Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

GSCodec Studio: A Modular Framework for Gaussian Splat Compression

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read GSCodec Studio claims a modular pipeline can combine existing Gaussian Splat compression techniques into codecs that are competitive on static scenes and state-of-the-art on dynamic ones.

desk verdict The framework is genuinely useful and the ablations are honest, but the paper never defines what is inside its reported file sizes, and the dynamic SOTA claim rests on one compressed baseline. read the letter →

arxiv 2506.01822 v1 pith:3CGXATZG submitted 2025-06-02 cs.CV cs.MM

Sicheng Li , Chengzhen Wu , Hao Li , Xiang Gao , Yiyi Liao , Lu Yu This is my paper
classification cs.CVcs.MM
keywords GaussianSplattingcompressionrate-distortiondynamicscenesentropyconstraintmodularframework3DGS4DGS
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

GSCodec Studio proposes that Gaussian Splat compression should be studied inside one modular pipeline that joins scene representation, training-time compression simulation, and post-training coding. The paper builds this framework and then assembles two concrete codecs, Static and Dynamic GSCodec, from components it identifies as best practices. On static benchmarks the resulting codec is competitive with leading baselines, and on the dynamic Neural 3D Video dataset it outperforms the CompactSTG baseline and cuts the bitrate of the underlying STG representation by roughly a factor of ten at similar quality. A central empirical finding is that adding an entropy constraint during training compacts the distribution of splat attributes, which makes subsequent lossless coding markedly more effective; the paper also shows that Gaussian positions, not appearance attributes, are now the dominant share of the compressed bitstream.

What carries the argument

The carrying mechanism is training with compression simulation: the training loop applies differentiable quantization (uniform noise or a straight-through estimator), an entropy constraint backed by a learned probability model (a shared per-channel factorized density model or a per-point Gaussian model with a hash-grid context, following HAC), and adaptive masking that drops unneeded higher-order SH coefficients or static points. This reshapes the splat parameters during optimization so that the later post-training stage—pruning low-opacity points, sorting attributes into 2D grids with a 3D-to-2D mapping, scalar and vector quantization, and lossless coding with image codecs or ANS—can achieve lower rates at equal distortion. The same components are ported to dynamic scenes by using STG's polynomial trajectory representation and time-varying opacity as the backbone, replacing per-frame parameter storage with a shared motion model.

What would settle it

Take the released code, compress the Truck scene from Tanks&Temples with Static GSCodec, and unpack the resulting bitstream; if the reported ~7.1 MB excludes the factorized density model parameters, ANS coder state, or quantization-step side information, the rate-distortion comparison against HAC is unsupported. Alternatively, retrain Dynamic GSCodec at GOF=50 on the Neural 3D Video scenes with a different random seed; if average PSNR at the same bitrate falls more than about 0.3 dB short of CompactSTG+PP, the dynamic state-of-the-art claim is not robust.

Watch

Extended reading notes

Core claim

The paper's central claim is that compression of 3D and 4D Gaussian Splats can be decomposed into reusable stages—preparation, training with compression simulation, post-training compression, and viewing—and that existing specialist methods from the literature map onto this pipeline as drop-in components. Using the framework for controlled comparisons, the authors select uniform-noise differentiable quantization, a factorized density model for the entropy constraint, learnable masking, pruning, 3D-to-2D sorting and image coding, and ANS-based entropy coding, and combine them into Static and Dynamic GSCodec. In their experiments these composite codecs reach rate-distortion quality comparable to HAC, IGS, and SOGS on Tanks&Temples, MipNeRF360, and DeepBlending, while on the Neural 3D Video dataset the dynamic codec outperforms CompactSTG at the same group-of-frames length and represents STG's dynamic scenes at about 5 Mbps (GOF=300) with similar PSNR and SSIM. The paper also claims that the entropy constraint is the load-bearing training-time ingredient: it visibly compacts attribute distributions and smooths sorted 2D attribute maps, improving both learned entropy coding and image coding, and it reports that spatial positions are the main remaining cost at over 40% of the bitstream.

Load-bearing premise

The benchmark comparisons are fair and comprehensive: every baseline is run in its best configuration, and every reported compressed size includes all side information such as entropy-model parameters, masks, and quantization steps.

