Pith. sign in

REVIEW 4 major objections 8 minor 1 cited by

Perceptual-GS: Scene-adaptive Perceptual Densification for Gaussian Splatting

T0 review · 4 major / 8 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 Perceptual-GS claims that steering Gaussian densification with a learned perceptual-sensitivity branch yields better LPIPS and SSIM using fewer primitives than vanilla 3DGS.

desk verdict A well-engineered 3DGS densification method whose edge-guidance mechanism works, but whose 'perceptual' framing is not yet supported by the experiments. read the letter →

arxiv 2506.12400 v2 pith:ZWDUAKBA submitted 2025-06-14 cs.CV

Hongbi Zhou , Zhangkai Ni This is my paper · ORCID
classification cs.CV
keywords 3DGaussianSplattingperceptualsensitivityadaptivedensificationnovelviewsynthesisdual-branchrenderinglearnablequality-efficiencytrade-offLPIPS
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

Perceptual-GS tries to establish that 3D Gaussian Splatting can be made both more accurate and cheaper by letting human visual sensitivity decide where Gaussian primitives are placed. The paper's claim is that a binarized, smoothed gradient-magnitude map extracted from each training view is a usable stand-in for perceptual importance, and that a learnable per-Gaussian sensitivity value rendered through a second branch can steer densification toward visually critical regions. If the claim is right, scene representations can hold fewer Gaussians while improving LPIPS and SSIM, and the perceptual branch can be bolted onto other splatting methods as a general efficiency upgrade. On the datasets tested, the paper reports a better quality-efficiency trade-off than vanilla 3DGS and the selected densification baselines, particularly in large-scale street scenes.

What carries the argument

The carrying mechanism is the dual-branch rendering of a learned per-Gaussian sensitivity parameter. Instead of accumulating pixel statistics over a Gaussian's footprint, the sensitivity branch renders $\sigma(\epsilon_i)$ through the same $\alpha$-blending weights used for color, so each Gaussian is forced to commit to one sensitivity value per viewpoint and the rendered map is compared to a binarized Sobel gradient map with BCE loss. That per-Gaussian $\epsilon_i$ then drives the densification rules: $\epsilon_i > \tau_h$ marks visually critical high-sensitivity Gaussians for splitting, $\tau_l \leq \epsilon_i \leq \tau_h$ marks viewpoint-conflicted medium-sensitivity Gaussians, and a maximum-over-views weight threshold $\omega^{\max}_i$ prevents over-densification. A scene-level sensitivity statistic $\beta$ decides whether clone or split is used, and a scene-adaptive depth reinitialization fires when the fraction of large medium-sensitivity Gaussians $\gamma$ exceeds a threshold; an opacity-decline transform on clones makes redundant primitives easier to prune.

What would settle it

On a held-out scene, compute the paper's binary sensitivity map from ground-truth training views and compare it with human fixation or eye-tracking data on the same views: if the map's high-sensitivity pixels do not agree with fixation densities, or if densifying according to fixation data instead improves LPIPS further, the perceptual proxy is not doing the claimed work. A cheaper test is to render a smooth, low-gradient wall with a faint shadow: the Sobel map marks it low-sensitivity, so Perceptual-GS should leave it sparse; if viewers reliably detect the shadow artifact, the premise that low gradient equals low perceptual need fails.

Watch

Extended reading notes

Core claim

The central discovery claimed is that perceptual sensitivity can be made a first-class training signal in 3DGS without new renderers: each Gaussian carries a scalar sensitivity $\epsilon_i$ that is optimized alongside color and geometry, rendered like an image via the same tile rasterizer, and supervised with binary cross-entropy against a precomputed binary sensitivity map. After warm-up, Gaussians whose learned sensitivity is high are split to add detail, those with intermediate sensitivity are split because a single primitive cannot reconcile conflicting viewpoints, and low-sensitivity regions are left with fewer primitives. The result reported is that on Mip-NeRF 360, Tanks & Temples, Deep Blending, and BungeeNeRF, Perceptual-GS achieves better average SSIM and LPIPS than vanilla 3DGS and the compared densification methods while using substantially fewer Gaussian primitives, and that it can wrap around existing splatting-based methods to improve their metrics.

Load-bearing premise

The load-bearing premise is that a thresholded and smoothed Sobel gradient map of the ground-truth image faithfully says where human viewers need more detail, and that the learned per-Gaussian sensitivities have converged to meaningful 0/1 values by the time densification starts.

