Pith. sign in

REVIEW 3 major objections 4 minor 71 references

Training 3D scenes on compressed Gaussians cuts peak memory 3.5x

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 ↗

ContraGS trains 3D Gaussian Splatting directly on codebook-compressed representations, cutting peak model memory ~3.5x with small quality loss.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A genuinely useful memory-compression trick for 3DGS training, but the MCMC/Bayesian justification does not survive contact with its own acceptance equations. the 3 major comments →

arxiv 2509.03775 v1 pith:IW4QUTVP submitted 2025-09-03 cs.GR cs.CV

ContraGS: Codebook-Condensed and Trainable Gaussian Splatting for Fast, Memory-Efficient Reconstruction

Sankeerth Durvasula , Sharanshangar Muhunthan , Zain Moustafa , Richard Chen , Ruofan Liang , Yushi Guan , Nilesh Ahuja , Nilesh Jain
show 2 more authors
Selvakumar Panneer Nandita Vijaykumar
This is my paper
classification cs.GR cs.CV
keywords 3D Gaussian Splattingcodebook compressionBayesian inferenceMarkov Chain Monte CarloMetropolis-Hastingsmemory-efficient trainingscene reconstructionradiance fields
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.

The reading

The paper claims that 3D Gaussian Splatting can be trained directly on a codebook-compressed representation instead of being compressed after training. It turns parameter estimation into Bayesian inference and uses Metropolis-Hastings sampling with update, split, and merge moves, which lets the optimizer handle the discrete Gaussian-to-codebook mappings that gradient descent cannot. In experiments, this reduces peak training memory by about 3.5x on average, speeds up training by 1.36x and rendering by 1.88x, and keeps reconstruction quality close to state-of-art 3DGS-MCMC. The practical point is to remove the memory wall that otherwise forces either fewer Gaussians or larger GPUs.

Core claim

ContraGS treats a compressed 3DGS scene as a state S={G,C}: each Gaussian stores only its position, opacity, and integer pointers into two codebooks—one for spherical-harmonic color coefficients (SH) and one for covariance scale and quaternion (SR). The posterior over states is p(G,C) ∝ exp(−L_recon − λ_sr|SR| − λ_sh|SH|), making codebook size a learned regularizer. Because the integer pointers are not differentiable, SGLD cannot learn them, so ContraGS uses Metropolis-Hastings with three proposal moves: a differentiable SGLD parameter update with 98% probability, a split move with 1%, and a merge move with 1%. The split and merge acceptance probabilities have closed forms driven by the code

What carries the argument

The engine is a codebook-compressed state with two codebooks, SH for spherical-harmonic colors and SR for scale+quaternion, plus a Metropolis-Hastings sampler whose proposal distribution is a mixture of 0.98 parameter update (SGLD), 0.01 split, and 0.01 merge. Split and merge moves change the discrete Gaussian-to-codebook mapping, and their acceptance probabilities are derived as min(1, e^{±λ} qsm(u)), where λ is the codebook-size penalty and qsm is a ratio of the split and merge proposal noises. This mechanism is what makes the otherwise non-differentiable codebook indices learnable.

Load-bearing premise

The sampling chain's validity rests on the assumption that splitting or merging a codebook vector leaves the reconstruction loss essentially unchanged, so the acceptance rule can ignore L_recon and depend only on the codebook-size penalty; the paper states this but does not prove or test it.

What would settle it

Instrument a ContraGS training run on a standard scene: for every proposed split or merge, compute the full posterior ratio including L_recon and compare it with the simplified e^{±λ} approximation used in Eqs. 19–20; if accepted moves systematically change PSNR or the average discrepancy is large, the chain is not sampling the stated posterior. A second check is to run with λ=0 and see whether the codebook collapses to a single vector.

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

If this is right

  • For a fixed memory budget, ContraGS claims the highest representation quality among compared training-time compression methods: 2M-Gaussian ContraGS peaks near 130 MB versus 473 MB for 3DGS-MCMC-2M, while 5M ContraGS uses 276–444 MB versus 1182 MB.
  • ContraGS-2M stays within about 0.3 PSNR of MCMC-2M on average, so users can keep high Gaussian counts without paying the full memory cost.
  • Training and rendering become faster (1.36x and 1.88x on average), with rendering FPS nearly doubling in reported scenes.
  • Unlike several prior efficient-training approaches, ContraGS trains from a random point cloud initialization and does not require an SfM point cloud.
  • The same codebook-plus-MCMC formulation is claimed to extend to other point-based reconstruction methods, including RadiantFoam, ADOP, Deformable Beta Splatting, and LinPrim.

