Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

FlexGaussian: Flexible and Cost-Effective Training-Free Compression for 3D Gaussian Splatting

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read FlexGaussian claims a training-free compression pipeline can shrink 3D Gaussian scenes by up to 96.4% with under 1 dB PSNR loss, in under a minute, and with 1-2 second reconfiguration when the target changes.

desk verdict Solid engineering paper with a real contribution, but the headline '<1 dB PSNR drop' is an in-sample number because FOA selects parameters on the same views used for reporting; needs a validation split before the claim is credible. read the letter →

arxiv 2507.06671 v1 pith:YOJSDPXM submitted 2025-07-09 cs.CV

Boyuan Tian , Qizhe Gao , Siran Xianyu , Xiaotong Cui , Minjia Zhang This is my paper
classification cs.CV
keywords 3DGaussiansplattingtraining-freecompressionmixed-precisionquantizationattribute-discriminativepruningnovelviewsynthesismobiledeploymentonlineadaptationrate-distortiontrade-off
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

FlexGaussian aims to show that a pre-trained 3D Gaussian Splatting model can be compressed by up to 96.4% with under 1 dB PSNR drop, without any retraining or fine-tuning, in under a minute. The method combines two cheap operations: pruning along both Gaussian and attribute dimensions, and mixed-precision quantization that gives more bits to sensitive attribute channels, then searches a small set of configurations to meet a target quality or file size. The payoff is flexibility: the same pipeline can be rerun in 1-2 seconds when the compression target changes, and it fits on a mobile GPU where training-involved methods cannot run. The paper positions training-free Gaussian compression as a reconfigurable post-processing step rather than a one-shot optimization.

What carries the argument

Three mechanisms work together. Channel-wise Mixed-Precision Quantization (MPQ) assigns each Gaussian attribute channel either INT4 or INT8 bits instead of one uniform bit-width, and splits each channel into sub-groups with their own quantization ranges, because attribute channels have very different value ranges. Attribute-Discriminative Pruning (ADP) ranks Gaussians by a global significance score, roughly opacity times volume times how often each Gaussian is hit by training rays, and then prunes either whole low-scoring Gaussians or only the view-dependent color (SH_adv) attributes of moderately important ones, so compression happens along both the Gaussian axis and the attribute axis without training a mask. Fast Online Adaptation (FOA) searches over a small candidate set of row-prune ratio, SH-prune ratio, and bit-width pairs, relying on the paper's empirical finding that quality loss versus pruning ratio stays linearly correlated across scenes and that the Pareto-optimal frontier shares the same parameter pairs; each search step prunes, quantizes, dequantizes, renders, and evaluates quality in about 1.5 seconds.

What would settle it

Take a held-out set of scenes not in the Mip-NeRF360, Tanks-and-Temples, or Deep Blending suites, enumerate all ADP and MPQ configurations, and check whether the configuration that achieves the under-1 dB PSNR constraint at maximum compression lies inside the small candidate set that FOA searches; if a substantial fraction of scenes require configurations outside that set, the cross-scene stability claim and the reported 1-2 second adaptation time would not generalize.

Watch

Extended reading notes

Core claim

The central claim is that attribute-channel sensitivity, not per-Gaussian sensitivity alone, is the right handle for training-free 3D Gaussian compression. The paper observes that geometric channels (position, rotation, scale) and texture channels (spherical-harmonic colors) have very different value ranges and quantization sensitivity, and that directly quantizing all channels to INT4 destroys quality. FlexGaussian therefore applies INT4 to insensitive channels and INT8 to sensitive ones, with subchannel groups sharing quantization ranges, and separately prunes whole Gaussians or just the view-dependent color attributes of less important Gaussians. A fast online adaptation procedure then evaluates a limited set of row-prune, attribute-prune, and bit-width combinations and selects one that meets a user-specified constraint, all without training. On the reported datasets, this yields 94.9-96.4% size reduction with under 1 dB PSNR drop, with compression times under 30 seconds on a desktop GPU and within minutes on a mobile platform.

Load-bearing premise

The argument depends on the empirical claim that the quality impact of mixed-precision quantization does not vary much between scenes and that the same small set of pruning-and-quantization choices is near-optimal for every scene; if a new scene needs a configuration outside that set, the seconds-scale search will miss it.

