Pith. sign in

REVIEW 2 major objections 5 minor 39 references

TemporalGS reuses geometry and color from the previous frame so 3D Gaussian Splatting only re-renders the tiles that actually changed, cutting work by up to 1.48 imes with no retraining.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 02:49 UTC pith:RRNNLYW6

load-bearing objection Solid training-free plug-in that reuses warped geometry/appearance buffers for 3DGS culling and selective tiles; real FPS gains on large scenes, quality trade-off transparent, but numbers are motion-regime dependent. the 2 major comments →

arxiv 2607.03390 v1 pith:RRNNLYW6 submitted 2026-07-03 cs.CV cs.GR

TemporalGS: Training-Free Plug-and-Play Acceleration for 3D Gaussian Splatting Rendering via Temporal Priors

Yuhongze Zhou , Zihao Yang , Xinxin Zuo , Juwei Lu This is my paper
classification cs.CV cs.GR
keywords 3D Gaussian Splattingtemporal coherencerendering accelerationtraining-freeplug-and-playtile-based rasterizationocclusion culling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

3D Gaussian Splatting already renders novel views quickly, but every frame still preprocesses, sorts, and alpha-blends every Gaussian that intersects the camera frustum. TemporalGS treats consecutive frames as highly overlapping: it stores a depth map, a depth-variance map, and a color image from a recent reference frame, warps those buffers into the new viewpoint, then discards Gaussians that lie outside an adaptive frustum or behind the warped surface. Only the tiles that cannot be safely approximated by the warped color image are fully re-rendered and stitched back in. Because the whole pipeline is a pure run-time algorithm, it can be dropped into any existing 3DGS codebase—software or hardware rasterizer—without fine-tuning or post-processing. On both small indoor scenes and city-scale reconstructions the method delivers speed-ups that match or exceed pruning-based competitors while keeping PSNR, SSIM and LPIPS competitive.

Core claim

A pair of temporal buffers (geometry and appearance) plus two online policies—temporal dynamic culling of Gaussians and selective tile rendering—suffice to accelerate any tile-based 3DGS pipeline by up to 1.48 imes without training or scene-specific post-processing, while image quality remains competitive with state-of-the-art pruning and visibility-filtering methods.

What carries the argument

Temporal priors: a temporal geometry buffer (depth + standard-deviation maps) and a temporal appearance buffer (color image) that are forward-warped into the next view and then used both to cull Gaussians and to decide which 16×16 tiles must be re-rasterized.

Load-bearing premise

Neighboring frames share enough geometric and appearance overlap that a simple warp of one reference depth-and-color pair will leave only a modest fraction of tiles needing full re-render.

What would settle it

Measure selected-tile percentage and PSNR drop on a continuous camera path whose inter-frame translation or rotation is large enough that more than roughly 40 % of tiles fail the warp test; if FPS gain collapses or quality falls far below the paper’s reported numbers, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any existing 3DGS or CityGS checkpoint can be accelerated at test time simply by inserting the TemporalGS module.
  • Hardware-rasterization pipelines (OpenGL/Vulkan/WebGPU) can reuse the same culling and tile-selection logic and still obtain measurable speed-ups.
  • Pruning or compression methods that already reduce Gaussian count remain complementary: TemporalGS can be stacked on top of them for further gains.
  • Adaptive reference-frame insertion keeps accumulated warp error bounded without requiring offline visibility precomputation.

Where Pith is reading between the lines

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

  • The same warp-and-select idea could be applied to 4D Gaussian sequences once temporal coherence is measured across time as well as viewpoint.
  • A lightweight residual network trained only on the stitched seams could recover the small quality gap that remains after pure geometric warping.
  • Because the algorithm never modifies the underlying Gaussian parameters, it is immediately usable for interactive multi-user roaming where each client has a slightly different camera path.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. TemporalGS is a training-free, plug-and-play acceleration layer for 3D Gaussian Splatting (and several of its variants). Instead of rendering each frame independently, it caches temporal geometry and appearance buffers (TGB/TAB) from a reference frame, then interleaves two strategies: temporal dynamic culling (adaptive frustum culling + temporal occlusion culling against a warped min-depth / max-σ surface band) and selective tile rendering (only tiles that fail occlusion or correspondence tests are sorted and rasterized; the rest are filled by warped TAB). An adaptive reference-frame insertion rule (η = 40 % selected tiles) bounds error accumulation. The method is implemented on top of tile-based software rasterization and also ported to two hardware-rasterization pipelines. On Mip-NeRF-360 and five large-scale aerial/street scenes the authors report up to 1.48× FPS gains relative to the host 3DGS method while keeping PSNR/SSIM/LPIPS competitive with (or better than) several post-training and post-processing baselines; ablations isolate the contribution of each module and motion-speed / hyper-parameter sweeps are provided.

