Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

No Redundancy, No Stall: Lightweight Streaming 3D Gaussian Splatting for Real-time Rendering

T0 review · 4 major / 4 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 A training-free pipeline renders 3D Gaussian scenes 5.41x faster on edge GPUs.

desk verdict Novel system with a real soundness gap in its depth-based culling, and an incomplete evaluation; worth a serious referee for revision. read the letter →

arxiv 2507.21572 v2 pith:UVI733LC submitted 2025-07-29 cs.AR

Linye Wei , Jiajun Tang , Fan Fei , Boxin Shi , Runsheng Wang , Meng Li This is my paper
classification cs.AR
keywords 3DGaussiansplattingreal-timerenderingedgecomputingsparseviewpointtransformationloadbalancinghardwareacceleratortraining-freeacceleration
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 claims that the computational redundancy and hardware stalls that slow down 3D Gaussian Splatting (3DGS) can be removed without any retraining of the scene model. It proposes LS-Gaussian, which reuses rendering results between nearby frames via tile-level viewpoint transformation, predicts where early stopping will occur in each tile to estimate true workloads, and uses a more accurate two-stage intersection test to discard spurious Gaussian-tile pairs. On top of that, a streaming hardware design balances tile assignments across rasterization units. If correct, this would make real-time 3DGS feasible on embedded devices and in edge applications like autonomous driving and AR/VR, with only a small visual quality cost.

What carries the argument

The load-bearing mechanism is the viewpoint transformation that yields both pixel reuse and a per-tile early-stopping depth. From a reference frame, pixels are reprojected to the target viewpoint, tiles with fewer than one-sixth missing pixels are inpainted by interpolation, and the maximum reprojected depth per tile becomes the truncation point: Gaussians beyond that depth are excluded from sorting and rasterization. This same depth prediction drives the load distribution that balances work across parallel rasterization blocks, while a coarse-to-fine intersection test removes false-positive Gaussian-tile pairs before sorting.

What would settle it

Render a video of a scene with fast-moving objects and depth discontinuities where the opacity-weighted depth estimate is wrong; if LS-Gaussian produces visible holes or artifacts in re-rendered tiles, or if measured per-tile workloads diverge from predictions, the early-stopping depth assumption fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that 3DGS rendering for continuous camera streams can be made largely redundancy-free and stall-free by exploiting inter-frame coherence at the tile level rather than the pixel level. Concretely, LS-Gaussian renders one full frame every six frames, warps the reference frame's pixels and depths to the new viewpoint, interpolates tiles that are mostly filled, fully re-renders only the remaining tiles, and uses the predicted early-stopping depth to cull Gaussians and to assign balanced workloads to rendering units. The paper reports an average 5.41x speedup on a Jetson AGX Orin GPU and up to 17.3x with a 16nm custom accelerator, with only minimal quality degradation (average 1.4 dB PSNR drop and 0.005 SSIM loss on Synthetic-NeRF).

Load-bearing premise

The depth map produced by opacity-weighted compositing is accurate enough that Gaussians beyond the predicted early-stopping depth never contribute to any pixel, so culling them changes neither image quality nor workload estimates.

Editorial extensions

If this is right

  • LS-Gaussian can be applied to pre-trained 3DGS models without retraining, so existing scene reconstructions benefit immediately.
  • The 5.41x speedup on Jetson AGX Orin brings 3DGS to 90 FPS on the Deep Blending dataset, making real-time edge rendering practical for AR/VR and robotics.
  • The accelerator design shows that dedicated hardware tailored to sparse, load-balanced 3DGS can outperform generic GPU execution by over 17x at a small area cost.
  • Indoor scenes with uniform color and smooth depth benefit most from sparse rendering, while load balancing gives the largest gains in outdoor scenes with varied detail.

