Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Decomposing Densification in Gaussian Splatting for Faster 3D Scene Reconstruction

T0 review · 4 major / 6 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 In 3D Gaussian splatting, split spreads Gaussians globally and clone refines locally, so scheduling split-only low-resolution training before full-resolution cloning roughly halves training time with comparable or better quality.

desk verdict A solid 2x speedup recipe for 3DGS with an overclaimed causal story; the split/clone analysis is confounded by the policy, but the engineering result stands. read the letter →

arxiv 2507.20239 v1 pith:DOPHC7Q7 submitted 2025-07-27 cs.CV

Binxiao Huang , Zhengwu Liu , Ngai Wong This is my paper
classification cs.CV
keywords 3DGaussianSplattingdensificationsplitoperationcloneglobal-to-localtrainingmulti-resolutionschedulingenergy-guidedresolutionopacitypruning
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

The paper dissects the two densification operations in 3D Gaussian Splatting training — split and clone — and argues they have distinct roles: split moves Gaussians out to cover the scene, clone packs them in to fit detail. It then builds a training recipe on that distinction: first train only with splits at low resolution to get global coverage, then switch to full resolution with clones for local refinement. An energy-guided scheduler sets how long each resolution is used, and an adaptive opacity threshold prunes redundant Gaussians. On three real-world datasets, the authors report roughly a 2x training speedup, about 40% fewer Gaussians than their accelerated baseline, and equal or better reconstruction quality. The point of the work is that understanding what densification actually does lets you spend computation where it matters.

What carries the argument

The load-bearing object is the pair of densification operations in 3D Gaussian Splatting — split, which replaces a large Gaussian with smaller ones, and clone, which copies a small Gaussian to its position. The paper's mechanism is the two-phase global-to-local schedule that decouples them: a split-only, low-resolution phase for scene-wide diffusion, followed by a full-resolution phase with both operations for detail. The energy-aware resolution scheduler computes how many training iterations to spend at each downsampled scale from the 2D image's Fourier energy density, and the adaptive opacity threshold prunes low-opacity Gaussians. Together these make densification spread first and refine later.

What would settle it

Train the same scenes with the operation assignments swapped — split small Gaussians and clone large ones whenever the gradient triggers — and measure final displacements; if clone-dominated Gaussians then show the large global displacements, the claim that splitting inherently spreads the scene is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the split operation takes charge of the global spread of Gaussian primitives while the clone operation governs local refinement. The paper's evidence is that, after training, split-dominated Gaussians end up roughly twenty times farther from their original structure-from-motion parent point than clone-dominated ones, and clone-dominated displacements are under 2% of the scene size, while about 80% of new primitives come from cloning. Based on this, the authors propose a global-to-local densification strategy that uses only splits during a low-resolution global spread phase and then reintroduces clones during full-resolution local refinement, together with an energy-guided multi-resolution scheduler and an adaptive opacity pruning mechanism. Compared with the accelerated 3DGS baseline, they report over 2x faster training with fewer Gaussian primitives and comparable or better PSNR, SSIM, and LPIPS.

Load-bearing premise

The account assumes that the larger displacements of split-dominated Gaussians come from what splitting does, rather than from the fact that the standard training rule only splits big Gaussians and only clones small ones.

Editorial extensions

If this is right

  • Training 3D Gaussian Splatting scenes can be roughly halved in wall-clock time while keeping or improving PSNR and SSIM and using fewer Gaussians, which lowers memory and rendering cost.
  • The energy-guided multi-resolution schedule sets the transition between global spread and local refinement per scene based on image content rather than by hand.
  • Adaptive opacity pruning with a percentile threshold removes redundant Gaussians, and the ablations show the full recipe, not any single component, delivers the largest speedup.
  • The split/clone decomposition provides an explanation for the cluster artifacts noted in prior work: early cloning over-concentrates Gaussians locally.
  • On a fixed 18K iteration budget, the method beats an aggressive-densification baseline in quality while training slightly faster.