Significance. If the reported speed/quality trade-off holds under realistic camera motion, TemporalGS supplies a genuinely training-free, drop-in acceleration path that is complementary to both Gaussian pruning and hardware-specific rasterization optimizations. The explicit construction of reusable geometry/appearance buffers, the adaptive reset rule, and the demonstration that the same algorithmic skeleton ports to both CUDA software and hardware rasterizers are concrete engineering contributions. The evaluation covers multiple public large-scale scenes and several independently re-implemented baselines, which strengthens the claim of generality. The work therefore has clear practical value for interactive 3DGS viewers and for any pipeline that already ships a trained Gaussian set.

major comments (2)
  1. Section III-C–E and Table V: the headline 1.48× figure (and the large-scale averages in Table I) rest on the premise that a single reference TGB/TAB plus a fixed surface band (λ1=3, λ2=0) and 3×3 depth fill keep the selected-tile fraction low. Table V already shows that larger interpolation intervals raise SR-tile % (Kitchen 33 %→64 %) and cut FPS while PSNR is preserved only because more pseudo-references are inserted. The limitation paragraph (Sec. V) acknowledges that the similarity assumption fails under fast motion and that high η can produce flickering. The manuscript should therefore either (a) report the same tables under a controlled high-motion protocol (e.g., real user trajectories with measured angular/linear velocity bins) or (b) qualify the abstract/claim language so that the acceleration is explicitly conditioned on moderate-overlap regimes.
  2. Section IV-C.1 / Tables I & III: quality metrics (PSNR/SSIM/LPIPS) are computed only on the original sparse test views, while FPS is measured on dense synthetic or user trajectories that interleave many TemporalGS frames. Because the adaptive reset inserts full (or AFC-only) reference frames precisely when selected-tile count exceeds η, the reported quality numbers do not fully capture the visual quality of the accelerated frames that dominate the FPS statistic. A trajectory-averaged quality metric (or at least a side-by-side of reference vs. TemporalGS frames along the same path) is needed to substantiate the “competitive rendering quality” claim for the operating regime that actually produces the speed-up.
minor comments (5)
  1. Eqs. (1)–(2) and the subsequent warping description: the notation for the warped buffers (−→Dmin, −→σmax) is dense; a short algorithmic box or pseudocode for the TGB construction and the two culling tests would improve reproducibility.
  2. Table II: web-splat fails on MatrixCity (>20 M Gaussians); the caption should state whether this is an out-of-memory or an API limitation so that readers can interpret the missing entries.
  3. Section III-E: the α-contribution threshold 1/255 for newly-visible Gaussians is stated without sensitivity analysis; a one-line ablation (or a pointer to the supplement) would be useful.
  4. Related-work placement of concurrent temporal methods (NeoG, GS-Cache) is adequate, but the text could more clearly distinguish algorithmic reuse (sorting reuse vs. geometry/appearance buffer reuse) so that the novelty claim is easier to locate.
  5. Minor typographical issues: “DISTW AR”, “F ormulation”, and occasional missing spaces after periods appear in the related-work and approach sections.

Circularity Check

0 steps flagged

No circularity: TemporalGS is an empirical engineering construction whose speed/quality claims are measured on external datasets, not derived by reducing outputs to fitted inputs or self-citation.

full rationale

The paper presents a training-free plug-and-play acceleration pipeline (temporal geometry/appearance buffers via Eqs. 1–3, adaptive frustum + occlusion culling, selective tile rendering with η=40% reset) whose claimed gains (up to 1.48× FPS, competitive PSNR/SSIM/LPIPS) are obtained by direct measurement against independently re-implemented baselines on public Mip-NeRF-360, MatrixCity, Residence, Building, Rubble, Small City and OccluScene3D trajectories. Hyper-parameters (λ1=3, λ2=0, λTDC=50, λSR=5, η=40%) are fixed once and never fitted to the reported metrics; ablations (Tables IV–VI) and motion-speed sweeps (Table V) simply quantify sensitivity rather than force the headline numbers by construction. No uniqueness theorem, self-citation chain, or renaming of a known result underpins the central claim. The limitation paragraph openly notes that the temporal-overlap premise can fail under fast motion, confirming that the results are conditional empirical observations rather than tautological predictions. Consequently the derivation chain contains no circular step.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 2 invented entities