Reading between the lines

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

  • If the depth-based culling is as reliable as claimed, the same early-stopping depth prediction could be used to skip sorting entirely for tiles that are nearly filled by warping, further reducing latency in scenes with high frame-to-frame overlap.
  • The quality-cost tradeoff suggests a tunable knob: the window size and the one-sixth tile threshold could be adapted at runtime to scene complexity, for example by lowering the threshold in high-motion sequences.
  • A natural stress test is transparent or reflective scenes where opacity-weighted depth is less reliable; the method's quality guarantee likely degrades there, which the current evaluation does not cover.
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 / 4 minor

Summary. The paper proposes LS-Gaussian, a training-free algorithm/hardware co-design framework that accelerates streaming 3D Gaussian Splatting (3DGS) rendering on edge platforms. The algorithmic side introduces tile-warping sparse rendering (TWSR), depth-prediction-based early stopping (DPES), and a two-stage accurate intersection test (TAIT), while the hardware side augments the GSCore architecture with a viewpoint transformation unit, a load distribution unit, and workload-aware tile scheduling. The main claimed results are an average 5.41x speedup over original 3DGS on Jetson AGX Orin and up to 17.3x with a custom 16nm accelerator, with minimal visual quality loss (1.4 dB PSNR and 0.005 SSIM on Synthetic-NeRF). The paper reports ablations isolating the contribution of each optimization and a hardware utilization study.

Significance. If the results are validated, LS-Gaussian would be a meaningful step toward real-time 3DGS on resource-constrained devices, especially because it is training-free and integrates with existing 3DGS models. The paper's strengths include a systematic bottleneck analysis that separates inter-frame and intra-frame redundancy from inter-block and intra-block stalls, a hardware architecture that reuses existing units to limit area overhead, and a clear ablation methodology that attributes speedup contributions to individual optimizations. The reported hardware utilization improvements (from 51.5% to 88.6% on average) are internally consistent with the load-balancing claims. However, several load-bearing evaluation choices currently prevent the central claims from being fully established: the quality metric on real-world scenes is self-consistency rather than ground truth, the DPES truncation rule lacks a correctness bound, and the hardware speedup normalization relies on an underdocumented area-speedup curve.

major comments (4)
  1. [Sec. IV-B, Algorithm 1] The early-stopping depth D_T for a re-rendered tile is defined as the maximum depth of valid reprojected pixels, but missing pixels in that tile (which trigger full re-rendering) may have a larger true early-stopping depth. In particular, when a foreground object translates and reveals background, the revealed pixels should composite Gaussians at depths greater than any inherited pixel depth. Culling all Gaussians beyond D_T can therefore drop visible background content, producing holes or wrong colors precisely in the low-reuse, disoccluded regions that Sec. IV-A already identifies as error-prone. The paper does not provide a bound that D_T safely upper-bounds the true truncation depth, nor does it mention a fallback for tiles with missing pixels. This is a correctness risk for the DPES load-balancing mechanism and directly threatens the 'minimal visual quality degradation' half of the central claim; the reported 1.4 dB PSNR loss on Synthetic-NeRF may not transfer to trajectories with significant disocclusion.
  2. [Sec. VI-B] For the real-world scenes, rendering quality is measured as the difference between outputs with and without the proposed viewpoint transformation, not against ground truth. The paper states this explicitly: 'we evaluate image quality by measuring the difference between rendered outputs w/ and w/o the proposed viewpoint transformation.' This metric measures self-consistency with the method's own full rendering, not fidelity to the actual scene. A method that introduces consistent errors in both branches could still score high on this metric while being visibly wrong. The claim of minimal visual degradation on real-world scenes is therefore not established; per-scene PSNR/SSIM against ground truth (or at least against the original 3DGS reconstruction) should be reported.
  3. [Sec. VI-D] The headline 17.3x accelerator speedup is obtained after normalizing GSCore and MetaSapiens to the same 1.45 mm2 area using the 'Speedup-Area Curve reported by MetaSapiens.' No details of this curve are given, and MetaSapiens is reported only as an average speedup, not per-scene. Scaling a speedup by area implicitly assumes a particular performance-area trade-off that may not hold for GSCore's architecture or for the reported scenes. Without the normalization curve or per-configuration data points, the hardware comparison is not reproducible and the relative ranking (17.3x vs. 9.1x vs. 14.5x) may be an artifact of the normalization rather than an intrinsic speedup. The ablation in Fig. 15 is useful, but it does not validate the normalization against an independent measurement.
  4. [Sec. VI] The paper reports averaged results without error bars, per-scene quality tables for real-world scenes, or confidence intervals. Fig. 12 shows only aggregate PSNR and FPS for two representative frames, and the GPU speedup numbers in Fig. 13a are given as averages across scenes without variance. No code or data are provided. The internal consistency of the ablations is plausible, but the central quantitative claims cannot be independently checked, and the reader cannot assess whether the 5.41x and 17.3x figures are stable across scenes and runs or dominated by a favorable subset.
minor comments (4)
  1. [Sec. IV-A heading] The heading 'Tile Warpping-based Sparse Rendering' contains a typo: 'Warpping' should be 'Warping' (also in the abstract, where 'TWSR' is defined).
  2. [Sec. IV-C] In Eq. (7), the symbol l is used both as the line segment connecting the tile center to the ellipse center and as its length; please clarify the notation (e.g., use |l| consistently with a definition).
  3. [Sec. IV-A / V-A] The threshold for tile interpolation is described as 'less than one-sixth of the total pixels' missing in Sec. IV-A and as 'exceeds 5/6 of the total pixels' valid in Sec. V-A; these are consistent, but the definition of N0 should be stated once in one place to avoid ambiguity.
  4. [Sec. II-A] The opacity notation is used both as o_i (Gaussian opacity) and as α_i (density after evaluation of the Gaussian); the relationship to the standard 3DGS formulation should be stated more explicitly, especially because Eq. (2) then uses α_i as the blending weight.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central speedup and quality claims are benchmarked against external baselines and independent datasets.

