Pith. sign in

REVIEW 5 major objections 7 minor 52 references

FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting

T0 review · 5 major / 7 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 FlexGS trains a 3D Gaussian splatting scene once, then at inference selects and reshapes a subset of Gaussians to hit any requested model size, without fine-tuning for each ratio.

desk verdict Solid practical idea with one overreaching claim: arbitrary-ratio elasticity is plausible but only tested inside the training span. read the letter →

arxiv 2506.04174 v1 pith:CSZNTHLN submitted 2025-06-04 cs.CV

Hengyu Liu , Yuehao Wang , Chenxin Li , Ruisi Cai , Kevin Wang , Wuyang Li , Pavlo Molchanov , Peihao Wang
show 1 more author
Zhangyang Wang
This is my paper
classification cs.CV
keywords 3DGaussiansplattingelasticinferencemodelcompressionGumbel-Softmaxselectionneuralfieldnovelviewsynthesisresource-adaptiverenderingmany-in-one
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

3D Gaussian splatting renders scenes by painting thousands of tiny ellipsoidal 'splats' whose positions, sizes, colors, and opacities are optimized per scene. The paper claims that such a model can be trained once and then deployed at any desired memory budget without per-ratio fine-tuning: at inference the user supplies a ratio, a small learned selector keeps that fraction of Gaussians, and a learned transform field shifts and reshapes the survivors to repair the holes left by pruning. Across three real-world datasets and ratios from 1% to 15%, the method matches or beats compression methods that retrain for each ratio, and it extends to ratios never seen in training (2%, 4%, 6%, 8%). If true, one trained scene model can serve phones, headsets, and workstations with no per-device retraining or per-ratio compression pipeline.

What carries the argument

The load-bearing object is the Gaussian Transform Field, a 4D volume of resolution 64x64x64x100 factorized into six bilinearly interpolated planes, whose query at the Gaussian position and the elastic ratio yields features fused by an MLP and fed to three light MLP heads that predict offsets for position, scale, and rotation. It repairs the geometry of a pruned model, which is what the ablation shows is indispensable. The selector is the second half of the machinery: GsNet produces a differentiable Gumbel-Softmax mask, regularized to match the Global Importance top-k mask and to hit the exact count, and at inference the continuous logits are ranked and the top entries are retained.

What would settle it

Take a scene not in the training set (for instance a large outdoor city model with several million Gaussians) and request a ratio of, say, 30% or 50%. If PSNR falls below the simple heuristic top-k pruning baseline or the interpolation breaks (for example, artifacts appear only at mid-ratios), the claim that FlexGS handles an arbitrary number of Gaussians is refuted. A sharper version: request ratio 0.03 (between the trained 1% and 5%) on each test scene; if the transform field produces discontinuous or degraded renderings rather than monotone quality, the smooth-interpolation premise fails.

Watch

Extended reading notes

Core claim

FlexGS's central claim is elastic inference for 3DGS: a model trained once with a random ratio sampled from a small discrete set can later render at arbitrary, newly requested Gaussian counts with no fine-tuning. This is achieved by jointly optimizing two modules. First, an adaptive Gaussian selector (GsNet) uses a Gumbel-Softmax relaxation to predict a binary mask conditioned on each Gaussian's attributes and the target ratio, with a Global Importance score serving as a stability-promoting guidance signal and a sparsity constraint enforcing the requested count. Second, a Gaussian Transform Field, a 4D spatial-ratio neural field factorized into six planes and queried at (position, ratio), predicts displacements in position, scale, and rotation for the selected Gaussians so that the reduced set re-covers geometry that naive pruning leaves empty. The two modules are co-adapted during training; the authors report that removing the transform field collapses PSNR to roughly 14.6 at every ratio, while the full model maintains competitive quality and extends to unseen ratios such as 2%, 4%, 6%, and 8%.

Load-bearing premise

The whole scheme rests on the assumption that a transform field trained on just four ratios (1%, 5%, 10%, 15%) can smoothly interpolate to any requested ratio on any new scene, since the transform field, not the selector, is what actually does the heavy lifting and it collapses without repair.

