Pith. sign in

REVIEW 3 major objections 5 minor 59 references

Smol-GS claims that storing splat coordinates in an entropy-coded octree and distilling appearance into 8-D abstract features compresses 3D Gaussian scenes by ~150x with no quality drop.

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

2026-08-03 19:19 UTC pith:3CE637GO

load-bearing objection Credible 3DGS compression with smallest reported sizes, but the octree coordinate codec needs a precise collision-handling story before the bitrate claims are solid. the 3 major comments →

arxiv 2512.00850 v3 pith:3CE637GO submitted 2025-11-30 cs.CV

Smol-GS: Compact Representations for Abstract 3D Gaussian Splatting

Haishan Wang , Mohammad Hassan Vali , Arno Solin This is my paper
classification cs.CV
keywords 3D Gaussian splattingscene compressionoccupancy octreeentropy codingquantizationneural renderingnovel view synthesiscompact representation
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.

Smol-GS claims that a 3D Gaussian splatting scene can be stored orders of magnitude more compactly by separating geometry from appearance: splat coordinates go into a fixed-depth occupancy octree whose byte stream is entropy-coded, while every other attribute is distilled into an 8-dimensional per-splat feature that a few tiny neural networks decode into color, opacity, rotation, and scale. The features are quantized with learned step sizes and arithmetic-coded using distributions predicted from a hash grid of spatial context. On standard benchmarks the method reports the smallest total model sizes among strong compression baselines—roughly 150x smaller than vanilla 3DGS—with PSNR, SSIM, and LPIPS about level with the strongest methods, and it still renders in real time. A sympathetic reader should care because it suggests the bulky per-splat attributes of 3DGS are redundant and can be abstracted without giving up the explicit, editable geometry that makes Gaussian splatting attractive.

Core claim

The paper's central claim is that in 3D Gaussian Splatting, the expensive per-splat attributes—coordinates, colors, opacities, rotations, scales, and view-dependent material cues—are over-parameterized. Smol-GS keeps coordinates explicit but packs them as occupancy bytes in a recursive octree (one splat per leaf at depth 16), so position storage is nearly lossless on a 2^-16 grid yet very small. Everything visual is compressed into an 8-dimensional abstract feature per splat, and a few tiny multilayer perceptrons synthesize opacity, color, rotation, and scale from that feature plus the viewing direction. A hash-based entropy model quantizes the features with learned step sizes and arithmetic

What carries the argument

Occupancy-octree coordinate coding, abstract splat features, and learned entropy coding work together. The octree recursively subdivides the scene's bounding box into eight children, emitting an 8-bit occupancy byte at each split so only non-empty children continue; the byte stream is entropy-coded to make coordinates cheap. Each splat carries an 8-dimensional abstract feature encoding color, opacity, transformation, and material cues, and tiny MLPs decode these into renderable attributes conditioned on viewing direction. A hash grid predicts per-dimension Gaussian distributions and quantization step sizes, and the negative log-likelihood of quantization bins becomes a rate loss, while arith

Load-bearing premise

The load-bearing premise is that snapping every splat coordinate to one of 65,536 grid positions along each axis costs almost no visual quality; the paper's support is a single-scene post-hoc quantization of a pretrained model, so scenes with very fine geometry are where this could break.

What would settle it

On a scene with dense thin structures such as leaves, railings, or hair, train Smol-GS at the default R=16 and at R=18 with matched regularization; if PSNR or LPIPS improves meaningfully at R=18 beyond the reported margins separating Smol-GS from the strongest baselines, the fixed 2^-16 grid is not universally sufficient. The paper's own ablation shows one dataset still improving from R=16 to R=17, so reproducing that improvement on a second dataset would settle the point.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • The smallest Smol-GS variant lands at roughly half the size of the previous smallest strong baseline while keeping PSNR within a few tenths of a decibel, and the base variant retains full PSNR parity on the main benchmark suite.
  • Coordinate storage stops being a bottleneck: on the reported large-scene suite, the octree-coded coordinates take about 1.4 MB of a 6.4 MB model, so geometry is no longer the main memory cost.
  • Rendering stays real-time at roughly 215–405 frames per second in the reported timings, so the compact representation does not trade away interactivity.
  • The recursion depth R is a direct size–quality dial: ablated quality saturates around R=16 on some suites but continues to improve at R=17 on another, making the default depth a tunable rather than fixed choice.
  • The discrete per-splat features and the parent–child octree structure could serve as tokens for editing, generation, or scene understanding without reconstructing full splat attributes, as the paper explicitly suggests.