full rationale

The paper's main claims are evaluated externally rather than derived from its own inputs. GPU speedups (5.41x average, up to 17.3x with hardware) are measured on Jetson AGX Orin against original 3DGS and compared with AdR-Gaussian, SeeLe, GSCore, and MetaSapiens. Rendering quality on Synthetic-NeRF is reported as PSNR/SSIM against original 3DGS outputs, not against LS-Gaussian's own rendering. The real-world 'w/ and w/o TWSR' comparison in Sec. VI-B is explicitly an internal ablation used to choose the warping window and to show the incremental cost of reuse; it is not the basis for the absolute quality claim. The algorithmic components (TWSR, DPES, TAIT) are heuristics with stated rules, and their contributions are isolated by ablation; none of the headline numbers is constructed by definition from a fitted parameter. The admitted limitation that reprojected depth errors amplify in low-reuse regions is a correctness or robustness concern, not a circularity: an unsound culling bound does not make the evaluation self-referential. Self-citations in the reference list are background citations and are not load-bearing for the speedup or quality conclusions. Hardware comparisons use a normalization curve reported by MetaSapiens, which weakens independent comparison but does not make LS-Gaussian's measured result equivalent to its inputs. Overall, the derivation chain is self-contained against external benchmarks, and no step reduces by construction to its own input.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

The ledger shows one empirically tuned threshold (N0), a window size chosen from a sensitivity sweep, and a workload-deferral slack heuristic. The key domain assumptions concern depth-map accuracy, tile smoothness for interpolation, and the validity of an external area-speed scaling curve. Two hardware units are introduced without independent validation.

