pith. sign in

arxiv: 2412.13547 · v3 · submitted 2024-12-18 · 💻 cs.CV

Turbo-GS: Accelerating 3D Gaussian Fitting for High-Quality Radiance Fields

Ankit Dhiman , Tao Lu , R Srinath , Emre Arslan , Angela Xing , Yuanbo Xiangli , R Venkatesh Babu , Srinath Sridhar This is my paper

Pith reviewed 2026-05-23 07:11 UTC · model grok-4.3

classification 💻 cs.CV
keywords 3D Gaussian Splattingnovel view synthesisradiance fieldsoptimization accelerationdensificationdilated renderinghigh-resolution fitting
0
0 comments X

The pith

Dilated rendering and dual-error densification speed up 3D Gaussian fitting for high-resolution radiance fields.

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

The paper tries to establish that 3D Gaussian Splatting models can be trained significantly faster for high-resolution images by reducing the number of pixels rendered per view and improving how new Gaussians are added during optimization. A dilated rendering technique processes only a subset of pixels, lowering computational costs. A convergence-aware budget control balances the addition of new Gaussians with the refinement of existing ones, while densification uses both positional and appearance errors to enhance efficiency and avoid gradient issues. If these changes work as intended, they enable quick 4K fitting with equal or better novel view quality than slower full-image methods. This would make high-quality 3D scene reconstruction more accessible for real-time applications.

Core claim

The central claim is that dilated rendering of only a subset of pixels, combined with a convergence-aware budget control mechanism and densification guided by both positional and appearance errors, accelerates the optimization of 3D Gaussian Splatting while preserving or improving rendering fidelity for high-resolution inputs.

What carries the argument

Dilated rendering technique that renders only a subset of pixels, along with convergence-aware budget control and dual positional-appearance error signals for densification.

If this is right

  • Optimization completes faster than prior 3DGS methods.
  • 4K-resolution scenes can be fitted quickly.
  • Novel view rendering quality stays the same or improves.
  • Densification avoids gradient vanishing through combined error signals.
  • Better balance between adding and optimizing Gaussians increases efficiency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The pixel subset approach could extend to other radiance field methods like NeRF variants.
  • Hardware acceleration might compound the speed gains in practical deployments.
  • Applying it to dynamic or very large scenes could show if context loss occurs in complex environments.

Load-bearing premise

The assumption that rendering only a dilated subset of pixels combined with dual positional-appearance error signals for densification will guide optimization to the same or better final model quality without introducing artifacts or missing scene details across varied inputs.

What would settle it

A comparison on benchmark datasets where the accelerated method produces lower PSNR or visible artifacts on test views compared to standard full-pixel 3DGS training.

Figures

Figures reproduced from arXiv: 2412.13547 by Angela Xing, Ankit Dhiman, Emre Arslan, R Srinath, R Venkatesh Babu, Srinath Sridhar, Tao Lu, Yuanbo Xiangli.

Figure 1
Figure 1. Figure 1: Turbo-GS accelerates 3DGS fitting significantly while preserving rendering quality. It proposes efficient densification strategy and innovative dilated rendering allow training on 4K images in minutes—significantly outperforming baseline methods. Notably, Turbo￾GS converges on the 4K bicycle scene in just 13 minutes—over 3×faster than Taming 3DGS (40 minutes), 14× faster than 3DGS (187 minutes) and Scaffol… view at source ↗
Figure 2
Figure 2. Figure 2: Effect of Densification Rate. This plot shows the effect of densification rate with Scaffold-GS [29] versus Turbo￾GS (Ours) on the Bicycle scene [1]. Scaffold-GS with densifi￾cation every 100 iterations (default, orange) takes time to con￾verge. An aggressive version of Scaffold-GS with densification every 20 iterations (green) initially shows improved convergence, but plateaus afterward. Ours (blue) produ… view at source ↗
Figure 4
Figure 4. Figure 4: Gradient Visualization. We rasterize the Gaussian gra￾dient into image plane and observe that: (a) Position Gradients fo￾cus only on certain regions in the scene, while (b) Color Gradients provide cues from overall regions. These are useful for regions such as grass and background structure. ing radiance field quality. Unlike other methods that aim to minimize the footprint of each optimization step [20, 3… view at source ↗
Figure 5
Figure 5. Figure 5: Loss analysis with power function fitting. For all scenes, the log(loss) is linear to the log(iterations) after the ini￾tial stage. Thus, the relation between iteration and convergence follows a power function. We design a power-law-based adaptive budget schedule based on these insights. the position and appearance gradient with τposition and τcolor respectively to determine whether densification is needed… view at source ↗
Figure 6
Figure 6. Figure 6: Dilated Rendering. Since each Gaussian affects mul￾tiple pixels in a view, dense pixel-wise supervision is redundant. Instead, we introduce a dilated rendering pipeline that selectively renders a subset of pixels in a chessboard pattern, which reduces the rendering burden while provide sufficient information for dif￾ferentiable training. Batched training To accelerate convergence in the final stage, we emp… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison with prior 3DGS-based methods and the corresponding ground truth images from testing views. We [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Convergence. We show “Number of primitives vs Step” and “PSNR vs Step” plots for scenes in MipNeRF-360 [1] dataset for with and without budget control in the optimization process. The proposed budgeting strategy prevents the number of primitives from increasing uncontrollably, while maintaining the overall quality. This is evident by the comparable PSNR plots, which demonstrate that the strategy maintains … view at source ↗
Figure 9
Figure 9. Figure 9: Impact of Dilated Rendering on time performance. We observe that dilated rendering significantly reduces the computational time required for both the (a) forward and (b) backward passes during the optimization process, compared to the without-dilated rendering approach. This highlights the efficiency of dilated rendering in accelerating the overall training process. The above results are shown for Bicycle … view at source ↗
read the original abstract