Where Pith is reading between the lines

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

  • The simplified split/merge acceptance rules assume that the reconstruction loss is essentially unchanged by structural moves; measuring the true posterior ratio during training would show how far the implemented chain drifts from the stated Bayesian target.
  • Because codebook vectors act as learned prototypes that split and merge over time, the same machinery could serve as a principled densification and pruning strategy for other 3D representation formats.
  • The compressed state could be pushed further with quantized codebook vectors or entropy-coded indices, compounding the memory savings without changing the MH formulation.
  • For dynamic or 4D scenes, where memory pressure is even higher, the method's ability to keep training memory near the codebook size may matter more than the modest quality gap reported here.
Share X Bluesky LinkedIn Reddit HN

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 / 4 minor

Summary. The paper proposes ContraGS, a codebook-compressed training method for 3D Gaussian Splatting. Instead of storing per-Gaussian parameters, ContraGS stores shared codebook vectors for spherical-harmonic and covariance quantities, with each Gaussian assigned integer indices. The authors argue that the discrete index assignments cannot be learned by SGD and instead formulate codebook training as Bayesian inference, using Metropolis–Hastings MCMC with split, merge, and SGLD-like parameter-update proposals. The paper reports reduced peak model memory (on average 3.49×), faster training (1.36×), and faster rendering (1.88×) with quality close to 3DGS-MCMC, and positions the method as the first to train directly on a compressed 3DGS representation.

Significance. If the Bayesian/MCMC formulation were correct, the paper would be a significant step toward memory-efficient 3DGS training on resource-constrained hardware. The empirical memory and speed gains are substantial, and the paper includes extensive comparisons across MipNeRF360, Deep Blending, Tanks and Temples, and Blender. However, the central theoretical claim—that ContraGS samples from the stated posterior via Metropolis–Hastings—is undermined by specific errors in the acceptance-probability derivation, and the empirical evaluation lacks statistical grounding. The practical results may still stand as a heuristic compression method, but the paper's stated contribution is not supported as written.

major comments (3)
  1. [Appendix A.1.2–A.1.3, Eqs. (19)–(20), (32)–(35)] The split and merge acceptance probabilities replace the posterior ratio p(S')/p(S) by e^{−λ Δ|C|}, omitting the reconstruction-loss ratio. The exact posterior (Eq. 12) gives p(S')/p(S) = exp[−(L_recon(S')−L_recon(S)) − λ Δ|C|]. A split changes the codebook vector for at least one Gaussian (c'' = c+u), and a merge changes the parameters of every Gaussian mapped to the child vector, so L_recon is not invariant under these transitions. The sentence in A.1.3 that merging 'does not affect the overall accuracy' is precisely the unverified premise. No proof, bound, or experiment is offered to justify dropping ΔL_recon. Consequently, the chain does not target the posterior of Eq. 12, and the Bayesian-inference claim for non-differentiable indices collapses.
  2. [Eqs. (21) and (36)] The Gaussian proposal-density ratio q_sm is defined only by the exponential factor exp(±||u||^2/2(...)), omitting the normalization constant (ϵ_split/ϵ_merge)^d from the ratio of multivariate normal densities. Since ϵ_split=0.1 and ϵ_merge=0.05, this omitted factor is 2^d (or its reciprocal), which is not negligible for the SH and SR vector dimensions. Thus, even under the loss-invariance assumption of the previous comment, the acceptance probabilities in Eqs. (19)–(20) and (32)–(35) are not the correct Metropolis–Hastings ratios, and the stationary distribution is consequently altered.
  3. [§5.1–§5.2, Tables 1–3] The quantitative claims (e.g., 'less than 0.3 PSNR drop' and the 3.49× average memory reduction) are based on a single training run per scene with no multiple seeds, error bars, or statistical tests. Because ContraGS uses random initialization and stochastic MCMC proposals, the reported averages may be unstable. The empirical comparison with baselines requires at least a small number of seeds with reported mean and variance to support the claimed advantages.
minor comments (4)
  1. [§4.2 and §5.1] Notation is inconsistent: Eq. (13) uses λ_sr and λ_sh, while Section 5.1 and later text use λ_SR and λ_SH. Please unify.
  2. [§5.2] The abstract states a 3.49× average memory reduction; Section 5.2 reports '3.78× and 3.5× on average respectively' for the 2M and 5M configurations. Clarify the relationship and which number is the headline.
  3. [Table 1] The header says 'Peak Mem' without units; the text states MB, but it should be stated in the table caption or header.
  4. [§4.3] The claim that sequences of split and merge operations generate all possible codebook-compressed representations is asserted without proof. This is plausible but should be argued more explicitly, especially regarding reachability of arbitrary index assignments.