free parameters (3)
  • Tile interpolation threshold N0 = Valid pixels greater than 5/6 of tile (missing less than 1/6)
    Empirically set in Sec. IV-A to decide which tiles are inpainted rather than re-rendered; controls the quality/speed tradeoff.
  • Warping window size n = 5 (full render every 6 frames)
    Default selected from the sensitivity sweep in Sec. VI-B and Fig. 12a; not derived from any criterion.
  • Workload balance slack factor = (1 + 1/N) * W
    Heuristic bound in Sec. V-B for deferring tiles to the next block; no optimality argument is given.
assumptions (5)
  • domain assumption Any Gaussian whose depth exceeds the maximum reprojected early-stopping depth of a tile contributes nothing to any pixel in that tile.
    Underpins DPES and all load estimates; asserted in Sec. IV-B without proof and presumes the estimated depth map is exact.
  • domain assumption Opacity-weighted depth compositing over Gaussians yields a depth map accurate enough for warping and culling.
    Used in Sec. IV-A to estimate depth in real time instead of using ground truth; Sec. IV-A admits these estimates carry errors that amplify in low-reuse regions.
  • ad hoc to paper Tiles with fewer than one-sixth missing pixels have smooth depth and color, making interpolation visually safe.
    Empirical observation in Sec. IV-A used to justify bypassing preprocessing, sorting, and rasterization for such tiles.
  • domain assumption Area scaling of speedups follows the MetaSapiens Speedup-Area curve.
    Used in Sec. VI-D to normalize GSCore and MetaSapiens to 1.45 mm^2; if the curve is not valid for these designs, the 17.3x vs 14.5x/9.1x comparison is unsupported.
  • domain assumption The accelerator speedup figures are correctly derived from the unstated performance model or simulation.
    Sec. VI-D reports speedups without describing cycle-accurate simulation, emulation, or an analytical performance model for the synthesized RTL.
invented entities (2)
  • Viewpoint Transformation Unit (VTU) augmentation
    purpose: Reprojects reference colors and depths, computes per-tile valid pixel counts, and feeds the sparse-rendering and load-distribution logic.
    New hardware unit described at block level; no RTL or silicon validation is released and no independent measurement is provided.
  • Load Distribution Unit (LDU) with counter buffer
    purpose: Assigns tiles to rasterization blocks by predicted workload and orders tiles light-to-heavy to reduce bubbles and idle time.
    New mechanism layered on GSCore; only synthesis area numbers are reported, with no disclosed performance-model details.

how reviews work

0 comments
Cite this review

Pith. "Pith review of No Redundancy, No Stall: Lightweight Streaming 3D Gaussian Splatting for Real-time Rendering." pith.science (2026). https://pith.science/paper/UVI733LC

@misc{pith2026250721572,
  author       = {Pith},
  title        = {Pith review of: No Redundancy, No Stall: Lightweight Streaming 3D Gaussian Splatting for Real-time Rendering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UVI733LC}},
  note         = {Machine review of arXiv:2507.21572}
}
read the original abstract

3D Gaussian Splatting (3DGS) enables high-quality rendering of 3D scenes and is getting increasing adoption in domains like autonomous driving and embodied intelligence. However, 3DGS still faces major efficiency challenges when faced with high frame rate requirements and resource-constrained edge deployment. To enable efficient 3DGS, in this paper, we propose LS-Gaussian, an algorithm/hardware co-design framework for lightweight streaming 3D rendering. LS-Gaussian is motivated by the core observation that 3DGS suffers from substantial computation redundancy and stalls. On one hand, in practical scenarios, high-frame-rate 3DGS is often applied in settings where a camera observes and renders the same scene continuously but from slightly different viewpoints. Therefore, instead of rendering each frame separately, LS-Gaussian proposes a viewpoint transformation algorithm that leverages inter-frame continuity for efficient sparse rendering. On the other hand, as different tiles within an image are rendered in parallel but have imbalanced workloads, frequent hardware stalls also slow down the rendering process. LS-Gaussian predicts the workload for each tile based on viewpoint transformation to enable more balanced parallel computation and co-designs a customized 3DGS accelerator to support the workload-aware mapping in real-time. Experimental results demonstrate that LS-Gaussian achieves 5.41x speedup over the edge GPU baseline on average and up to 17.3x speedup with the customized accelerator, while incurring only minimal visual quality degradation.