Novel-view synthesis plays a crucial role in computer vision with applications in 3D reconstruction, mixed reality, and robotics. Recent approaches, such as 3D Gaussian Splatting (3DGS), have emerged as state-of-the-art solutions, offering high-quality novel view synthesis in real time. However, training 3DGS models remains slow, particularly for high-resolution images, often requiring hours to fit a scene with 200 views. In this work, we aim to accelerate the fitting process by reducing computational overhead and improving learning efficiency. Specifically, we introduce a dilated rendering technique that renders only a subset of pixels instead of the full image, significantly reducing computational costs. To enhance learning efficiency, we develop a convergence-aware budget control mechanism that balances the addition of new Gaussians with the optimization of existing ones. Additionally, to improve densification efficiency and prevent gradient vanishing, we incorporate both positional and appearance errors to improve the effectiveness of densification. With these improvements, we achieve fast 4K-resolution fitting while maintaining, or even improving, novel view rendering quality. Extensive experiments demonstrate that our method achieves significantly faster optimization than existing approaches while preserving high rendering fidelity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes Turbo-GS to accelerate 3D Gaussian Splatting (3DGS) training for novel-view synthesis. It introduces dilated rendering (rendering only a subset of pixels), a convergence-aware budget control mechanism to balance Gaussian addition and optimization, and dual positional-appearance error signals for densification to avoid gradient vanishing. The central claim is that these changes enable fast 4K-resolution fitting while maintaining or improving rendering quality, with significantly faster optimization than existing methods across extensive experiments.

Significance. If the empirical results hold, the work would be significant for practical high-resolution radiance field applications in mixed reality and robotics by addressing the hours-long training bottleneck of 3DGS. The modifications target computational overhead and densification efficiency directly. Credit is due for focusing on engineering improvements that could scale 3DGS to 4K without new primitives or architectures.

major comments (2)
  1. [Abstract] Abstract: The central performance claim (fast 4K fitting with maintained or improved quality and significantly faster optimization) is stated without any quantitative results, error bars, ablation details, dataset descriptions, or baseline comparisons, preventing evaluation of whether the claim holds.
  2. [Abstract] The claim that dilated rendering plus dual-error densification recovers all visible high-frequency detail (fine textures, specular highlights, thin structures) rests on the unverified assumption that the chosen pixel subset and error signals supply complete gradients; no direct evidence, ablation, or failure-case analysis is supplied to confirm completeness across scene types.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments on the abstract below and will revise accordingly to improve clarity and support for the claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central performance claim (fast 4K fitting with maintained or improved quality and significantly faster optimization) is stated without any quantitative results, error bars, ablation details, dataset descriptions, or baseline comparisons, preventing evaluation of whether the claim holds.

    Authors: We agree the abstract would be stronger with quantitative support. In revision we will add concise numerical highlights (e.g., training-time speed-ups and PSNR/SSIM on Mip-NeRF 360 and Tanks & Temples) together with the main baselines and a brief note on the ablation studies, while respecting the word limit. revision: yes

  2. Referee: [Abstract] The claim that dilated rendering plus dual-error densification recovers all visible high-frequency detail (fine textures, specular highlights, thin structures) rests on the unverified assumption that the chosen pixel subset and error signals supply complete gradients; no direct evidence, ablation, or failure-case analysis is supplied to confirm completeness across scene types.