Where Pith is reading between the lines

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

  • The motivating evidence that coordinate precision is cheap comes from post-hoc quantization of a pretrained vanilla 3DGS model, not from Smol-GS trained end-to-end under octree quantization; a head-to-head comparison of those two paths would sharpen the claim.
  • Because abstract features dominate the bit budget—roughly half the total size—future compression gains are more likely to come from a stronger feature entropy model, such as conditioning on neighboring splats or octree ancestry, than from further coordinate coding.
  • The octree's parent–child relations imply a natural progressive-transmission or level-of-detail scheme: coarse nodes could be sent before refinements, a capability the paper mentions but does not quantify.
  • A testable extension is to choose the octree depth per scene or per region rather than globally, which could retain the size advantage on scenes with very thin or high-frequency geometry.

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

3 major / 5 minor

Summary. Smol-GS is a 3D Gaussian splatting compression method. Each splat stores only a low-dimensional (nf=8) abstract feature and a scale controller; opacity, color, rotation, and scale are decoded by tiny MLPs from the feature plus viewing direction/distance. Coordinates are quantized to leaf centers of an occupancy octree and Huffman-coded; features and scale controllers are quantized with learned step sizes and arithmetic-coded using a hash-conditioned Gaussian entropy model. Training proceeds in five stages (warm-up, densification, compaction, feature compression, coordinate compression). On MIP-NeRF 360, Tanks&Temples, and DeepBlending, the authors report total sizes as low as 4.75 MB on MIP-NeRF 360, compared with 8.74 MB for HAC++, at comparable PSNR/SSIM/LPIPS.

Significance. If the reported bitstreams are actually decodable at the stated sizes, this is a strong practical result: it would be the smallest published 3DGS representation at comparable quality on these benchmarks, about 1.5–2x smaller than HAC++/HEMGS. The method is clearly motivated, the stage-wise training is coherent, and the ablations (R, nf, λo, λq) cover the main design choices. The paper also provides per-component size and encoding/decoding time breakdowns. The principal risk is the under-specified octree collision handling, which directly affects the coordinate bitrate and hence the central size claim. Reproducibility would be improved by releasing code/bitstreams; baseline numbers are imported from a survey rather than re-run.

major comments (3)
  1. [§3.2, §3.4, §3.5] §3.2 states that the occupancy octree stores coordinates with 'only one splat in each leaf', and this assumption underlies the coordinate sizes in Table 2 (e.g., 1.412 MB on MIP-NeRF 360). However, nothing in the pipeline enforces this invariant. In stage 5 (§3.5), all splat coordinates are quantized to leaf centers; during densification (§3.4) new splats are created near parents, so multiple splats can fall into the same leaf. The manuscript never defines how such collisions are encoded. If duplicates are retained, the decoder needs per-leaf counts (or per-splat indices), which would increase the bitrate beyond Table 2. If duplicates are dropped, the rendered model differs from the trained one, so the reported PSNR/SSIM are not for the compressed representation. Please specify a collision-resolution scheme, report collision statistics on all benchmark scenes, and validate the actual dec
  2. [Table 1, Fig. 7] Table 1 and Fig. 7 compare Smol-GS with baseline sizes/quality taken from the 3DGS.zip survey [1], not re-run under the same protocol, and Smol-GS numbers appear to be single runs with no error bars. The advantage over the strongest baselines is not large on every dataset: on Tanks & Temples, Smol-GS-base has PSNR 23.98 vs HEMGS 24.42 and size 5.525 MB vs HAC++ 5.427 MB; only the small variant is actually smallest. To support the 'smallest at comparable quality' claim, please provide per-scene baseline numbers from [1] (or re-run the top baselines) and report variance/seed variability for Smol-GS across at least the three datasets.
  3. [Fig. 4 and Table A7] Fig. 4 motivates octree coordinate quantization using a pre-trained vanilla 3DGS model on one scene (Garden), not the actual Smol-GS pipeline. While the R-ablation in Table A7 is more relevant and reassuring, it does not report how many leaves contain multiple splats after stage 5. Since the coordinate bitrate and the quality of the compressed model both depend on this, please add an end-to-end analysis of the trained Smol-GS model: number of splats before and after quantization, distribution of splats per occupied leaf, and quality measured after decoding the exact stored bitstream.
minor comments (5)
  1. [Table 1 caption] 'Smol-GS has the smallest sizes' is only true for the small/tiny variants; Smol-GS-base is not smaller than HAC++ on Tanks & Temples (5.525 vs 5.427 MB). Please clarify which variant the claim refers to.
  2. [Supplementary A.1] 'python=2.11' should presumably be 'Python 3.11'.
  3. [§4 Implementation Details] 'scalling controllers' should be 'scaling controllers'.
  4. [Eq. (8)] The NLL is described as a 'theoretical lower bound' for bits; more precisely it is the ideal code length under the assumed Gaussian model, not an information-theoretic lower bound.
  5. [Abstract] The abstract says 'orders-of-magnitude reduction in storage'; the reduction vs vanilla 3DGS is large, but vs HAC++/HEMGS it is about 1.5–2x. Consider using precise numbers.