Editorial extensions

If this is right

  • If the perceptual proxy is right, Gaussian counts can drop by tens of percent with no perceptual loss, which directly cuts memory and streaming costs for radiance-field scenes.
  • Densification becomes scene-adaptive: texture-rich regions get more primitives, flat or defocused regions stay sparse, benefiting large-scale outdoor scenes where fixed-gradient densification blows up primitive counts.
  • Because the sensitivity branch is a wrapper on the training loss and density control, it can be combined with existing splatting methods to improve their LPIPS while reducing their Gaussian counts.
  • The opacity-decline clone rule should make trained models easier to prune and compress, since redundant cloned Gaussians are actively pushed toward low opacity.

Reading between the lines

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

  • Extension: if the Sobel-plus-threshold proxy is a crude stand-in, the method is likely to under-allocate Gaussians to smooth, low-gradient regions that still carry semantic meaning (faces, skies, signage); a perceptual metric weighted by human fixations might reallocate primitives differently.
  • Extension: the learned $\epsilon_i$ distribution after warm-up is itself a free saliency prior for the scene; it could be reused for foveated rendering, level-of-detail selection, or compression priority without retraining.
  • Extension: a testable improvement is to replace the hand-tuned thresholds $\tau_h$, $\tau_l$, $\tau_\omega$, and $\tau_\beta$ with per-scene calibration or a learned policy, since the paper's ablations show several thresholds trade LPIPS against Gaussian count and the operating point is tunable per application.
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

4 major / 8 minor

Summary. The paper proposes Perceptual-GS, a training-time densification strategy for 3D Gaussian Splatting. A per-Gaussian learnable parameter epsilon_i is introduced; it is supervised through a second 'sensitivity' rendering branch whose target is a binarized and smoothed Sobel gradient map of the ground-truth images. The learned sensitivities are then used to select Gaussians for high- and medium-sensitivity densification, supplemented by scene-adaptive depth reinitialization and an opacity-decline mechanism for clone operations. Experiments are reported on Mip-NeRF 360, Tanks & Temples, Deep Blending, and BungeeNeRF, with per-scene tables and integration results with 3DGS, Pixel-GS, and CoR-GS.

Significance. If the reported gains hold, Perceptual-GS offers a simple, training-only modification of 3DGS that improves SSIM and LPIPS while reducing the number of Gaussians, with public code and unusually detailed per-scene tables. The integration experiments with Pixel-GS and CoR-GS are a useful practical contribution. However, the central 'perceptual' attribution is not yet isolated from generic edge-density-guided densification, and the evaluation protocol has several load-bearing gaps: hyperparameters are tuned on the same benchmark used for the headline results, and the BungeeNeRF averages compare different scene subsets. These issues need to be addressed before the state-of-the-art claim is supported.