    Authors: The manuscript already contains quantitative results (Section 4) and ablations (Section 4.3) showing that quality is preserved or improved, with visual examples of fine-detail recovery. We nevertheless accept that a more explicit discussion of gradient completeness and potential failure cases would be valuable; we will add a short analysis paragraph and, if space permits, a supplementary figure addressing this point. revision: partial

Circularity Check

0 steps flagged

No circularity: engineering modifications presented without self-referential derivations

full rationale

The paper proposes three algorithmic changes (dilated rendering of pixel subsets, convergence-aware budget control, and dual positional-appearance error for densification) to accelerate 3DGS fitting. These are described as independent engineering decisions whose correctness is asserted via experiments, not via any derivation chain, uniqueness theorem, or fitted parameter renamed as prediction. No equations, self-citations, or ansatzes are shown that reduce the claimed quality preservation to the inputs by construction. The reader's assessment of score 1.0 is consistent with the absence of load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; techniques are described as modifications to the existing 3DGS pipeline.

pith-pipeline@v0.9.0 · 5767 in / 1066 out tokens · 42073 ms · 2026-05-23T07:11:10.899774+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

54 extracted references · 54 canonical work pages

  1. [1]

    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. 1, 2, 5, 6, 7, 8, 3

    work page 2022

  2. [2]

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

    David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. 2024 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 19457–19467, 2023. 3

    work page 2024

  3. [3]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InEuropean con- ference on computer vision, pages 333–350. Springer, 2022. 2

    work page 2022

  4. [4]

    Lara: Efficient large-baseline radiance fields.ArXiv, abs/2407.04699, 2024

    Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields.ArXiv, abs/2407.04699, 2024. 3

    work page arXiv 2024

  5. [5]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images.ArXiv, abs/2403.14627, 2024

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images.ArXiv, abs/2403.14627, 2024. 3

    work page arXiv 2024

  6. [6]

    Graspnerf: Multiview-based 6-dof grasp detection for transparent and specular objects using gener- alizable nerf

    Qiyu Dai, Yan Zhu, Yiran Geng, Ciyu Ruan, Jiazhao Zhang, and He Wang. Graspnerf: Multiview-based 6-dof grasp detection for transparent and specular objects using gener- alizable nerf. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1757–1763. IEEE,

    work page

  7. [7]

    Fov-nerf: Foveated neural radiance fields for virtual reality.IEEE Transactions on Visualization and Computer Graphics, 28(11):3854–3864, 2022

    Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. Fov-nerf: Foveated neural radiance fields for virtual reality.IEEE Transactions on Visualization and Computer Graphics, 28(11):3854–3864, 2022. 2

    work page 2022

  8. [8]

    Distwar: Fast differentiable rendering on raster-based ren- dering pipelines.ArXiv, abs/2401.05345, 2023

    Sankeerth Durvasula, Adrian Zhao, Fan Chen, Ruofan Liang, Pawan Kumar Sanjaya, and Nandita Vijaykumar. Distwar: Fast differentiable rendering on raster-based ren- dering pipelines.ArXiv, abs/2401.05345, 2023. 3

    work page arXiv 2023

  9. [9]

    Lightgaus- sian: 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, abs/2311.17245, 2023. 3

    work page arXiv 2023

  10. [10]

    Instantsplat: Un- bounded sparse-view pose-free gaussian splatting in 40 sec- onds

    Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, B. Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaus- sian splatting in 40 seconds.ArXiv, abs/2403.20309, 2024. 3

    work page arXiv 2024

  11. [11]

    Instantsplat: Unbounded sparse-view pose-free gaus- sian splatting in 40 seconds, 2024

    Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaus- sian splatting in 40 seconds, 2024. 2

    work page 2024

  12. [12]

    Mini-splatting: Represent- ing scenes with a constrained number of gaussians.Euro- pean Conference on Computer Vision, 2024

    Guangchi Fang and Bing Wang. Mini-splatting: Represent- ing scenes with a constrained number of gaussians.Euro- pean Conference on Computer Vision, 2024. 3, 7, 8, 2

    work page 2024

  13. [13]

    Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering.ArXiv, abs/2408.07967, 2024

    Guofeng Feng, Siyan Chen, Rong Fu, Zimu Liao, Yi Wang, Tao Liu, Zhiling Pei, Hengjie Li, Xingcheng Zhang, and Bo Dai. Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering.ArXiv, abs/2408.07967, 2024. 3

    work page arXiv 2024

  14. [14]