Circularity Check

1 steps flagged

Split/merge acceptance probabilities drop the reconstruction-loss ratio, so the 'Bayesian posterior sampling' claim reduces to a hand-set codebook-size penalty.

specific steps
  1. other [Appendix A.1.2-A.1.3 (Eqs. 31-35); main text Eq. 19-20]
    "The ratio p(Ssplit)/p(S) ≈ e−λSH if we choose to split the SH codebook, and e−λSR if we choose to split the SR codebook. We thus have the acceptance probability given by: A(S → Ssplit) = min(1, e−λSH/qsm(u)) ... p(S)/p(Smerge) ≈ eλSH if it leads to a reduction in the number of rows, or 1 otherwise, as merging a small set of rows of codebook vectors does not affect the overall accuracy."

    By the paper's own posterior (Eq. 12), p(S')/p(S) = exp(−[L_recon(S')−L_recon(S)] − λΔ|C|). The appendix replaces this with exp(−λΔ|C|) by asserting that the reconstruction-loss difference is negligible. Thus the MH acceptance probabilities (Eqs. 19/20, 32/35) are not derived from the stated posterior; they are constructed to be exactly the hand-specified codebook-size penalty ratio. The claimed result that MCMC 'samples over a posterior distribution of compressed representations' therefore reduces to enforcing the λ_SH/λ_SR penalties the authors chose, with the L_recon term that defines the posterior playing no role in structural moves. The final codebook size is consequently the direct output of the user-set hyperparameters rather than an emergent Bayesian estimate.

full rationale

The empirical memory/quality/speed comparisons (Tables 1-3) are independently measured against 3DGS-MCMC and other baselines, so the engineering claims are not circular. However, the paper's central methodological claim—posing codebook learning as Bayesian inference over Eq. 12—is partially circular/unsupported: the split/merge acceptance derivation omits the ΔL_recon term, so the chain does not target Eq. 12 and the compression incentive is effectively the hand-tuned λ penalty. This is a load-bearing approximation presented as a derivation; it should be scored as partial circularity rather than a mere missing proof. There is no problematic self-citation chain; the flawed step is internal to the paper's equations.

Axiom & Free-Parameter Ledger

6 free parameters · 3 axioms · 0 invented entities

The central claims rest on several hand-tuned hyperparameters (lambda, epsilon, mixture weights) and on an unproven approximation in the acceptance derivation. No new physical entities are introduced; the codebooks are standard compression structures.

free parameters (6)
  • lambda_SH = 2.3
    Penalty weight on SH codebook size in the posterior L (Eq. 13), hand-tuned to balance quality and compression.
  • lambda_SR = 3
    Penalty weight on SR codebook size in the posterior L (Eq. 13), hand-tuned.
  • epsilon_split = 0.1
    Noise scale for the split proposal (Eq. 17), hand-tuned.
  • epsilon_merge = 0.05
    Noise scale for the merge proposal (Eq. 18), hand-tuned.
  • Proposal mixture probabilities = 0.98 update, 0.01 split, 0.01 merge
    Mixing probabilities in Eq. 14, hand-tuned.
  • SGLD step sizes epsilon_p, epsilon_o, epsilon_SH, epsilon_SR
    Step sizes in Eqs. 25-28 are not given numerical values in the paper; they are inherited or tuned from 3DGS-MCMC but unspecified, making the method partially un-reproducible.
axioms (3)
  • ad hoc to paper The posterior ratio for split/merge transitions is approximately e^{-lambda * Delta(|codebook|)}, ignoring the reconstruction-loss change.
    Used in Appendix A.1.2 and A.1.3 to derive Eqs. 19-20. It assumes structural moves do not affect reconstruction quality, which is not proven and is required for a valid MH acceptance.
  • domain assumption Split/merge transitions are ergodic over the state space of codebook-compressed models.
    Stated in Section 4.3 without proof. The merge transition only rejoins a vector to its previously split parent, so reachable mappings form a hierarchical tree, not the full space of arbitrary codebook assignments.
  • standard math SGLD update with acceptance A=1 is a valid approximation of Metropolis-Hastings.
    Inherited from 3DGS-MCMC [26]; this is a known approximation, but it applies only to the continuous parameter update step, not to the discrete split/merge moves.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ContraGS: Codebook-Condensed and Trainable Gaussian Splatting for Fast, Memory-Efficient Reconstruction." pith.science (2026). https://pith.science/paper/IW4QUTVP