major comments (4)
  1. [Section 3.3, Eqs. (5)-(7), Table 4] The 'perceptual' component of the method is under-identified. The perceptual sensitivity map is a binarized and smoothed Sobel gradient magnitude; the only ablation ('w/o PE') replaces the binarization/smoothing with raw Sobel magnitudes. Because LPIPS and SSIM are themselves most sensitive at high-gradient pixels, a non-perceptual binary edge mask with matched density would likely produce similar quality-efficiency gains. To support the central claim, add an ablation that uses a generic binary edge map (e.g., Canny edges or a fixed-percentile gradient threshold) with the same downstream densification rules, and report whether the LPIPS/SSIM improvements persist. If they do, the 'perceptual' wording should be softened; if they do not, that should be stated explicitly.
  2. [Section 4.1, Appendix D, Tables 1 and 11] All newly introduced hyperparameters (tau_e, tau_s, lambda_S, Iter_h, Iter_m, tau_h, tau_l, tau_omega_h, tau_omega_m, tau_beta, tau_gamma, and k) are tuned on Mip-NeRF 360, and the final results in Tables 2-4 are reported on the same benchmark. This test-set selection makes the reported margins difficult to interpret. Please either reserve a validation split (e.g., a subset of scenes or a separate dataset) for parameter selection and report test results on the remainder, or demonstrate that the conclusions are insensitive to the parameter choices across all datasets.
  3. [Tables 2 and 3, BungeeNeRF rows] The BungeeNeRF averages for Pixel-GS* and Taming-3DGS are computed over only 7 and 6 of the 8 scenes, respectively, because of CUDA out-of-memory, while 'Ours' is averaged over all 8 scenes. Presenting these numbers in the same averaged row is misleading. Please report averages over the common subset of scenes that all methods can run, or clearly mark the number of scenes contributing to each average; the per-scene tables already provided would allow this.
  4. [Section 3.4, Eqs. (8)-(10), Figure 13] The densification rules in Eqs. (11)-(12) assume that after the 500-iteration warm-up, the per-Gaussian sensitivity epsilon_i has converged to values close to 0 or 1, so that thresholds tau_h and tau_l are meaningful. The paper shows rendered sensitivity maps but provides no quantitative convergence analysis, such as histograms of epsilon_i over training, final binarization accuracy against the ground-truth sensitivity map, or sensitivity of the downstream densification to the warm-up length. This is especially important for the large-scale BungeeNeRF scenes where the claimed robustness gains are largest.
minor comments (8)
  1. [Abstract and Section 4.2] The abstract's 'state-of-the-art performance' claim is stronger than the tables support: on Tanks & Temples and Deep Blending, Taming-3DGS reports higher PSNR (24.04 vs 23.90; 30.14 vs 29.94). Please qualify the claim as state-of-the-art on perceptual metrics and efficiency, or rephrase to 'competitive'.
  2. [Section 3.1 and Section 3.3] The symbol G is used both for the set of Gaussian primitives in Eq. (1) and for the Sobel response map in Eq. (6), which is confusing; please use distinct notation, e.g., M for the response map.
  3. [Section 3.3] The 'average pooling with threshold tau_s' step is underspecified: the kernel size, stride, and exact binarization rule after pooling are not given. Please specify these details in the appendix.
  4. [Section 3.1 and Section 3.4] There are typos: 'Gussian primitive' in Section 3.1 and 'Gassians' in Section 3.4 should be 'Gaussian primitive' and 'Gaussians'.
  5. [Eq. (13) and Table 10] Eq. (13) uses tau_omega without a subscript, while the text and Table 10 use tau_omega_h and tau_omega_m; please define both forms in the equation.
  6. [Table 6] The caption says 'We present metrics averaged on the dataset and from three single scenes,' but the columns are BungeeNeRF, Pompidou, Chicago, and Amsterdam; please clarify that the first column is the 8-scene average and the other columns are individual scenes.
  7. [Figure 4] The label 'Ours Taming-3DGS' appears to be a formatting mistake; it should list 'Ours' and 'Taming-3DGS' as separate labels.
  8. [Eq. (21)] The QEB metric's units and scaling are unclear: please state explicitly that #G is in millions and justify the factor of 100 in the definition.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Perceptual-GS is an externally supervised empirical method whose reported metrics are held-out measurements, not fitted outputs.

full rationale

The claimed derivation chain is constructive rather than circular. Perceptual sensitivity maps are extracted from ground-truth training images by Sobel convolution (Eqs. 5-7); per-Gaussian sensitivity parameters epsilon_i are trained with a BCE loss (Eqs. 8-10) to reproduce those maps; and densification decisions are thresholded functions of the fitted epsilon (Eqs. 11-14). None of the reported headline quantities (PSNR, SSIM, LPIPS, #G, FPS) is a fitted constant: quality metrics are computed on held-out viewpoints, and #G and FPS are training and rendering outcomes rather than values forced by the construction of the loss. The perception proxy may be under-identified experimentally (the ablations compare binarized vs raw Sobel maps but do not test a matched-density non-perceptual binary edge mask), and the self-citations (Ni et al. 2016; Shen et al. 2020; Ren et al. 2025) are motivational rather than load-bearing; neither issue makes the central claim reduce to its own inputs. The paper is self-contained against external benchmarks and baselines, so the circularity score is 0.

Assumptions & free parameters 13 free parameters · 5 assumptions · 1 invented entities

The central method introduces a large set of hand-chosen thresholds and learned per-Gaussian parameters. The main epistemic load is carried by the assumption that thresholded Sobel maps capture human perceptual sensitivity, and by hyperparameters tuned on the same benchmark used for final evaluation. There are no new physical entities in the usual sense, but the per-Gaussian epsilon_i is a new latent variable whose external validity is unverified.

