Pith. sign in

REVIEW 3 major objections 5 minor 42 references

A Bag of Tricks for Efficient Implicit Neural Point Clouds

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read INPC training time drops 25%, rendering doubles, and VRAM falls 20% while quality holds.

desk verdict A practical, honestly reported speedup for INPC at 8M samples; the extension to 33M is asserted, not shown, and that's the part I'd want fixed before publication. read the letter →

arxiv 2508.19140 v1 pith:K3S2KSWJ submitted 2025-08-26 cs.GR cs.CVcs.LG

Florian Hahlbohm , Linus Franke , Leon Overkämping , Paula Wespe , Susana Castillo , Martin Eisemann , Marcus Magnor This is my paper
classification cs.GRcs.CVcs.LG
keywords implicitneuralpointcloudsnovelviewsynthesispoint-basedrenderingdifferentiablerasterizationtiledradiancefields3DGaussiansplattinghole-fillingCNN
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

Implicit Neural Point Clouds (INPC) achieve high image quality by representing a scene as a probability field of point positions plus a neural appearance field, but the original implementation is slow to train and renders at only a few frames per second. This paper argues that most of that cost is not fundamental: sorting fragments in the rasterizer, resampling point clouds for every frame, starting the hole-filling CNN from scratch, and evaluating low-frequency background with an MLP all waste time and memory. The paper replaces the per-pixel sort with a two-stage tiled sort, reuses point clouds across frames with a ring buffer, pre-trains the hole-filling CNN on other scenes, and renders points as small Gaussians during inference. On the 8M-point configuration, the combined changes cut training time by up to 25%, double rendering speed, reduce peak VRAM by about 20%, and slightly improve standard image-quality metrics. If these gains transfer to the heavier default configuration, INPC becomes practical on consumer GPUs and competitive with explicit particle methods.

What carries the argument

The load-bearing mechanism is the rasterizer's sorting strategy. INPC's original renderer creates four copies of every point (one per covered pixel), keyed with a 64-bit combination of depth and pixel index, and sorts all of them, costing about O(28n). The paper adapts an 8x8-pixel tiled rendering scheme: points are first depth-sorted once on 32-bit keys, then only the roughly 1.27 tiles each splat touches get a shallow 16-bit sort, reducing complexity to about O(6.54n) and shrinking key memory by 18%. Around this core sit three supporting mechanisms: the ring buffer that reuses view-specific point clouds across consecutive frames, the max-over-training-views PDF for global pre-extraction wi

What would settle it

Train the optimized pipeline and the original at the 16M and 33M sample counts on the same scenes using a GPU large enough for both; if per-iteration time and peak VRAM do not improve by roughly the 8M-measured margins (about 25% faster, 20% less VRAM), the central efficiency claim is limited to the low point budget.

Watch

Extended reading notes

Core claim

The central claim is that INPC's practicality bottleneck can be removed without changing the underlying representation. The paper's pipeline keeps the implicit probability field and hash-grid appearance model, but reworks four things around it: a tiled rasterizer that sorts points once by depth and then sorts only per-tile 16-bit keys, cutting the sort complexity from roughly O(28n) to O(6.54n) and key memory by 18%; a ring buffer that reuses the previous frames' view-specific point clouds during inference, halving per-frame sampling cost and improving temporal stability; a global pre-extraction strategy that weights voxels by their maximum weight across training views, producing a 33M-point

Load-bearing premise

All headline efficiency numbers come from the 8-million-point training configuration, and the paper assumes the larger 16M and 33M configurations benefit by the same margins without measuring them at those settings.

Editorial extensions

If this is right

  • At the 8M-point configuration, per-frame view-specific rendering drops from about 114 ms to 46 ms with the ring buffer, crossing into interactive frame rates.
  • The 33M-point global pre-extraction matches or beats the original 67M-point clouds on quality metrics while using roughly half the point memory, and outperforms 3D Gaussian Splatting on the Mip-NeRF360 test set.
  • Fused weight decay and fused rejection sampling remove the VRAM spikes that previously forced low-resolution training; higher-resolution Tanks and Temples scenes become trainable on a 24 GB GPU where original INPC runs out of memory.
  • Pre-trained CNN weights cut early-training instability and reduce active voxels, so a fixed training budget yields higher quality and more effective empty-space pruning.
  • Inference-time Gaussian splats improve extrapolated close-up views and reduce aliasing without any change to the trained scene representation.