Editorial extensions

If this is right

  • If the framework is adopted, new GS compression ideas become directly comparable against all representative existing paths in a single controlled codebase, so claims of improvement can be tested rather than asserted.
  • Entropy-constrained training should be treated as a default component in future GS codecs, since the paper shows it improves compressibility of every constrained attribute, not just the ones that are explicitly entropy-coded.
  • The position-data bottleneck (over 40% of the bitstream) redirects research effort toward geometry coding, such as predictive or hierarchical position models, instead of further tuning appearance attributes.
  • The dynamic result—around 5 Mbps for 300-frame segments at quality close to CompactSTG—suggests that streamable free-viewpoint video of real scenes is approaching practical bitrates for distribution.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's ablation shows only a marginal gain from learned entropy coding over PNG-style image coding, so a natural next experiment is to measure the cross-entropy gap between the learned factorized model and the actual quantized symbol frequencies; closing that gap would likely convert the learned coder from a small win into a substantial one.
  • Because the factorized density model outperforms HAC's per-point Gaussian model despite the latter's richer context, the paper's own evidence hints that the rate loss's effect on parameter distribution matters more than the entropy model's capacity; a testable extension is to ablate the hash-grid context while keeping the same rate loss weight.
  • The state-of-the-art claim for dynamic compression rests on a single dataset (Neural 3D Video) with CompactSTG and STG as the only baselines; extending the same Dynamic GSCodec to other dynamic datasets and more 4D baselines would show whether the advantage generalizes.
  • A direct way to test the framework's modularity claim is to swap the STG backbone for an MLP- or triplane-based motion representation and re-run the dynamic ablation; the paper says these are easy to integrate but does not provide those numbers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper presents GSCodec Studio, an open-source modular framework for compressing static and dynamic 3D Gaussian Splatting representations. The framework integrates several reconstruction methods, training-time compression simulations (differentiable quantization, entropy constraints, adaptive masking), and post-training compression tools (pruning, quantization, 3D-to-2D mapping, image coding, learned entropy coding). The authors propose two composed pipelines, Static GSCodec and Dynamic GSCodec, and report rate-distortion comparisons against SOGS, HAC, IGS, gsplat compression on static scenes and against STG and CompactSTG on the Neural 3D Video dataset. The paper also reports ablation studies on compression-simulation modules, post-training modules, and representation choices. The abstract claims competitive rate-distortion performance; the conclusion further claims that Dynamic GSCodec achieves state-of-the-art dynamic compression performance.

Significance. If the quantitative claims are supported, the framework is a useful community resource: it is open-source, modular, and provides controlled comparisons and ablations that are largely internally consistent. The paper ships reproducible code, which is a concrete strength. The rate-distortion measurements and component ablations are the appropriate evidence for the 'competitive' claim. However, the strength of that claim depends on two things that the manuscript does not currently establish: what exactly is included in the reported compressed sizes, and whether the baseline set is sufficiently comprehensive and fairly configured. The dynamic state-of-the-art claim rests on a single compression baseline and is therefore not yet supported. These issues are fixable without changing the framework itself, but they affect the central quantitative conclusions.

major comments (3)
  1. [Section VI-A, Section IV-D, Table V] The rate axis is not defined precisely. The text states only that 'we report the compressed file size' for static scenes and average bitrate for dynamic scenes, but it never specifies whether auxiliary data is included in those numbers. The post-training pipeline of Section IV-D involves a learned factorized density model, quantized attributes, an adaptive binary mask, quantization-step tables, possible VQ codebooks, and sorted 2D-map dimensions; the dynamic codec additionally has per-GOF motion bases/coefficients and static/dynamic masks. All of these must reach the decoder. The quantitative importance of this omission is shown by Table V: replacing learned-distribution entropy coding with plain image coding changes the final size from 7.10 MB to 7.16 MB, a gain of only 0.06 MB. If even a small portion of the auxiliary data is stored outside the reported file, the measured advantage disappears and every rate point in Tables II and III is understated relative to baselines that do include their context-model or codebook overhead. Please state explicitly what is included in the reported file sizes/bitrates, and provide a per-component breakdown, extending Table IX to include all side information.
  2. [Section VII, Section VI-C, Table III] The conclusion's claim that 'Dynamic GSCodec achieves state-of-the-art rate-distortion performance in dynamic GS compression' is not supported by the evidence. Table III compares against only one compression baseline, CompactSTG+PP, plus uncompressed STG. Several dynamic Gaussian compression methods discussed in the related work (e.g., references [52], [54], and [55]) are not included, so the state-of-the-art claim overreaches. The data do support a more modest statement: at GOF=50, the low-rate variant is competitive with CompactSTG, and at GOF=300 it achieves lower rate at similar PSNR. Please either restrict the claim to competitiveness with the tested baselines or add the missing dynamic compression baselines.
  3. [Section VI-D, Section VI-E, Tables IV/V and II/III] The final module choices of Static and Dynamic GSCodec are selected using ablation experiments conducted on the same test benchmarks that are then used for the headline comparisons in Tables II and III. For example, the decision to prefer uniform-noise differentiable quantization over STE, the factorized density model over the per-point Gaussian model, and the learnable mask over gradient thresholding is made from Table IV on Tanks&Temples, and the same dataset appears in the main static comparison of Table II. This is in-sample model selection: the 'best practices' may reflect overfitting to these particular benchmarks rather than general principles. The manuscript should disclose this limitation and ideally validate the selected configuration on a held-out set or external benchmark. This does not invalidate the measured numbers, but it is load-bearing for the claim that the paper identifies general best practices.
minor comments (4)
  1. [Table I] The entries in Table I render as placeholder symbols ('♂lightbulb') instead of check marks, making the table unreadable. Please replace these with proper markers.
  2. [Section IV-D, Variable Bit-width Quantization] The sentence 'Only with 8-bit quantization-aware training, PSNR improves to 24.02dB and 23.65dB, yielding gains of 0.24dB and 0.87dB' is confusing because 24.02 dB and 23.65 dB appear to correspond to 6-bit and 5-bit evaluation, not to 8-bit training alone. Please clarify which bit widths are used in training versus evaluation.
  3. [Section VI-B] The sentence 'Compared to gsplat compression, our method reduces the file size to 8 MB' is imprecise: Table II reports 7.16 MB for Ours and 15.32 MB for gsplat on Tanks&Temples, whereas on DeepBlending the sizes are 4.76 MB and 7.89 MB. Please state the exact dataset and configuration being referenced.
  4. [Table IX] The memory consumption breakdown sums to 6372 KB, while the final sizes in Tables II and V are about 7.1 MB. Please clarify whether Table IX reports a single test scene, whether bytes are measured before or after the final entropy coding stage, and whether side information such as the factorized density model is included.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims rest on measured rate-distortion comparisons against external baselines, not on derivation from its own assumptions.