Editorial extensions

If this is right

  • A single trained 3DGS model can serve multiple hardware tiers, from standalone VR headsets to desktop GPUs, with no per-device fine-tuning.
  • Deployment can become progressive: a coarse model renders immediately, and additional Gaussians are loaded incrementally as bandwidth allows, because the requested ratio can change at any moment.
  • Ratios never seen in training still work: on ZipNeRF, FlexGS generalizes to 2%, 4%, 6%, and 8% with quality above every compared baseline, supporting the claim that the ratio conditioning is continuous.
  • At the extreme 1% retention rate, FlexGS exceeds the quality of fine-tuned per-ratio compression such as LightGaussian and C3DGS on the tested datasets, which is the regime most relevant to weak mobile GPUs.

Reading between the lines

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

  • The same select-and-repair idea could transfer to other primitive-based renderers (for example 2D Gaussians or mesh splats) by swapping the transform field's coordinate conditioning, a step the paper does not take.
  • Because the transform field is only trained on ratios 1-15%, testing at 30%, 50%, or 90% would reveal whether 'arbitrary ratio' truly holds or only holds inside a neighborhood of the training band.
  • On very large scenes with millions of Gaussians, the per-Gaussian query cost of the 4D volume and the per-scene memory of the field itself could offset some of the memory savings, a trade-off the paper does not quantify.
  • Conditioning the transform field on the actual surviving mask (which specific Gaussians remain) rather than only on the ratio could strengthen repair when the selector's predictions deviate from the exact count, an idea worth testing.
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

5 major / 7 minor

Summary. FlexGS proposes a many-in-one compression scheme for 3D Gaussian Splatting. A single model is trained with elastic ratios e randomly sampled from {0.01, 0.05, 0.10, 0.15}; at inference the user supplies an arbitrary ratio e, and two learned modules act on the trained Gaussians: (i) GsNet, a Gumbel-Softmax selector guided by a Global Importance metric borrowed from LightGaussian, and (ii) a Gaussian Transform Field, a factorized 4D (x, y, z, e) volume with an MLP head that predicts displacements for the selected Gaussians' positions, scales, and rotations. The pipeline is trained with a rendering loss on both the selected subset and the full model, plus GI-guidance and sparsity regularizers. The paper reports results on MipNeRF360, Tanks and Temples, and ZipNeRF at 1-15% budgets, an ablation study over six configurations, a 'generalization' test on ZipNeRF at ratios {0.02, 0.04, 0.06, 0.08} that were not used in training, and per-scene tables in the appendix. The central claim is that a single trained model can be deployed at arbitrary memory budgets at inference time without fine-tuning.

Significance. The idea is well-motivated: existing pruning methods require per-ratio fine-tuning, and the paper's two-module design (learned selector plus ratio-conditioned transform field) is a reasonable way to obtain elastic inference. Within the trained ratio range, the reported numbers are competitive with per-ratio fine-tuned baselines (LightGaussian, C3DGS, EAGLES), and the supplementary per-scene tables and six-way ablation are careful. The Gumbel-Softmax gradient derivation in the appendix is a useful formal complement. If the arbitrary-ratio generalization is validated outside the training span, the contribution is significant for deployment-oriented 3DGS research. At present, however, the evidence for the central 'arbitrary ratio' claim is limited to ratios inside the training interval and near training points, and the ablation reveals that the selector alone performs far worse than a fixed heuristic, so the demonstration of the load-bearing interpolation property of the transform field remains to be supplied. These are addressable with additional experiments rather than fundamental flaws.