Circularity Check

0 steps flagged

No significant circularity: reported sizes are measured zipped files, rate terms are standard training objectives, and self-citations are peripheral.

full rationale

The paper's load-bearing claims—smaller model size at comparable rendering quality—are supported by externally measurable outputs: actual zip-file sizes (Sec. 4, Tables 1–2) and PSNR/SSIM/LPIPS on held-out test views. The NLL rate term in Eqs. (8)–(9) is an entropy-model training loss; optimizing it does not by itself determine the arithmetic-coded bit count, so the final measured sizes are an independent result rather than a restatement of the loss. The occupancy-octree coordinate codec (Sec. 3.2) is a quantizing transform applied to trained splat coordinates, and its bitrate is separately measured; the Fig. 4 quantization experiment motivates the design but is not the target prediction. Self-citations [43,45] appear only in related-work discussions of vector quantization and prior 3DGS compression; no central equation, benchmark number, or uniqueness argument is imported from them. The hash-grid conditional entropy model is credited to external works (HAC [4], InstantNGP [31]). The main weakness—under-specification of the one-splat-per-leaf invariant when multiple splats collapse into the same octree leaf during stage-5 quantization—is a correctness/reproducibility risk, not a circularity: it does not make any result true by definition or reduce a prediction to a fitted input.

Axiom & Free-Parameter Ledger

10 free parameters · 6 axioms · 1 invented entities

The central empirical claims rest mainly on domain assumptions: that R=16 octree quantization is perceptually lossless, that the Gaussian bin-NLL rate model matches actual arithmetically coded sizes, that imported baseline sizes are comparable, and that the chosen hyperparameters generalize. No fundamental mathematical axioms are introduced; no new physical entities are required beyond the learned 8-d splat feature, which has no independent evidence outside the benchmark metrics.

free parameters (10)
  • Feature dimension nf = 8
    Ablation Table A8 chose 8 as the size/quality trade-off; it directly sets storage and MLP input dimension.
  • Octree recursion depth R = 16
    Ablation Table A7 chose 16; Tanks and Temples prefers 17. Directly sets coordinate quantization precision and size.
  • Opacity regularization weight λo = 2e-7 (base), 3e-7 (small), 4e-7 (tiny)
    Ablation Table A9 on the same test sets used for reporting; controls splat count and model size.
  • Quantization regularization weight λq = 1e-3
    Ablation Table A10; controls the bit-rate penalty on features and scaling controllers.
  • SSIM loss weight λs = 0.2
    Implementation detail in Sec. 4; not ablated.
  • Densification gradient thresholds τgk = 2^k × 10^-4, k=1..5
    Chosen thresholds controlling how many new splats are generated; affects final splat count.
  • Opacity pruning threshold τo = 5e-3, with linear schedule in Eq. 13
    Determines pruning aggressiveness; strongly affects size and quality.
  • Learning rates and schedulers = Table A5 values
    Optimization hyperparameters per parameter group; not justified by theory.
  • Hash table sizes/resolutions = 2D 2^15, 3D 2^13, 4-dim features, resolution lists in Sec. A.4
    Entropy-model capacity; copied from HAC/InstantNGP style, not ablated.
  • MLP hidden dim/layers = 3 layers, 128 hidden units
    Architecture choices that set decoder capacity and MLP parameter count.
axioms (6)
  • domain assumption Quantized splat features follow a Gaussian distribution N(f; μ_f, Σ_f) with diagonal covariance predicted by the hash-grid MLP (Eq. 8).
    Used to compute the NLL rate term; if the true distribution deviates, the estimated bit rate and entropy-coding efficiency may be optimistic (Sec. 3.3).
  • domain assumption Occupancy octree with R=16 and one splat per leaf provides sufficient coordinate precision for perceptually lossless reconstruction.
    Coordinate compression depends on this; only directly validated on the Garden scene with vanilla 3DGS, not on Smol-GS across all scenes (Fig. 4, Sec. 3.2).
  • standard math Arithmetic coding achieves the NLL lower bound, and Huffman coding suitably compresses occupancy bytes.
    Standard information-theoretic result assumed for the reported sizes (Secs. 3.2-3.3).
  • domain assumption 3DGS alpha-blending rasterization (Eqs. 4-5) is a valid differentiable renderer, and test-view metrics are meaningful.
    Foundation inherited from 3DGS [18]; needed for all experiments.
  • domain assumption SfM point clouds provide adequate initialization for densification and octree construction.
    Follows 3DGS; if initialization is poor, octree structure and densification may be suboptimal (Sec. 3.4).
  • domain assumption Baselines collected from the 3DGS.zip survey are comparable to the authors' Smol-GS results.
    All baseline quality/size numbers are imported from [1], not re-run; cross-method comparability depends on the survey methodology (Sec. 4.1, Table 1).
invented entities (1)
  • Abstract splat feature f (nf=8) no independent evidence
    purpose: Encodes color, opacity, transformation, and material cues compactly; decoded by tiny MLPs.
    Learned per-scene latent representation with no external falsifiable handle beyond reconstruction quality on benchmarks. Listed for completeness because it is a new representational object, not a physical entity.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Smol-GS: Compact Representations for Abstract 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/3CE637GO