@misc{pith2026250903775,
  author       = {Pith},
  title        = {Pith review of: ContraGS: Codebook-Condensed and Trainable Gaussian Splatting for Fast, Memory-Efficient Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IW4QUTVP}},
  note         = {Machine review of arXiv:2509.03775}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

3D Gaussian Splatting (3DGS) is a state-of-art technique to model real-world scenes with high quality and real-time rendering. Typically, a higher quality representation can be achieved by using a large number of 3D Gaussians. However, using large 3D Gaussian counts significantly increases the GPU device memory for storing model parameters. A large model thus requires powerful GPUs with high memory capacities for training and has slower training/rendering latencies due to the inefficiencies of memory access and data movement. In this work, we introduce ContraGS, a method to enable training directly on compressed 3DGS representations without reducing the Gaussian Counts, and thus with a little loss in model quality. ContraGS leverages codebooks to compactly store a set of Gaussian parameter vectors throughout the training process, thereby significantly reducing memory consumption. While codebooks have been demonstrated to be highly effective at compressing fully trained 3DGS models, directly training using codebook representations is an unsolved challenge. ContraGS solves the problem of learning non-differentiable parameters in codebook-compressed representations by posing parameter estimation as a Bayesian inference problem. To this end, ContraGS provides a framework that effectively uses MCMC sampling to sample over a posterior distribution of these compressed representations. With ContraGS, we demonstrate that ContraGS significantly reduces the peak memory during training (on average 3.49X) and accelerated training and rendering (1.36X and 1.88X on average, respectively), while retraining close to state-of-art quality.

Figures

Figures reproduced from arXiv: 2509.03775 by Nandita Vijaykumar, Nilesh Ahuja, Nilesh Jain, Richard Chen, Ruofan Liang, Sankeerth Durvasula, Selvakumar Panneer, Sharanshangar Muhunthan, Yushi Guan, Zain Moustafa.

Figure 1
Figure 1. Figure 1: ContraGS: A 3D scene is modeled using 3D Gaussian Splatting [ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Impact of number of Gaussians on representation quality [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Gaussians mapped to one vector in a codebook of quater [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Codebook layout used for ContraGS. ContraGS al￾locates two codebooks: (1) SH stores spherical harmonics coef￾ficientes of RGB colors, and (2) SR stores scaling + quaternion parameters concatenated together as vectors. Each Gaussian has g2sh, g2sr to index the two codebooks. 4.2. Formulating The Posterior Distribution We define a posterior probability distribution over the Con￾traGS model state space as fol… view at source ↗
Figure 5
Figure 5. Figure 5: ContraGS performs one of the 3 steps for each 3DG: (1) [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Split and Merge transitions. We now derive the proposal and acceptance distributions. 4.4. Proposal and Acceptance Distributions We define the proposal distribution for MH sampling tran￾sition from state as: q(S → S ′ ) = 0.98qupdate + 0.01qsplit + 0.01qmerge (14) Where qupdate is the parameter update transition, qsplit is the split transition and qmerge is the merge transition. In other words, at each ste… view at source ↗
Figure 7
Figure 7. Figure 7: The impact of Gaussian count for 3DGS-MCMC and ContraGS on training quality (PSNR), model memory, and training time. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative results of ContraGS compared to ground truth reconstruction [PITH_FULL_IMAGE:figures/full_fig_p016_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

71 extracted references · 42 canonical work pages · 6 internal anchors

  1. [1]

    Physically embodied gaussian splatting: A re- altime correctable world model for robotics

    Jad Abou-Chakra, Krishan Rana, Feras Dayoub, and Niko S¨underhauf. Physically embodied gaussian splatting: A re- altime correctable world model for robotics. arXiv preprint arXiv:2406.10788, 2024. 2

    Pith/arXiv arXiv 2024

  2. [2]

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

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 7

    work page 2022

  3. [3]

    Dge: Di- rect gaussian 3d editing by consistent multi-view editing

    Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Di- rect gaussian 3d editing by consistent multi-view editing. In European Conference on Computer Vision, pages 74–92. Springer, 2024. 2

    work page 2024

  4. [4]