free parameters (13)
  • Enhancement threshold tau_e = 0.05
    Binarizes the Sobel gradient magnitude map; chosen by hand with no sensitivity analysis reported.
  • Smoothing threshold tau_s = 0.3
    Threshold applied after average pooling to produce the final binary sensitivity map; chosen by hand.
  • Sensitivity loss weight lambda_S = 0.1
    Balances RGB and sensitivity branch losses; selected via ablation on Mip-NeRF 360 and used for final scores on the same dataset.
  • High-sensitivity densification interval Iter_h = 1000
    Densification interval for high-sensitivity Gaussians; tuned on Mip-NeRF 360 according to Table 11.
  • Medium-sensitivity densification interval Iter_m = 1500
    Densification interval for medium-sensitivity Gaussians; tuned on Mip-NeRF 360 according to Table 11.
  • High-sensitivity threshold tau_h = 0.9
    Learned sensitivity above which a Gaussian is treated as high-sensitivity; chosen by hand.
  • Low-sensitivity threshold tau_l = 0.3
    Lower boundary of the medium-sensitivity band; chosen by hand.
  • Weight threshold for high-sensitivity tau_omega_h = 25
    Limits how many high-sensitivity Gaussians get densified; selected through ablation on Mip-NeRF 360.
  • Weight threshold for medium-sensitivity tau_omega_m = 10
    Limits how many medium-sensitivity Gaussians get densified; selected through ablation on Mip-NeRF 360.
  • Scene sensitivity threshold tau_beta = 0.85
    Determines whether high-sensitivity Gaussians are split or cloned; chosen by hand with no reported ablation.
  • Sparse-point-cloud threshold tau_gamma = 0.55
    Triggers scene-adaptive depth reinitialization; chosen by hand with no reported ablation.
  • Opacity decline exponent k = 1.2
    Power function exponent for reducing cloned Gaussian opacity; selected from Table 1 on Mip-NeRF 360, with PSNR differences between k values within 0.01 dB.
  • Per-Gaussian sensitivity epsilon_i = learned, sigmoid-constrained to [0,1]
    Auxiliary latent variable trained with BCE loss against binary sensitivity maps; its learned value drives all densification decisions but is never validated against independent human perception data.
assumptions (5)
  • domain assumption Binarized Sobel gradient magnitude maps are a valid proxy for human perceptual sensitivity.
    Section 3.3 derives the sensitivity map from Sobel gradients, thresholding, and smoothing, but no human-study validation or comparison to established perceptual models is provided.
  • domain assumption Densifying more Gaussians in high-sensitivity regions improves perceptual quality more than densifying elsewhere.
    This is the core motivation of Perceptual-GS, related to SSIM and visual attention, but it is not proven for this setting.
  • domain assumption After 500 warm-up iterations the per-Gaussian sensitivity epsilon_i is well learned and can be thresholded to guide densification.
    Section 3.5 assumes the sensitivity branch reaches reliable 0 or 1 values; the appendix shows rendered maps but no quantitative convergence analysis.
  • domain assumption The scene-level thresholds tau_beta and tau_gamma generalize across datasets.
    These heuristics are fixed without a validation procedure independent of the test sets, so their transferability is assumed.
  • standard math Alpha-compositing formula A = alpha_hat + (1 - alpha_hat) * alpha_hat.
    Used in Section 3.7 to derive the opacity decline transformation; this is standard alpha blending math and is correct.
invented entities (1)
  • Per-Gaussian perceptual sensitivity parameter epsilon_i
    purpose: Encodes the perceptual sensitivity of the 3D region represented by each Gaussian; rendered into a sensitivity map and used to decide which Gaussians to densify.
    It is a learned latent variable supervised only by the binary edge maps and the overall scene losses; no external human or perceptual validation links epsilon_i to actual human sensitivity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Perceptual-GS: Scene-adaptive Perceptual Densification for Gaussian Splatting." pith.science (2026). https://pith.science/paper/ZWDUAKBA

@misc{pith2026250612400,
  author       = {Pith},
  title        = {Pith review of: Perceptual-GS: Scene-adaptive Perceptual Densification for Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZWDUAKBA}},
  note         = {Machine review of arXiv:2506.12400}
}
read the original abstract

3D Gaussian Splatting (3DGS) has emerged as a powerful technique for novel view synthesis. However, existing methods struggle to adaptively optimize the distribution of Gaussian primitives based on scene characteristics, making it challenging to balance reconstruction quality and efficiency. Inspired by human perception, we propose scene-adaptive perceptual densification for Gaussian Splatting (Perceptual-GS), a novel framework that integrates perceptual sensitivity into the 3DGS training process to address this challenge. We first introduce a perception-aware representation that models human visual sensitivity while constraining the number of Gaussian primitives. Building on this foundation, we develop a perceptual sensitivity-adaptive distribution to allocate finer Gaussian granularity to visually critical regions, enhancing reconstruction quality and robustness. Extensive evaluations on multiple datasets, including BungeeNeRF for large-scale scenes, demonstrate that Perceptual-GS achieves state-of-the-art performance in reconstruction quality, efficiency, and robustness. The code is publicly available at: https://github.com/eezkni/Perceptual-GS

Figures

Figures reproduced from arXiv: 2506.12400 by the authors.