full rationale

GSCodec Studio is a systems/framework paper whose central claims are empirical: the proposed Static and Dynamic GSCodec pipelines achieve competitive or state-of-the-art rate-distortion performance. These claims are supported by measured file sizes, PSNR/SSIM/LPIPS values, and comparisons to external baselines (HAC, IGS, SOGS, gsplat, STG, CompactSTG) on standard datasets. There is no mathematical derivation chain in which an output is constructed from an input definition, and no fitted parameter is renamed as a prediction. The method choices (noise-based differentiable quantization, factorized density entropy model, learnable mask, pruning, 3D-to-2D mapping, image coding, learned entropy coding) are presented as modular components selected by ablation, and the ablations themselves are measured comparisons rather than circular validations. The only potentially load-bearing assumptions are about benchmark fairness, such as whether reported compressed sizes include all side information (entropy model parameters, masks, codebooks). That is a measurement-validity concern, not a circularity concern: it does not make the conclusion equivalent to its inputs by construction. There is also no load-bearing self-citation: the paper does not invoke a prior uniqueness theorem or an author-defined ansatz to force its design choices. Accordingly, the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities or free parameters beyond standard codec hyperparameters. The listed hand-chosen knobs set the rate-distortion operating points, while the main assumptions are domain-standard: learned entropy models approximate true entropy, compression simulation transfers to test time, and the benchmark baselines are representative.

free parameters (4)
  • rate loss weight lambda = 0.01, 0.006, 0.002 (static); 0.02, 0.01 (dynamic)
    Chosen by hand to produce rate-distortion operating points; not derived or fitted to a target PSNR.
  • quantization bit-width = 8-bit default; 5-bit and 6-bit variants tested
    Selected manually; lower bit-width trades quality for size, as shown in Table V.
  • adaptive mask target ratio = Not reported in the paper
    The learnable mask uses a KL loss toward an unspecified target active ratio; this hyperparameter affects final size and quality.
  • pruning thresholds (opacity and scale) = Not specified numerically
    Post-training pruning relies on opacity and scale thresholds; the paper mentions at least 5% of points are pruned with negligible impact but does not give threshold values.
assumptions (4)
  • standard math A rate-distortion Lagrangian L = D + lambda * R is a valid objective for jointly optimizing distortion and rate.
    Used in the entropy constraint training in Section IV-C; standard in compression literature.
  • domain assumption Compression simulation during training transfers to actual post-training compression, so entropy estimates from learned density models approximate the real bitstream cost.
    Load-bearing for the training-time components; tested indirectly by ablations in Table VI but not proven.
  • domain assumption PSNR, SSIM, and LPIPS on test views are adequate proxies for perceptual quality of compressed Gaussian Splats.
    Standard in the field; the paper relies on these metrics throughout Section VI.
  • domain assumption The chosen baselines (HAC, IGS, SOGS, gsplat, CompactSTG) are representative of the state of the art and are run in their best configurations.
    Underpins the 'competitive' and 'state-of-the-art' claims in Sections VI-B and VI-C; not verified externally in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GSCodec Studio: A Modular Framework for Gaussian Splat Compression." pith.science (2026). https://pith.science/paper/3CGXATZG