Reading between the lines

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

  • Editorial inference: if the causal reading holds, the split/clone distinction predicts that any training schedule postponing local densification should achieve similar speedups, making the two-phase split/clone choice one instance of a broader curriculum principle.
  • Editorial inference: the displacement statistics are collected under the standard policy that splits large Gaussians and clones small ones, so the observed roles may reflect the policy rather than the operations; a randomized assignment test would separate the two.
  • Editorial inference: the energy-density resolution scheduler depends only on 2D image statistics, so it could transfer to other primitive-based radiance field methods or to streaming reconstruction with minimal changes.
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 / 6 minor

Summary. This paper studies the split and clone densification operations in 3D Gaussian Splatting. It claims that split is responsible for global spatial diffusion while clone is responsible for local refinement, based on a post-hoc displacement table. It then proposes a two-phase 'global-to-local' training schedule: split-only densification at low resolution, followed by clone-enabled full-resolution refinement, plus an energy-guided coarse-to-fine resolution schedule and an adaptive opacity pruning rule. Experiments on MipNeRF-360, Deep Blending, and Tanks & Temples report about 2x training speedup over an accelerated 3DGS baseline with fewer Gaussians and approximately equal or better PSNR/SSIM.

Significance. If the causal decomposition were established, the paper would offer a simple and practical acceleration recipe for 3DGS training, and its three-dataset evaluation with ablations and hyperparameter studies is a useful engineering contribution. The main speedup and Gaussian-count reductions are consistently reported across datasets. However, the central conceptual claim is currently supported only by an uncontrolled post-hoc comparison, and the ablation table shows that the headline quality is restored only when all components are combined. The paper is therefore valuable as an empirical system paper, but its explanatory narrative outruns the evidence.

major comments (4)
  1. [Methodology, 'Spatial diffusion' (Table 1)] The central claim that split operations drive global spread and clone operations drive local refinement is not established by Table 1. Under the standard adaptive densification policy described in the Preliminary, a primitive is split precisely when its maximum scale exceeds a threshold and cloned otherwise. Split-dominated Gaussians are therefore descendants of initially large Gaussians, which project to many pixels, accumulate larger position gradients, and are moved by the optimizer, while clone-dominated Gaussians descend from small Gaussians. The roughly twenty-fold displacement difference may be a selection effect of the triggering policy rather than an intrinsic property of the operations. A controlled experiment is needed, for example comparing displacements when split and clone are applied to matched parent populations, or measuring displacement under split-only versus clone-only policies from identical initializations. Without this, the two-phase split-only design, whose stated motivation is this causal claim, lacks quantitative support.
  2. [Experiments, 'Ablation studies' (Table 4)] Table 4 shows that adding the global-to-local strategy alone to the backbone degrades quality: SSIM falls from 0.8213 to 0.8066, PSNR from 27.57 to 27.47, and LPIPS from 0.2095 to 0.2235. The paper describes this as a 'slight degradation', but it is a substantial drop on SSIM and LPIPS, and it undermines the claim that G2L is a quality-neutral accelerator. The full-method quality parity with the baseline is produced only in combination with the coarse-to-fine schedule and pruning. The authors should either temper the causal claims or provide a controlled version of G2L that does not sacrifice quality.
  3. [Experiments, 'Ablation studies' (Table 4)] The full factorial interaction is missing. C2F alone improves SSIM/PSNR (0.8246/27.84) over the backbone, while G2L+C2F is worse in SSIM and LPIPS (0.8176/0.2203 vs 0.8246/0.2202), and the full model adds pruning to recover quality (0.8257/27.79). This pattern suggests that G2L's contribution is mostly computational and may even hurt quality unless compensated. The paper should report the contribution of each component to the reported 2x speedup separately (e.g., wall-clock time per component), and discuss the negative interaction explicitly.
  4. [Coarse-to-fine multi-resolution densification, Eq. (6)] The energy-guided schedule is not validated against alternative schedules. The paper does not compare Eq. (6) with a fixed or uniformly spaced resolution schedule, nor does it show that the energy ratios E_r/E_1 produce the intended phase boundary. In addition, E(I) as defined is the magnitude of the Fourier spectrum, not an 'energy density' in the standard signal-processing sense, and the r^2 normalization in Eq. (5) is introduced without derivation. Please clarify the terminology and provide a comparison to simpler schedules to justify this mechanism.
minor comments (6)
  1. [Throughout] Section cross-references appear empty in the text ('In Sec. ,', 'cf. Sec. '); these should be filled or removed.
  2. [Introduction, Contributions] The contributions paragraph uses 'global-to-fine densification' while the rest of the paper uses 'global-to-local'; please unify the terminology.
  3. [Table 2] The arrow on the Time column is 'up' although lower training time is better; this is inconsistent with NGS down and should be corrected.
  4. [Fig. 3] The axes and the definition of the clone/split ratio are not described in the caption; please add units and a precise definition.
  5. [Methodology, 'Split-clone ratio'] The statement that approximately 80% of new primitives originate from cloning should specify whether this is averaged over iterations and scenes and how it is computed from Fig. 3.
  6. [Conclusion] There are several typographical issues: 'Subsequenctly', 'We first revealed that', 'we claims two statements', and 'preview researches' should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the 2x speedup claim is established by external benchmarking, and the split/clone analysis is an empirical observation rather than a fitted input.