@misc{pith2026251200850,
  author       = {Pith},
  title        = {Pith review of: Smol-GS: Compact Representations for Abstract 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3CE637GO}},
  note         = {Machine review of arXiv:2512.00850}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present Smol-GS, a novel method for learning compact representations for 3D Gaussian Splatting (3DGS). Our approach learns highly efficient splat-wise features to model 3D space, which capture abstracted cues, including color, opacity, transformation, and material properties. We propose octree-derived positional encoding, which explicitly models spatial locality and enhances representation efficiency. We further apply entropy-based compression to exploit feature redundancy and compress splat coordinates using a recursive voxel hierarchy. This design enables orders-of-magnitude reduction in storage while preserving representation flexibility. Smol-GS achieves state-of-the-art compression performance on standard benchmarks with high-level rendering quality.

Figures

Figures reproduced from arXiv: 2512.00850 by Arno Solin, Haishan Wang, Mohammad Hassan Vali.

Figure 1
Figure 1. Figure 1: Smol-GS learns a compact representation of the 3D scene that (i) stores coordinates in an efficient octree-like structure and (ii) abstracts away view-dependent splat features such as color, shape, and material cues. We visualize the abstract 8-dimensional features f in the second and third panel by coloring RGB = sigmoid(PCA(f)), which reveals structure in the encodings. On the MIP-NERF 360 benchmark, the… view at source ↗
Figure 3
Figure 3. Figure 3: Method overview of Smol-GS: We train (trained parameters in blue) neural splats with tiny MLP-decoders for view-dependent rendering (Sec. 3.1). The coordinates are compressed with occupancy-octree coordinate coding (Sec. 3.2). We also learn the quantization and arithmetic coding of splat features with NLL rate terms (Sec. 3.3), and employ adaptive density control with stage-wise training (loss-terms in red… view at source ↗
Figure 4
Figure 4. Figure 4: Occupancy-octree depth vs. PSNR. We experiment with quantizing the coordinates of a trained 3DGS-30k model to be the ‘Quantized-30k’ models. These models are finetuned as ‘Quantized￾45k’ for an additional 15k iterations with quantized coordinates fixed. Top: Quantitative metrics (PSNR) and splat number ratio after quantization vs. quantization recursion on the Garden scene. Bottom: Qualitative results of Q… view at source ↗
Figure 5
Figure 5. Figure 5: Occupancy-octree coordinate coding: Given a point cloud (left), we recursively divide the bounding box into eight sub-boxes. Only the non-empty sub-boxes (gray) are further divided. Each division is represented by an 8-bit binary code (1=non-empty, 0=empty). Cri denotes the code of the j th division at r th recursion. The occupancy octree (right) is constructed by arranging all bits in a breadth-first mann… view at source ↗
Figure 6
Figure 6. Figure 6: Training curves over different stages: We log the number of splats (left) and PSNR (right) over different stages of training, showing how the compaction and compression do not de￾grade the overall quality. The splat amount increases stably during densification (0.5k–15k), then decreases rapidly during compaction (15k–20k). truth, the opacity regularization loss Lo, and the quantization regularization losse… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of top methods on MIP-NERF 360. Each point is a method; the x-axis is total model size in MBs (smaller is better) and the y-axis reports PSNR/SSIM (higher is better) or LPIPS (lower is better). Smol-GS (tiny–small–base) consistently lies on or beyond the Pareto frontier, achieving markedly smaller sizes at comparable quality to strong baselines collected via the 3DGS.zip benchmark. See Fig. A13 … view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative results on the TANKS AND TEMPLES and DEEP BLENDING data sets. Smol-GS has fewer artefacts and better preservation of scene details under challenging lighting and materials. The learned splat features in Smol-GS capture view-dependent and material cues more compactly than SH, aligning with its smaller size in MBs and competitive metrics in [PITH_FULL_IMAGE:figures/full_fig_p008_8.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