@misc{pith2026250601822,
  author       = {Pith},
  title        = {Pith review of: GSCodec Studio: A Modular Framework for Gaussian Splat Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3CGXATZG}},
  note         = {Machine review of arXiv:2506.01822}
}
read the original abstract

3D Gaussian Splatting and its extension to 4D dynamic scenes enable photorealistic, real-time rendering from real-world captures, positioning Gaussian Splats (GS) as a promising format for next-generation immersive media. However, their high storage requirements pose significant challenges for practical use in sharing, transmission, and storage. Despite various studies exploring GS compression from different perspectives, these efforts remain scattered across separate repositories, complicating benchmarking and the integration of best practices. To address this gap, we present GSCodec Studio, a unified and modular framework for GS reconstruction, compression, and rendering. The framework incorporates a diverse set of 3D/4D GS reconstruction methods and GS compression techniques as modular components, facilitating flexible combinations and comprehensive comparisons. By integrating best practices from community research and our own explorations, GSCodec Studio supports the development of compact representation and compression solutions for static and dynamic Gaussian Splats, namely our Static and Dynamic GSCodec, achieving competitive rate-distortion performance in static and dynamic GS compression. The code for our framework is publicly available at https://github.com/JasonLSC/GSCodec_Studio , to advance the research on Gaussian Splats compression.

Figures

Figures reproduced from arXiv: 2506.01822 by the authors.

Figure 1
Figure 1. It incorporates a variety of 3D/4D GS representations [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. GSCodec Studio is a modular framework that unifies various Gaussian Splatting reconstruction and compression techniques to push the frontier of compression efficiency of Gaussian Splatting. GSCodec, achieving advanced rate-distortion performance in 3D/4D Gaussian Splat compression. iii) Our extensive exper￾iments offer valuable insights and highlight key challenges of GS coding. A notable finding is that entropy con… view at source ↗
Figure 2
Figure 2. Abstraction and Code Design of the Framework. The framework consists of four main parts: Preparation (beige), Training with Compression Simulation (pink), Post-Training Compression (blue), and Viewing (green). The preparation stage processes input data and initializes scene representations. The training with compression simulation stage handles the training of Gaussian Splats and introduces compression simulation du… view at source ↗
Figures from the paper (7 more)
Figure 3
Figure 3. Figure 3: Motion Representations of dynamic Gaussians. We implement two representative representations for the time-variant motion Tt i ∈ SE(3) with i denoting the Gaussian index and t the timestamps. Left: Motion parameters are shared over the time dimension for each individual…
Figure 4
Figure 4. Figure 4: Entropy models. We illustrate the probability estimation for one channel across multiple Gaussian Splats. Left: Factorized density model uses a shallow MLP to estimate a non-parametric distribution shared for all Gaussian Splats. Right: Per-point Gaussian distribution …
Figure 5
Figure 5. Figure 5: Real-time 4D Player supports free-viewpoint, free-time, and seamless switching between color/depth modalities for rendering. V. REPRESENTATIVE COMPOSITIONAL METHODS Our modular framework covers almost all the major GS compression methodologies currently available. This…
Figure 6
Figure 6. Figure 6: Rate-distortion comparisons on static datasets. The three rows in this figure, from top to bottom, represent the results on Tanks&Temples, MipNeRF360, and Deep Blending datasets [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Rate-distortion comparisons on the dynamic dataset, Neural 3D Video dataset. GT HAC gsplat comp. Ours [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Qualitative comparisons of dynamic GS compression methods on Neural 3D video dataset. reaches 14 MB—nearly double the final version—though quality improves slightly by 0.39 dB. This highlights its contribution to the final compression ratio. Next, we provide an in-dept…
Figure 10
Figure 10. Figure 10: Visualization on the effectiveness of entropy constraint. As an example, we select the quaternion attribute from the attribute set for visual￾ization. The left column visualizes the parameter distribution of quaternions, while the right column shows the quaternion map…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Efficient Geometry Compression and Communication for 3D Gaussian Splatting Point Clouds

    cs.MM 2025-09 conditional novelty 3.0 of 10

    Integrating AVS PCRM geometry coding into the i3DV Gaussian platform, with Morton-code alignment, saves 10-25% total bitrate without changing rendering quality.