major comments (5)
  1. [§4.3, Table 4; §B.1] The evidence for the paper's central claim that FlexGS can be deployed 'with an arbitrary number of Gaussians' (§3) is restricted to the interval spanned by the training ratios. Table 4 tests the unseen ratios {0.02, 0.04, 0.06, 0.08} on ZipNeRF, all of which lie inside [0.01, 0.15] and are within 0.02–0.04 of a training ratio; no evaluation is reported for ratios above 0.15 (e.g., 0.20, 0.50, 0.80) or below 0.01. This matters because the Gaussian Transform Field is a 4D volume whose e-axis is sampled at resolution 100 and is queried during training only at e values in {0.01, 0.05, 0.10, 0.15}; the portion of the e-axis above 0.15, including the e = 1 state relative to which the displacement predictors are defined, is never trained or tested. The sentence in §4.3 that the results 'indicate the continuity and smoothness of the Gaussian Transform Field' is an extrapolation from four nearby interpolation points. Please add evaluations at out-of-span ratios, ideally with per-ratio plots of the interpolation error against training ratios.
  2. [Table 3, §4.3] The ablation 'w/o Transform Field' drops to a roughly constant PSNR of 14.6 at 5%, 10%, and 15% ratios, which is far below the 'w/o Adaptive Selector' row (the GI heuristic, matching LightGS* at 19.5/21.7/23.6 PSNR on the same ratios). The paper reports this result but does not explain why the learned selector alone is substantially worse than the fixed GI heuristic it is trained to imitate through L_GI in Eq. (11). The natural reading is that the selector and the Transform Field are strongly co-adapted: the selected Gaussians are only viable after geometric transformation, so the robustness of the whole method at untested ratios depends on both modules generalizing together. Please provide a diagnostic of this co-adaptation, for example the agreement rate between the learned mask and the GI top-k mask during inference, or a cross-combination experiment (learned mask with heuristic-selected transforms).
  3. [Table 1, §4.2] Several entries in Table 1 appear to be duplicated or near-identical between Ours and LightGS, and the accompanying text overstates the comparison. At the 1% row, Ours and LightGS share exactly the same SSIM (0.6128) and LPIPS (0.4384); at 15%, Ours (27.035/0.8072/0.2087) and LightGS (27.239/0.8079/0.2070) are nearly identical. The text in §4.2 claims FlexGS 'consistently outperforms' across all metrics at 10%, 5%, and 1%, but at 5% LightGS achieves a substantially higher SSIM (0.8381 vs. 0.7475) and at 15% LightGS has a higher PSNR. Please verify the reported numbers and revise the comparison claims accordingly.
  4. [§3.2–3.3, §4.1, Fig. 6] The elastic ratio e controls only the number of active Gaussians, not the total deployed model size. The deployment pipeline also includes GsNet, the Spatial-Ratio Neural Field (six planes at two resolutions over a {64, 64, 64, 100} grid), and the three MLP heads, whose byte footprint is never reported. At a 1% budget these auxiliary modules may be a non-negligible fraction of the compressed Gaussian storage, so the statement that the method operates at a user-desired 'fraction of the memory footprint of the original 3DGS models' (§4) is not yet supported as an end-to-end memory claim. Please report the actual model size in bytes or MB for the full pipeline, and clarify whether the 0.13s inference time in Fig. 6 refers to the one-time elastic adaptation or to per-frame rendering.
  5. [§3.2, §3.4, §B.2] The description of how the target count ⌊eN⌋ is enforced during training is internally inconsistent. Section 3.2 states that binary masks M̂ are sampled from the Gumbel-Softmax categorical distribution (Eq. 7), with L_spar in Eq. (11) only encouraging the average count; Section 3.4 then states that 'the top e most significant Gaussians are selected' from the output masks; and Section B.2 notes that the predicted mask count can deviate from the desired ratio (19.5% selected for a target of 20%), which is why inference uses a top-⌊eN⌋ operation on continuous softmax outputs. This makes the exact training-time selection procedure ambiguous: it is not clear whether the rendering in Eq. (15) uses the sampled binary mask, the top-k of logits, or the GI guidance mask, and the train/inference mismatch is not discussed. Please specify the exact procedure and justify the mismatch.
minor comments (7)
  1. [Fig. 4 caption] The caption lists the elastic ratios as {0.1, 0.5, 1.0, 1.5}, which appears to be a typo for {0.01, 0.05, 0.10, 0.15}; the figure panels and Tables 1–2 use percentages 1%, 5%, 10%, 15%.
  2. [§3.2] There are typos in the text: 'the Gaussian model can be inferenced with arbitrary remaining raio' should read 'the Gaussian model can be inferred with an arbitrary remaining ratio.'
  3. [Eq. (16) vs §4.1] The final loss in Eq. (16) uses weighting coefficients β_1 and β_2, but the implementation details in §4.1 refer to β_s (set to 1.0) and β_d (set to 0.01); please align the notation.
  4. [§2] The sentence 'Recent advances in compressing 3DGS can be categorized into three main approaches scene [1]:' is garbled and should be revised.
  5. [Tables 1–4] No error bars, standard deviations, or multi-seed results are reported, so margins such as 'improvements of at least 0.5 in PSNR' (§4.3) cannot be assessed for statistical significance; please report variance across runs or at least state the number of seeds.
  6. [Fig. 6] The figure reports an inference time of 0.13s and a PSNR of 23.0, but the text does not specify which scene, ratio, or hardware configuration this corresponds to, nor whether 0.13s covers the full elastic adaptation or only rendering.
  7. [Abstract] The statement 'Code is available at https://flexgs.github.io' points to a project page rather than a code repository; please provide the actual repository link or state the availability status explicitly.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: FlexGS validates its elastic-ratio behavior against external baselines, and the Global Importance guidance borrowed from LightGaussian is non-load-bearing because the full model clearly improves over that heuristic.