Figure 1
Figure 1. The quality-efficiency trade-off and robustness in large￾scale scenes of Perceptual-GS are quantified by LPIPS and the number of Gaussians (millions). sian primitives in 3DGS dynamically evolves during training through adaptive density control, using the average position gradient of Gaussians to determine the need for additional primitives, enhancing the model’s capacity to capture fine details in local regions. Whi… view at source ↗
Figure 2
Figure 2. Overview of the proposed Perceptual-GS. We first construct a perception-aware representation of the scene, enabling each Gaussian primitive to adapt to the perceptual sensitivity of its represented region while constraining the number of Gaussians through perceptual sensitivity extraction and dual-branch rendering. Subsequently, we propose a perceptual sensitivity-adaptive distribution, allocating more Gaussians to … view at source ↗
Figure 3
Figure 3. Specifically, we use the Sobel operator to extract the local structure of the original RGB image I, and the horizontal and vertical gradient convolution kernels Gx and Gy are defined as: Gx =   −1 0 1 −2 0 2 −1 0 1   , Gy =   −1 −2 −1 0 0 0 1 2 1   . (5) Local Structure Extraction Perception-oriented Smoothing Perception-oriented Enhancement [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: A qualitative comparison of Bilbao in BungeeNeRF. with the baselines, and the settings for the newly introduced hyperparameters in Perceptual-GS are provided in the Ap￾pendix. To achieve a better balance between quality and efficiency, we use different weight threshold…
Figure 5
Figure 5. Figure 5: A qualitative comparison of Perceptual-GS with other methods on Stump and Treehill in Mip-NeRF 360 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: A qualitative comparison of Perceptual-GS with other methods on Amsterdam and Rome in BungeeNeRF. Ground Truth Ours Pixel-GS 3DGS Ground Truth Ours (3.89, 162) LPIPS: 0.189 3DGS (5.78, 100) LPIPS: 0.215 Pixel-GS (8.46, 59) LPIPS: 0.191 [PITH_FULL_IMAGE:figures/full_fi…
Figure 7
Figure 7. Figure 7 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: A qualitative comparison of the rendering depth between Perceptual-GS and Pixel-GS on Mip-NeRF 360 [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: The qualitative result of the proposed method is based on the vanilla 3DGS on Mip-NeRF 360. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: The qualitative result of the proposed method is based on Pixel-GS on Mip-NeRF 360 [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: The qualitative result of the proposed method is based on CoR-GS on 24-view Mip-NeRF 360. to prioritize quality. (c) Densification Interval Iterh and Iterm: To determine the optimal densification intervals, we experiment with different values of Iterh and Iterm, as sh…
Figure 12
Figure 12. Figure 12: The visualization of the effect of perceptual sensitivity map in different spatial regions. Perceptual-GS distributes more primitives to perceptually sensitive regions [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: The visualization of perceptual sensitivity maps rendered during the training process. E. Per Scene Quantitative Comparisons with State-of-the-art We present per-scene quantitative comparisons with existing methods to further illustrate the improvements in quality, ef…
Figure 14
Figure 14. Figure 14: Qualitative comparison of the reconstruction quality of low-sensitive regions between 3DGS and the proposed method. terms of both efficiency and perceptual quality, we introduce a new metric, QEB: QEB = 100 × #G × LPIPS FPS , (21) which jointly considers rendering qua…
Figure 15
Figure 15. Figure 15: Visual results of the ablation study, highlighting the impact of each module on reconstruction quality [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]

Discussion (0). Continue with ORCID 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. You Only Gaussian Once: Controllable 3D Gaussian Splatting for Ultra-Densely Sampled Scenes

    cs.CV 2026-04 unverdicted novelty 6.0 of 10

    YOGO reformulates stochastic 3D Gaussian Splatting into a deterministic budget-aware system and supplies an ultra-dense dataset to enforce physical fidelity over viewpoint interpolation.

Reference graph

Works this paper leans on

47 extracted references · 41 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    T., Mildenhall, B., Verbin, D., Srinivasan, P

    Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., and Hedman, P. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 5470--5479, 2022

    work page 2022

  3. [3]

    Campbell, F. W. and Robson, J. G. Application of Fourier Analysis to The Visibility of Gratings . The Journal of Physiology, 197 0 (3): 0 551, 1968

    work page 1968

  4. [4]

    HAC: Hash-Grid Assisted Context for 3D Gaussian Splatting Compression

    Chen, Y., Wu, Q., Lin, W., Harandi, M., and Cai, J. HAC: Hash-Grid Assisted Context for 3D Gaussian Splatting Compression . In Proceedings of the European Conference on Computer Vision, pp.\ 422--438, 2025

    work page 2025

  5. [5]

