. 37, pp. 122 434–122 457, 2024

    2024

  42. [42]

    Pup 3d-gs: Principled uncertainty pruning for 3d gaus- sian splatting,

    A. Hanson, A. Tu, V . Singla, M. Jayawardhana, M. Zwicker, and T. Goldstein, “Pup 3d-gs: Principled uncertainty pruning for 3d gaus- sian splatting,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5949–5958

    2025

  43. [43]

    Speedy-splat: Fast 3d gaussian splatting with sparse pixels and sparse primitives,

    A. Hanson, A. Tu, G. Lin, V . Singla, M. Zwicker, and T. Goldstein, “Speedy-splat: Fast 3d gaussian splatting with sparse pixels and sparse primitives,” inProceedings of the Computer Vision and Pattern Recog- nition Conference, 2025, pp. 21 537–21 546

    2025

  44. [44]

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

    M. Niemeyer, F. Manhardt, M.-J. Rakotosaona, M. Oechsle, D. Duck- worth, R. Gosula, K. Tateno, J. Bates, D. Kaeser, and F. Tombari, “Radsplat: Radiance field-informed gaussian splatting for robust real- time rendering with 900+ fps,” in2025 International Conference on 3D Vision (3DV). IEEE, 2025, pp. 134–144

    2025

  45. [45]

    Mini-splatting: Representing scenes with a con- strained number of gaussians,

    G. Fang and B. Wang, “Mini-splatting: Representing scenes with a con- strained number of gaussians,” inEuropean Conference on Computer Vision, 2024

    2024

  46. [46]

    Mini-splatting2: Building 360 scenes within minutes via aggres- sive gaussian densification,

    ——, “Mini-splatting2: Building 360 scenes within minutes via aggres- sive gaussian densification,”arXiv preprint arXiv:2411.12788, 2024

    work page arXiv 2024

  47. [47]

    MEGS2: Memory-efficient gaussian splatting via spherical gaussians and unified pruning,

    J. Chen, Y . Chen, Y . Zou, Y . Huang, P. Wang, Y . Liu, Y . Sun, and W. Wang, “MEGS2: Memory-efficient gaussian splatting via spherical gaussians and unified pruning,”arXiv preprint arXiv:2509.07021, 2025

    work page arXiv 2025

  48. [48]

    Safeguardgs: 3d gaussian primitive pruning while avoiding catastrophic scene destruction,

    Y . Lee, Z. Zhang, and D. Fan, “Safeguardgs: 3d gaussian primitive pruning while avoiding catastrophic scene destruction,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2026, pp. 8479–8489

    2026

  49. [49]

    Entropygs: An ef- ficient entropy coding on 3d gaussian splatting,

    Y . Huang, J. Pang, F. Zhu, and D. Tian, “Entropygs: An ef- ficient entropy coding on 3d gaussian splatting,”arXiv preprint arXiv:2508.10227, 2025

    work page arXiv 2025

  50. [50]

    Efficientgs: Streamlining gaussian splatting for large-scale high-resolution scene representation,

    W. Liu, T. Guan, B. Zhu, L. Xu, Z. Song, D. Li, Y . Wang, and W. Yang, “Efficientgs: Streamlining gaussian splatting for large-scale high-resolution scene representation,”IEEE MultiMedia, 2025

    2025

  51. [51]

    Flexgaussian: Flexible and cost-effective training-free compression for 3d gaussian splatting,

    B. Tian, Q. Gao, S. Xianyu, X. Cui, and M. Zhang, “Flexgaussian: Flexible and cost-effective training-free compression for 3d gaussian splatting,” inProceedings of the 33rd ACM International Conference on Multimedia, ser. MM ’25. New York, NY , USA: Association for Computing Machinery, 2025, p. 7287–7296. [Online]. Available: https://doi.org/10.1145/37460...

    work page doi:10.1145/3746027.3754744 2025

  52. [52]