full rationale

FlexGS's central claim is that a single 3D Gaussian Splatting model can render at arbitrary retention ratios through a learned Gumbel-Softmax selector plus a ratio-conditioned transform field, and this is verified by actually rendering against ground-truth images and comparing with external baselines (LightGS, C3DGS, EAGLES) across trained ratios in Tables 1 and 2. No predicted quantity is defined in terms of a fitted quantity: the quality metric at each ratio is an independent rendering loss, not the training objective's optimum. The one self-citation of note is the Global Importance heuristic taken from LightGaussian (Fan et al., a paper with overlapping authors, including Kevin Wang and Zhangyang Wang). GI enters only as a soft supervision target (Eq. 11, L_GI = ||M_hat - M_GI||), and the paper's own ablation (Table 3) shows the full model (23.090 PSNR at 1%) beats both the pure GI top-k heuristic (LightGS*, 14.653 PSNR) and the selector without GI supervision, so the learned components are not equivalent to their guidance input; the citation is real, externally falsifiable support rather than a forced premise. The 'arbitrary ratio' ability is implemented through bilinear interpolation along the e-axis of the factorized 4D transform volume, and Table 4 evaluates ratios {0.02, 0.04, 0.06, 0.08} that were not in the training set {0.01, 0.05, 0.10, 0.15}, so the results are produced by an actual interpolation mechanism rather than by re-fitting per ratio. A genuine limitation remains: all tested 'unseen' ratios lie strictly inside the training span [0.01, 0.15], so extrapolation to e > 0.15 or e < 0.01 is asserted but not demonstrated, and the ablation in Table 3 shows the transform field is load-bearing (removing it collapses PSNR to about 14.6), which makes the untested e-axis interpolation a correctness and robustness risk. However, an untested inductive property is a coverage gap, not a circular reduction: no equation in the paper defines the reported prediction in terms of the fitted parameters by construction, and no uniqueness theorem or ansatz is smuggled in via self-citation. Accordingly, the appropriate circularity verdict is essentially clean, with only a minor overlapping-author citation that does not carry the argument.

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

The central claim rests on several hand-set hyperparameters (loss weights, training ratios, GI threshold, field resolution) and on the assumption that the co-adapted selector and transform field generalize beyond the training ratios. The GI guidance is borrowed from prior work, so the learned selector is partly anchored to an existing heuristic.

free parameters (4)
  • Loss weights beta_1 and beta_2 = 1.0 and 0.01
    Chosen by hand in Sec. 3.4 and Implementation Details; they control the strength of GI guidance and the sparsity ratio constraint, directly shaping the learned selector.
  • Training ratio set E = {0.01, 0.05, 0.10, 0.15}
    The model is trained only on these four ratios; generalization to other ratios is assumed via interpolation.
  • GI volume threshold = V90%
    Used in Eq. 9 for gamma(s_i), inherited from LightGaussian; the 90th percentile is a hand-set threshold.
  • Spatial-Ratio field resolution = {64, 64, 64, 100}
    Resolution of the 4D volume across x, y, z, and e; chosen in Implementation Details B.1, it affects the capacity to represent transformations.
assumptions (5)
  • domain assumption 3DGS differentiable rendering (Eq. 1-2) is a valid scene representation and the rasterizer gradients are correct.
    The whole method builds on the standard 3DGS framework [17].
  • standard math Gumbel-Softmax with the straight-through estimator provides usable gradients for the discrete selection.
    Gumbel-Softmax is a standard reparameterization trick; the supplement derives the gradient formulas.
  • domain assumption Global Importance score is a meaningful proxy for rendering contribution.
    GI (Eq. 8-10) is a heuristic inspired by LightGaussian; the ablation shows that removing GI guidance degrades PSNR by about 3.6 at 1% ratio (Table 3), so the central result depends on this assumption.
  • ad hoc to paper The 4D voxel field can smoothly interpolate transformations across ratios.
    The Transform Field is a novel module; the paper assumes it generalizes to unseen ratios, and it is tested only on 0.02 to 0.08 on ZipNeRF (Table 4).
  • domain assumption Two-stage training, 15k standard 3DGS iterations followed by 20k elastic iterations, is sufficient to learn the selector and transform field.
    No ablation on training length is provided, so the adequacy of this schedule is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/CSZNTHLN