full rationale

The paper's headline claim—about 2x training speedup with fewer Gaussians and comparable or better quality—is established by direct benchmark comparisons in Table 2 against external baselines, not derived from the analysis of split and clone. The energy-guided iteration allocation Tr in Eq. 6 is computed from image energy spectra (Er/E1), with no fitted dependence on the quality or timing metrics reported later. The split/clone behavioral claim in Table 1 is an empirical trajectory statistic under the default 3DGS densification policy; although the scale-based trigger (split large Gaussians, clone small ones) may confound the causal attribution, the claim is not identical to its input definitions and does not by itself determine the measured speedup. The two-phase design is validated by the ablation in Table 4, which separately tests G2L, C2F, pruning, and the full combination against the 3DGS-accel backbone. There are no load-bearing self-citations: none of the cited works are authored by the present authors, and the cited DashGaussian is used only as design inspiration, not as evidence for the method's correctness. Hyperparameter selection (Tables 5 and 6) is standard ablation practice and does not make the benchmarked result a tautology. The conclusion's limitation statement about blur is an honest scope restriction. Overall, no step in the derivation reduces to its own inputs by construction.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the empirical split/clone attribution and on the energy-based resolution scheduler. The energy scheduler adds hand-chosen hyperparameters (K, p, tau_u, Tdensify) and a domain assumption that downsampled image energy predicts training difficulty; the split/clone analysis adds an ad hoc assumption that final displacement measures causal role. No new physical entities are introduced.

free parameters (4)
  • Max scale factor K = 8
    From implementation details; controls how many downsampling levels are used in the energy-based scheduler; chosen without a principled derivation.
  • Pruning ratio p = 0.03
    From implementation details and Table 6; controls adaptive opacity threshold; tuned on MipNeRF-360.
  • Pruning upper limit tau_u = 0.05
    From implementation details and Table 6; caps the adaptive pruning threshold; tuned for speed/quality trade-off.
  • Densification iterations Tdensify = 25K
    From implementation details and Table 5; sets total densification length; tuned for speed/quality trade-off.
assumptions (3)
  • domain assumption Bilinear downsampling with anti-aliasing (Eq. 4) preserves the energy information needed to choose training resolution schedule.
    Used in Sec. 4.3 to compute Er and allocate iterations; no validation that this proxy tracks optimization difficulty.
  • domain assumption The accumulated gradient norm thresholding from 3DGS remains appropriate when training is done at downsampled resolutions.
    The method keeps the standard gradient-based densification trigger while changing image resolution and phase; not re-derived.
  • ad hoc to paper Final Gaussian displacement relative to parent points is a valid measure of an operation's role in spatial diffusion.
    Introduced in Sec. 4.1 'Spatial diffusion'; the classification into split/clone-dominated and the displacement comparison assume this attribution is meaningful, but selection effects are not controlled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decomposing Densification in Gaussian Splatting for Faster 3D Scene Reconstruction." pith.science (2026). https://pith.science/paper/DOPHC7Q7