    Evaluating alternatives to sfm point cloud ini- tialization for gaussian splatting

    Yalda Foroutan, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasacchi. Evaluating alternatives to sfm point cloud ini- tialization for gaussian splatting. 2024. 3

    work page 2024

  15. [15]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2, 3, 8

    work page 2022

  16. [16]

    Ea- gles: Efficient accelerated 3d gaussians with lightweight en- codings.European Conference on Computer Vision, 2024

    Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Ea- gles: Efficient accelerated 3d gaussians with lightweight en- codings.European Conference on Computer Vision, 2024. 2, 3, 7, 8

    work page 2024

  17. [17]

    Antoine Gu’edon and Vincent Lepetit. Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354–5363, 2023. 3

    work page 2024

  18. [18]

    Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering

    Antoine Gu ´edon and Vincent Lepetit. Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354–5363, 2024. 2

    work page 2024

  19. [19]

    Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (ToG), 37(6):1–15, 2018

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (ToG), 37(6):1–15, 2018. 6, 8, 1, 2, 3

    work page 2018

  20. [20]

    3dgs-lm: Faster gaussian-splatting opti- mization with levenberg-marquardt.ArXiv, abs/2409.12892,

    Lukas H ¨ollein, Aljavz Bovzivc, Michael Zollhofer, and Matthias Nießner. 3dgs-lm: Faster gaussian-splatting opti- mization with levenberg-marquardt.ArXiv, abs/2409.12892,

    work page arXiv

  21. [21]

    2d gaussian splatting for geometrically ac- curate radiance fields.ArXiv, abs/2403.17888, 2024

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac- curate radiance fields.ArXiv, abs/2403.17888, 2024. 3

    work page arXiv 2024

  22. [22]

    Relaxing accurate initialization constraint for 3d gaussian splatting.ArXiv, abs/2403.09413, 2024

    Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Seungryong Kim. Relaxing accurate initialization constraint for 3d gaussian splatting.ArXiv, abs/2403.09413, 2024. 3

    work page arXiv 2024

  23. [23]

    Relu fields: The little non-linearity that could.ACM SIGGRAPH 2022 Conference Proceedings,

    Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Jyoti Mitra. Relu fields: The little non-linearity that could.ACM SIGGRAPH 2022 Conference Proceedings,

    work page 2022

  24. [24]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

    work page

  25. [25]

    A hierarchical 3d gaussian representation for real-time ren- dering of very large datasets.ACM Trans

    Bernhard Kerbl, Andr’eas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time ren- dering of very large datasets.ACM Trans. Graph., 43:62:1– 62:15, 2024. 3

    work page 2024

  26. [26]

    3d gaussian splatting as markov chain monte carlo

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Wei- wei Sun, Jeff Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo.ArXiv, abs/2404.09591, 2024. 3

    work page arXiv 2024

  27. [27]

    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. 1, 2, 3

    work page 2017

  28. [28]

    Scaling laws for diffusion transformers.arXiv preprint arXiv:2410.08184, 2024

    Zhengyang Liang, Hao He, Ceyuan Yang, and Bo Dai. Scaling laws for diffusion transformers.arXiv preprint arXiv:2410.08184, 2024. 5

    work page arXiv 2024

  29. [29]

    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, pages 20654–20664, 2024. 2, 3, 4, 5, 6, 7, 8

    work page 2024

  30. [30]

    Taming 3dgs: High-quality radiance fields with limited re- sources

    Mallick and Goel, Bernhard Kerbl, Francisco Vicente Car- rasco, Markus Steinberger, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited re- sources. InSIGGRAPH Asia 2024 Conference Papers, 2024. 2, 3, 4, 6, 7, 8

    work page 2024

  31. [31]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf. Communications of the ACM, 65:99 – 106, 2020. 1, 2, 3

    work page 2020

  32. [32]

    Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM Transactions on Graphics (TOG), 41:1 – 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), 41:1 – 15, 2022. 3

    work page 2022

  33. [33]

    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, 8

    work page 2022

  34. [34]

    K. L. Navaneet, Kossar Pourahmadi Meibodi, Soroush Ab- basi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantiza- tion. 2023. 3

    work page 2023

  35. [35]

    Coherentgs: Sparse novel view synthesis with coherent 3d gaussians.ArXiv, abs/2403.19495, 2024

    Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalan- tari. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians.ArXiv, abs/2403.19495, 2024. 3

    work page arXiv 2024

  36. [36]