The paper is an engineering algorithm paper. Its load-bearing content consists of a handful of hand-chosen hyper-parameters that control culling aggressiveness and reference-frame frequency, plus the domain assumption that consecutive camera poses produce highly overlapping Gaussian sets and warpable surfaces. No new physical entities are postulated; the buffers and culling rules are pure algorithmic constructs.

free parameters (5)
  • λ_TDC (TDC frame interval) = 50
    Fixed to 50; controls how many frames use temporal dynamic culling before a new reference.
  • λ_SR (SR frame interval) = 5
    Fixed to 5; controls interleaving of selective rendering.
  • η (max selected-tile percentage) = 40%
    Fixed to 40 %; triggers adaptive insertion of a pseudo-reference frame when too many tiles need re-rendering.
  • λ1, λ2 (surface band multipliers on σ) = 3 and 0
    Set to 3 and 0 to define the occlusion/visibility band around warped depth; chosen by authors.
  • α contribution threshold for newly-visible Gaussians = 1/255
    Fixed to 1/255 to discard trivial Gaussians during occluded-tile detection.
axioms (3)
  • domain assumption Neighboring frames share substantial geometric and appearance overlap so that a single reference depth+color buffer can be warped to approximate most tiles.
    Stated as the key motivation in the abstract and Section III-B; underpins both TDC and SR. Explicitly noted as failing under fast motion in Section V.
  • ad hoc to paper A simple min-depth / max-σ surface band plus 3×3 neighborhood fill is a sufficient geometric prior for 3DGS occlusion and correspondence tests.
    Introduced in Section III-C without theoretical justification beyond “we find that this simple surface formulation works well.”
  • domain assumption Tile-based software (or hardware) rasterization of 3DGS is the target pipeline; acceleration is measured by wall-clock FPS on that pipeline.
    Standard in the 3DGS literature; assumed throughout Sections III–IV.
invented entities (2)
  • Temporal Geometry Buffer (TGB) and Temporal Appearance Buffer (TAB) no independent evidence
    purpose: Cache depth statistics and color of a reference frame so later frames can cull Gaussians and approximate tiles by warping.
    Defined in Section III-C; pure algorithmic data structures with no independent physical existence.
  • Temporal Dynamic Culling (AFC + TOC) and Selective Rendering (SR) no independent evidence
    purpose: Two interleaved strategies that reduce preprocessed Gaussians and fully-rendered tiles.
    Core algorithmic contribution of Sections III-D and III-E; evaluated only inside this paper’s experiments.

pith-pipeline@v1.1.0-grok45 · 21880 in / 3121 out tokens · 26824 ms · 2026-07-12T02:49:47.676996+00:00 · methodology

0 comments
read the original abstract

3D Gaussian Splatting (3DGS) has revolutionized novel-view synthesis with its fast and high-fidelity rendering. However, rendering at high FPS and low latency across various scenes remains a challenge, especially when large amounts of 3D Gaussian ellipsoids appear in the scene. To address this issue, we introduce TemporalGS, to the best of our knowledge, the first training-free plug-and-play algorithmic approach to accelerate 3DGS rendering without any post-training or post-processing, implemented on top of tile-based software rasterization. The key idea is that, instead of rendering frames independently as 3DGS, we leverage the temporal priors, represented by novel geometry and appearance buffers, etc., to reduce redundancy of Gaussian preprocessing, sorting, and rasterization operations of consecutive frames. Specifically, we propose two acceleration strategies: (1) temporal dynamic culling, which filters out Gaussians that contribute less to current frame rendering; (2) selective rendering, which renders only a small portion of tiles that cannot be approximated by the temporal priors. By adapting and interleaving these two strategies, TemporalGS yields a simple but effective plug-and-play solution for 3DGS rendering speed-up without any training. Extensive experiments show that TemporalGS achieves comparable or even better performance compared to existing state-of-the-art post-training or post-processing-based 3DGS rendering acceleration approaches. TemporalGS can significantly enhance the rendering speed of various 3DGS methods, achieving up to $1.48\times$ acceleration, while maintaining competitive rendering quality. We further extend our TemporalGS to hardware rasterization-based 3DGS to show the portability of our algorithm.

Figures

Figures reproduced from arXiv: 2607.03390 by Juwei Lu, Xinxin Zuo, Yuhongze Zhou, Zihao Yang.