@misc{pith2026250720239,
  author       = {Pith},
  title        = {Pith review of: Decomposing Densification in Gaussian Splatting for Faster 3D Scene Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DOPHC7Q7}},
  note         = {Machine review of arXiv:2507.20239}
}
read the original abstract

3D Gaussian Splatting (GS) has emerged as a powerful representation for high-quality scene reconstruction, offering compelling rendering quality. However, the training process of GS often suffers from slow convergence due to inefficient densification and suboptimal spatial distribution of Gaussian primitives. In this work, we present a comprehensive analysis of the split and clone operations during the densification phase, revealing their distinct roles in balancing detail preservation and computational efficiency. Building upon this analysis, we propose a global-to-local densification strategy, which facilitates more efficient growth of Gaussians across the scene space, promoting both global coverage and local refinement. To cooperate with the proposed densification strategy and promote sufficient diffusion of Gaussian primitives in space, we introduce an energy-guided coarse-to-fine multi-resolution training framework, which gradually increases resolution based on energy density in 2D images. Additionally, we dynamically prune unnecessary Gaussian primitives to speed up the training. Extensive experiments on MipNeRF-360, Deep Blending, and Tanks & Temples datasets demonstrate that our approach significantly accelerates training,achieving over 2x speedup with fewer Gaussian primitives and superior reconstruction performance.

Figures

Figures reproduced from arXiv: 2507.20239 by the authors.

Figure 1
Figure 1. Pipeline of the global-to-local and coarse-to-fine [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the distribution of Gaussian primitives (left) and the rendered images (right) after optimization . [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The number of Gaussian primitives generated through split and clone operations and the ratio of clone to split during [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison between our method and prior 3DGS approaches, along with the corresponding ground truth [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results for small and distant object re [PITH_FULL_IMAGE:figures/full_fig_p007_5.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. Geometry and Gradient-based Partitioning for Panoramic Outdoor Reconstruction

    cs.CV 2026-07 accept novelty 6.0 of 10

    A geometry and gradient-based partitioning strategy enables scalable block-wise 3D Gaussian Splatting for large-scale panoramic outdoor scenes.

Reference graph

Works this paper leans on

39 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

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

    Barron, J. T.; Mildenhall, B.; Verbin, D.; Srinivasan, P. P.; and Hedman, P. 2022. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 5470--5479

    2022

  2. [2]

    Cha, H.; Lee, I.; and Joo, H. 2024. PERSE: Personalized 3D Generative Avatars from A Single Portrait. arXiv preprint arXiv:2412.21206

    arXiv 2024

  3. [3]

    Chao, B.; Tseng, H.-Y.; Porzi, L.; Gao, C.; Li, T.; Li, Q.; Saraf, A.; Huang, J.-B.; Kopf, J.; Wetzstein, G.; et al. 2024. Textured Gaussians for Enhanced 3D Scene Appearance Modeling. arXiv preprint arXiv:2411.18625

    arXiv 2024

  4. [4]

    Chen, Y.; Jiang, J.; Jiang, K.; Tang, X.; Li, Z.; Liu, X.; and Nie, Y. 2025. DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds. arXiv preprint arXiv:2503.18402

    work page Pith review arXiv 2025

  5. [5]

    M.; Ivanovic, B.; Litany, O.; Gojcic, Z.; Fidler, S.; Pavone, M.; et al

    Chen, Z.; Yang, J.; Huang, J.; de Lutio, R.; Esturo, J. M.; Ivanovic, B.; Litany, O.; Gojcic, Z.; Fidler, S.; Pavone, M.; et al. 2024. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760

    arXiv 2024

  6. [6]

    Cheng, K.; Long, X.; Yang, K.; Yao, Y.; Yin, W.; Ma, Y.; Wang, W.; and Chen, X. 2024. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning

    work page 2024

  7. [7]