59 extracted references · 6 linked inside Pith

  1. [1]

    3DGS.zip: A survey on 3D Gaussian splatting compression methods

    Milena T Bagdasarian, Paul Knoll, Y Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgenstern. 3DGS.zip: A survey on 3D Gaussian splatting compression methods. InComputer Graphics Forum, page e70078. Wiley Online Library, 2025. 6, 7

    2025

  2. [2]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470–5479, 2022. 6

    2022

  3. [3]

    GaussianEditor: Swift and controllable 3D editing with Gaussian splatting

    Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. GaussianEditor: Swift and controllable 3D editing with Gaussian splatting. InProceedings of the IEEE/CVF conference on Computer Vision and Pattern Recog- nition (CVPR), pages 21476–21485, 2024. 2

    2024

  4. [4]

    HAC: Hash-grid assisted context for 3D Gaus- sian splatting compression

    Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: Hash-grid assisted context for 3D Gaus- sian splatting compression. InProceedings of the European Conference on Computer Vision (ECCV), pages 422–438. Springer, 2024. 1, 4, 6, 7

    2024

  5. [5]

    HAC++: Towards 100x compression of 3D Gaussian splatting.arXiv preprint arXiv:2501.12255, 2025

    Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC++: Towards 100x compression of 3D Gaussian splatting.arXiv preprint arXiv:2501.12255, 2025. 1, 2, 6, 7

    Pith/arXiv arXiv 2025

  6. [6]

    Octformer: Efficient octree-based transformer for point cloud compression with local enhancement

    Mingyue Cui, Junhua Long, Mingjian Feng, Boyang Li, and Huang Kai. Octformer: Efficient octree-based transformer for point cloud compression with local enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 470–478, 2023. 3, 4

    2023

  7. [7]

    LightGaussian: Unbounded 3D Gaus- sian compression with 15x reduction and 200+ fps

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. LightGaussian: Unbounded 3D Gaus- sian compression with 15x reduction and 200+ fps. InAd- vances in Neural Information Processing Systems (NeurIPS), pages 140138–140158. Curran Associates, Inc., 2024. 1, 2

    2024

  8. [8]

    Mini-splatting: Representing scenes with a constrained number of Gaussians

    Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of Gaussians. InProceed- ings of the European Conference on Computer Vision (ECCV), pages 165–181. Springer, 2024. 1, 2

    2024

  9. [9]

    Oc- tattention: Octree-based large-scale contexts model for point cloud compression

    Chunyang Fu, Ge Li, Rui Song, Wei Gao, and Shan Liu. Oc- tattention: Octree-based large-scale contexts model for point cloud compression. InProceedings of the AAAI conference on artificial intelligence, pages 625–633, 2022. 3, 4

    2022

  10. [10]

    Springer Science & Business Media,

    Allen Gersho and Robert M Gray.Vector Quantization and Signal Compression. Springer Science & Business Media,

  11. [11]

    EA- GLES: Efficient accelerated 3D Gaussians with lightweight encodings

    Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. EA- GLES: Efficient accelerated 3D Gaussians with lightweight encodings. InProceedings of the European Conference on Computer Vision (ECCV), pages 54–71. Springer, 2024. 1

    2024

  12. [12]

    Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (ToG), 37(6), 2018

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (ToG), 37(6), 2018. 6

    2018

  13. [13]