Figure 1
Figure 1. Figure 1: Our TemporalGS can speed up rendering across various-scale scenes [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) We present the high-level comparison about rendering a sequence of frames between vanilla 3DGS and TemporalGS. TemporalGS leverages [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: We visualize, at frame t + i, (a) 3DGS near-plane culling, (b) adaptive frustum culling, (c) temporal occlusion culling, and (d) selective rendering regions. All Gaussians within the pink region are preprocessed at frame t + i. standard deviation to obtain high recall. Note that before warping, the default warped depth and standard deviation are initialized by zfar+ϵ and 0, respectively. We denote the warp… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

39 extracted references · 4 linked inside Pith

  1. [1]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  2. [2]

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

    Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang, “Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ FPS,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. [Online]. Available: https://openreview.net/forum?id=6AeIDnrTN2

    2024

  3. [3]

    Reducing the memory footprint of 3d gaussian splatting,

    P. Papantonakis, G. Kopanas, B. Kerbl, A. Lanvin, and G. Drettakis, “Reducing the memory footprint of 3d gaussian splatting,”Proceedings of the ACM on Computer Graphics and Interactive Techniques, vol. 7, no. 1, pp. 1–17, 2024

    2024

  4. [4]

    Compact 3d gaussian representation for radiance field,

    J. C. Lee, D. Rho, X. Sun, J. H. Ko, and E. Park, “Compact 3d gaussian representation for radiance field,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 719–21 728

    2024

  5. [5]

    Speedy-splat: Fast 3d gaussian splatting with sparse pixels and sparse primitives,

    A. Hanson, A. Tu, G. Lin, V . Singla, M. Zwicker, and T. Goldstein, “Speedy-splat: Fast 3d gaussian splatting with sparse pixels and sparse primitives,” inProceedings of the Computer Vision and Pattern Recog- nition Conference, 2025, pp. 21 537–21 546

    2025

  6. [6]

    Radsplat: Radiance field-informed gaussian splatting for robust real- time rendering with 900+ fps,

    M. Niemeyer, F. Manhardt, M.-J. Rakotosaona, M. Oechsle, D. Duck- worth, R. Gosula, K. Tateno, J. Bates, D. Kaeser, and F. Tombari, “Radsplat: Radiance field-informed gaussian splatting for robust real- time rendering with 900+ fps,” in2025 International Conference on 3D Vision (3DV). IEEE, 2025, pp. 134–144

    2025

  7. [7]

    Occlugaussian: Occlusion-aware gaussian splatting for large scene reconstruction and rendering,

    S. Liu, X. Tang, Z. Li, Y . He, C. Ye, J. Liu, B. Huang, S. Zhou, and X. Wu, “Occlugaussian: Occlusion-aware gaussian splatting for large scene reconstruction and rendering,”arXiv preprint arXiv:2503.16177, 2025

    arXiv 2025

  8. [8]

    GScore: Efficient radiance field rendering via architectural support for 3d gaussian splatting,

    J. Lee, S. Lee, J. Lee, J. Park, and J. Sim, “GScore: Efficient radiance field rendering via architectural support for 3d gaussian splatting,” in Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, V olume 3, 2024, pp. 497–511

    2024

  9. [9]

    Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering,

    G. Feng, S. Chen, R. Fu, Z. Liao, Y . Wang, T. Liu, B. Hu, L. Xu, Z. Pei, H. Liet al., “Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 26 652–26 662

    2025

  10. [10]

    Tc-gs: A faster gaussian splatting module utilizing tensor cores,

    Z. Liao, J. Ding, S. Cui, R. Gong, B. Hu, Y . Wang, H. Li, H. Wang, X. Zhang, and R. Fu, “Tc-gs: A faster gaussian splatting module utilizing tensor cores,” inProceedings of the SIGGRAPH Asia 2025 Conference Papers, 2025, pp. 1–9

    2025

  11. [11]

    Stochasticsplats: Stochastic ras- terization for sorting-free 3d gaussian splatting,

    S. Kheradmand, D. Vicini, G. Kopanas, D. Lagun, K. M. Yi, M. Matthews, and A. Tagliasacchi, “Stochasticsplats: Stochastic ras- terization for sorting-free 3d gaussian splatting,”arXiv preprint arXiv:2503.24366, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Taming 3dgs: High-quality radiance fields with limited resources,

