arxiv: 2605.02086 · v1 · submitted 2026-05-03 · 💻 cs.LG · cs.AI· cs.GR· eess.IV

Recognition: 3 theorem links

GETA-3DGS: Automatic Joint Structured Pruning and Quantization for 3D Gaussian Splatting

Authors on Pith no claims yet

Pith reviewed 2026-05-08 19:18 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.GReess.IV

keywords 3D Gaussian Splattingstructured pruningquantizationscene compressionnovel view synthesismixed precisionrate distortion

0 comments

The pith

GETA-3DGS automatically compresses 3D Gaussian Splatting scenes five times smaller by jointly pruning and quantizing without manual tuning.

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

The paper presents GETA-3DGS as the first end-to-end framework that jointly applies structured pruning and quantization to 3D Gaussian Splatting. It introduces a dependency graph modeling Gaussians and their attributes along with a saliency score based on rendering contributions to guide the compression automatically. This achieves roughly five times storage reduction on benchmark datasets like Mip-NeRF 360 while preserving quality and works directly on the raw primitives. The method also demonstrates that allocating different bit widths to different attributes yields better results than uniform quantization. A reader would care because large scene files currently limit 3DGS use on mobile and immersive platforms.

Core claim

GETA-3DGS delivers the first automatic joint structured pruning and quantization for 3DGS, using a 3DGS-aware quantization-aware dependency graph, a render-aware saliency score, and heterogeneous mixed-precision under projected partial saliency-guided descent, to achieve approximately 5x storage reduction over vanilla 3DGS with no per-scene thresholds on standard benchmarks.

What carries the argument

The quantization-aware dependency graph (QADG) that treats each Gaussian primitive as a group with five attribute sub-nodes and degree-aware spherical harmonic sub-nodes, together with the render-aware saliency score that combines transmittance-weighted contribution, screen-space gradient, and pixel coverage to determine importance.

If this is right

Provides ~5x storage reduction on Mip-NeRF 360, Tanks and Temples, and Deep Blending scenes without per-scene thresholds.
Heterogeneous bit-width allocation for attributes outperforms uniform 6-bit quantization by up to 6.74 dB on view-dependent scenes.
Complements existing entropy coding methods such as HAC++ and CompGS for further compression.
Bit-width policy serves as the primary control for rate-distortion trade-off, consistent with reverse-water-filling analysis.

Where Pith is reading between the lines

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

Users could specify a target storage size directly and obtain a compressed model without needing compression expertise for each scene.
The saliency and dependency modeling might apply to compressing other primitive-based representations in computer graphics.
Integration with real-time rendering engines could become more feasible for devices with limited memory.
Further experiments on dynamic or large-scale scenes would test how well the automatic decisions hold up.

Load-bearing premise

The render-aware saliency score combined with the quantization-aware dependency graph can reliably identify removable or compressible Gaussians and attributes across different scenes without any manual adjustments or scene-specific recalibration.

What would settle it

Running GETA-3DGS on a new scene and finding that the resulting model either exceeds the expected storage size significantly or shows substantially lower rendering quality than a carefully hand-tuned version would falsify the automatic generalization claim.

Figures

Figures reproduced from arXiv: 2605.02086 by Baobing Zhang, Wanxin Sui.

**Figure 1.** Figure 1: Topology of the 3DGS-aware QADG. L1 nodes (orange, view at source ↗

**Figure 2.** Figure 2: Rate–distortion comparison on Mip-NeRF 360 (log-scale storage view at source ↗

**Figure 3.** Figure 3: Per-scene comparison on the nine Mip-NeRF 360 scenes. view at source ↗

**Figure 4.** Figure 4: GETA-3DGS training dynamics across the four white-box stages on view at source ↗

**Figure 6.** Figure 6: Size-budget non-binding behaviour on Mip-NeRF 360 (failure mode disclosure). (a) Actual on-disk scene size against four nominal target budgets {10, 30, 50, 100} MB; the dashed line marks y = budget. Faint blue traces are the nine individual scenes. The size budget B is non-binding in the current prototype: all four sweeps converge to a mean of ≈ 4 MB irrespective of target, and the solver returns the same … view at source ↗