Reference graph

Works this paper leans on

73 extracted references · 51 canonical work pages · cited by 1 Pith paper

  1. [52]

    V3: Viewing volumetric videos on mobiles via streamable 2d dynamic gaussians,

    P. Wang, Z. Zhang, L. Wang, K. Yao, S. Xie, J. Yu, M. Wu, and L. Xu, “V3: Viewing volumetric videos on mobiles via streamable 2d dynamic gaussians,”ACM Transactions on Graphics (TOG), vol. 43, no. 6, p. 13, Dec 2024. [Online]. Available: https://doi.org/10.1145/3687935

    work page doi:10.1145/3687935 2024

  2. [54]

    MEGA: Memory-Efficient 4D Gaussian Splatting for Dynamic Scenes

    X. Zhang, Z. Liu, Y . Zhang, X. Ge, D. He, T. Xu, Y . Wang, Z. Lin, S. Yan, and J. Zhang, “Mega: Memory-efficient 4d gaussian splatting for dynamic scenes,”arXiv preprint arXiv:2410.13613, 2024

    work page Pith review arXiv 2024

  3. [55]

    Modec-gs: Global-to-local motion decomposition and temporal interval adjustment for compact dynamic 3d gaussian splatting,

    S. Kwak, J. Kim, J. Y . Jeong, W.-S. Cheong, J. Oh, and M. Kim, “Modec-gs: Global-to-local motion decomposition and temporal interval adjustment for compact dynamic 3d gaussian splatting,”arXiv preprint arXiv:2501.03714, 2025

    arXiv 2025

  4. [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. on Graphics, vol. 42, no. 4, July 2023. [Online]. Available: https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

    work page 2023

  5. [2]

    Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis,

    J. Luiten, G. Kopanas, B. Leibe, and D. Ramanan, “Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis,” in3DV, 2024

    2024

  6. [3]

    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 Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  7. [4]

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

    Z. Yang, X. Gao, W. Zhou, S. Jiao, Y . Zhang, and X. Jin, “Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction,” Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  8. [5]

    Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting,

    Z. Yang, H. Yang, Z. Pan, and L. Zhang, “Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting,” inProc. of the International Conf. on Learning Representations (ICLR), 2024

    work page 2024

Show all 73 references
  1. [6]

    Spacetime gaussian feature splatting for real-time dynamic view synthesis,

    Z. Li, Z. Chen, Z. Li, and Y . Xu, “Spacetime gaussian feature splatting for real-time dynamic view synthesis,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  2. [7]

    Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting,

    A. Kratimenos, J. Lei, and K. Daniilidis, “Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting,”arXiV, 2023

    2023

  3. [8]

    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 Transactions on Circuits and Systems for Video Technology, 2025

    2025

  4. [9]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction,

    D. Charatan, S. Li, A. Tagliasacchi, and V . Sitzmann, “pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  5. [10]

    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 Conf. on Computer Vision and Pattern Recognition (CVPR), June 2024

    2024

  6. [11]

    3d gaussian splatting as markov chain monte carlo,

    S. Kheradmand, D. Rebain, G. Sharma, W. Sun, J. Tseng, H. Isack, A. Kar, A. Tagliasacchi, and K. M. Yi, “3d gaussian splatting as markov chain monte carlo,”arXiv preprint arXiv:2404.09591, 2024

    2024 arXiv

  7. [12]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images,

    Y . Chen, H. Xu, C. Zheng, B. Zhuang, M. Pollefeys, A. Geiger, T.-J. Cham, and J. Cai, “Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images,”Proc. of the European Conf. on Computer Vision (ECCV), 2024

    2024

  8. [13]

    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,” inACM Trans. on Graphics, 2024

    2024

  9. [14]

    Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians,

    L. Hu, H. Zhang, Y . Zhang, B. Zhou, B. Liu, S. Zhang, and L. Nie, “Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  10. [15]

    Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians,

    S. Qian, T. Kirschstein, L. Schoneveld, D. Davoli, S. Giebenhain, and M. Nießner, “Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians,”Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  11. [16]

    Relightable gaussian codec avatars,

    S. Saito, G. Schwartz, T. Simon, J. Li, and G. Nam, “Relightable gaussian codec avatars,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 130–141

    2024

  12. [17]

    Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real- time human novel view synthesis,

    S. Zheng, B. Zhou, R. Shao, B. Liu, S. Zhang, L. Nie, and Y . Liu, “Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real- time human novel view synthesis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 19 680–19 690

    2024

  13. [18]

    Humanref-gs: Image-to-3d human generation with reference-guided diffusion and 3d gaussian splatting,

    J. Zhang, X. Li, H. Zhong, Q. Zhang, Y . Cao, Y . Shan, and J. Liao, “Humanref-gs: Image-to-3d human generation with reference-guided diffusion and 3d gaussian splatting,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  14. [19]

    Dreamgaussian: Generative gaussian splatting for efficient 3d content creation,

    J. Tang, J. Ren, H. Zhou, Z. Liu, and G. Zeng, “Dreamgaussian: Generative gaussian splatting for efficient 3d content creation,”Proc. of the International Conf. on Learning Representations (ICLR), 2024