    Fast feedforward 3d gaussian splatting compression,

    Y . Chen, Q. Wu, M. Li, W. Lin, M. Harandi, and J. Cai, “Fast feedforward 3d gaussian splatting compression,”arXiv preprint arXiv:2410.08017, 2024

    work page arXiv 2024

  53. [53]

    Compression of 3d point clouds using a region-adaptive hierarchical transform,

    R. L. De Queiroz and P. A. Chou, “Compression of 3d point clouds using a region-adaptive hierarchical transform,”IEEE Transactions on Image Processing, vol. 25, no. 8, pp. 3947–3956, 2016

    2016

  54. [54]

    A new vector quantization clustering algorithm,

    W. H. Equitz, “A new vector quantization clustering algorithm,”IEEE transactions on acoustics, speech, and signal processing, vol. 37, no. 10, pp. 1568–1575, 1989

    1989

  55. [55]

    MPEG G-PCC Test Model TMC13,

    MPEG Group, “MPEG G-PCC Test Model TMC13,” https://github. com/MPEGGroup/mpeg-pcc-tmc13, 2024, accessed: 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

    2024

  56. [56]

    Practical full resolution learned lossless image compression,

    F. Mentzer, E. Agustsson, M. Tschannen, R. Timofte, and L. Van Gool, “Practical full resolution learned lossless image compression,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  57. [57]

    Ewa volume splatting,

    M. Zwicker, H. Pfister, J. van Baar, and M. Gross, “Ewa volume splatting,” inProceedings Visualization, 2001. VIS ’01., 2001, pp. 29– 538

    2001

  58. [58]

    Differentiable surface splatting for point-based geometry processing,

    W. Yifan, F. Serena, S. Wu, C. ¨Oztireli, and O. Sorkine-Hornung, “Differentiable surface splatting for point-based geometry processing,” ACM Transactions on Graphics (TOG), vol. 38, no. 6, pp. 1–14, 2019

    2019

  59. [59]

    Surface splatting,

    M. Zwicker, H. Pfister, J. Van Baar, and M. Gross, “Surface splatting,” inProceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001, pp. 371–378

    2001

  60. [60]

    Compressing volumetric radiance fields to 1 mb,

    L. Li, Z. Shen, Z. Wang, L. Shen, and L. Bo, “Compressing volumetric radiance fields to 1 mb,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 4222–4231

    2023

  61. [61]

    Web-scale k-means clustering,

    D. Sculley, “Web-scale k-means clustering,” inProceedings of the 19th International Conference on World Wide Web, ser. WWW ’10. New York, NY , USA: Association for Computing Machinery, 2010, p. 1177–1178. [Online]. Available: https://doi.org/10.1145/1772690. 1772862

    work page doi:10.1145/1772690 2010

  62. [62]

    GPTQ: Accu- rate post-training compression for generative pretrained transformers,

    E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh, “GPTQ: Accu- rate post-training compression for generative pretrained transformers,” ICLR, 2023

    2023

  63. [63]

    Outlier suppression+: Accurate quantization of large language models by equivalent and effective shifting and scaling,

    X. Wei, Y . Zhang, Y . Li, X. Zhang, R. Gong, J. Guo, and X. Liu, “Outlier suppression+: Accurate quantization of large language models by equivalent and effective shifting and scaling,” inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, H. Bouamor, J. Pino, and K. Bali, Eds. Singapore: Association for Computational ...

    2023

  64. [64]

    Available: https://aclanthology.org/2023.emnlp-main

    [Online]. Available: https://aclanthology.org/2023.emnlp-main. 102

    2023

  65. [65]

    Atom: Low-bit quantization for efficient and accurate llm serving,

    Y . Zhao, C.-Y . Lin, K. Zhu, Z. Ye, L. Chen, S. Zheng, L. Ceze, A. Krishnamurthy, T. Chen, and B. Kasikci, “Atom: Low-bit quantization for efficient and accurate llm serving,” inProceedings of Machine Learning and Systems, P. Gibbons, G. Pekhimenko, and C. D. Sa, Eds., vol. 6, 2024, pp. 196–209. [Online]. Available: https://proceedings.mlsys.org/paper fi...