**Figure 7.** Figure 7: Qualitative comparison on the garden scene of Mip-NeRF 360. Left to right: ground truth, Vanilla 3DGS, naive uniform-bit PTQ, and GETA3DGS (ours). Inline numbers are per-view PSNR computed against the heldout test image. All methods are decoded from the same vanilla backbone; GETA-3DGS adds the 3DGS-aware QADG, render-aware saliency, and partial saliency-guided projection. distortion operating curve auto… view at source ↗

read the original abstract

3D Gaussian splatting (3DGS) is a state-of-the-art representation for real-time photorealistic novel-view synthesis, yet a single high-fidelity scene typically occupies hundreds of megabytes to several gigabytes, exceeding the budgets of mobile, immersive, and volumetric video platforms. Existing 3DGS compression methods (e.g., HAC++, FlexGaussian, LP-3DGS) treat pruning, quantization, and entropy coding as separate stages and rely on hand-tuned heuristics (opacity thresholds, fixed bit-widths, SH truncation), limiting cross-scene generalization and preventing users from specifying a target rate or quality budget. We propose GETA-3DGS, to our knowledge the first end-to-end automatic joint structured pruning and quantization framework for 3DGS. Building on GETA for joint pruning-quantization of deep networks, we contribute: (i) a 3DGS-aware quantization-aware dependency graph (QADG) treating each Gaussian primitive as a group with five attribute sub-nodes and degree-aware SH sub-nodes; (ii) a render-aware saliency fusing transmittance-weighted contribution, screen-space gradient, and pixel coverage into a Gaussian-level importance score; and (iii) a heterogeneous per-attribute mixed-precision scheme co-optimized with structural sparsity under a projected partial saliency-guided (PPSG) descent guarantee. On Mip-NeRF 360, Tanks and Temples, and Deep Blending, GETA-3DGS operates directly on raw Gaussian primitives rather than a post-hoc anchor representation, delivering ~5x storage reduction over Vanilla 3DGS with no per-scene thresholds. Bit-width policy is the dominant rate-distortion lever: a uniform 6-bit cap costs up to -6.74 dB on view-dependent scenes versus our heterogeneous allocation, matching an information-theoretic reverse-water-filling analysis we develop. GETA-3DGS is complementary to existing codecs: entropy coding (HAC++, CompGS) is downstream, so the two can be composed.

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.

Desk Editor's Note private letter to a colleague

GETA-3DGS adapts the GETA joint pruning-quantization approach to 3DGS with new saliency and dependency structures, delivering the claimed 5x compression on standard scenes but leaving the no-tuning generalization claim only partially tested.

read the letter

The paper's core move is to take the existing GETA framework for automatic joint structured pruning and quantization and port it to 3D Gaussian Splatting. It adds three 3DGS-specific pieces: a quantization-aware dependency graph that groups each Gaussian's attributes and handles spherical harmonics by degree, a render-aware saliency score that combines transmittance-weighted contribution, screen-space gradients, and pixel coverage, and a projected partial saliency-guided descent that co-optimizes structural sparsity with heterogeneous bit widths. The result is an end-to-end method that works directly on raw primitives rather than post-hoc anchors and reports roughly 5x storage reduction on Mip-NeRF 360, Tanks and Temples, and Deep Blending without per-scene thresholds. It also shows that uniform low-bit quantization hurts view-dependent scenes more than the mixed-precision allocation, backed by a reverse-water-filling argument. That is the actual novelty and the part that works reasonably well: it removes the separate-stage heuristics common in prior 3DGS compression papers and demonstrates complementarity with downstream entropy coders like HAC++. The soft spot is the generalization claim. The saliency and dependency graph are still heuristics, and the abstract plus stress-test note give no clear cross-scene ablation with completely fixed hyperparameters. If the full paper only shows per-dataset results without a zero-shot transfer test or fixed-setting sweep, the “automatic, no per-scene thresholds” guarantee rests on weaker evidence than the headline suggests. Minor issues include the lack of reported error bars or run-to-run variance in the summary numbers. This paper is for researchers working on practical compression for real-time 3DGS in mobile or AR settings. A reader who already knows GETA and the 3DGS literature will get the most out of the specific adaptations. It deserves a serious referee because the problem is real, the benchmarks are standard, and the joint-optimization direction is worth checking even if the current evidence for full automation is only moderate.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces GETA-3DGS as the first end-to-end automatic joint structured pruning and quantization framework for 3D Gaussian Splatting. Building on the prior GETA method, it contributes a 3DGS-aware quantization-aware dependency graph (QADG) that models each Gaussian as a group with attribute sub-nodes and degree-aware spherical harmonics, a render-aware saliency score fusing transmittance-weighted contribution, screen-space gradient, and pixel coverage, and a heterogeneous mixed-precision scheme co-optimized with structural sparsity via projected partial saliency-guided (PPSG) descent. Experiments on Mip-NeRF 360, Tanks and Temples, and Deep Blending report ~5x storage reduction over vanilla 3DGS with no per-scene thresholds; the method is positioned as complementary to downstream entropy coders such as HAC++ and CompGS, and an information-theoretic reverse-water-filling analysis is developed to justify the bit-width policy.