@misc{pith2026250604174,
  author       = {Pith},
  title        = {Pith review of: FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CSZNTHLN}},
  note         = {Machine review of arXiv:2506.04174}
}
read the original abstract

3D Gaussian splatting (3DGS) has enabled various applications in 3D scene representation and novel view synthesis due to its efficient rendering capabilities. However, 3DGS demands relatively significant GPU memory, limiting its use on devices with restricted computational resources. Previous approaches have focused on pruning less important Gaussians, effectively compressing 3DGS but often requiring a fine-tuning stage and lacking adaptability for the specific memory needs of different devices. In this work, we present an elastic inference method for 3DGS. Given an input for the desired model size, our method selects and transforms a subset of Gaussians, achieving substantial rendering performance without additional fine-tuning. We introduce a tiny learnable module that controls Gaussian selection based on the input percentage, along with a transformation module that adjusts the selected Gaussians to complement the performance of the reduced model. Comprehensive experiments on ZipNeRF, MipNeRF and Tanks\&Temples scenes demonstrate the effectiveness of our approach. Code is available at https://flexgs.github.io.

Figures

Figures reproduced from arXiv: 2506.04174 by the authors.

Figure 1
Figure 1. Comparison between compression methods for 3D [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall framework of FlexGS: (a) Adaptive Gaussian Selector: utilizes a GsNet to output the differentiable mask and Global Important as guidance for adaptive selection; (b) Gaussian Transform Field: queries a Spatial-Ratio Neural Field and outputs the transforms of Gaussian attributes under the given elastic ratio. All the modules are jointly optimized to minimize the rendering loss. ratio e, we model the categorica… view at source ↗
Figure 3
Figure 3. Visualization of the point distribution for masked 3DGS [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Visual results compared with other methods on various elastic ratios: [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: PSNR vs. inference time: Our method achieves the [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 5
Figure 5. Figure 5: Visualization of FlexGS for generalization explo [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Visual results compared with other methods on various elastic ratios: [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Potential use of elastic inference in the application scenario of incremental scene loading. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 30 canonical work pages

  1. [1]

    Milena T Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgen- stern. 3dgs. zip: A survey on 3d gaussian splatting compres- sion methods.arXiv preprint arXiv:2407.09510, 2024. 2

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 6, 7, 8, 3, 5

    work page 2022

  3. [3]

    Zip-nerf: Anti-aliased grid-based neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023. 2, 6, 7, 8

    work page 2023

  4. [4]

    Once-for-all: Train one network and specialize it for efficient deployment.arXiv preprint arXiv:1908.09791,

    Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment.arXiv preprint arXiv:1908.09791,

    arXiv 1908

  5. [5]

    Flextron: Many-in-one flexible large language model.arXiv preprint arXiv:2406.10260, 2024

    Ruisi Cai, Saurav Muralidharan, Greg Heinrich, Hongxu Yin, Zhangyang Wang, Jan Kautz, and Pavlo Molchanov. Flextron: Many-in-one flexible large language model.arXiv preprint arXiv:2406.10260, 2024. 3

    arXiv 2024

  6. [6]

    Hexplane: A fast representa- tion for dynamic scenes

    Ang Cao and Justin Johnson. Hexplane: A fast representa- tion for dynamic scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130–141, 2023. 5

    work page 2023

  7. [7]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InEuropean Conference on Computer Vision, pages 333–350. Springer,

  8. [8]

    How far can we compress instant-ngp-based nerf? In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 2024

    Yihang Chen, Qianyi Wu, Mehrtash Harandi, and Jianfei Cai. How far can we compress instant-ngp-based nerf? In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 2024. 2

    work page 2024

Show all 52 references
  1. [9]

    Compressing ex- plicit voxel grid representations: fast nerfs become also small

    Chenxi Lola Deng and Enzo Tartaglione. Compressing ex- plicit voxel grid representations: fast nerfs become also small. InProceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pages 1236–1245,

  2. [10]

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

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, De- jia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. arXiv preprint arXiv:2311.17245, 2023. 1, 2, 3, 4, 6, 7, 8, 5

    2023 arXiv

  3. [11]