    2024

  15. [20]

    Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation,

    Y . Xu, Z. Shi, W. Yifan, H. Chen, C. Yang, S. Peng, Y . Shen, and G. Wetzstein, “Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation,”arXiv preprint arXiv:2403.14621, 2024

    2024 arXiv

  16. [21]

    Gs-slam: Dense visual slam with 3d gaussian splatting,

    C. Yan, D. Qu, D. Xu, B. Zhao, Z. Wang, D. Wang, and X. Li, “Gs-slam: Dense visual slam with 3d gaussian splatting,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  17. [22]

    Splatam: Splat, track & map 3d gaussians for dense rgb-d slam,

    N. Keetha, J. Karhade, K. M. Jatavallabhula, G. Yang, S. Scherer, D. Ramanan, and J. Luiten, “Splatam: Splat, track & map 3d gaussians for dense rgb-d slam,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  18. [23]

    Tele-aloha: A low-budget and high- authenticity telepresence system using sparse rgb cameras,

    H. Tu, R. Shao, X. Dong, S. Zheng, H. Zhang, L. Chen, M. Wang, W. Li, S. Ma, S. Zhanget al., “Tele-aloha: A low-budget and high- authenticity telepresence system using sparse rgb cameras,”ACM Trans. on Graphics, 2024

    2024

  19. [24]

    Streetsurfgs: Scalable urban street surface reconstruction with planar- based gaussian splatting,

    X. Cui, W. Ye, Y . Wang, G. Zhang, W. Zhou, T. He, and H. Li, “Streetsurfgs: Scalable urban street surface reconstruction with planar- based gaussian splatting,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  20. [25]

    Physgaussian: Physics-integrated 3d gaussians for generative dynam- ics,

    T. Xie, Z. Zong, Y . Qiu, X. Li, Y . Feng, Y . Yang, and C. Jiang, “Physgaussian: Physics-integrated 3d gaussians for generative dynam- ics,”Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  21. [26]

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

    Y . Jiang, C. Yu, T. Xie, X. Li, Y . Feng, H. Wang, M. Li, H. Lau, F. Gao, Y . Yang, and C. Jiang, “Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality,”ACM Trans. on Graphics, 2024

    2024

  22. [27]

    Physically embodied gaussian splatting: A realtime correctable world model for robotics,

    J. Abou-Chakra, K. Rana, F. Dayoub, and N. S ¨underhauf, “Physically embodied gaussian splatting: A realtime correctable world model for robotics,” in8th Annual Conference on Robot Learning, 2024

    2024

  23. [28]

    Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis,

    J. Luiten, G. Kopanas, B. Leibe, and D. Ramanan, “Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis,” inProc. of the International Conf. on 3D Vision (3DV), 2024

    2024

  24. [29]

    3dgstream: On- the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos,

    J. Sun, H. Jiao, G. Li, Z. Zhang, L. Zhao, and W. Xing, “3dgstream: On- the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  25. [30]

    Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction,

    Z. Guo, W. Zhou, L. Li, M. Wang, and H. Li, “Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  26. [31]

    Frpgs: Fast, robust, and photorealistic monocular dynamic scene reconstruction with deformable 3d gaussians,

    W. Li, X. Pan, J. Lin, P. Lu, D. Feng, and W. Shi, “Frpgs: Fast, robust, and photorealistic monocular dynamic scene reconstruction with deformable 3d gaussians,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  27. [32]

    Sc- gs: Sparse-controlled gaussian splatting for editable dynamic scenes,

    Y .-H. Huang, Y .-T. Sun, Z. Yang, X. Lyu, Y .-P. Cao, and X. Qi, “Sc- gs: Sparse-controlled gaussian splatting for editable dynamic scenes,” in Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

    2024

  28. [33]

    Shape of motion: 4d reconstruction from a single video,

    Q. Wang, V . Ye, H. Gao, J. Austin, Z. Li, and A. Kanazawa, “Shape of motion: 4d reconstruction from a single video,” 2024

    2024

  29. [34]

    4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes,

    Y . Duan, F. Wei, Q. Dai, Y . He, W. Chen, and B. Chen, “4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes,” inACM Trans. on Graphics, 2024, pp. 1–11

    2024

  30. [35]

    Scaffold- gs: Structured 3d gaussians for view-adaptive rendering,

    T. Lu, M. Yu, L. Xu, Y . Xiangli, L. Wang, D. Lin, and B. Dai, “Scaffold- gs: Structured 3d gaussians for view-adaptive rendering,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  31. [36]