    Gaussianeditor: Swift and control- lable 3d editing with gaussian splatting

    Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xi- aofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and control- lable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21476–21485, 2024. 2

    work page 2024

  5. [5]

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

    Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. Hac: Hash-grid assisted context for 3d gaussian splatting compression. In European Conference on Computer Vision, pages 422–438. Springer, 2024. 4

    work page 2024

  6. [6]

    Hac++: Towards 100x compression of 3d gaussian splatting

    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

    Pith/arXiv arXiv 2025

  7. [7]

    Compact 3d gaussian splatting for dense visual slam

    Tianchen Deng, Yaohui Chen, Leyan Zhang, Jianfei Yang, Shenghai Yuan, Jiuming Liu, Danwei Wang, Hesheng Wang, and Weidong Chen. Compact 3d gaussian splatting for dense visual slam. arXiv preprint arXiv:2403.11247, 2024. 2, 4

    Pith/arXiv arXiv 2024

  8. [8]

    Distwar: Fast differentiable rendering on raster-based ren- dering pipelines

    Sankeerth Durvasula, Adrian Zhao, Fan Chen, Ruofan Liang, Pawan Kumar Sanjaya, and Nandita Vijaykumar. Distwar: Fast differentiable rendering on raster-based ren- dering pipelines. arXiv preprint arXiv:2401.05345, 2023. 5

    Pith/arXiv arXiv 2023

  9. [9]

    Arc: Warp-level adaptive atomic reduction in gpus to accelerate differentiable render- ing

    Sankeerth Durvasula, Adrian Zhao, Fan Chen, Ruofan Liang, Pawan Kumar Sanjaya, Yushi Guan, Christina Gian- noula, and Nandita Vijaykumar. Arc: Warp-level adaptive atomic reduction in gpus to accelerate differentiable render- ing. In Proceedings of the 30th ACM International Confer- ence on Architectural Support for Programming Languages and Operating Syst...

    work page 2025

  10. [10]

    Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, Zhangyang Wang, et al. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. Advances in neural information processing systems , 37: 140138–140158, 2025. 2, 4, 5

    work page 2025

  11. [11]

    Mini-splatting: Repre- senting scenes with a constrained number of gaussians

    Guangchi Fang and Bing Wang. Mini-splatting: Repre- senting scenes with a constrained number of gaussians. In European Conference on Computer Vision, pages 165–181. Springer, 2024. 4

    work page 2024

  12. [12]

    Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering

    Guofeng Feng, Siyan Chen, Rong Fu, Zimu Liao, Yi Wang, Tao Liu, Zhilin Pei, Hengjie Li, Xingcheng Zhang, and Bo Dai. Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering. arXiv preprint arXiv:2408.07967, 2024. 5

    Pith/arXiv arXiv 2024

  13. [13]

    Gaussianflow: Splatting gaussian dynamics for 4d content creation

    Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wen- chao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splatting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. 2

    Pith/arXiv arXiv 2024

  14. [14]

    Cat3d: Create anything in 3d with multi-view diffusion models

    Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2

    Pith/arXiv arXiv 2024

  15. [15]

    Ea- gles: Efficient accelerated 3d gaussians with lightweight en- codings

    Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Ea- gles: Efficient accelerated 3d gaussians with lightweight en- codings. In European Conference on Computer Vision, pages 54–71. Springer, 2024. 2, 4, 7

    work page 2024

  16. [16]

    Radiant foam: Real-time differen- tiable ray tracing

    Shrisudhan Govindarajan, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasacchi. Radiant foam: Real-time differen- tiable ray tracing. arXiv preprint arXiv:2502.01157 , 2025. 3

    Pith/arXiv arXiv 2025

  17. [17]

    Speedy-splat: Fast 3d gaus- sian splatting with sparse pixels and sparse primitives

    Alex Hanson, Allen Tu, Geng Lin, Vasu Singla, Matthias Zwicker, and Tom Goldstein. Speedy-splat: Fast 3d gaus- sian splatting with sparse pixels and sparse primitives. arXiv preprint arXiv:2412.00578, 2024. 2, 4, 7

    arXiv 2024

  18. [18]

    Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splatting

    Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Goldstein. Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splatting. arXiv preprint arXiv:2406.10219, 2024. 2, 4

    arXiv 2024

  19. [19]

    W. K. Hastings. Monte carlo sampling methods using markov chains and their applications. Biometrika, 57(1):97– 109, 1970. 4

    work page 1970

  20. [20]