    K-planes: Explicit radiance fields in space, time, and appearance

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12479–12488, 2023. 2

    2023

  4. [12]

    K-planes: Explicit radiance fields in space, time, and appearance

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12479–12488, 2023. 5, 3

    2023

  5. [13]

    Ea- gles: Efficient accelerated 3d gaussians with lightweight en- codings.arXiv preprint arXiv:2312.04564, 2023

    Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Ea- gles: Efficient accelerated 3d gaussians with lightweight en- codings.arXiv preprint arXiv:2312.04564, 2023. 6, 7, 8, 5

    2023 arXiv

  6. [14]

    Shacira: Scalable hash-grid compression for implicit neural representations

    Sharath Girish, Abhinav Shrivastava, and Kamal Gupta. Shacira: Scalable hash-grid compression for implicit neural representations. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 17513–17524,

  7. [15]

    Dragon: Drone and ground gaussian splatting for 3d building reconstruction

    Yujin Ham, Mateusz Michalkiewicz, and Guha Balakrish- nan. Dragon: Drone and ground gaussian splatting for 3d building reconstruction. In2024 IEEE International Confer- ence on Computational Photography (ICCP), pages 1–12. IEEE, 2024. 1

    2024

  8. [16]

    Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144, 2016

    Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144, 2016. 3, 4, 8, 2

    2016 arXiv

  9. [17]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023. 2

    2023

  10. [18]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4):1–14, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4):1–14, 2023. 1, 3, 6

    2023

  11. [19]

    Color-cued efficient densification method for 3d gaussian splatting

    Sieun Kim, Kyungjin Lee, and Youngki Lee. Color-cued efficient densification method for 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 775–783, 2024. 2

    2024

  12. [20]

    Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36 (4):1–13, 2017

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36 (4):1–13, 2017. 6

    2017

  13. [21]

    Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36 (4):1–13, 2017

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36 (4):1–13, 2017. 2, 6, 7

    2017

  14. [22]

    Matformer: Nested transformer for elastic inference.arXiv preprint arXiv:2310.07707, 2023

    Sneha Kudugunta, Aditya Kusupati, Tim Dettmers, Kaifeng Chen, Inderjit Dhillon, Yulia Tsvetkov, Hannaneh Hajishirzi, Sham Kakade, Ali Farhadi, Prateek Jain, et al. Matformer: Nested transformer for elastic inference.arXiv preprint arXiv:2310.07707, 2023. 3

    arXiv

  15. [23]

    Matryoshka representation learning.Advances in Neu- ral Information Processing Systems, 35:30233–30249, 2022

    Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al. Matryoshka representation learning.Advances in Neu- ral Information Processing Systems, 35:30233–30249, 2022. 3

    2022

  16. [24]

    Pulsar: Effi- cient sphere-based neural rendering

    Christoph Lassner and Michael Zollhofer. Pulsar: Effi- cient sphere-based neural rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1440–1449, 2021. 2

    2021

  17. [25]

    Compact 3d gaussian representation for radiance field

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719– 21728, 2024. 2

    2024

  18. [26]

    Compressing volumetric radiance fields to 1 mb

    Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Liefeng Bo. Compressing volumetric radiance fields to 1 mb. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4222–4231, 2023. 2

    2023

  19. [27]

    Compact real-time radiance fields with neural code- book

    Lingzhi Li, Zhongshu Wang, Zhen Shen, Li Shen, and Ping Tan. Compact real-time radiance fields with neural code- book. InICME, 2023. 2

    2023

  20. [28]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction

    Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, You- liang Yan, et al. Vastgaussian: Vast 3d gaussians for large scene reconstruction. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recogni...

    2024

  21. [29]

    Citygaussian: Real-time high-quality large-scale scene rendering with gaussians

    Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Jun- ran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In European Conference on Computer Vision, pages 265–282. Springer, 2025. 1

    2025

  22. [30]

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

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2

    2024

  23. [31]

    The concrete distribution: A continuous relaxation of discrete random variables

    Connor J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. InInternational Conference on Learning Representations (ICLR), 2017. 3

    2017

  24. [32]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    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. 2

    2021

  25. [33]

    Compact 3d scene representation via self- organizing gaussian grids.arXiv preprint arXiv:2312.13299,

    Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3d scene representation via self- organizing gaussian grids.arXiv preprint arXiv:2312.13299,

    arXiv

  26. [34]

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

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM Transactions on Graphics (ToG), 41(4):1–15, 2022. 2