    Sags: Structure-aware 3d gaussian splatting,

    E. Ververas, R. A. Potamias, J. Song, J. Deng, and S. Zafeiriou, “Sags: Structure-aware 3d gaussian splatting,”arXiv preprint arXiv:2404.19149, 2024

    2024 arXiv

  32. [37]

    Reducing the memory footprint of 3d gaussian splatting,

    P. Papantonakis, G. Kopanas, B. Kerbl, A. Lanvin, and G. Drettakis, “Reducing the memory footprint of 3d gaussian splatting,”Proceedings of the ACM on Computer Graphics and Interactive Techniques, vol. 7, no. 1, pp. 1–17, 2024

    2024

  33. [38]

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

    Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang, “Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps,” arXiv preprint arXiv:2311.17245, 2023

    2023 arXiv

  34. [39]

    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 Conf. on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 21 719–21 728

    2024

  35. [40]

    Compact 3d gaussian splatting for static and dynamic radiance fields,

    J. C. Lee, D. Rho, X. Sun, J. H. Ko, and E. B. Park, “Compact 3d gaussian splatting for static and dynamic radiance fields,”arXiv preprint arXiv:2408.03822, 2024

    2024 arXiv

  36. [41]

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

    G. Fang and B. Wang, “Mini-splatting: Representing scenes with a constrained number of gaussians,”arXiv preprint arXiv:2403.14166, 2024

    2024 arXiv

  37. [42]

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

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

    2024

  38. [43]

    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,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  39. [44]

    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,”Proc. of the European Conf. on Computer Vision (ECCV), 2024

    2024

  40. [45]

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

    W. Morgenstern, F. Barthel, A. Hilsmann, and P. Eisert, “Compact 3d scene representation via self-organizing gaussian grids,”Proc. of the European Conf. on Computer Vision (ECCV), 2024

    2024

  41. [46]

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

    S. Xie, W. Zhang, C. Tang, Y . Bai, R. Lu, S. Ge, and Z. Wang, “Mesongs: Post-training compression of 3d gaussians via efficient attribute trans- formation,” inEuropean Conference on Computer Vision. Springer, 2025, pp. 434–452

    2025

  42. [47]

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

    Y . Wang, Z. Li, L. Guo, W. Yang, A. C. Kot, and B. Wen, “Contextgs: Compact 3d gaussian splatting with anchor level context model,”arXiv preprint arXiv:2405.20721, 2024

    2024 arXiv

  43. [48]

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

    H. Wang, H. Zhu, T. He, R. Feng, J. Deng, J. Bian, and Z. Chen, “End- to-end rate-distortion optimized 3d gaussian representation,”Proc. of the European Conf. on Computer Vision (ECCV), 2024

    2024

  44. [49]

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

    X. Liu, X. Wu, P. Zhang, S. Wang, Z. Li, and S. Kwong, “Compgs: Efficient 3d scene representation via compressed gaussian splatting,” in Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 2936–2944

    2024

  45. [50]

    Hifi4g: High-fidelity human performance rendering via compact gaussian splatting,

    Y . Jiang, Z. Shen, P. Wang, Z. Su, Y . Hong, Y . Zhang, J. Yu, and L. Xu, “Hifi4g: High-fidelity human performance rendering via compact gaussian splatting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 19 734–19 745

    2024

  46. [51]

    Robust dual gaussian splatting for immersive human-centric volumetric videos,

    Y . Jiang, Z. Shen, Y . Hong, C. Guo, Y . Wu, Y . Zhang, J. Yu, and L. Xu, “Robust dual gaussian splatting for immersive human-centric volumetric videos,”ACM Transactions on Graphics (TOG), vol. 43, no. 6, p. 15, Dec 2024. [Online]. Available: https://doi.org/10.1145/3687926

    2024 doi

  47. [53]

    Per-gaussian embedding-based deformation for deformable 3d gaussian splatting,

    J. Bae, S. Kim, Y . Yun, H. Lee, G. Bang, and Y . Uh, “Per-gaussian embedding-based deformation for deformable 3d gaussian splatting,” arXiv preprint arXiv:2404.03613, 2024

    2024 arXiv

  48. [56]

    Queen: Quantized efficient encoding of dynamic gaussians for stream- ing free-viewpoint videos,

    S. Girish, T. Li, A. Mazumdar, A. Shrivastava, S. De Melloet al., “Queen: Quantized efficient encoding of dynamic gaussians for stream- ing free-viewpoint videos,”Advances in Neural Information Processing Systems (NeurIPS), vol. 37, pp. 43 435–43 467, 2024

    2024

  49. [57]

    Evolvinggs: High-fidelity streamable volumetric video via evolving 3d gaussian representation,