Figures

Figures reproduced from arXiv: 2507.21572 by the authors.

Figure 1
Figure 1. Rendering with 3DGS on Jetson AGX Orin and LS-Gaussian. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the 3DGS pipeline. 3DGS pipeline optimized for real-time rendering without retraining, leveraging frame continuity, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. (a) Proportion of overlap pixels between consecutive frames [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: Distribution of the covered Gaussian numbers on different [PITH_FULL_IMAGE:figures/full_fig_p003_5.png]
Figure 6
Figure 6. Figure 6: An example of viewpoint transformation and sparse rendering [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 8
Figure 8. Figure 8: Two-stage accurate intersection test [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 9
Figure 9. Figure 9: Comparison of our two-stage accurate intersection test with [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]
Figure 10
Figure 10. Figure 10: Overall design of LS-Gaussian architecture. Components inherited from GSCore [30] are shown in gray, while newly introduced [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 11
Figure 11. Figure 11: (a) Comparison of the rendering quality between LS [PITH_FULL_IMAGE:figures/full_fig_p007_11.png]
Figure 12
Figure 12. Figure 12: (a) Speedup and PSNR (w/ and w/o TWSR) under different [PITH_FULL_IMAGE:figures/full_fig_p007_12.png]
Figure 14
Figure 14. Figure 14: Speedup comparison between our hardware and prior archi [PITH_FULL_IMAGE:figures/full_fig_p008_14.png]
Figure 13
Figure 13. Figure 13: (a) Comparison of LS-Gaussian performance against prior [PITH_FULL_IMAGE:figures/full_fig_p008_13.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. Cooperative Perception: A Resource-Efficient Framework for Multi-Drone 3D Scene Reconstruction Using Federated Diffusion and NeRF

    cs.AI 2025-08 reject novelty 4.0 of 10

    The framework claims drone swarms can reconstruct 3D scenes by sharing semantic labels and poses, with a federated diffusion model generating missing views for NeRF training.

Reference graph

Works this paper leans on

40 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

    NeRF-VINS: A real-time neural radiance field map- based visual-inertial navigation system,

    S. Katragadda, W. Lee, Y . Peng, P. Geneva, C. Chen, C. Guo, M. Li, and G. Huang, “NeRF-VINS: A real-time neural radiance field map- based visual-inertial navigation system,” in Proc. of IEEE International Conference on Robotics and Automation (ICRA) , 2024, pp. 10 230– 10 237

    work page 2024

  2. [2]

    VR-GS: A physical dynamics-aware interactive Gaussian splatting system in virtual reality,

    Y . Jiang, C. Yu, T. Xie, X. Li, Y . Feng, H. Wang, M. Li, H. Y . K. Lau, F. Gao, Y . Yang, and C. Jiang, “VR-GS: A physical dynamics-aware interactive Gaussian splatting system in virtual reality,” in Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH) , 2024, p. 78

    work page 2024

  3. [3]

    Lightning NeRF: Efficient hybrid scene representation for autonomous driving,

    J. Cao, Z. Li, N. Wang, and C. Ma, “Lightning NeRF: Efficient hybrid scene representation for autonomous driving,” in Proc. of IEEE International Conference on Robotics and Automation (ICRA) , 2024, pp. 16 803–16 809

    work page 2024

  4. [4]

    NeuRAD: Neural rendering for autonomous driving,

    A. Tonderski, C. Lindstr ¨om, G. Hess, W. Ljungbergh, L. Svensson, and C. Petersson, “NeuRAD: Neural rendering for autonomous driving,” in Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2024, pp. 14 895–14 904

    work page 2024

  5. [5]

    AIR-Embodied: An Efficient Active 3DGS-based Interaction and Reconstruction Framework with Embodied Large Language Model

    Z. Qi, S. Yuan, F. Liu, H. Cao, T. Deng, J. Yang, and L. Xie, “AIR- Embodied: An efficient active 3DGS-based interaction and reconstruc- tion framework with embodied large language model,” arXiv preprint arXiv:2409.16019, 2024

    work page Pith review arXiv 2024

  6. [6]