    Efficient Density Control for 3D Gaussian Splatting

    Deng, X., Diao, C., Li, M., Yu, R., and Xu, D. Efficient Density Control for 3D Gaussian Splatting . arXiv preprint arXiv:2411.10133, 2024

    arXiv 2024

  6. [6]

    MVGS: Multi-view-regulated Gaussian Splatting for Novel View Synthesis

    Du, X., Wang, Y., and Yu, X. MVGS: Multi-view-regulated Gaussian Splatting for Novel View Synthesis . arXiv preprint arXiv:2410.02103, 2024

    arXiv 2024

  7. [7]

    and Wang, B

    Fang, G. and Wang, B. Mini-Splatting: Representing Scenes with a Constrained Number of Gaussians . In Proceedings of the European Conference on Computer Vision, pp.\ 165--181, 2024

    work page 2024

  8. [8]

    VR-Splatting: Foveated Radiance Field Rendering via 3D Gaussian Splatting and Neural Points

    Franke, L., Fink, L., and Stamminger, M. VR-Splatting: Foveated Radiance Field Rendering via 3D Gaussian Splatting and Neural Points . In Proceedings of the ACM on Computer Graphics and Interactive Techniques, pp.\ 1--21, 2025

    work page 2025

Show all 47 references
  1. [9]

    Saliency-guided Quality Assessment of Screen Content Images

    Gu, K., Wang, S., Yang, H., Lin, W., Zhai, G., Yang, X., and Zhang, W. Saliency-guided Quality Assessment of Screen Content Images . IEEE Transactions on Multimedia, 18 0 (6): 0 1098--1110, 2016

    2016

  2. [10]

    Deep Blending for Free-viewpoint Image-based Rendering

    Hedman, P., Philip, J., Price, T., Frahm, J.-M., Drettakis, G., and Brostow, G. Deep Blending for Free-viewpoint Image-based Rendering . ACM Transactions on Graphics, 37 0 (6): 0 1--15, 2018

    2018

  3. [11]

    GeoTexDensifier: Geometry-Texture-Aware Densification for High-Quality Photorealistic 3D Gaussian Splatting

    Jiang, H., Xiang, X., Sun, H., Li, H., Zhou, L., Zhang, X., and Zhang, G. GeoTexDensifier: Geometry-Texture-Aware Densification for High-Quality Photorealistic 3D Gaussian Splatting . arXiv preprint arXiv:2412.16809, 2024

    2024

  4. [12]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering

    Kerbl, B., Kopanas, G., Leimk \"u hler, T., and Drettakis, G. 3D Gaussian Splatting for Real-Time Radiance Field Rendering . ACM Transactions on Graphics , 0 (4): 0 1--14, 2023

    2023

  5. [13]

    Kheradmand, S., Rebain, D., Sharma, G., Sun, W., Tseng, Y.-C., Isack, H., Kar, A., Tagliasacchi, A., and Yi, K. M. 3D Gaussian Splatting as Markov Chain Monte Carlo . In Proceedings of the Annual Conference on Neural Information Processing Systems, pp.\ 80965--80986, 2024

    2024

  6. [14]

    Color-cued Efficient Densification Method for 3D Gaussian Splatting

    Kim, S., Lee, K., and Lee, Y. Color-cued Efficient Densification Method for 3D Gaussian Splatting . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp.\ 775--783, 2024

    2024

  7. [15]

    Tanks and Temples: Benchmarking Large-scale Scene Reconstruction

    Knapitsch, A., Park, J., Zhou, Q.-Y., and Koltun, V. Tanks and Temples: Benchmarking Large-scale Scene Reconstruction . ACM Transactions on Graphics, 36 0 (4): 0 1--13, 2017

    2017

  8. [16]

    Deblurring 3D Gaussian Splatting

    Lee, B., Lee, H., Sun, X., Ali, U., and Park, E. Deblurring 3D Gaussian Splatting . In Proceedings of the European Conference on Computer Vision, pp.\ 127--143, 2025

    2025

  9. [17]

    F., Yue, Y., Lim, E

    Li, Z., Yao, S., Chu, Y., Garcia-Fernandez, A. F., Yue, Y., Lim, E. G., and Zhu, X. MVG-Splatting: Multi-View Guided Gaussian Splatting with Adaptive Quantile-Based Geometric Consistency Densification . arXiv preprint arXiv:2407.11840, 2024

    2024 arXiv

  10. [18]

    MetaSapiens: Real-Time Neural Rendering with Efficiency-Aware Pruning and Accelerated Foveated Rendering

    Lin, W., Feng, Y., and Zhu, Y. 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, pp.\ 669-...