    2022

  27. [35]

    Compact3d: Com- pressing gaussian splat radiance field models with vector quantization.arXiv preprint arXiv:2311.18159, 2023

    KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compact3d: Com- pressing gaussian splat radiance field models with vector quantization.arXiv preprint arXiv:2311.18159, 2023. 2

    2023 arXiv

  28. [36]

    Compgs: Smaller and faster gaussian splatting with vector quantization

    KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. InEuro- pean Conference on Computer Vision, 2024. 2

    2024

  29. [37]

    Compressed 3d gaussian splatting for accelerated novel view synthesis

    Simon Niedermayr, Josef Stumpfegger, and R ¨udiger West- ermann. Compressed 3d gaussian splatting for accelerated novel view synthesis. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10349–10358, 2024. 1, 2, 3, 4, 6, 7, 8, 5

    2024

  30. [38]

    Reducing the memory footprint of 3d gaussian splatting.Proceedings of the ACM on Computer Graphics and Interactive Tech- niques, 7(1):1–17, 2024

    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. 2

    2024

  31. [39]

    Masked wavelet representation for compact neural radiance fields

    Daniel Rho, Byeonghyeon Lee, Seungtae Nam, Joo Chan Lee, Jong Hwan Ko, and Eunbyung Park. Masked wavelet representation for compact neural radiance fields. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20680–20690, 2023. 2

    2023

  32. [40]

    Binary radiance fields.Ad- vances in neural information processing systems, 2023

    Seungjoo Shin and Jaesik Park. Binary radiance fields.Ad- vances in neural information processing systems, 2023. 2

    2023

  33. [41]

    Spc-nerf: Spatial predictive compression for voxel based radiance field.arXiv preprint arXiv:2402.16366, 2024

    Zetian Song, Wenhong Duan, Yuhuai Zhang, Shiqi Wang, Siwei Ma, and Wen Gao. Spc-nerf: Spatial predictive compression for voxel based radiance field.arXiv preprint arXiv:2402.16366, 2024. 2

    2024 arXiv

  34. [42]

    Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction

    Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459– 5469, 2022. 2

    2022

  35. [43]

    Compressible-composable nerf via rank-residual de- composition.Advances in Neural Information Processing Systems, 35:14798–14809, 2022

    Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable nerf via rank-residual de- composition.Advances in Neural Information Processing Systems, 35:14798–14809, 2022. 2

    2022

  36. [44]

    End-to-end rate- distortion optimized 3d gaussian representation.arXiv preprint arXiv:2406.01597, 2024

    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.arXiv preprint arXiv:2406.01597, 2024. 2

    2024 arXiv

  37. [45]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310–20320, ...

    2024

  38. [46]

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

    Wang Yifan, Felice Serena, Shihao Wu, Cengiz ¨Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing.ACM Transactions on Graphics (TOG), 38(6):1–14, 2019. 3

    2019

  39. [47]

    Universally slimmable net- works and improved training techniques

    Jiahui Yu and Thomas S Huang. Universally slimmable net- works and improved training techniques. InProceedings of the IEEE/CVF international conference on computer vision, pages 1803–1811, 2019. 3

    2019

  40. [48]

    Slimmable neural networks.arXiv preprint arXiv:1812.08928, 2018

    Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. Slimmable neural networks.arXiv preprint arXiv:1812.08928, 2018. 3

    2018 arXiv

  41. [49]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 6

    2018

  42. [50]

    Gs-hider: Hiding messages into 3d gaussian splatting.Advances in Neural Information Processing Sys- tems, 37:49780–49805, 2024

    Xuanyu Zhang, Jiarui Meng, Runyi Li, Zhipei Xu, Jian Zhang, et al. Gs-hider: Hiding messages into 3d gaussian splatting.Advances in Neural Information Processing Sys- tems, 37:49780–49805, 2024. 1

    2024

  43. [51]

    Securegs: Boosting the security and fidelity of 3d gaussian splatting steganography.arXiv preprint arXiv:2503.06118, 2025

    Xuanyu Zhang, Jiarui Meng, Zhipei Xu, Shuzhou Yang, Yanmin Wu, Ronggang Wang, and Jian Zhang. Securegs: Boosting the security and fidelity of 3d gaussian splatting steganography.arXiv preprint arXiv:2503.06118, 2025

    2025 arXiv

  44. [52]

    chain rule

    Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaus- sian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2...

    2024

Pith tools

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