Editorial extensions

If this is right

  • Any pre-trained 3D Gaussian model can be compressed immediately without access to the original training pipeline, though computing the importance score still needs training-view camera poses.
  • A single saved model can serve many deployment targets: rerun FOA with a different quality or size constraint and get a new compression-quality point in about 1-2 seconds.
  • Because no optimizer state or training images are held in GPU memory, peak memory during compression stays close to model size, which is what makes mobile deployment possible.
  • If the cross-scene stability of the Pareto frontier holds, the same small candidate set can be ported to new scenes with only a few seconds of search overhead.
  • The reported rate-distortion span from 8x to 256x compression makes the method a candidate for adaptive streaming or storage scaling without retraining.

Reading between the lines

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

  • The method's reliance on a fixed global-significance score means it inherits any blind spots of that score; the paper's comparison with an alternative score suggests the role of the score is compatible but not fully characterized.
  • Because MPQ and ADP operate non-iteratively on the attribute matrix, the same pipeline could plausibly extend to other explicit scene representations with column-wise attributes, such as 2D Gaussian splatting variants, provided a meaningful per-attribute importance can be defined.
  • The claimed 1-2 second reconfiguration implies a practical adaptive-streaming use case: a sender could recompress a scene on the fly as network bandwidth changes, something training-involved methods cannot do.
  • A testable extension is to replace the fixed candidate set with a lightweight predictor of the optimal configuration from cheap per-scene statistics, which would remove the reliance on cross-scene stability entirely.
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 / 5 minor

Summary. The paper proposes FlexGaussian, a training-free compression method for 3D Gaussian splatting that combines channel-wise mixed-precision quantization (MPQ), attribute-discriminative pruning (ADP), and a fast online adaptation (FOA) module that searches over compression parameters. The authors report up to 96.4% model size reduction with less than 1 dB PSNR loss, compression times under 30 seconds on a desktop and minutes on a Jetson Xavier, and faster reconfiguration than the training-free baseline FCGS. The method is evaluated on Mip-NeRF360, Tanks&Temples, and Deep Blending, with ablations for each component and a large-scale study using Grendel-GS.

Significance. If the reported numbers hold under a proper evaluation protocol, FlexGaussian would be a practical contribution: it is the first training-free 3DGS compression method demonstrated on a mobile-class device, and its per-adjustment cost of 1-2 seconds is a genuine advantage for adaptive compression. The paper includes extensive experiments, ablations that isolate the contributions of ADP and MPQ, and a large-scale study with Grendel-GS. However, the evaluation-protocol issue described below prevents the headline results from being accepted at face value, and the cross-scene invariance assumption needs stronger support.