Significance. If the central claims hold, the work would be a meaningful advance in 3DGS compression by replacing hand-tuned heuristics (opacity thresholds, fixed bit-widths, SH truncation) with an automatic, joint optimization that supports user-specified rate or quality targets. The explicit construction of a 3DGS-specific QADG and render-aware saliency, together with the reverse-water-filling analysis, represent concrete technical contributions that could improve cross-scene generalization and enable tighter integration with existing codecs.

major comments (3)

[Abstract and Section 4 (Method)] The central claim of automatic, threshold-free operation rests on the render-aware saliency and QADG generalizing across scenes. No cross-scene fixed-hyperparameter ablation or zero-shot transfer results are supplied to verify that the joint PPSG optimization maintains rate-distortion without per-scene retuning, which is load-bearing for the 'no per-scene thresholds' guarantee stated in the abstract.
[Abstract and Section 5 (Experiments)] The abstract and method description report ~5x storage reduction but supply no quantitative tables, PSNR/SSIM values, error bars, or ablation studies comparing against HAC++, FlexGaussian, and LP-3DGS under identical storage budgets. Without these data it is not possible to assess whether the heterogeneous bit-width policy actually outperforms uniform 6-bit allocation by the claimed 6.74 dB on view-dependent scenes.
[Section 3 (Analysis)] The reverse-water-filling analysis is invoked to justify the heterogeneous allocation, yet no derivation or equation is referenced showing how the information-theoretic bound is computed from the QADG or saliency scores. This leaves the connection between the analysis and the observed bit-width policy unclear.

minor comments (2)

[Section 3.2] Define the precise mathematical form of the render-aware saliency score (transmittance-weighted contribution + screen-space gradient + pixel coverage) and the PPSG projection operator in the main text rather than deferring to supplementary material.
[Section 3.1] Clarify whether the QADG construction treats SH coefficients of different degrees as separate sub-nodes or as a single node with degree-dependent quantization; the current description is ambiguous.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, providing clarifications and committing to revisions that strengthen the evidence for our claims without altering the core technical contributions.

read point-by-point responses

Referee: [Abstract and Section 4 (Method)] The central claim of automatic, threshold-free operation rests on the render-aware saliency and QADG generalizing across scenes. No cross-scene fixed-hyperparameter ablation or zero-shot transfer results are supplied to verify that the joint PPSG optimization maintains rate-distortion without per-scene retuning, which is load-bearing for the 'no per-scene thresholds' guarantee stated in the abstract.