    C. Zhang, Y . Zhou, S. Wang, W. Li, D. Wang, Y . Xu, and S. Jiao, “Evolvinggs: High-fidelity streamable volumetric video via evolving 3d gaussian representation,”arXiv preprint arXiv:2503.05162, 2025

    2025 arXiv

  50. [58]

    gsplat: An open-source library for Gaussian splatting,

    V . Ye, R. Li, J. Kerr, M. Turkulainen, B. Yi, Z. Pan, O. Seiskari, J. Ye, J. Hu, M. Tancik, and A. Kanazawa, “gsplat: An open-source library for Gaussian splatting,”arXiv preprint arXiv:2409.06765, 2024. [Online]. Available: https://arxiv.org/abs/2409.06765

    2024 arXiv

  51. [59]

    Gaustudio: A modular framework for 3d gaussian splatting and beyond,

    C. Ye, Y . Nie, J. Chang, Y . Chen, Y . Zhi, and X. Han, “Gaustudio: A modular framework for 3d gaussian splatting and beyond,”arXiv preprint arXiv:2403.19632, 2024

    2024 arXiv

  52. [60]

    Easyvolcap: Accelerating neural volumetric video research,

    Z. Xu, T. Xie, S. Peng, H. Lin, Q. Shuai, Z. Yu, G. He, J. Sun, H. Bao, and X. Zhou, “Easyvolcap: Accelerating neural volumetric video research,” inSIGGRAPH Asia 2023 Technical Communications, 2023, pp. 1–4

    2023

  53. [61]

    Nerfstudio: A modular framework for neural radiance field develop- ment,

    M. Tancik, E. Weber, E. Ng, R. Li, B. Yi, J. Kerr, T. Wang, A. Kristof- fersen, J. Austin, K. Salahi, A. Ahuja, D. McAllister, and A. Kanazawa, “Nerfstudio: A modular framework for neural radiance field develop- ment,”arXiv preprint arXiv:2302.04264, 2023

    2023 arXiv

  54. [62]

    Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures,

    Z. Chen, T. Funkhouser, P. Hedman, and A. Tagliasacchi, “Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures,”arXiv.org, 2022

    2022

  55. [63]

    A white paper on neural network quantization,

    M. Nagel, M. Fournarakis, R. A. Amjad, Y . Bondarenko, M. Van Baalen, and T. Blankevoort, “A white paper on neural network quantization,” arXiv preprint arXiv:2106.08295, 2021

    2021 arXiv

  56. [64]

    Variational image compression with a scale hyperprior,

    J. Ball ´e, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,” inProc. of the International Conf. on Learning Representations (ICLR), 2018

    2018

  57. [65]

    Compressai: a pytorch library and evaluation platform for end-to-end compression research,

    J. B ´egaint, F. Racap ´e, S. Feltman, and A. Pushparaja, “Compressai: a pytorch library and evaluation platform for end-to-end compression research,”arXiv preprint arXiv:2011.03029, 2020

    2011 arXiv

  58. [66]

    The jpeg ai standard: Providing efficient human and machine visual data consumption,

    J. Ascenso, E. Alshina, and T. Ebrahimi, “The jpeg ai standard: Providing efficient human and machine visual data consumption,”IEEE Multimedia, vol. 30, no. 1, pp. 100–111, 2023

    2023

  59. [67]

    Pranckevi ˇcius

    A. Pranckevi ˇcius. (2023) Making gaussian splats more smaller. [Online]. Available: https://aras-p.info/blog/2023/09/27/ Making-Gaussian-Splats-more-smaller/

    2023

  60. [68]

    (2022) Torchpq

    DeMoriarty. (2022) Torchpq. [Online]. Available: https://github.com/ DeMoriarty/TorchPQ

    2022

  61. [69]

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

    M. Wu and T. Tuytelaars, “Implicit gaussian splatting with efficient multi-level tri-plane representation,”arXiv preprint arXiv:2408.10041, 2024

    2024 arXiv

  62. [70]

    Understanding entropy coding with asymmetric nu- meral systems (ans): a statistician’s perspective,

    R. Bamler, “Understanding entropy coding with asymmetric nu- meral systems (ans): a statistician’s perspective,”arXiv preprint arXiv:2201.01741, 2022

    2022 arXiv

  63. [71]

    Asymmetric numeral systems,

    J. Duda, “Asymmetric numeral systems,”arXiv preprint arXiv:0902.0271, 2009

    2009 arXiv

  64. [72]

    Learned image compres- sion with discretized gaussian mixture likelihoods and attention mod- ules,

    Z. Cheng, H. Sun, M. Takeuchi, and J. Katto, “Learned image compres- sion with discretized gaussian mixture likelihoods and attention mod- ules,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 7939–7948

    2020

  65. [73]

    “gsplat,” https://github.com/nerfstudio-project/gsplat

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.