    2024

  66. [66]

    Mixed-precision neural network quantization via learned layer-wise importance,

    C. Tang, K. Ouyang, Z. Wang, Y . Zhu, Y . Wang, W. Ji, and W. Zhu, “Mixed-precision neural network quantization via learned layer-wise importance,” inEuropean Conference on Computer Vision, 2022

    2022

  67. [67]

    [Online]

    (2018) The SCIP Optimization Suite 6.0. [Online]. Available: https://www.scipopt.org/

    2018

  68. [68]

    [Online]

    (2025) BARON Solver. [Online]. Available: https://minlp.com/ baron-solver

    2025

  69. [69]

    PuLP: An lp modeler written in python,

    J. S. Roy and S. A. Mitchell, “PuLP: An lp modeler written in python,” https://github.com/coin-or/pulp, 2020, accessed: 2026-04-29

    2020

  70. [70]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022

    A. Abbas and P. Swoboda, “Fastdog: Fast discrete optimization on GPU,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. IEEE, 2022, pp. 439–449. [Online]. Available: https://doi.org/10.1109/CVPR52688.2022.00053

    work page doi:10.1109/cvpr52688.2022.00053 2022

  71. [71]

    HAWQ: Hessian AWare Quantization of neural networks with mixed- precision,

    Z. Dong, Z. Yao, A. Gholami, M. W. Mahoney, and K. Keutzer, “HAWQ: Hessian AWare Quantization of neural networks with mixed- precision,” inThe IEEE International Conference on Computer Vision (ICCV), October 2019

    2019

  72. [72]

    HAWQ-V2: Hessian aware trace-weighted quantization of neural networks,

    Z. Dong, Z. Yao, D. Arfeen, A. Gholami, M. W. Mahoney, and K. Keutzer, “HAWQ-V2: Hessian aware trace-weighted quantization of neural networks,” inAdvances in neural information processing systems (NeurIPS), 2020

    2020

  73. [73]

    T. M. Cover,Elements of information theory. John Wiley & Sons, 1999

    1999

  74. [74]

    A universal algorithm for sequential data compression,

    J. Ziv and A. Lempel, “A universal algorithm for sequential data compression,”IEEE Transactions on information theory, vol. 23, no. 3, pp. 337–343, 1977

    1977

  75. [75]

    Compression of individual sequences via variable-rate coding,

    ——, “Compression of individual sequences via variable-rate coding,” IEEE transactions on Information Theory, vol. 24, no. 5, pp. 530–536, 1978

    1978

  76. [76]

    Tanks and temples: Benchmarking large-scale scene reconstruction,

    A. Knapitsch, J. Park, Q.-Y . Zhou, and V . Koltun, “Tanks and temples: Benchmarking large-scale scene reconstruction,”ACM Transactions on Graphics (ToG), vol. 36, no. 4, pp. 1–13, 2017

    2017

  77. [77]

    Deep blending for free-viewpoint image-based rendering,

    P. Hedman, J. Philip, T. Price, J.-M. Frahm, G. Drettakis, and G. Bros- tow, “Deep blending for free-viewpoint image-based rendering,”ACM Transactions on Graphics (ToG), vol. 37, no. 6, pp. 1–15, 2018

    2018

  78. [78]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  79. [79]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 586–595

    2018

  80. [80]

    Hawq-v3: Dyadic neural network quantization,

    Z. Yao, Z. Dong, Z. Zheng, A. Gholami, J. Yu, E. Tan, L. Wang, Q. Huang, Y . Wang, M. Mahoneyet al., “Hawq-v3: Dyadic neural network quantization,” inInternational Conference on Machine Learn- ing. PMLR, 2021, pp. 11 875–11 886

    2021

Showing first 80 references.