Reading between the lines

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

  • The two-stage tiled sort is renderer-agnostic: any point-based renderer with per-pixel alpha-blended splats should be able to adopt it, so the speedup is likely to generalize beyond INPC even though the paper only evaluates it there.
  • Ring-buffer reuse suggests a streaming variant where view-specific point clouds are sampled once per camera-motion segment rather than per frame, trading a small accuracy loss for constant-time rendering in interactive viewers.
  • The invertible tonemapper used during CNN pre-training could let future work pre-train hole-filling networks on much larger RGB-only datasets, since ground-truth HDR targets can be recovered from ordinary images.
  • If the appearance model learned Gaussian parameters during training instead of only at inference, the close-up quality gain seen here might become part of the representation; the paper notes that such changes currently disrupt sampling-weight convergence, so this remains an open question.
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

3 major / 5 minor

Summary. The paper presents a collection of engineering and algorithmic optimizations for Implicit Neural Point Clouds (INPC), a hybrid radiance-field representation. The main modifications are: a tiled rasterizer with a two-stage depth/tile sorting scheme; fused CUDA kernels for spherical contraction, Cauchy loss, weight-decay gradient, and rejection sampling; a ring buffer that reuses view-specific point clouds across frames; an improved global pre-extraction method that halves the point budget; distillation of the background MLP into an environment map; rendering points as small isotropic Gaussians at inference; and a three-stage pre-training pipeline for the hole-filling CNN. Experiments on Mip-NeRF360 and Tanks and Temples at the 8M-sample configuration report up to 25% faster training, 2x faster rendering, and about 20% lower VRAM usage, with image quality roughly preserved. An ablation in Table 2 isolates the contribution of each modification.

Significance. If the results hold, this is a timely and practically useful systems contribution: INPC's rendering speed is a known bottleneck, and the proposed modifications are modular, clearly described, and evaluated with a structured ablation. The tiled rasterizer and fused CUDA kernels are transferable to other point-based neural renderers, and the pre-training procedure with the invertible tonemapper is an interesting contribution in its own right. The main limitation is that the quantitative claims are established only for the 8M-sample configuration, not for INPC's default 33M configuration, and the Tanks and Temples baseline is missing because the original implementation runs out of memory. With additional evidence at higher sample counts or more careful qualification of the claims, the paper would be a solid contribution.

major comments (3)
  1. [Sec. 2 ('A Note on Sample Count')] The paper limits all evaluation to the 8M-sample configuration and extrapolates to the default 33M configuration with the statement 'we are confident that the more expensive configurations would benefit equally' (Sec. 2). This is an assertion, not a measurement. The optimizations interact with sample count non-trivially: fused rejection sampling removes VRAM spikes whose magnitude grows with the sample budget; the tiled rasterizer's copy reduction depends on per-point tile overlaps that may change with occupancy and cache behavior; and pre-trained CNN initialization affects early convergence differently for denser point clouds. Since the abstract's headline figures (up to 25% faster training, 2x rendering, 20% VRAM reduction) are the paper's central practical claim, please provide experiments at 16M/33M on at least a subset of scenes, or explicitly state that the advertised numbers apply
  2. [Table 1 (Tanks and Temples row)] For Tanks and Temples, the original INPC implementation runs out of memory, so there is no direct baseline for training time, VRAM, or quality on this dataset. The text says 'Our bag of tricks enables considerably faster training with similar or better quality,' but this comparison is only possible on Mip-NeRF360. On Tanks and Temples, the reader sees only the proposed method's numbers (24.54 PSNR, 5h26m, 21.1 GiB) with no reference point to attribute the gains. Please provide a baseline on at least a subset of Tanks and Temples (e.g., a smaller scene or lower-resolution images) or explicitly state that the Tanks and Temples results are standalone rather than comparative.
  3. [Sec. 3.2, 'Rendering Points as Small Gaussians'] The paper identifies close-up and extrapolated views as the main motivation for rendering points as Gaussians, and Figure 1 shows a qualitative example. However, all quantitative results in Tables 3 and 4 are averaged over the standard test views and show slightly lower PSNR with Gaussians (27.87 vs 27.95 in Table 3; 27.35 vs 27.47 in Table 4). To support the claim of improved fidelity in extrapolated views, please include quantitative metrics on out-of-distribution views (e.g., zoomed-in crops or held-out camera trajectories). Without this, the benefit of the Gaussian modification is not conclusively demonstrated.
minor comments (5)
  1. [Figure 1 caption] Typo: 'Applying our bag to tricks' should be 'Applying our bag of tricks'.
  2. [Sec. 3.3] 'Training Datasubset' should be 'Training data subset'.
  3. [Table 4 caption] The notation is confusing: 'Ours' vs 'Ours w/ Gaussians' suggests that the row labeled 'Ours' uses bilinear splatting, but this is not stated explicitly. Please clarify the splatting method used for each row in the caption or table header.
  4. [Table 2] For the 'Cached Pixel Directions' row, the VRAM delta is reported as '± 0.0 GiB'; this is odd notation. Suggest '0.0 GiB'.
  5. [General] The paper reports single-run results for small quality differences (e.g., 0.1 dB PSNR). Reporting variance across multiple seeds or runs, even for a subset of scenes, would strengthen the comparisons.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical optimization with direct measurements and disjoint pretraining data; the 8M-to-33M extrapolation is a limitation, not circular reasoning.