    EnerVerse: Envisioning embodied future space for robotics manipulation,

    S. Huang, L. Chen, P. Zhou, S. Chen, Z. Jiang, Y . Hu, P. Gao, H. Li, M. Yao, and G. Ren, “EnerVerse: Envisioning embodied future space for robotics manipulation,” arXiv preprint arXiv:2501.01895 , 2025

    arXiv 2025

  7. [7]

    Splat-Nav: Safe real-time robot navigation in Gaussian splatting maps,

    T. Chen, O. Shorinwa, W. Zeng, J. Bruno, P. M. Dames, and M. Schwa- ger, “Splat-Nav: Safe real-time robot navigation in Gaussian splatting maps,” IEEE Transactions on Robotics , 2025

    work page 2025

  8. [8]

    DriveDreamer4D: World models are effective data machines for 4d driving scene repre- sentation,

    G. Zhao, C. Ni, X. Wang, Z. Zhu, X. Zhang, Y . Wang, G. Huang, X. Chen, B. Wang, Y . Zhang, W. Mei, and X. Wang, “DriveDreamer4D: World models are effective data machines for 4d driving scene repre- sentation,” arXiv preprint arXiv:2410.13571 , 2024

    arXiv

Show all 40 references
  1. [9]

    ReconDreamer: Crafting world models for driving scene reconstruction via online restoration,

    C. Ni, G. Zhao, X. Wang, Z. Zhu, W. Qin, G. Huang, C. Liu, Y . Chen, Y . Wang, X. Zhang, Y . Zhan, K. Zhan, P. Jia, X. Lang, X. Wang, and W. Mei, “ReconDreamer: Crafting world models for driving scene reconstruction via online restoration,” arXiv preprint arXiv:2411.19548 , 2024

    2024 arXiv

  2. [10]

    NeRF: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “NeRF: Representing scenes as neural radiance fields for view synthesis,” in Proc. of European Conference on Computer Vision (ECCV), 2020

    2020

  3. [11]

    RT-NeRF: Real-time on- device neural radiance fields towards immersive AR/VR rendering,

    C. Li, S. Li, Y . Zhao, W. Zhu, and Y . Lin, “RT-NeRF: Real-time on- device neural radiance fields towards immersive AR/VR rendering,” in Proc. of IEEE/ACM International Conference on Computer-Aided Design (ICCAD) , 2022, pp. 132:1–132:9

    2022

  4. [12]

    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 Transactions on Graphics (TOG), vol. 42, no. 4, pp. 139:1–139:14, 2023

    2023

  5. [13]

    A survey on 3D Gaussian splatting,

    G. Chen and W. Wang, “A survey on 3D Gaussian splatting,” arXiv preprint arXiv:2401.03890, 2024

    2024 arXiv

  6. [14]

    Recent advances in 3D Gaussian splatting,

    T. Wu, Y . Yuan, L. Zhang, J. Yang, Y . Cao, L. Yan, and L. Gao, “Recent advances in 3D Gaussian splatting,” Computational Visual Media, vol. 10, no. 4, pp. 613–642, 2024

    2024

  7. [15]

    Instant-3D: Instant neural radiance field training towards on-device AR/VR 3D reconstruction,

    S. Li, C. Li, W. Zhu, B. T. Yu, Y . K. Zhao, C. Wan, H. You, H. Shi, and Y . C. Lin, “Instant-3D: Instant neural radiance field training towards on-device AR/VR 3D reconstruction,” in Proc. of Annual International Symposium on Computer Architecture (ISCA) , 2023, pp. 6:1–6:13

    2023

  8. [16]

    Cicero: Addressing algorithmic and architectural bottlenecks in neural rendering by radiance warping and memory optimizations,