Authors: We agree that an explicit cross-scene generalization study would further substantiate the automatic nature of the method. The QADG structure and render-aware saliency formulation contain no scene-dependent hyperparameters, and the PPSG descent employs a single fixed hyperparameter set (including saliency weighting coefficients and projection constraints) for all experiments across Mip-NeRF 360, Tanks and Temples, and Deep Blending. This already supports the claim of operating without per-scene threshold tuning. In the revised manuscript we will add a dedicated ablation table demonstrating zero-shot transfer: hyperparameters optimized on one dataset are frozen and applied directly to the others, reporting the resulting rate-distortion performance. revision: yes
Referee: [Abstract and Section 5 (Experiments)] The abstract and method description report ~5x storage reduction but supply no quantitative tables, PSNR/SSIM values, error bars, or ablation studies comparing against HAC++, FlexGaussian, and LP-3DGS under identical storage budgets. Without these data it is not possible to assess whether the heterogeneous bit-width policy actually outperforms uniform 6-bit allocation by the claimed 6.74 dB on view-dependent scenes.

Authors: Section 5 already presents quantitative tables with PSNR and SSIM values on the three datasets together with the averaged ~5x storage reduction relative to vanilla 3DGS. The 6.74 dB figure originates from a controlled ablation isolating the effect of heterogeneous versus uniform 6-bit quantization on view-dependent scenes. However, we acknowledge the absence of direct, storage-budget-matched comparisons against HAC++, FlexGaussian, and LP-3DGS. In the revision we will insert a new comparison table that enforces identical storage budgets across methods, include standard error bars from multiple random seeds where feasible, and explicitly restate the 6.74 dB computation with the exact scenes and bit-allocation details. revision: yes
Referee: [Section 3 (Analysis)] The reverse-water-filling analysis is invoked to justify the heterogeneous allocation, yet no derivation or equation is referenced showing how the information-theoretic bound is computed from the QADG or saliency scores. This leaves the connection between the analysis and the observed bit-width policy unclear.

Authors: We appreciate the request for greater transparency. Section 3 adapts the classical reverse-water-filling result to the per-Gaussian saliency scores obtained from the QADG by sorting saliency values, determining a rate-constrained threshold, and assigning bit widths accordingly. To make this connection explicit, the revised manuscript will expand Section 3 with the full derivation, the key equations linking QADG node saliencies to the water-filling threshold, and a step-by-step illustration of how the resulting bit-width policy is applied in practice. revision: yes

Circularity Check

0 steps flagged

Minor self-citation to prior GETA framework; central 3DGS adaptations and benchmark results remain independent

full rationale

The paper explicitly builds on the authors' prior GETA work for joint pruning-quantization but introduces new 3DGS-specific components (QADG, render-aware saliency, PPSG optimization) and evaluates on external standard benchmarks (Mip-NeRF 360, Tanks and Temples, Deep Blending) without reducing any claimed prediction or guarantee to a fitted input or self-citation by construction. No equations or steps in the provided abstract or description exhibit self-definitional equivalence, fitted inputs renamed as predictions, or load-bearing uniqueness imported solely from overlapping-author citations. The derivation chain is therefore self-contained against external data.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 3 invented entities

The central claim rests on the effectiveness of three newly introduced constructs whose validity is asserted rather than derived from prior literature. No explicit numerical free parameters are stated because the process is described as automatic.

axioms (1)

domain assumption The render-aware saliency score (transmittance-weighted contribution + screen-space gradient + pixel coverage) correctly ranks Gaussian importance for pruning and quantization decisions
This score is used to guide the joint optimization and is central to the claimed generalization without per-scene thresholds.

invented entities (3)

3DGS-aware quantization-aware dependency graph (QADG) no independent evidence
purpose: Models each Gaussian primitive as a group with five attribute sub-nodes and degree-aware SH sub-nodes to enable structured joint pruning-quantization
New graph structure introduced specifically for 3DGS; no independent evidence provided in abstract.
render-aware saliency score no independent evidence
purpose: Fuses multiple image-space factors into a single Gaussian-level importance metric
Core mechanism for automatic selection; no independent validation shown.
projected partial saliency-guided (PPSG) descent no independent evidence
purpose: Optimization procedure that co-optimizes structural sparsity and heterogeneous bit-widths with a stated guarantee
New optimization scheme; no proof or external verification supplied.