major comments (3)
  1. [Section 4.2.3 / Section 5.2] FOA selects compression parameters by evaluating PSNR on the same 'complete evaluation sets' that are later used for the reported quality numbers (Section 5.2 states: 'Only complete evaluation sets are needed for quality loss calculation'). Because the final per-scene PSNR values in Tables 1 and 3 are the result of selecting (Row-P, SH-P, bit-width) on those same evaluation views, the claimed '<1 dB PSNR drop' is an in-sample optimization score, not an unbiased estimate on unseen views. This also makes the comparison with baselines (which do not tune on the test views) unfair. Please either restrict the FOA search to training views and report on held-out test views, or introduce an explicit validation split and describe the protocol in the paper.
  2. [Section 4.2.3] The assertion that 'the Pareto-optimal frontier shares the same set of parameter pairs' across scenes is based on extending Fig. 6 to only three additional scenes. If this invariance fails for a new scene, FOA would need to enlarge its candidate set, which would break the claimed '1-2 second adjustment' and 'under 30 seconds' compression times. The paper should report the candidate-set size used per scene, the runtime dependence on candidate-set size, and ideally verify the invariance across all 13 test scenes rather than a subset.
  3. [Table 1 / Section 5.2 (Comparison to FCGS)] The main comparison with FCGS uses FCGS-Opt, a customized variant, because FCGS-Raw runs out of memory on 7 of 13 scenes. The paper should explicitly verify on the scenes where FCGS-Raw does run that FCGS-Opt reproduces the same PSNR, size, and time (Table 3 provides some data but does not reconcile the differences, e.g., the Bicycle row appears inconsistent). It should also state whether any of the FCGS-Opt results in Table 1 differ from FCGS-Raw on those scenes, and argue that the customization does not disadvantage the baseline relative to FlexGaussian.
minor comments (5)
  1. [Section 4.2.2] The definitions of Row-P and SH-P are ambiguous: the text first defines 𝛼% as the retained top fraction and 𝛽% as the discarded bottom fraction, but then says Row-P is 𝛼% and SH-P is 1-𝛼%-𝛽%, which would make SH-P the partially retained fraction rather than a pruning ratio. Please clarify the exact meaning of both parameters and how they map to the axes in Fig. 6.
  2. [Equation (1)] The summation index 'MHW' is unclear; please write it as a triple sum over training views, image height, and width, or define M, H, W explicitly.
  3. [Table 3] The FCGS-Raw row for Bicycle appears to show '1𝑒−4' in the PSNR column, which is likely a formatting error (possibly intended as OOM or a numerical value); please check the table rendering.
  4. [Section 5.1 / Table 1] The table caption refers to colors (yellow, gray, pink, green, cyan, orange) for method categories and ranking; if the paper is read in grayscale, the legend is lost. Please add explicit textual labels or a legend.
  5. [Abstract / Section 5.3] The claim of mobile deployability is based on a single platform (Jetson Xavier); consider adding a second mobile-class device or tempering the claim to 'Jetson-class platforms'.

Circularity Check

1 steps flagged · score 6.0 of 10

FOA selects compression parameters using the same evaluation views whose PSNR is later reported, so the claimed '<1 dB drop' is enforced by the search's own stopping rule rather than independently measured.

  1. fitted input called prediction [Section 4.2.3 (FOA); Section 5.2 (Main Results)]
    "At each search step, FOA duplicates the input Gaussians, then prunes, quantizes, dequantizes, and evaluates quality using standard metrics like PSNR, repeating until the best compression parameters are found. ... Only complete evaluation sets are needed for quality loss calculation, while training camera poses suffice for importance computation. ... FlexGaussian reduces 94.9%, 96.1%, and 96.4% of the data across three datasets, with quality losses of 0.8 dB, 0.7 dB, and 0.8 dB, respectively — all well below the 1 dB PSNR drop constraint."

    The reported PSNR drop is computed on the same 'complete evaluation sets' that FOA uses to select Row-P, SH-P, and bit-widths. Because the search's stopping rule is to keep configurations whose PSNR loss satisfies the user-specified <1 dB constraint, the headline quality loss is, by construction, an in-sample optimized value rather than an independent estimate on unseen views. The 'maintaining high rendering quality' component of the central claim therefore reduces to the FOA objective itself; the 96.4% compression ratio is a real search outcome, but the <1 dB quality guarantee is not independently verified.

full rationale

The paper does not derive its rate-distortion behavior from first principles; it is an engineering search method. The only circularity-like step is the evaluation-protocol leak: FOA optimizes PSNR on the test views and then reports PSNR on those same views, so the quality component of the headline is statistically forced. There is no load-bearing self-citation (the importance score is taken from LightGaussian/MesonGS, which are external works), no imported uniqueness theorem, and no ansatz smuggled in by citation. Speed, memory, and mobile-deployment comparisons against FCGS, LightGaussian, Compressed3D, and CompGS provide independent grounding for the efficiency claims, though the quality comparison is contaminated by the selection-on-test-set issue. Hence a moderate score of 6: one central quantitative claim reduces by construction to the search objective, while the rest of the contribution remains substantially empirical and externally compared.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central performance claim rests on empirically calibrated components: per-channel bit-width assignment chosen from sensitivity measurements, sub-group quantization sizes chosen by hand, and per-scene pruning ratios selected by FOA to meet the quality constraint. It also assumes that attribute sensitivity and the Pareto-optimal parameter set generalize across scenes, and that the borrowed LightGaussian importance score remains valid without retraining. No new physical entities are introduced.