    Cong, W.; Zhu, H.; Wang, K.; Lei, J.; Stearns, C.; Cai, Y.; Wang, D.; Ranjan, R.; Feiszli, M.; Guibas, L.; et al. 2025. VideoLifter: Lifting Videos to 3D with Fast Hierarchical Stereo Alignment. arXiv preprint arXiv:2501.01949

    arXiv 2025

  8. [8]

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

    arXiv 2024

Show all 39 references
  1. [9]

    Fang, G.; and Wang, B. 2024 a . Mini-Splatting: Representing Scenes with a Constrained Number of Gaussians. In Leonardis, A.; Ricci, E.; Roth, S.; Russakovsky, O.; Sattler, T.; and Varol, G., eds., Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September...

    2024

  2. [10]

    Fang, G.; and Wang, B. 2024 b . Mini-Splatting2: Building 360 Scenes within Minutes via Aggressive Gaussian Densification. arXiv preprint arXiv:2411.12788

    2024

  3. [11]

    Girish, S.; Gupta, K.; and Shrivastava, A. 2023. EAGLES: Efficient Accelerated 3D Gaussians with Lightweight EncodingS. CoRR, abs/2312.04564

    2023 arXiv

  4. [12]

    Hanson, A.; Tu, A.; Lin, G.; Singla, V.; Zwicker, M.; and Goldstein, T. 2024. Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives. arXiv preprint arXiv:2412.00578

    2024

  5. [13]

    Hedman, P.; Philip, J.; Price, T.; Frahm, J.; Drettakis, G.; and Brostow, G. J. 2018. Deep blending for free-viewpoint image-based rendering. ACM Trans. Graph. , 37(6): 257

    2018

  6. [14]

    Hess, G.; Lindstr \"o m, C.; Fatemi, M.; Petersson, C.; and Svensson, L. 2024. SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving. arXiv preprint arXiv:2411.16816

    2024 arXiv

  7. [15]

    J.; Dumery, C.; Zhao, C.; and Salzmann, M

    Javed, S.; Khan, A. J.; Dumery, C.; Zhao, C.; and Salzmann, M. 2024. Temporally Compressed 3D Gaussian Splatting for Dynamic Scenes. arXiv preprint arXiv:2412.05700

    2024

  8. [16]

    Jiang, Y.; Shen, Z.; Hong, Y.; Guo, C.; Wu, Y.; Zhang, Y.; Yu, J.; and Xu, L. 2024. Robust dual gaussian splatting for immersive human-centric volumetric videos. ACM Transactions on Graphics (TOG), 43(6): 1--15

    2024

  9. [17]

    Kerbl, B.; Kopanas, G.; Leimk \" u hler, T.; and Drettakis, G. 2023. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Trans. Graph. , 42(4): 139:1--139:14

    2023

  10. [18]

    Knapitsch, A.; Park, J.; Zhou, Q.-Y.; and Koltun, V. 2017. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4): 1--13

    2017

  11. [19]

    Lei, X.; Wang, M.; Zhou, W.; and Li, H. 2025. Gaussnav: Gaussian splatting for visual navigation. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2025

  12. [20]

    Lyu, W.; Zhou, Y.; Yang, M.-H.; and Shu, Z. 2024. FaceLift: Single Image to 3D Head with View Generation and GS-LRM. arXiv preprint arXiv:2412.17812

    2024 arXiv

  13. [21]

    S.; Goel, R.; Kerbl, B.; Carrasco, F

    Mallick, S. S.; Goel, R.; Kerbl, B.; Carrasco, F. V.; Steinberger, M.; and la Torre, F. D. 2024. Taming 3DGS: High-Quality Radiance Fields with Limited Resources. CoRR, abs/2406.15643

    2024 arXiv

  14. [22]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2020. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In Vedaldi, A.; Bischof, H.; Brox, T.; and Frahm, J., eds., Computer Vision - ECCV 2020 - 16th European Confe...