    2025

  11. [19]

    HQGS: High-quality Novel View Synthesis with Gaussian Splatting in Degraded Scenes

    Lin, X., Luo, S., Shan, X., Zhou, X., Ren, C., Qi, L., Yang, M.-H., and Vasconcelos, N. HQGS: High-quality Novel View Synthesis with Gaussian Splatting in Degraded Scenes . In Proceedings of the International Conference on Learning Representations, pp.\ 1--17, 2025 b

    2025

  12. [20]

    LGS: A Light-Weight 4D Gaussian Splatting for Efficient Surgical Scene Reconstruction

    Liu, H., Liu, Y., Li, C., Li, W., and Yuan, Y. LGS: A Light-Weight 4D Gaussian Splatting for Efficient Surgical Scene Reconstruction . In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Intervention, pp.\ 660--670, 2024

    2024

  13. [21]

    Efficientgs: Streamlining Gaussian Splatting for Large-scale High-resolution Scene Representation

    Liu, W., Guan, T., Zhu, B., Xu, L., Song, Z., Li, D., Wang, Y., and Yang, W. Efficientgs: Streamlining Gaussian Splatting for Large-scale High-resolution Scene Representation . IEEE MultiMedia, 32 0 (1): 0 61--71, 2025

    2025

  14. [22]

    Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering

    Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., and Dai, B. Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 20654--20664, 2024

    2024

  15. [23]

    A Human Vision System Model for Objective Picture Quality Measurements

    Lubin, J. A Human Vision System Model for Objective Picture Quality Measurements . In Proceedings of the International Broadcasting Conference, pp.\ 498--503, 1997

    1997

  16. [24]

    ResGS: Residual Densification of 3D Gaussian for Efficient Detail Recovery

    Lyu, Y., Cheng, K., Kang, X., and Chen, X. ResGS: Residual Densification of 3D Gaussian for Efficient Detail Recovery . arXiv preprint arXiv:2412.07494, 2024

    2024

  17. [25]

    S., Goel, R., Kerbl, B., Steinberger, M., Carrasco, F

    Mallick, S. S., Goel, R., Kerbl, B., Steinberger, M., Carrasco, F. V., and De La Torre, F. Taming 3DGS: High-Quality Radiance Fields with Limited Resources . In SIGGRAPH Asia 2024 Conference Papers, pp.\ 1--11, 2024

    2024

  18. [26]

    P., Tancik, M., Barron, J

    Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R., and Ng, R. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis . In Proceedings of the European Conference on Computer Vision, pp.\ 405--421, 2020

    2020

  19. [27]

    Gradient Direction for Screen Content Image Quality Assessment

    Ni, Z., Ma, L., Zeng, H., Cai, C., and Ma, K.-K. Gradient Direction for Screen Content Image Quality Assessment . IEEE Signal Processing Letters, 23 0 (10): 0 1394--1398, 2016

    2016

  20. [28]

    ColNeRF: Collaboration for Generalizable Sparse Input Neural Radiance Field

    Ni, Z., Yang, P., Yang, W., Wang, H., Ma, L., and Kwong, S. ColNeRF: Collaboration for Generalizable Sparse Input Neural Radiance Field . In Proceedings of the AAAI Conference on Artificial Intelligence, pp.\ 4325--4333, 2024

    2024

  21. [29]

    T., and Mildenhall, B

    Poole, B., Jain, A., Barron, J. T., and Mildenhall, B. DreamFusion: Text-to-3D using 2D Diffusion . In Proceedings of the International Conference on Learning Representations, pp.\ 1--18, 2023

    2023

  22. [30]

    Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians

    Ren, K., Jiang, L., Lu, T., Yu, M., Xu, L., Ni, Z., and Dai, B. Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians . IEEE Transactions on Pattern Analysis and Machine Intelligence, pp.\ 1--15, 2025

    2025

  23. [31]

    and Speed, H

    Ross, J. and Speed, H. D. Contrast Adaptation and Contrast Masking in Human Vision . Proceedings of the Royal Society of London. Series B: Biological Sciences, 246 0 (1315): 0 61--70, 1991

    1991

  24. [32]

    Revising Densification in Gaussian Splatting

    Rota Bul \`o , S., Porzi, L., and Kontschieder, P. Revising Densification in Gaussian Splatting . In Proceedings of the European Conference on Computer Vision, pp.\ 347--362, 2025

    2025

  25. [33]

    Just Noticeable Distortion Profile Inference: A Patch-level Structural Visibility Learning Approach

    Shen, X., Ni, Z., Yang, W., Zhang, X., Wang, S., and Kwong, S. Just Noticeable Distortion Profile Inference: A Patch-level Structural Visibility Learning Approach . IEEE Transactions on Image Processing, 30: 0 26--38, 2020