    S. S. Mallick, R. Goel, B. Kerbl, M. Steinberger, F. V . Carrasco, and F. De La Torre, “Taming 3dgs: High-quality radiance fields with limited resources,” inSIGGRAPH Asia 2024 Conference Papers, 2024, pp. 1– 11

    2024

  13. [13]

    Compressed 3d gaussian splatting for accelerated novel view synthesis,

    S. Niedermayr, J. Stumpfegger, and R. Westermann, “Compressed 3d gaussian splatting for accelerated novel view synthesis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 10 349–10 358

    2024

  14. [14]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction,

    J. Lin, Z. Li, X. Tang, J. Liu, S. Liu, J. Liu, Y . Lu, X. Wu, S. Xu, Y . Yan et al., “Vastgaussian: Vast 3d gaussians for large scene reconstruction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 5166–5175

    2024

  15. [15]

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

    Y . Liu, C. Luo, L. Fan, N. Wang, J. Peng, and Z. Zhang, “Citygaussian: Real-time high-quality large-scale scene rendering with gaussians,” in European Conference on Computer Vision. Springer, 2024, pp. 265– 282

    2024

  16. [16]

    A hierarchical 3d gaussian representation for real-time rendering of very large datasets,

    B. Kerbl, A. Meuleman, G. Kopanas, M. Wimmer, A. Lanvin, and G. Drettakis, “A hierarchical 3d gaussian representation for real-time rendering of very large datasets,”ACM Transactions on Graphics (TOG), vol. 43, no. 4, pp. 1–15, 2024

    2024

  17. [17]

    DOGS: Distributed-oriented gaussian splatting for large-scale 3d reconstruction via gaussian consensus,

    Y . Chen and G. H. Lee, “DOGS: Distributed-oriented gaussian splatting for large-scale 3d reconstruction via gaussian consensus,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. [Online]. Available: https://openreview.net/forum?id= HAocQ9dSAX

    2024

  18. [18]

    On scaling up 3d gaussian splatting training,

    H. Zhao, H. Weng, D. Lu, A. Li, J. Li, A. Panda, and S. Xie, “On scaling up 3d gaussian splatting training,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview.net/forum?id=pQqeQpMkE7

    2025

  19. [19]

    Retinags: Scal- able training for dense scene rendering with billion-scale 3d gaussians,

    B. Li, S. Chen, L. Wang, K. Liao, S. Yan, and Y . Xiong, “Retinags: Scal- able training for dense scene rendering with billion-scale 3d gaussians,” arXiv preprint arXiv:2406.11836, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Distwar: Fast differentiable rendering on raster-based rendering pipelines,

    S. Durvasula, A. Zhao, F. Chen, R. Liang, P. K. Sanjaya, and N. Vijayku- mar, “Distwar: Fast differentiable rendering on raster-based rendering pipelines,”arXiv preprint arXiv:2401.05345, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Layered depth images,

    J. Shade, S. Gortler, L.-w. He, and R. Szeliski, “Layered depth images,” inProceedings of the 25th annual conference on Computer graphics and interactive techniques, 1998, pp. 231–242

    1998

  22. [22]

    Iterative image warping,

    H. Bowles, K. Mitchell, R. W. Sumner, J. Moore, and M. Gross, “Iterative image warping,” inComputer graphics forum, vol. 31, no. 2pt1. Wiley Online Library, 2012, pp. 237–246

    2012

  23. [23]

    Temporal coherence in image-based visual hull rendering,

    S. Hauswiesner, M. Straka, and G. Reitmayr, “Temporal coherence in image-based visual hull rendering,”IEEE transactions on visualization and computer graphics, vol. 19, no. 10, pp. 1758–1767, 2013

    2013

  24. [24]

    Real-time neural radiance caching for path tracing,

    T. M ¨uller, F. Rousselle, J. Nov ´ak, and A. Keller, “Real-time neural radiance caching for path tracing,”ACM Transactions on Graphics (TOG), vol. 40, no. 4, pp. 1–16, 2021

    2021

  25. [25]

    Temporal coherence-based distributed ray tracing of massive scenes,

    X. Xu, L. Wang, A. P ´erard-Gayot, R. Membarth, C. Li, C. Yang, and P. Slusallek, “Temporal coherence-based distributed ray tracing of massive scenes,”IEEE Transactions on Visualization and Computer Graphics, vol. 30, no. 2, pp. 1489–1501, 2022

    2022

  26. [26]