    3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt

    Lukas H ¨ollein, Alja ˇz Bo ˇziˇc, Michael Zollh ¨ofer, and Matthias Nießner. 3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiv preprint arXiv:2409.12892, 2024. 5

    Pith/arXiv arXiv 2024

  21. [21]

    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. In European Confer- ence on Computer Vision, pages 93–112. Springer, 2024. 2

    work page 2024

  22. [22]

    SplatMAP: Online Dense Monocular SLAM with 3D Gaussian Splatting

    Yue Hu, Rong Liu, Meida Chen, Andrew Feng, and Peter Beerel. Splatmap: Online dense monocular slam with 3d gaussian splatting. arXiv preprint arXiv:2501.07015, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    Brightdreamer: Generic 3d gaussian genera- tive framework for fast text-to-3d synthesis

    Lutao Jiang, Xu Zheng, Yuanhuiyi Lyu, Jiazhou Zhou, and Lin Wang. Brightdreamer: Generic 3d gaussian genera- tive framework for fast text-to-3d synthesis. arXiv preprint arXiv:2403.11273, 2024. 2

    Pith/arXiv arXiv 2024

  24. [24]

    Deformable 3d gaussian splatting for animat- able human avatars

    HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez- Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Ben- jamin Busam. Deformable 3d gaussian splatting for animat- able human avatars. arXiv preprint arXiv:2312.15059, 2023. 2

    Pith/arXiv arXiv 2023

  25. [25]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1,

  26. [26]

    3d gaussian splatting as markov chain monte carlo

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Wei- wei Sun, Jeff Tseng, Hossam Isack, Abhishek Kar, An- drea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. arXiv preprint arXiv:2404.09591, 2024. 2, 3, 5, 7

    Pith/arXiv arXiv 2024

  27. [27]

    Tanks and temples: Benchmarking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 2, 7

    work page 2017

  28. [28]

    Hugs: Human gaussian splats

    Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 505–515, 2024. 2

    work page 2024

  29. [29]

    3DGS$^2$: Near Second-order Converging 3D Gaussian Splatting

    Lei Lan, Tianjia Shao, Zixuan Lu, Yu Zhang, Chenfanfu Jiang, and Yin Yang. 3dgs 2: Near second-order converg- ing 3d gaussian splatting. arXiv preprint arXiv:2501.13975,

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    Gscore: Efficient radiance field rendering via architectural support for 3d gaussian splatting

    Junseo Lee, Seokwon Lee, Jungi Lee, Junyong Park, and Jae- woong Sim. Gscore: Efficient radiance field rendering via architectural support for 3d gaussian splatting. In Proceed- ings of the 29th ACM International Conference on Architec- tural Support for Programming Languages and Operating Systems, Volume 3, pages 497–511, 2024. 5

    work page 2024

  31. [31]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719– 21728, 2024. 2, 4, 5

    work page 2024

  32. [32]

    Robogsim: A real2sim2real robotic gaussian splatting simulator

    Xinhai Li, Jialin Li, Ziheng Zhang, Rui Zhang, Fan Jia, Tiancai Wang, Haoqiang Fan, Kuo-Kun Tseng, and Ruiping Wang. Robogsim: A real2sim2real robotic gaussian splatting simulator. arXiv preprint arXiv:2411.11839, 2024. 2

    Pith/arXiv arXiv 2024

  33. [33]

    Rtgs: Enabling real- time gaussian splatting on mobile devices using efficiency- guided pruning and foveated rendering

    Weikai Lin, Yu Feng, and Yuhao Zhu. Rtgs: Enabling real- time gaussian splatting on mobile devices using efficiency- guided pruning and foveated rendering. arXiv preprint arXiv:2407.00435, 2024. 5

    Pith/arXiv arXiv 2024

  34. [34]

    Atomgs: Atomizing gaussian splatting for high-fidelity radi- ance field

    Rong Liu, Rui Xu, Yue Hu, Meida Chen, and Andrew Feng. Atomgs: Atomizing gaussian splatting for high-fidelity radi- ance field. arXiv preprint arXiv:2405.12369, 2024. 2

    Pith/arXiv arXiv 2024

  35. [35]

    Deformable Beta Splatting

    Rong Liu, Dylan Sun, Meida Chen, Yue Wang, and An- drew Feng. Deformable beta splatting. arXiv preprint arXiv:2501.18630, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  36. [36]