    2020

  26. [34]

    Mega-NERF: Scalable Construction of Large-Scale NeRFs for Virtual Fly-Throughs

    Turki, H., Ramanan, D., and Satyanarayanan, M. Mega-NERF: Scalable Construction of Large-Scale NeRFs for Virtual Fly-Throughs . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12922--12931, 2022

    2022

  27. [35]

    C., Sheikh, H

    Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image Quality Assessment: from Error Visibility to Structural Similarity . IEEE Transactions on Image Processing, 13 0 (4): 0 600--612, 2004

    2004

  28. [36]

    GaussianRoom: Improving 3D Gaussian Splatting with SDF Guidance and Monocular Cues for Indoor Scene Reconstruction

    Xiang, H., Li, X., Cheng, K., Lai, X., Zhang, W., Liao, Z., Zeng, L., and Liu, X. GaussianRoom: Improving 3D Gaussian Splatting with SDF Guidance and Monocular Cues for Indoor Scene Reconstruction . arXiv preprint arXiv:2405.19671, 2024

    2024 arXiv

  29. [37]

    BungeeNeRF: Progressive Neural Radiance Field for Extreme Multi-scale Scene Rendering

    Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., and Lin, D. BungeeNeRF: Progressive Neural Radiance Field for Extreme Multi-scale Scene Rendering . In Proceedings of the European Conference on Computer Vision, pp.\ 106--122, 2022

    2022

  30. [38]

    CityDreamer: Compositional Generative Model of Unbounded 3D Cities

    Xie, H., Chen, Z., Hong, F., and Liu, Z. CityDreamer: Compositional Generative Model of Unbounded 3D Cities . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9666--9675, 2024

    2024

  31. [39]

    Pushing Rendering Boundaries: Hard Gaussian Splatting

    Xu, Q., Cui, J., Yi, X., Wang, Y., Zhou, Y., Ong, Y.-S., and Zhang, H. Pushing Rendering Boundaries: Hard Gaussian Splatting . arXiv preprint arXiv:2412.04826, 2024

    2024 arXiv

  32. [40]

    Xue, W., Zhang, L., Mou, X., and Bovik, A. C. Gradient Magnitude Similarity Deviation: A Highly Efficient Perceptual Image Quality Index . IEEE Transactions on Image Processing, 23 0 (2): 0 684--695, 2014

    2014

  33. [41]

    GSDF : 3 DGS Meets SDF for Improved Neural Rendering and Reconstruction

    Yu, M., Lu, T., Xu, L., Jiang, L., Xiangli, Y., and Dai, B. GSDF : 3 DGS Meets SDF for Improved Neural Rendering and Reconstruction . In Proceedings of the Annual Conference on Neural Information Processing Systems, pp.\ 129507--129530, 2024 a

    2024

  34. [42]

    Mip-Splatting: Alias-free 3D Gaussian Splatting

    Yu, Z., Chen, A., Huang, B., Sattler, T., and Geiger, A. Mip-Splatting: Alias-free 3D Gaussian Splatting . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 19447--19456, 2024 b

    2024

  35. [43]

    CoR-GS: Sparse-view 3D Gaussian Splatting via Co-regularization

    Zhang, J., Li, J., Yu, X., Huang, L., Gu, L., Zheng, J., and Bai, X. CoR-GS: Sparse-view 3D Gaussian Splatting via Co-regularization . In Proceedings of the European Conference on Computer Vision, pp.\ 335--352, 2024 a

    2024

  36. [44]

    FreGS: 3D Gaussian Splatting with Progressive Frequency Regularization

    Zhang, J., Zhan, F., Xu, M., Lu, S., and Xing, E. FreGS: 3D Gaussian Splatting with Progressive Frequency Regularization . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 21424--21433, 2024 b

    2024

  37. [45]

    A., Shechtman, E., and Wang, O

    Zhang, R., Isola, P., Efros, A. A., Shechtman, E., and Wang, O. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric . In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 586--595, 2018

    2018

  38. [46]

    Pixel-GS: Density Control with Pixel-Aware Gradient for 3D Gaussian Splatting

    Zhang, Z., Hu, W., Lao, Y., He, T., and Zhao, H. Pixel-GS: Density Control with Pixel-Aware Gradient for 3D Gaussian Splatting . In Proceedings of the European Conference on Computer Vision, pp.\ 326--342, 2025

    2025

  39. [47]

    Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields

    Zhou, S., Chang, H., Jiang, S., Fan, Z., Zhu, Z., Xu, D., Chari, P., You, S., Wang, Z., and Kadambi, A. Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ...

    2024

Pith tools

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