    Y . Feng, Z. Liu, J. Leng, M. Guo, and Y . Zhu, “Cicero: Addressing algorithmic and architectural bottlenecks in neural rendering by radiance warping and memory optimizations,” in Proc. of Annual International Symposium on Computer Architecture (ISCA) , 2024, pp. 1293–1308

    2024

  9. [17]

    Fusion-3D: Integrated acceleration for instant 3D reconstruc- tion and real-time rendering,

    S. Li, Y . Zhao, C. Li, B. Guo, J. Zhang, W. Zhu, Z. Ye, C. Wan, and Y . C. Lin, “Fusion-3D: Integrated acceleration for instant 3D reconstruc- tion and real-time rendering,” in Proc. of International Symposium on Microarchitecture (MICRO), 2024, pp. 78–91

    2024

  10. [18]

    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 Proc. of European Conference on Computer Vision (ECCV) , 2024

    2024

  11. [19]

    Momentum-GS: Momentum Gaussian self-distillation for high-quality large scene reconstruction,

    J. Fan, W. Li, Y . Han, and Y . Tang, “Momentum-GS: Momentum Gaussian self-distillation for high-quality large scene reconstruction,” arXiv preprint arXiv:2412.04887 , 2024

    2024 arXiv

  12. [20]

    MetaSapiens: Real-time neural rendering with efficiency-aware pruning and accelerated foveated rendering,

    W. Lin, Y . Feng, and Y . Zhu, “MetaSapiens: Real-time neural rendering with efficiency-aware pruning and accelerated foveated rendering,” in Proc. of International Conference on Architectural Support for Program- ming Languages and Operating Systems (ASPLOS) , 2025

    2025

  13. [21]

    CompGS: Smaller and faster Gaussian splatting with vector quantiza- tion,

    K. L. Navaneet, K. P. Meibodi, S. A. Koohpayegani, and H. Pirsiavash, “CompGS: Smaller and faster Gaussian splatting with vector quantiza- tion,” in Proc. of European Conference on Computer Vision (ECCV) , 2024

    2024

  14. [22]

    GoDe: Gaussians on demand for progressive level of detail and scalable compression,

    F. Di Sario, R. Renzulli, M. Grangetto, A. Sugimoto, and E. Tartaglione, “GoDe: Gaussians on demand for progressive level of detail and scalable compression,” arXiv preprint arXiv:2501.13558 , 2025

    2025

  15. [23]

    Mini-Splatting: Representing scenes with a constrained number of Gaussians,

    G. Fang and B. Wang, “Mini-Splatting: Representing scenes with a constrained number of Gaussians,” in Proc. of European Conference on Computer Vision (ECCV) , 2024

    2024

  16. [24]

    3D Gaussian rendering can be sparser: Efficient rendering via learned fragment pruning,

    Z. Ye, C. Wan, C. Li, J. Hong, S. Li, L. Li, Y . Zhang, and Y . C. Lin, “3D Gaussian rendering can be sparser: Efficient rendering via learned fragment pruning,” in Proc. of Neural Information Processing Systems (NeurIPS), 2024

    2024

  17. [25]

    GaussianSpa: An “optimizing- sparsifying

    Y . Zhang, W. Jia, W. Niu, and M. Yin, “GaussianSpa: An “optimizing- sparsifying” simplification framework for compact and high-quality 3D Gaussian splatting,” arXiv preprint arXiv:2411.06019 , 2024

    2024 arXiv

  18. [26]

    Scaffold- GS: Structured 3D Gaussians for view-adaptive rendering,

    T. Lu, M. Yu, L. Xu, Y . Xiangli, L. Wang, D. Lin, and B. Dai, “Scaffold- GS: Structured 3D Gaussians for view-adaptive rendering,” in Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 20 654–20 664

    2024

  19. [27]

    Octree- GS: Towards consistent real-time rendering with LOD-structured 3D Gaussians,

    K. Ren, L. Jiang, T. Lu, M. Yu, L. Xu, Z. Ni, and B. Dai, “Octree- GS: Towards consistent real-time rendering with LOD-structured 3D Gaussians,” arXiv preprint arXiv:2403.17898 , 2024