    Turbo-GS: Accelerating 3D Gaussian Fitting for High-Quality Radiance Fields

    Tao Lu, Ankit Dhiman, R Srinath, Emre Arslan, Angela Xing, Yuanbo Xiangli, R Venkatesh Babu, and Srinath Srid- har. Turbo-gs: Accelerating 3d gaussian fitting for high- quality radiance fields. arXiv preprint arXiv:2412.13547 ,

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654–20664, 2024. 4, 7

    work page 2024

  38. [38]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654–20664, 2024. 2

    2024

  39. [39]

    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. In SIGGRAPH Asia 2024 Con- ference Papers, pages 1–11, 2024. 4, 5, 7

    work page 2024

  40. [40]

    Splatfields: Neural gaussian splats for sparse 3d and 4d re- construction

    Marko Mihajlovic, Sergey Prokudin, Siyu Tang, Robert Maier, Federica Bogo, Tony Tung, and Edmond Boyer. Splatfields: Neural gaussian splats for sparse 3d and 4d re- construction. In European Conference on Computer Vision, pages 313–332. Springer, 2024. 2

    work page 2024

  41. [41]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65(1):99–106, 2021. 7

    work page 2021

  42. [42]

    Expressive whole-body 3d gaussian avatar

    Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3d gaussian avatar. In European Conference on Computer Vision , pages 19–35. Springer,

    work page

  43. [43]

    Human gaussian splatting: Real-time rendering of animatable avatars

    Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo P ´erez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 788–798, 2024. 2

    work page 2024

  44. [44]

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

    Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto- saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splat- ting for robust real-time rendering with 900+ fps. arXiv preprint arXiv:2403.13806, 2024. 4

    Pith/arXiv arXiv 2024

  45. [45]

    Gsedit: Efficient text-guided edit- ing of 3d objects via gaussian splatting

    Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodol `a. Gsedit: Efficient text-guided edit- ing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2

    Pith/arXiv arXiv 2024

  46. [46]

    Reducing the memory footprint of 3d gaussian splatting

    Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Tech- niques, 7(1):1–17, 2024. 4, 7

    work page 2024

  47. [47]

    C3dgs: Compressing 3d gaussian model for surface reconstruction of large-scale scenes based on multi-view uav images

    Jiating Qian, Yiming Yan, Fengjiao Gao, Baoyu Ge, Maosheng Wei, Boyi Shangguan, and Guangjun He. C3dgs: Compressing 3d gaussian model for surface reconstruction of large-scale scenes based on multi-view uav images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2025. 2, 3, 4, 5

    work page 2025

  48. [48]

    3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting

    Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5020–5030, 2024. 2

    work page 2024

  49. [49]

    Drag your gaussian: Effective drag-based editing with score distillation for 3d gaussian splatting

    Yansong Qu, Dian Chen, Xinyang Li, Xiaofan Li, Shengchuan Zhang, Liujuan Cao, and Rongrong Ji. Drag your gaussian: Effective drag-based editing with score distillation for 3d gaussian splatting. arXiv preprint arXiv:2501.18672, 2025. 2

    Pith/arXiv arXiv 2025

  50. [50]

    Adop: Approximate differentiable one-pixel point rendering

    Darius R ¨uckert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1–14, 2022. 3

    work page 2022

  51. [51]

    Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting

    Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1606–1616, 2024. 2

    work page 2024

  52. [52]

    Haha: Highly articulated gaussian human avatars with textured mesh prior

    David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. Haha: Highly articulated gaussian human avatars with textured mesh prior. In Proceedings of the Asian Con- ference on Computer Vision, pages 4051–4068, 2024. 2

    work page 2024

  53. [53]

    Dreamgaussian: Generative gaussian splatting for effi- cient 3d content creation

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for effi- cient 3d content creation. arXiv preprint arXiv:2309.16653,

    Pith/arXiv arXiv

  54. [54]

    Linprim: Lin- ear primitives for differentiable volumetric rendering

    Nicolas von L ¨utzow and Matthias Nießner. Linprim: Lin- ear primitives for differentiable volumetric rendering. arXiv preprint arXiv:2501.16312, 2025. 3

    work page arXiv 2025

  55. [55]

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

    Henan Wang, Hanxin Zhu, Tianyu He, Runsen Feng, Jia- jun Deng, Jiang Bian, and Zhibo Chen. End-to-end rate- distortion optimized 3d gaussian representation. InEuropean Conference on Computer Vision , pages 76–92. Springer,

    work page

  56. [56]

    4d gaussian splatting for real-time dynamic scene rendering

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 20310–20320, 2024. 2

    2024

  57. [57]