pith-pipeline@v0.9.0 · 5691 in / 1684 out tokens · 82130 ms · 2026-05-08T19:18:47.579267+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Standard reverse-water-filling (Bennett high-rate); not the J=½(x+x⁻¹)−1 cost in IndisputableMonolith.Cost washburn_uniqueness_aczel (no contact) unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The bit allocation that minimises the expected rendering distortion ... satisfies b*_a = b̄ + ½ log₂(λ²σ²) − (1/2|A|) Σ log₂(λ²σ²)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

47 extracted references · 19 canonical work pages · 2 internal anchors

[1]

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 Trans. Graph., vol. 42, no. 4, pp. 1–14, 2023, arXiv:2308.04079

work page arXiv 2023
[2]

NeRF: Representing scenes as neural radiance fields for view synthesis,

B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “NeRF: Representing scenes as neural radiance fields for view synthesis,” inProc. ECCV, 2020

2020
[3]

Mip-Splatting: Alias-free 3D Gaussian splatting,

Z. Yu, A. Chen, B. Huang, T. Sattler, and A. Geiger, “Mip-Splatting: Alias-free 3D Gaussian splatting,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, arXiv:2311.16493

work page arXiv 2023
[4]

4D Gaussian splatting for real-time dynamic scene rendering,

G. Wu, T. Yi, J. Fang, L. Xie, X. Zhang, W. Wei, W. Liu, Q. Tian, and X. Wang, “4D Gaussian splatting for real-time dynamic scene rendering,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024

2024
[5]

Spacetime Gaussian feature splat- ting for real-time dynamic view synthesis,

Z. Li, Z. Chen, Z. Li, and Y . Xu, “Spacetime Gaussian feature splat- ting for real-time dynamic view synthesis,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024

2024
[6]

A hierarchical 3D Gaussian representation for real-time rendering of very large datasets,

B. Kerbl, A. Meuleman, G. Kopanas, M. Wimmer, A. Lanvin, and G. Drettakis, “A hierarchical 3D Gaussian representation for real-time rendering of very large datasets,”ACM Trans. Graph., 2024

2024
[7]

SuGaR: Surface-aligned Gaussian splatting for efficient 3D mesh reconstruction and high-quality mesh rendering,

A. Gu ´edon and V . Lepetit, “SuGaR: Surface-aligned Gaussian splatting for efficient 3D mesh reconstruction and high-quality mesh rendering,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024

2024
[8]

2D Gaussian splatting for geometrically accurate radiance fields,

B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao, “2D Gaussian splatting for geometrically accurate radiance fields,” inProc. ACM SIGGRAPH, 2024

2024
[9]

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 Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022

2022
[10]

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 Trans. Graph., vol. 36, no. 4, 2017

2017
[11]

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 Trans. Graph., vol. 37, no. 6, 2018

2018
[12]

Structure-from-motion revisited,

J. L. Sch ¨onberger and J.-M. Frahm, “Structure-from-motion revisited,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016

2016
[13]

HAC: Hash-grid assisted context for 3D Gaussian splatting compression,

Y . Chen, Q. Wu, J. Cai, M. Harandi, and W. Lin, “HAC: Hash-grid assisted context for 3D Gaussian splatting compression,” inProc. ECCV, 2024

2024
[14]

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

Y . Chen, Q. Wu, W. Lin, M. Harandi, and J. Cai, “HAC++: Towards 100×compression of 3D Gaussian splatting,”IEEE Trans. Pattern Anal. Mach. Intell., 2025, arXiv:2501.12255

work page arXiv 2025
[15]

FlexGaussian: Flex- ible and cost-effective training-free compression for 3D Gaussian splat- ting,

B. Tian, Q. Gao, S. Xianyu, X. Cui, and M. Zhang, “FlexGaussian: Flex- ible and cost-effective training-free compression for 3D Gaussian splat- ting,” inProc. ACM Multimedia (ACM MM), 2025, arXiv:2507.06671. IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY , VOL. XX, NO. X, MONTH YEAR 14

work page arXiv 2025
[16]

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,” inProc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2024

2024
[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,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, arXiv:2401.02436

work page arXiv 2024
[18]