free parameters (4)
  • Channel-wise bit-width assignment (INT4/INT8 per attribute channel) = Not disclosed; chosen from sensitivity analysis in Fig. 4
    The set of bit-widths per attribute channel is selected by measuring PSNR drop when a channel is quantized to INT4, then reused across scenes. This is a tuned configuration of the method.
  • Sub-group size for grouped quantization = 1000 sub-groups (example in Section 4.2.1)
    The paper buckets sequential attributes into 1000 sub-groups per channel; the number is given as an example and affects quantization accuracy.
  • ADP pruning ratios (Row-P alpha, SH-P beta) = Per-scene values found by FOA; not listed in the paper
    The fraction of Gaussians pruned entirely and the fraction of SH_adv attributes pruned from remaining Gaussians are searched per scene to meet quality and compression targets.
  • FOA candidate parameter set and step sizes = Not specified
    The limited set of candidate (Row-P, SH-P, bit-width) configurations and the search step size determine speed and optimality; the paper does not enumerate them.
assumptions (5)
  • domain assumption Attribute channels have different quantization sensitivity, and this sensitivity is similar across scenes.
    Motivates sharing one channel-wise bit-width assignment across scenes; asserted from Fig. 4 and Section 4.2.1.
  • domain assumption The Pareto-optimal frontier for ADP parameters is shared across scenes and convex.
    FOA relies on this to search a small candidate set monotonically; stated in Section 4.2.3.
  • domain assumption PSNR/SSIM/LPIPS computed on rendered images is a valid proxy for compression quality, and FOA can use rendered reference images from the uncompressed model.
    Used throughout evaluation; FOA evaluates parameter settings by PSNR on rendered views (Section 4.2.3, Section 5.1).
  • domain assumption LightGaussian's global significance score remains a valid importance metric for pruning without retraining.
    ADP sorts Gaussians by this score (Eq. 1, Section 4.2.2); if the score is miscalibrated after pruning, quality would drop.
  • domain assumption Sequential sub-group bucketing for quantization ignores attribute correlation without hurting quality significantly.
    The paper states the method does not consider correlation between channels but finds it robust (Section 4.2.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlexGaussian: Flexible and Cost-Effective Training-Free Compression for 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/YOJSDPXM

@misc{pith2026250706671,
  author       = {Pith},
  title        = {Pith review of: FlexGaussian: Flexible and Cost-Effective Training-Free Compression for 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YOJSDPXM}},
  note         = {Machine review of arXiv:2507.06671}
}
read the original abstract

3D Gaussian splatting has become a prominent technique for representing and rendering complex 3D scenes, due to its high fidelity and speed advantages. However, the growing demand for large-scale models calls for effective compression to reduce memory and computation costs, especially on mobile and edge devices with limited resources. Existing compression methods effectively reduce 3D Gaussian parameters but often require extensive retraining or fine-tuning, lacking flexibility under varying compression constraints. In this paper, we introduce FlexGaussian, a flexible and cost-effective method that combines mixed-precision quantization with attribute-discriminative pruning for training-free 3D Gaussian compression. FlexGaussian eliminates the need for retraining and adapts easily to diverse compression targets. Evaluation results show that FlexGaussian achieves up to 96.4% compression while maintaining high rendering quality (<1 dB drop in PSNR), and is deployable on mobile devices. FlexGaussian delivers high compression ratios within seconds, being 1.7-2.1x faster than state-of-the-art training-free methods and 10-100x faster than training-involved approaches. The code is being prepared and will be released soon at: https://github.com/Supercomputing-System-AI-Lab/FlexGaussian

Figures

Figures reproduced from arXiv: 2507.06671 by the authors.

Figure 1
Figure 1. The trend of training 3D-GS for larger, more detailed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Left: Quality of quantized 3D-GS with varying attribute [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of FlexGaussian. FlexGaussian first applies attribute-discriminative pruning (ADP) to obtain semi-structurally [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Attribute channels ex￾hibit divergent impact to ren￾dering quality, evidencing the need for mixed-precision quantization that accounts for attribute sensitivity [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 10
Figure 10. Figure 10: Time breakdown on desktop and mobile devices. Gaussian-intensive ops, like Imp. Score Calc., are more im￾pacted by computing power. 8 16 32 64 128 256 Compression Ratio (X) 16 18 20 22 24 26 PSNR (dB) Baseline [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]

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. AtlasLC: Fast Codec-Ready Compression of Object-Centric 3D Gaussian Splatting

    cs.GR 2026-07 conditional novelty 6.0 of 10

    A training-free pipeline prunes object-centric 3D Gaussian splats by local competition and packs them into deterministic codec-ready atlases, cutting preparation time and payload with modest quality loss.