    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. In European Conference on Computer Vision, pages 55–

    work page

  58. [58]

    Physgaussian: Physics- integrated 3d gaussians for generative dynamics

    Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics- integrated 3d gaussians for generative dynamics. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389–4398, 2024. 2

    work page 2024

  59. [59]

    Gaussianproperty: Integrat- ing physical properties to 3d gaussians with lmms

    Xinli Xu, Wenhang Ge, Dicong Qiu, ZhiFei Chen, Dongyu Yan, Zhuoyun Liu, Haoyu Zhao, Hanfeng Zhao, Shunsi Zhang, Junwei Liang, et al. Gaussianproperty: Integrat- ing physical properties to 3d gaussians with lmms. arXiv preprint arXiv:2412.11258, 2024. 2

    Pith/arXiv arXiv 2024

  60. [60]

    DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization

    Yueming Xu, Haochen Jiang, Zhongyang Xiao, Jianfeng Feng, and Li Zhang. Dg-slam: Robust dynamic gaussian splatting slam with hybrid pose optimization. arXiv preprint arXiv:2411.08373, 2024. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  61. [61]

    FGS-SLAM: Fourier-based Gaussian Splatting for Real-time SLAM with Sparse and Dense Map Fusion

    Yansong Xu, Junlin Li, Wei Zhang, Siyu Chen, Shengyong Zhang, Yuquan Leng, and Weijia Zhou. Fgs-slam: Fourier- based gaussian splatting for real-time slam with sparse and dense map fusion. arXiv preprint arXiv:2503.01109, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  62. [62]

    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. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 19595–19604, 2024. 2

    work page 2024

  63. [63]

    3dsce- needitor: Controllable 3d scene editing with gaussian splat- ting

    Ziyang Yan, Lei Li, Yihua Shao, Siyu Chen, Wuzong Kai, Jenq-Neng Hwang, Hao Zhao, and Fabio Remondino. 3dsce- needitor: Controllable 3d scene editing with gaussian splat- ting. arXiv preprint arXiv:2412.01583, 2024. 2

    arXiv 2024

  64. [64]

    Deformable 3d gaussians for high- fidelity monocular dynamic scene reconstruction

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high- fidelity monocular dynamic scene reconstruction. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20331–20341, 2024. 2

    work page 2024

  65. [65]

    Gaussian grouping: Segment and edit anything in 3d scenes

    Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In European Conference on Computer Vision , pages 162–

  66. [66]

    Gavatar: Animatable 3d gaussian avatars with implicit mesh learning

    Ye Yuan, Xueting Li, Yangyi Huang, Shalini De Mello, Koki Nagano, Jan Kautz, and Umar Iqbal. Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 896–905, 2024. 2

    work page 2024

  67. [67]

    Gaussian in the wild: 3d gaussian splatting for unconstrained image collections

    Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. In European Conference on Computer Vision, pages 341–359. Springer, 2024. 2

    2024

  68. [68]

    Gs-lrm: Large recon- struction model for 3d gaussian splatting

    Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large recon- struction model for 3d gaussian splatting. In European Con- ference on Computer Vision, pages 1–19. Springer, 2024. 2

    work page 2024

  69. [69]

    Gaussianimage: 1000 fps image representation and compres- sion by 2d gaussian splatting

    Xinjie Zhang, Xingtong Ge, Tongda Xu, Dailan He, Yan Wang, Hongwei Qin, Guo Lu, Jing Geng, and Jun Zhang. Gaussianimage: 1000 fps image representation and compres- sion by 2d gaussian splatting. In European Conference on Computer Vision, 2024. 4

    work page 2024

  70. [70]

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

    Zhaoliang Zhang, Tianchen Song, Yongjae Lee, Li Yang, Cheng Peng, Rama Chellappa, and Deliang Fan. Lp-3dgs: Learning to prune 3d gaussian splatting. Advances in Neural Information Processing Systems, 37:122434–122457, 2025. 4

    work page 2025

  71. [71]

    Generative object insertion in gaussian splat- ting with a multi-view diffusion model

    Hongliang Zhong, Can Wang, Jingbo Zhang, and Jing Liao. Generative object insertion in gaussian splat- ting with a multi-view diffusion model. arXiv preprint arXiv:2409.16938, 2024. 2 A. Acceptance Probability Derivation A.1. Proposal and Acceptance Probabilities To determine whether a split, merge or parameter update step proposed is accepted to be taken...

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.