CompGS: Smaller and faster Gaussian splatting with vector quantization,

K. L. Navaneet, K. Pourahmadi Meibodi, S. Abbasi Koohpayegani, and H. Pirsiavash, “CompGS: Smaller and faster Gaussian splatting with vector quantization,” inProc. ECCV, 2024, arXiv:2311.18159

work page arXiv 2024
[19]

Compact 3D Gaussian representation for radiance field,

J. C. Lee, D. Rho, X. Sun, J. H. Ko, and E. Park, “Compact 3D Gaussian representation for radiance field,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Highlight, 2024, arXiv:2311.13681

work page arXiv 2024
[20]

3DGS.zip: A survey on 3D Gaus- sian splatting compression methods,

M. T. Bagdasarian, P. Knoll, Y .-H. Li, F. Barthel, A. Hilsmann, P. Eisert, and W. Morgenstern, “3DGS.zip: A survey on 3D Gaus- sian splatting compression methods,”Computer Graphics Forum, 2024, arXiv:2407.09510

work page arXiv 2024
[21]

Radsplat: Radiance field-informed gaussian splat- ting for robust real-time rendering with 900+ fps.arXiv preprint arXiv:2403.13806, 2024

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,” inProc. Int. Conf. 3D Vision (3DV), 2024, arXiv:2403.13806

work page arXiv 2024
[22]

LightGaussian: Unbounded 3D Gaussian compression with15×reduction and200+ FPS,

Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang, “LightGaussian: Unbounded 3D Gaussian compression with15×reduction and200+ FPS,” inProc. NeurIPS, 2024

2024
[23]

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,”Proc. ACM Comput. Graph. Interact. Tech., vol. 7, no. 1, 2024

2024
[24]

Mobile-GS: Real-time Gaussian splatting for mobile devices,

X. Du, Y . Wang, K. Zhan, and X. Yu, “Mobile-GS: Real-time Gaussian splatting for mobile devices,”arXiv preprint arXiv:2603.11531, 2026

work page arXiv 2026
[25]

ContextGS: Compact 3D Gaussian splatting with anchor level context model,

Y . Wanget al., “ContextGS: Compact 3D Gaussian splatting with anchor level context model,” inProc. NeurIPS, 2024

2024
[26]

EAGLES: Efficient accelerated 3D Gaussians with lightweight encodings,

S. Girish, K. Gupta, and A. Shrivastava, “EAGLES: Efficient accelerated 3D Gaussians with lightweight encodings,” inProc. ECCV, 2023, arXiv:2312.04564

work page arXiv 2023
[27]

Compression in 3D Gaussian splatting: A survey of methods, trends, and future directions,

M. S. Ali, C. Zhang, M. Cagnazzo, G. Valenzise, E. Tartaglione, and S.-H. Bae, “Compression in 3D Gaussian splatting: A survey of methods, trends, and future directions,”ACM Computing Surveys, 2025, arXiv:2502.19457

work page arXiv 2025
[28]

A Survey on 3D Gaussian Splatting

G. Chen and W. Wang, “A survey on 3D Gaussian splatting,”ACM Computing Surveys, 2024, arXiv:2401.03890

work page internal anchor Pith review Pith/arXiv arXiv 2024
[29]

Automatic joint structured pruning and quantization for efficient neural network training and compression,

X. Qu, T. Aponte, C. Banbury, D. P. Robinson, T. Ding, K. Koishida, I. Zharkov, and T. Chen, “Automatic joint structured pruning and quantization for efficient neural network training and compression,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025. arXiv:2502.16638

work page arXiv 2025
[30]

HESSO: Towards automatic efficient and user-friendly any neural network training and pruning,

T. Chen, X. Qu, D. Aponte, C. Banbury, J. Ko, T. Ding, Y . Ma, V . Lyapunov, I. Zharkov, and L. Liang, “HESSO: Towards automatic efficient and user-friendly any neural network training and pruning,” arXiv preprint arXiv:2409.09085, 2024.(Note: T. Chen here is Tianyi Chen at Microsoft, also senior author of GETA [29].)