    Mob-fgsr: Frame generation and super resolution for mobile real-time rendering,

    S. Yang, Q. Zhu, J. Zhuge, Q. Qiu, C. Li, Y . Yan, H. Xu, L.-Q. Yan, and X. Jin, “Mob-fgsr: Frame generation and super resolution for mobile real-time rendering,” inACM SIGGRAPH 2024 Conference Papers, 2024, pp. 1–11

    2024

  27. [27]

    Gffe: G-buffer free frame extrapolation for low-latency real-time rendering,

    S. Wu, D. Vembar, A. Sochenov, S. Panneer, S. Kim, A. Kaplanyan, and L.-Q. Yan, “Gffe: G-buffer free frame extrapolation for low-latency real-time rendering,”ACM Transactions on Graphics (TOG), vol. 43, no. 6, pp. 1–15, 2024

    2024

  28. [28]

    Interactive cloud rendering using temporally coherent photon mapping,

    O. Elek, T. Ritschel, A. Wilkie, and H.-P. Seidel, “Interactive cloud rendering using temporally coherent photon mapping,”Computers & Graphics, vol. 36, no. 8, pp. 1109–1118, 2012

    2012

  29. [29]

    Real-time denoising of volumetric path tracing for direct volume rendering,

    J. A. Iglesias-Guitian, P. Mane, and B. Moon, “Real-time denoising of volumetric path tracing for direct volume rendering,”IEEE Transactions on Visualization and Computer Graphics, vol. 28, no. 7, pp. 2734–2747, 2020

    2020

  30. [30]

    Progressive real-time rendering of one billion points without hierarchical accelera- tion structures,

    M. Sch ¨utz, G. Mandlburger, J. Otepka, and M. Wimmer, “Progressive real-time rendering of one billion points without hierarchical accelera- tion structures,” inComputer Graphics F orum, vol. 39, no. 2. Wiley Online Library, 2020, pp. 51–64

    2020

  31. [31]

    Steernerf: Accelerating nerf rendering via smooth viewpoint trajectory,

    S. Li, H. Li, Y . Wang, Y . Liao, and L. Yu, “Steernerf: Accelerating nerf rendering via smooth viewpoint trajectory,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 20 701–20 711

    2023

  32. [32]

    Nerfbuff: Fast neural rendering via inter-frame feature buffering,

    A. Liu, Y . Liu, X. Long, P. Wang, C. Lin, P. Luo, and W. Wang, “Nerfbuff: Fast neural rendering via inter-frame feature buffering,”IEEE Transactions on Visualization and Computer Graphics, 2024

    2024

  33. [33]

    Frustum volume caching for accelerated nerf rendering,

    M. Steiner, T. K ¨ohler, L. Radl, and M. Steinberger, “Frustum volume caching for accelerated nerf rendering,”Proceedings of the ACM on Computer Graphics and Interactive Techniques, vol. 7, no. 3, pp. 1–22, 2024

    2024

  34. [34]

    Neo: Real-time on-device 3d gaussian splatting with reuse-and-update sorting acceleration,

    C. Oh, S. Oh, J. Hwang, Y . Kim, H. Sharma, and J. Park, “Neo: Real-time on-device 3d gaussian splatting with reuse-and-update sorting acceleration,” inProceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, V olume 2, 2026, pp. 1268–1284

    2026

  35. [35]

    Gs-cache: A gs-cache inference framework for large-scale gaussian splatting models,

    M. Tao, Y . Zhou, H. Xu, Z. He, Z. Yang, Y . Zhang, Z. Su, L. Xu, Z. Ma, R. Fuet al., “Gs-cache: A gs-cache inference framework for large-scale gaussian splatting models,”arXiv preprint arXiv:2502.14938, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Mip-NeRF 360: Unbounded anti-aliased neural radiance fields,

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Mip-NeRF 360: Unbounded anti-aliased neural radiance fields,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5470–5479

    2022

  37. [37]

    Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond,

    Y . Li, L. Jiang, L. Xu, Y . Xiangli, Z. Wang, D. Lin, and B. Dai, “Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 3205–3215. JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 9

    2023

  38. [38]

    Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs,

    H. Turki, D. Ramanan, and M. Satyanarayanan, “Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2022, pp. 12 922–12 931

    2022

  39. [39]

    Repre- senting long volumetric video with temporal gaussian hierarchy,

    Z. Xu, Y . Xu, Z. Yu, S. Peng, J. Sun, H. Bao, and X. Zhou, “Repre- senting long volumetric video with temporal gaussian hierarchy,”ACM Transactions on Graphics (TOG), vol. 43, no. 6, pp. 1–18, 2024

    2024