    CG-SLAM: Efficient dense RGB-D slam in a consistent uncertainty-aware 3D Gaussian field

    Jiarui Hu, Xianhao Chen, Boyin Feng, Guanglin Li, Liangjing Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. CG-SLAM: Efficient dense RGB-D slam in a consistent uncertainty-aware 3D Gaussian field. InProceedings of the European Conference on Computer Vision (ECCV), pages 93–112. Springer, 2024. 2

    2024

  14. [14]

    Photo-SLAM: Real-time simultaneous localization and photo- realistic mapping for monocular stereo and RGB-D cameras

    Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-SLAM: Real-time simultaneous localization and photo- realistic mapping for monocular stereo and RGB-D cameras. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21584–21593, 2024. 2

    2024

  15. [15]

    A method for the construction of minimum-redundancy codes.Proceedings of the IRE, 40 (9):1098–1101, 2007

    David A Huffman. A method for the construction of minimum-redundancy codes.Proceedings of the IRE, 40 (9):1098–1101, 2007. 3, 4

    2007

  16. [16]

    VR-GS: A physical dynamics-aware interactive Gaussian splatting system in virtual reality

    Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, and Jiang Chenfanfu. VR-GS: A physical dynamics-aware interactive Gaussian splatting system in virtual reality. In ACM SIGGRAPH (Conference Paper Track), 2024. 2

    2024

  17. [17]

    SplaTAM: Splat track & map 3D Gaussians for dense RGB-D SLAM

    Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. SplaTAM: Splat track & map 3D Gaussians for dense RGB-D SLAM. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21357–21366, 2024. 2

    2024

  18. [18]

    3D Gaussian splatting for real-time radi- ance field rendering.ACM Transactions on Graphics (ToG), 42(4), 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨uhler, and George Drettakis. 3D Gaussian splatting for real-time radi- ance field rendering.ACM Transactions on Graphics (ToG), 42(4), 2023. 1, 4, 5, 6, 7

    2023

  19. [19]

    Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014. 6, 1

    Pith/arXiv arXiv 2014

  20. [20]

    Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36(4),

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36(4),

  21. [21]

    Point-based neural rendering with per-view optimization

    Georgios Kopanas, Julien Philip, Thomas Leimk ¨uhler, and George Drettakis. Point-based neural rendering with per-view optimization. InComputer Graphics Forum, pages 29–43. Wiley Online Library, 2021. 4

    2021

  22. [22]

    Pulsar: Effi- cient sphere-based neural rendering

    Christoph Lassner and Michael Zollhofer. Pulsar: Effi- cient sphere-based neural rendering. InProceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, pages 1440–1449, 2021. 1

    2021

  23. [23]

    Compact 3D Gaussian representation for radiance field

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3D Gaussian representation for radiance field. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21719–21728, 2024. 1, 2, 6, 7

    2024

  24. [24]

    An algorithm for vector quantizer design.IEEE Transactions on Communi- cations, 28(1):84–95, 2003

    Yoseph Linde, Andres Buzo, and Robert Gray. An algorithm for vector quantizer design.IEEE Transactions on Communi- cations, 28(1):84–95, 2003. 2

    2003

  25. [25]

    HEMGS: A hybrid entropy model for 3D Gaussian splatting data compression.arXiv preprint arXiv:2411.18473, 2024

    Lei Liu, Zhenghao Chen, Wei Jiang, Wei Wang, and Dong Xu. HEMGS: A hybrid entropy model for 3D Gaussian splatting data compression.arXiv preprint arXiv:2411.18473, 2024. 1, 2, 6, 7

    Pith/arXiv arXiv 2024

  26. [26]

    Least squares quantization in pcm.IEEE Transactions on Information Theory, 28(2):129–137, 1982

    Stuart Lloyd. Least squares quantization in pcm.IEEE Transactions on Information Theory, 28(2):129–137, 1982. 2

    1982

  27. [27]

    Scaffold-GS: Structured 3D Gaussians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-GS: Structured 3D Gaussians for view-adaptive rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20654–20664, 2024. 1, 2, 5

    2024

  28. [28]

    Taming 3DGS: High-quality radiance fields with limited resources

    Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fer- nando De La Torre. Taming 3DGS: High-quality radiance fields with limited resources. InSIGGRAPH Asia 2024 Con- ference Papers, pages 1–11, 2024. 1, 2

    2024

  29. [29]