work page arXiv 2024
[31]

Bayesian Bits: Unifying quantization and pruning,

M. van Baalen, C. Louizos, M. Nagel, R. A. Amjad, Y . Wang, T. Blankevoort, and M. Welling, “Bayesian Bits: Unifying quantization and pruning,” inProc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2020, arXiv:2005.07093

work page arXiv 2020
[32]

Differentiable joint pruning and quantization for hardware efficiency,

Y . Wang, Y . Lu, and T. Blankevoort, “Differentiable joint pruning and quantization for hardware efficiency,” inProc. ECCV, 2020

2020
[33]

CLIP-Q: Deep network compression learning by in-parallel pruning-quantization,

F. Tung and G. Mori, “CLIP-Q: Deep network compression learning by in-parallel pruning-quantization,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018

2018
[34]

Optimal brain compression: A framework for accurate post-training quantization and pruning,

E. Frantar and D. Alistarh, “Optimal brain compression: A framework for accurate post-training quantization and pruning,” inProc. NeurIPS, 2022

2022
[35]

Learning efficient convolutional networks through network slimming,

Z. Liu, J. Li, Z. Shen, G. Huang, S. Yan, and C. Zhang, “Learning efficient convolutional networks through network slimming,” inProc. IEEE Int. Conf. Comput. Vis. (ICCV), 2017

2017
[36]

Up or down? Adaptive rounding for post-training quantization,

M. Nagel, R. A. Amjad, M. van Baalen, C. Louizos, and T. Blankevoort, “Up or down? Adaptive rounding for post-training quantization,” in Proc. ICML, 2020

2020
[37]

A survey of quantization methods for efficient neural network infer- ence,

A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, and K. Keutzer, “A survey of quantization methods for efficient neural network infer- ence,”Low-Power Computer Vision: Improvement of the Efficiency of Artificial Intelligence, CRC Press, 2022

2022
[38]

Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation

Y . Bengio, N. L ´eonard, and A. Courville, “Estimating or propagating gradients through stochastic neurons for conditional computation,”arXiv preprint arXiv:1308.3432, 2013

work page internal anchor Pith review arXiv 2013
[39]

Importance estimation for neural network pruning,

P. Molchanov, A. Mallya, S. Tyree, I. Frosio, and J. Kautz, “Importance estimation for neural network pruning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019

2019
[40]

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 Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018

2018
[41]

TorchMetrics — measuring reproducibility in PyTorch,

N. S. Detlefsenet al., “TorchMetrics — measuring reproducibility in PyTorch,”J. Open Source Softw., vol. 7, no. 70, p. 4101, 2022

2022
[42]

gsplat: An open-source library for Gaussian splatting,

V . Yeet al., “gsplat: An open-source library for Gaussian splatting,” arXiv preprint arXiv:2409.06765, 2024

work page arXiv 2024
[43]

Calculation of average PSNR differences between RD- curves,

G. Bjontegaard, “Calculation of average PSNR differences between RD- curves,” ITU-T VCEG document VCEG-M33, 2001

2001
[44]

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 Trans. Image Process., vol. 13, no. 4, pp. 600–612, 2004

2004
[45]

3D Gaussian splatting: Survey, technologies, challenges, and opportunities,

Y . Bao, T. Ding, J. Huo, Y . Liu, Y . Li, W. Li, Y . Gao, and J. Luo, “3D Gaussian splatting: Survey, technologies, challenges, and opportunities,”IEEE Trans. Circuits Syst. Video Technol., 2025, doi:10.1109/TCSVT.2025.3538684, arXiv:2407.17418

work page doi:10.1109/tcsvt.2025.3538684 2025
[46]

Boyd and L

S. Boyd and L. Vandenberghe,Convex Optimization. Cambridge, U.K.: Cambridge Univ. Press, 2004

2004
[47]

Bennett’s integral for vector quantizers,

S. Na and D. L. Neuhoff, “Bennett’s integral for vector quantizers,” IEEE Trans. Inf. Theory, vol. 58, no. 6, pp. 3654–3679, 2012

2012