Reference graph

Works this paper leans on

48 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Muhammad Salman Ali, Maryam Qamar, Sung-Ho Bae, and Enzo Tartaglione

    work page

  2. [2]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. 2022. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 . IEEE, 5460–5469

    work page 2022

  3. [3]

    Yu Chen and Gim Hee Lee. 2025. DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus. Advances in Neural Information Processing Systems 37 (2025), 34487–34512

    work page 2025

  4. [4]

    Yihang Chen, Qianyi Wu, Mengyao Li, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. 2024. Fast feedforward 3d gaussian splatting compression. arXiv preprint arXiv:2410.08017 (2024)

    arXiv 2024

  5. [5]

    Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. 2025. Hac: Hash-grid assisted context for 3d gaussian splatting compression. In Euro- pean Conference on Computer Vision . Springer, 422–438

    work page 2025

  6. [6]

    Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer

  7. [7]

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. 2023. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. arXiv preprint arXiv:2311.17245 (2023)

    arXiv 2023

  8. [8]

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2022. GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers. CoRR abs/2210.17323 (2022)

    arXiv 2022

Show all 48 references
  1. [9]

    Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. 2023. Eagles: Effi- cient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564 (2023)

    2023 arXiv

  2. [10]

    Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Goldstein. 2024. PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting. CoRR abs/2406.10219 (2024)

    2024

  3. [11]

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. 2018. Deep blending for free-viewpoint image-based ren- dering. ACM Transactions on Graphics (ToG) 37, 6 (2018), 1–15

    2018

  4. [12]

    Liang Hou, Zehuan Yuan, Lei Huang, Huawei Shen, Xueqi Cheng, and Changhu Wang. 2021. Slimmable generative adversarial networks. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 35. 7746–7753

    2021

  5. [13]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis

  6. [14]

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. 2017. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36, 4 (2017), 1–13

    2017

  7. [15]

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park

  8. [16]

    Bingling Li, Shengyi Chen, Luchao Wang, Kaimin Liao, Sijie Yan, and Yuanjun Xiong. 2024. Retinags: Scalable training for dense scene rendering with billion- scale 3d gaussians. arXiv preprint arXiv:2406.11836 (2024)

    2024 arXiv

  9. [17]

    Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, and Xiao- jun Chang. 2021. Dynamic slimmable network. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition . 8607–8617

    2021

  10. [18]

    Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. 2023. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 3205–3215

    2023

  11. [19]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Compact 3d gaussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 21719–21728

  12. [20]

    Weikai Lin, Yu Feng, and Yuhao Zhu. 2025. MetaSapiens: Real-Time Neural Rendering with Efficiency-Aware Pruning and Accelerated Foveated Rendering. In Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems,...

    2025

  13. [21]

    Tao Lu, Ankit Dhiman, R Srinath, Emre Arslan, Angela Xing, Yuanbo Xiangli, R Venkatesh Babu, and Srinath Sridhar. 2024. Turbo-GS: Accelerating 3D Gaussian Fitting for High-Quality Radiance Fields. arXiv preprint arXiv:2412.13547 (2024)

    2024 arXiv

  14. [22]

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. 2024. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20654–20664

    2024

  15. [23]

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. 2024. AWQ: Activation-aware Weight Quantization for On-Device LLM Compression and Acceleration. Proceedings of Machine Learning and Systems 6 (2...

    2024

  16. [24]

    Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. 2022. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph. 41, 4, Article 102 (July 2022), 15 pages. doi:10.1145/3528223.3530127

    2022

  17. [25]

    KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. 2023. Compact3d: Compressing gaussian splat radiance field models with vector quantization. arXiv preprint arXiv:2311.18159 (2023)

    2023 arXiv

  18. [26]

    KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. 2024. Compgs: Smaller and faster gaussian splatting with vector quantization. In European Conference on Computer Vision

    2024

  19. [27]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Commun. ACM 65, 1 (2021), 99–106. ACM MM, October 27–31, 2025, Dublin, Ireland Boyuan Tian, Qizhe...