    Compact 3D scene representation via self- organizing Gaussian grids

    Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3D scene representation via self- organizing Gaussian grids. InProceedings of the Euro- pean Conference on Computer Vision (ECCV), pages 18–34. Springer, 2024. 1, 6, 7

    2024

  30. [30]

    International Busi- ness Machines Company, 1966

    Guy M Morton.A Computer Oriented Geodetic Data Base and a New Technique in File Sequencing. International Busi- ness Machines Company, 1966. 4

    1966

  31. [31]

    Instant neural graphics primitives with a multireso- lution hash encoding.ACM Transactions on Graphics (ToG), 41(4):1–15, 2022

    Thomas M¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a multireso- lution hash encoding.ACM Transactions on Graphics (ToG), 41(4):1–15, 2022. 4

    2022

  32. [32]

    CompGS: Smaller and faster Gaussian splatting with vector quantization

    KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Ab- basi Koohpayegani, and Hamed Pirsiavash. CompGS: Smaller and faster Gaussian splatting with vector quantization. InPro- ceedings of the European Conference on Computer Vision (ECCV), pages 330–349. Springer, 2024. 1, 2, 6, 7

    2024

  33. [33]

    Compressed 3D Gaussian splatting for accelerated novel view synthesis

    Simon Niedermayr, Josef Stumpfegger, and R ¨udiger West- ermann. Compressed 3D Gaussian splatting for accelerated novel view synthesis. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pages 10349–10358, 2024. 6, 7

    2024

  34. [34]

    Reducing the mem- ory footprint of 3D Gaussian splatting.Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7 (1):1–17, 2024

    Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the mem- ory footprint of 3D Gaussian splatting.Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7 (1):1–17, 2024. 1, 2, 6, 7

    2024

  35. [35]

    Pytorch: An imperative style, high-performance deep learning library.Ad- vances in Neural Information Processing Systems (NeurIPS), 32, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library.Ad- vances in Neural Information Processing Systems (NeurIPS), 32, 2019. 6, 1

    2019

  36. [36]

    BOGausS: Better optimized Gaussian splatting.arXiv preprint arXiv:2504.01844, 2025

    St´ephane Pateux, Matthieu Gendrin, Luce Morin, Th ´eo Ladune, and Xiaoran Jiang. BOGausS: Better optimized Gaussian splatting.arXiv preprint arXiv:2504.01844, 2025. 1, 2

    Pith/arXiv arXiv 2025

  37. [37]

    V oxelcontext-net: An octree based framework for point cloud compression

    Zizheng Que, Guo Lu, and Dong Xu. V oxelcontext-net: An octree based framework for point cloud compression. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6042–6051, 2021. 3, 4

    2021

  38. [38]

    Octree-GS: Towards consistent real-time rendering with lod-structured 3D Gaussians.arXiv preprint arXiv:2403.17898, 2024

    Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-GS: Towards consistent real-time rendering with lod-structured 3D Gaussians.arXiv preprint arXiv:2403.17898, 2024. 1, 2

    Pith/arXiv arXiv 2024

  39. [39]

    Octree-based point- cloud compression

    Ruwen Schnabel and Reinhard Klein. Octree-based point- cloud compression. InProceedings of the 3rd Eurographics / IEEE VGTC Conference on Point-Based Graphics, page 111–121, Goslar, DEU, 2006. Eurographics Association. 2, 3, 4

    2006

  40. [40]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InProceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 4104–4113, 2016. 2, 5

    2016

  41. [41]

    Control4D: Efficient 4D portrait editing with text

    Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4D: Efficient 4D portrait editing with text. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4556–4567, 2024. 2

    2024

  42. [42]

    DreamGaussian: Generative Gaussian splatting for ef- ficient 3D content creation.arXiv preprint arXiv:2309.16653,

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative Gaussian splatting for ef- ficient 3D content creation.arXiv preprint arXiv:2309.16653,

    Pith/arXiv arXiv

  43. [43]

    Stochastic optimization of vector quantization methods in application to speech and image processing

    Mohammad Hassan Vali and Tom B ¨ackstr¨om. Stochastic optimization of vector quantization methods in application to speech and image processing. InInternational Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2023. 2

    2023

  44. [44]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. InAdvances in Neu- ral Information Processing Systems 30 (NIPS). Curran Asso- ciates, Inc., 2017. 2

    2017

  45. [45]