    2024 arXiv

  20. [28]

    FlashGS: Efficient 3D Gaussian splatting for large-scale and high-resolution rendering,

    G. Feng, S. Chen, R. Fu, Z. Liao, Y . Wang, T. Liu, Z. Pei, H. Li, X. Zhang, and B. Dai, “FlashGS: Efficient 3D Gaussian splatting for large-scale and high-resolution rendering,” arXiv preprint arXiv:2408.07967, 2024

    2024 arXiv

  21. [29]

    SeeLe: A unified acceleration framework for real-time Gaussian splatting,

    X. Huang, H. Zhu, Z. Liu, W. Lin, X. Liu, Z. He, J. Leng, M. Guo, and Y . Feng, “SeeLe: A unified acceleration framework for real-time Gaussian splatting,” arXiv preprint arXiv:2503.05168 , 2025

    2025

  22. [30]

    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 Proc. of International Conference on Architectural Support for Program- ming Languages and Operating Systems (ASPLOS) , 2024

    2024

  23. [31]

    Structure-from-Motion revisited,

    J. L. Schonberger and J.-M. Frahm, “Structure-from-Motion revisited,” in Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2016

    2016

  24. [32]

    Potamoi: Accelerating neural rendering via a unified streaming architecture,

    Y . Feng, W. Lin, Z. Liu, J. Leng, M. Guo, H. Zhao, X. Hou, J. Zhao, and Y . Zhu, “Potamoi: Accelerating neural rendering via a unified streaming architecture,” ACM Transactions on Architecture and Code Optimization, vol. 21, no. 4, pp. 80:1–80:25, 2024

    2024

  25. [33]

    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. Fu et al., “GS-Cache: A gs-cache inference framework for large-scale Gaussian splatting models,” arXiv preprint arXiv:2502.14938 , 2025

    2025 arXiv

  26. [34]

    AdR-Gaussian: Accelerating Gaussian splatting with adaptive radius,

    X. Wang, R. Yi, and L. Ma, “AdR-Gaussian: Accelerating Gaussian splatting with adaptive radius,” in Proc. of the ACM SIGGRAPH Confer- ence and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia) , 2024

    2024

  27. [35]

    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,” arXiv preprint arXiv:2412.00578 , 2024

    2024

  28. [36]

    GauRast: Enhancing GPU triangle rasterizers to accelerate 3D Gaussian splatting,

    S. Li, B. Keller, Y . C. Lin, and B. Khailany, “GauRast: Enhancing GPU triangle rasterizers to accelerate 3D Gaussian splatting,” arXiv preprint arXiv:2503.16681, 2025

    2025 arXiv

  29. [37]

    Gaussian blending unit: An edge GPU plug- in for real-time Gaussian-based rendering in AR/VR,

    Z. Ye, Y . Fu, J. Zhang, L. Li, Y . Zhang, S. Li, C. Wan, C. Wan, C. Li, S. Prathipati et al., “Gaussian blending unit: An edge GPU plug- in for real-time Gaussian-based rendering in AR/VR,” arXiv preprint arXiv:2503.23625, 2025

    2025 arXiv

  30. [38]

    Tanks and temples: Benchmarking large-scale scene reconstruction,

    A. Knapitsch, J. Park, Q.-Y . Zhou, and V . Koltun, “Tanks and temples: Benchmarking large-scale scene reconstruction,” ACM Transactions on Graphics (TOG), vol. 36, no. 4, pp. 1–13, 2017

    2017

  31. [39]

    Deep blending for free-viewpoint image-based rendering,

    P. Hedman, J. Philip, T. Price, J.-M. Frahm, G. Drettakis, and G. Bros- tow, “Deep blending for free-viewpoint image-based rendering,” ACM Transactions on Graphics (TOG) , vol. 37, no. 6, pp. 1–15, 2018

    2018

  32. [40]

    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 Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2022

    2022

Pith tools

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