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

    work page 2024

  37. [37]

    Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians.arXiv preprint arXiv:2403.17898, 2024

    Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians.arXiv preprint arXiv:2403.17898, 2024. 3

    work page arXiv 2024

  38. [38]

    Pixelwise View Selection for Un- structured Multi-View Stereo

    Johannes Lutz Sch ¨onberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise View Selection for Un- structured Multi-View Stereo. InEuropean Conference on Computer Vision (ECCV), 2016. 2

    work page 2016

  39. [39]

    Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5449–5459,

    work page 2022

  40. [40]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 6

    work page 2004

  41. [41]

    Grid-guided neural radiance fields for large urban scenes

    Linning Xu, Yuanbo Xiangli, Sida Peng, Xingang Pan, Nanxuan Zhao, Christian Theobalt, Bo Dai, and Dahua Lin. Grid-guided neural radiance fields for large urban scenes. 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 8296–8306, 2023. 3

    work page 2023

  42. [42]

    Point-nerf: Point- based neural radiance fields.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5428–5438, 2022

    Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point- based neural radiance fields.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5428–5438, 2022. 3

    work page 2022

  43. [43]

    Bakedsdf: Meshing neural sdfs for real- time view synthesis

    Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real- time view synthesis. InACM SIGGRAPH 2023 Conference Proceedings, pages 1–9, 2023. 2

    work page 2023

  44. [44]

    gsplat: An open-source library for gaussian splatting.ArXiv, abs/2409.06765, 2024

    Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jef- frey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for gaussian splatting.ArXiv, abs/2409.06765, 2024. 3

    work page arXiv 2024

  45. [45]

    Plenoctrees for real-time rendering of neural radiance fields

    Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752– 5761, 2021. 2

    work page 2021

  46. [46]

    Gsdf: 3dgs meets sdf for improved rendering and reconstruction.ArXiv, abs/2403.16964, 2024

    Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction.ArXiv, abs/2403.16964, 2024. 3

    work page arXiv 2024

  47. [47]

    Mip-splatting: Alias-free 3d gaussian splat- ting.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447–19456, 2023

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splat- ting.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447–19456, 2023. 3, 8, 2

    work page 2024

  48. [48]

    Gnfactor: Multi-task real robot learning with generalizable neural feature fields

    Yanjie Ze, Ge Yan, Yueh-Hua Wu, Annabella Macaluso, Yuying Ge, Jianglong Ye, Nicklas Hansen, Li Erran Li, and Xiaolong Wang. Gnfactor: Multi-task real robot learning with generalizable neural feature fields. InConference on Robot Learning, pages 284–301. PMLR, 2023. 2

    work page 2023

  49. [49]

    Gs-lrm: Large reconstruction model for 3d gaussian splatting.ArXiv, abs/2404.19702, 2024

    Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting.ArXiv, abs/2404.19702, 2024. 3

    work page arXiv 2024

  50. [50]

    Gs-lrm: Large recon- struction model for 3d gaussian splatting.European Confer- ence on Computer Vision, 2024

    Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large recon- struction model for 3d gaussian splatting.European Confer- ence on Computer Vision, 2024. 2

    work page 2024

  51. [51]

    The unreasonable effectiveness of 10 deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of 10 deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 6

    work page 2018

  52. [52]

    Long-lrm: Long- sequence large reconstruction model for wide-coverage gaussian splats, 2024

    Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yi- cong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long- sequence large reconstruction model for wide-coverage gaussian splats, 2024. 2 11 Turbo-GS: Accelerating 3D Gaussian Fitting for High-Quality Radiance Fields Supplementary Material

    work page 2024

  53. [53]

    More Implementation Details For all the dataset with a resolution below 4K, we train it for10kiterations

    Implementation Details 7.1. More Implementation Details For all the dataset with a resolution below 4K, we train it for10kiterations. The maximum budget is set to300kor 500kfor low resolution dataset,700kfor 4K and higher res- olution dataset. The batched training is activated in the last 50 iterations, with a batch size of 4. We calculate the aver- age l...

    work page

  54. [54]

    Number of primitives vs Step

    More Experiments and Results Per-scene ResultsHere we list the error metrics used in our evaluation in Sec.4 across all considered methods and scenes, as shown in Tab 5- 8.drjohnson-playroom[19] belongs to the deep blending dataset;train-truckcome from the Tanks and Temple [27] dataset;bicycle-boonsai are from MipNeRF360 [1]. Dilated RenderingThe effectiv...

    work page