    2020

  15. [23]

    Oechsle, M.; Peng, S.; and Geiger, A. 2021. UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021 , 5569--5579. IEEE

    2021

  16. [24]

    T.; Bouaziz, S.; Goldman, D

    Park, K.; Sinha, U.; Barron, J. T.; Bouaziz, S.; Goldman, D. B.; Seitz, S. M.; and Martin-Brualla, R. 2021. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, 5865--5874

    2021

  17. [25]

    Tan, B.; Yu, R.; Shen, Y.; and Xue, N. 2024. PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes. arXiv preprint arXiv:2412.03451

    2024 arXiv

  18. [26]

    K.; Tabor, J.; and Spurek, P

    Waczynska, J.; Borycki, P.; Tadeja, S. K.; Tabor, J.; and Spurek, P. 2024. GaMeS: Mesh-Based Adapting and Modification of Gaussian Splatting. CoRR, abs/2402.01459

    2024 arXiv

  19. [27]

    Wang, P.; Liu, L.; Liu, Y.; Theobalt, C.; Komura, T.; and Wang, W. 2021. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction. In Ranzato, M.; Beygelzimer, A.; Dauphin, Y. N.; Liang, P.; and Vaughan, J. W., eds., Advances in Neural Informat...

    2021

  20. [28]

    C.; Sheikh, H

    Wang, Z.; Bovik, A. C.; Sheikh, H. R.; and Simoncelli, E. P. 2004. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4): 600--612

    2004

  21. [29]

    Xie, T.; Chen, X.; Xu, Z.; Xie, Y.; Jin, Y.; Shen, Y.; Peng, S.; Bao, H.; and Zhou, X. 2024. Envgs: Modeling view-dependent appearance with environment gaussian. arXiv preprint arXiv:2412.15215

    2024 arXiv

  22. [30]

    Xu, J.; Mei, Y.; and Patel, V. 2024. Wild-gs: Real-time novel view synthesis from unconstrained photo collections. Advances in Neural Information Processing Systems, 37: 103334--103355

    2024

  23. [31]

    Yan, Y.; Lin, H.; Zhou, C.; Wang, W.; Sun, H.; Zhan, K.; Lang, X.; Zhou, X.; and Peng, S. 2024. Street Gaussians for Modeling Dynamic Urban Scenes. CoRR, abs/2401.01339

    2024 arXiv

  24. [32]

    Yang, C.; Li, S.; Fang, J.; Liang, R.; Xie, L.; Zhang, X.; Shen, W.; and Tian, Q. 2024. GaussianObject: Just Taking Four Images to Get A High-Quality 3D Object with Gaussian Splatting. CoRR, abs/2402.10259

    2024 arXiv

  25. [33]

    Yariv, L.; Gu, J.; Kasten, Y.; and Lipman, Y. 2021. Volume Rendering of Neural Implicit Surfaces. In Ranzato, M.; Beygelzimer, A.; Dauphin, Y. N.; Liang, P.; and Vaughan, J. W., eds., Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information...

    2021

  26. [34]

    Ye, Z.; Li, W.; Liu, S.; Qiao, P.; and Dou, Y. 2024. Absgs: Recovering fine details in 3d gaussian splatting. In Proceedings of the 32nd ACM International Conference on Multimedia, 1053--1061

    2024

  27. [35]

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

    Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, 586--595

    2018

  28. [36]

    Zhou, H.; Shao, J.; Xu, L.; Bai, D.; Qiu, W.; Liu, B.; Wang, Y.; Geiger, A.; and Liao, Y. 2024. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21336--21345

    2024

  29. [37]

    J.; Lattas, A.; Kopanas, G.; Gotardo, P.; Beeler, T.; Thies, J.; and Bolkart, T

    Zielonka, W.; Garbin, S. J.; Lattas, A.; Kopanas, G.; Gotardo, P.; Beeler, T.; Thies, J.; and Bolkart, T. 2025. Synthetic Prior for Few-Shot Drivable Head Avatar Inversion. arXiv preprint arXiv:2501.06903

    2025 arXiv

  30. [38]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  31. [39]

    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 gl...

Pith tools

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