    Compressing 3D Gaussian splatting by noise-substituted vec- tor quantization

    Haishan Wang, Mohammad Hassan Vali, and Arno Solin. Compressing 3D Gaussian splatting by noise-substituted vec- tor quantization. InScandinavian Conference on Image Anal- ysis (SCIA), pages 338–352. Springer, 2025. 1

    2025

  46. [46]

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

    Yufei Wang, Zhihao Li, Lanqing Guo, Wenhan Yang, Alex Kot, and Bihan Wen. ContextGS: Compact 3D Gaussian splatting with anchor level context model. InAdvances in Neural Information Processing Systems (NeurIPS), pages 51532–51551. Curran Associates, Inc., 2024. 1, 2, 6, 7

    2024

  47. [47]

    The University of North Carolina at Chapel Hill, 1991

    Lee Alan Westover.Splatting: A Parallel, Feed-Forward Vol- ume Rendering Algorithm. The University of North Carolina at Chapel Hill, 1991. 2

    1991

  48. [48]

    Arithmetic coding for data compression.Communications of the ACM, 30(6):520–540, 1987

    Ian H Witten, Radford M Neal, and John G Cleary. Arithmetic coding for data compression.Communications of the ACM, 30(6):520–540, 1987. 5

    1987

  49. [49]

    GaussCtrl: Multi-view consistent text-driven 3D Gaussian splatting edit- ing

    Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. GaussCtrl: Multi-view consistent text-driven 3D Gaussian splatting edit- ing. InProceedings of the European Conference on Computer Vision (ECCV), pages 55–71. Springer, 2024. 2

    2024

  50. [50]

    GS-SLAM: Dense visual SLAM with 3D Gaussian splatting

    Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. GS-SLAM: Dense visual SLAM with 3D Gaussian splatting. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pages 19595–19604, 2024. 2

    2024

  51. [51]

    gsplat: An open-source library for Gaussian splatting.Journal of Machine Learning Research (JMLR), 26(34):1–17, 2025

    Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for Gaussian splatting.Journal of Machine Learning Research (JMLR), 26(34):1–17, 2025. 1

    2025

  52. [52]

    Differentiable surface splatting for point-based geometry processing.ACM Transactions on Graphics (ToG), 38(6):1–14, 2019

    Wang Yifan, Felice Serena, Shihao Wu, Cengiz ¨Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing.ACM Transactions on Graphics (ToG), 38(6):1–14, 2019. 4

    2019

  53. [53]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2018. 6

    2018

  54. [54]

    PhysDreamer: Physics-based interaction with 3d objects via video generation

    Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. PhysDreamer: Physics-based interaction with 3d objects via video generation. InProceedings of the European Conference on Computer Vision (ECCV), pages 388–406. Springer, 2024. 2

    2024

  55. [55]

    optimizing-sparsifying

    Yangming Zhang, Wenqi Jia, Wei Niu, and Miao Yin. Gaus- sianSpa: An “optimizing-sparsifying” simplification frame- work for compact and high-quality 3D Gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26673–26682,

  56. [56]

    Tip-editor: An accurate 3D editor follow- ing both text-prompts and image-prompts.ACM Transactions on Graphics (ToG), 43(4):1–12, 2024

    Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. Tip-editor: An accurate 3D editor follow- ing both text-prompts and image-prompts.ACM Transactions on Graphics (ToG), 43(4):1–12, 2024. 2

    2024

  57. [57]

    Driv- able 3D Gaussian avatars

    Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollh¨ofer, Justus Thies, and Javier Romero. Driv- able 3D Gaussian avatars. InInternational Conference on 3D Vision (3DV), pages 979–990. IEEE, 2025. 2

    2025

  58. [58]

    Ewa splatting.IEEE Transactions on Visual- ization and Computer Graphics, 8(3):223–238, 2002

    Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting.IEEE Transactions on Visual- ization and Computer Graphics, 8(3):223–238, 2002. 2 Smol-GS: Compact Representations for Abstract 3D Gaussian Splatting Supplementary Material A. Implementation Details A.1. Hardware and Software Our experiments are conducted on NVIDIA H200 ...

    2002

  59. [59]

    The Gaussian rasterization module is built on top of the Inria 3DGS repository [18]

    with CUDA 12.4. The Gaussian rasterization module is built on top of the Inria 3DGS repository [18]. The arithmetic coding and hash encoding functions modules are built upon the implementations from HAC [4]. A.2. Optimization Details Learning rates and schedulersThe model param- eters are optimized using the Adam optimizer [ 19]. We use an exponential dec...

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.