    2021

  20. [28]

    Nvidia. 2018. Jetson AGX Xavier Series, https://tinyurl.com/jetsonxavier. https://www.nvidia.com/en-sg/autonomous- machines/embedded-systems/jetson-agx-xavier/

    2018

  21. [29]

    Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. 2024. Reducing the Memory Footprint of 3D Gaussian Splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques 7, 1 (2024), 1–17

    2024

  22. [30]

    Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. 2022. Compressible- composable nerf via rank-residual decomposition.Advances in Neural Information Processing Systems 35 (2022), 14798–14809

    2022

  23. [31]

    Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. 2024. Com- pressed 3D Gaussian Splatting for Accelerated Novel View Synthesis. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 . IEEE, 10349– 10358

    2024

  24. [32]

    Henan Wang, Hanxin Zhu, Tianyu He, Runsen Feng, Jiajun Deng, Jiang Bian, and Zhibo Chen. 2024. End-to-End Rate-Distortion Optimized 3D Gaussian Representation. arXiv preprint arXiv:2406.01597 (2024)

    2024 arXiv

  25. [33]

    Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. 2019. Haq: Hardware- aware automated quantization with mixed precision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8612–8620

    2019

  26. [34]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing 13, 4 (2004), 600–612

    2004

  27. [35]

    Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. 2022. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 12922–12931

    2022

  28. [36]

    Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. 2023. Smoothquant: Accurate and efficient post-training quantization for large language models. In International Conference on Machine Learning . PMLR, 38087–38099

    2023

  29. [37]

    Shuzhao Xie, Weixiang Zhang, Chen Tang, Yunpeng Bai, Rongwei Lu, Shijia Ge, and Zhi Wang. 2025. MesonGS: Post-training Compression of 3D Gaussians via Efficient Attribute Transformation. In European Conference on Computer Vision . Springer, 434–452

    2025

  30. [38]

    Jiahui Yu and Thomas S Huang. 2019. Universally slimmable networks and improved training techniques. In Proceedings of the IEEE/CVF international con- ference on computer vision . 1803–1811

    2019

  31. [39]

    Bichen Wu, Yanghan Wang, Peizhao Zhang, Yuandong Tian, Peter Vajda, and Kurt Keutzer. 2018. Mixed precision quantization of convnets via differentiable neural architecture search. arXiv preprint arXiv:1812.00090 (2018)

    2018 arXiv

  32. [40]

    Shiran Yuan and Hao Zhao. 2024. SlimmeRF: Slimmable Radiance Fields. In 2024 International Conference on 3D Vision (3DV) . IEEE, 64–74

    2024

  33. [41]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  34. [42]

    Zhaoliang Zhang, Tianchen Song, Yongjae Lee, Li Yang, Cheng Peng, Rama Chellappa, and Deliang Fan. 2024. LP-3DGS: Learning to Prune 3D Gaussian Splatting. arXiv preprint arXiv:2405.18784 (2024)

    2024 arXiv

  35. [43]

    Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. 2018. Slimmable neural networks. arXiv preprint arXiv:1812.08928 (2018)

    2018 arXiv

  36. [48]

    Hexu Zhao, Haoyang Weng, Daohan Lu, Ang Li, Jinyang Li, Aurojit Panda, and Saining Xie. 2024. On scaling up 3d gaussian splatting training. arXiv preprint arXiv:2406.18533 (2024)

    2024

  37. [2018]

    In Proceedings of the IEEE conference on computer vision and pattern recognition

    The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition . 586–595

  38. [2019]

    In Proceedings of the IEEE/CVF international conference on computer vision

    Hawq: Hessian aware quantization of neural networks with mixed- precision. In Proceedings of the IEEE/CVF international conference on computer vision. 293–302

  39. [2023]

    ACM Trans

    3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Trans. Graph. 42, 4 (2023), 139–1

    2023

  40. [2024]

    arXiv preprint arXiv:2406.18214 (2024)

    Trimming the fat: Efficient compression of 3d gaussian splats through pruning. arXiv preprint arXiv:2406.18214 (2024)

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.