full rationale

The paper is an empirical systems contribution. Its central claims are per-modification speedups (Table 2) and end-to-end training/inference comparisons (Tables 1, 3, 4) against the authors' own INPC baseline, which is the natural control rather than a self-referential validation. No quantity is defined in terms of the prediction it is supposed to validate: the 'up to 25% faster training, 2x faster rendering, and 20% reduced VRAM' figures are direct runtime and memory measurements, not fitted constants. The pre-trained CNN is trained on a disjoint set of scenes (Tanks and Temples Training subset plus CO3D) and evaluated on Mip-NeRF360 and Tanks and Temples Intermediate; the 'Truck' scene used for warm-starting is in the Training subset and not in the evaluation sets. The improved global extraction computes a PDF from the model's own voxel weights, but that is an algorithmic input-output relation, not a prediction made from a fitted parameter. The one genuinely unsupported statement is the extrapolation from 8M to 33M samples ('we are confident that the more expensive configurations would benefit equally', Sec. 2), and Table 1 lacks a Tanks-and-Temples baseline for the original INPC due to OOM. However, an untested generalization is a correctness/robustness concern, not a circular derivation. There is no equation that reduces to its own input, no fitted parameter renamed as a prediction, and no load-bearing argument that rests solely on a self-citation. Hence score 0.

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

The central claim is empirical; no free parameters are fitted to the test benchmarks. The listed parameters are hand-chosen design choices in the optimizations, and the axioms are standard rendering and transfer assumptions.

free parameters (4)
  • Gaussian world-space scale rule = standard deviation of 5 pixels at the near plane
    Hand-chosen in Sec. 3.2 to balance close-up coverage against aliasing; the small-Gaussian inference quality depends on it.
  • Isotropic 2D Gaussian dilation variance = 0.16
    Chosen smaller than 3DGS's 0.3 to preserve detail for distant points (Sec. 3.2).
  • Global pre-extraction point budget = 33 million (down from 67 million)
    Halved based on experiments in Sec. 3.1; the quality improvement of the new global sampler is shown at this budget.
  • Pre-training dataset composition = 7 Tanks and Temples Training scenes + 4 CO3D scenes
    Hand-selected in Sec. 3.3; cross-scene transfer results depend on this set.
assumptions (4)
  • domain assumption Alpha blending of depth-sorted fragments (Eq. 1) correctly composites point splats into feature images
    Used throughout Sec. 2 and Appendix A; standard in point-based neural rendering but not derived here.
  • domain assumption A 2x2 pixel splat contributes on average to 1.27 tiles of an 8x8 tiling
    Sec. 3.2 uses this expected value to justify the tiled rasterizer; it depends on screen-space point distribution.
  • ad hoc to paper Small isotropic Gaussians can replace bilinear splats at inference without harming quality
    Modeling choice in Sec. 3.2, empirically validated only on the tested scenes.
  • domain assumption CNN weights pretrained on 11 scenes transfer to unseen scenes
    Sec. 3.3 assumes cross-scene transfer; the FFC residual block is excluded from transfer, indicating the assumption is partial.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Bag of Tricks for Efficient Implicit Neural Point Clouds." pith.science (2026). https://pith.science/paper/K3S2KSWJ

@misc{pith2026250819140,
  author       = {Pith},
  title        = {Pith review of: A Bag of Tricks for Efficient Implicit Neural Point Clouds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K3S2KSWJ}},
  note         = {Machine review of arXiv:2508.19140}
}
read the original abstract

Implicit Neural Point Cloud (INPC) is a recent hybrid representation that combines the expressiveness of neural fields with the efficiency of point-based rendering, achieving state-of-the-art image quality in novel view synthesis. However, as with other high-quality approaches that query neural networks during rendering, the practical usability of INPC is limited by comparatively slow rendering. In this work, we present a collection of optimizations that significantly improve both the training and inference performance of INPC without sacrificing visual fidelity. The most significant modifications are an improved rasterizer implementation, more effective sampling techniques, and the incorporation of pre-training for the convolutional neural network used for hole-filling. Furthermore, we demonstrate that points can be modeled as small Gaussians during inference to further improve quality in extrapolated, e.g., close-up views of the scene. We design our implementations to be broadly applicable beyond INPC and systematically evaluate each modification in a series of experiments. Our optimized INPC pipeline achieves up to 25% faster training, 2x faster rendering, and 20% reduced VRAM usage paired with slight image quality improvements.

Figures

Figures reproduced from arXiv: 2508.19140 by the authors.

Figure 1
Figure 1. Applying our bag to tricks to the original INPC [HFK∗ 25] pipeline doubles rendering speed, reduces training time and memory usage, and improves image quality, enhancing its practical usability. We show our results alongside those of the original INPC implementation with the frame rates inset. Abstract Implicit Neural Point Cloud (INPC) is a recent hybrid representation that combines the expressiveness of neural fie… view at source ↗
Figure 2
Figure 2. Convergence plots showing the number of voxels in P and PSNR on the testset for the Bicycle (top) and Bonsai (bottom) scenes from Mip-NeRF360 [BMV∗ 22] over the duration of training. The reduction in active voxels indicates more efficient empty space pruning, which we find to increase robustness of the optimization. We consistently observed this effect across all tested scenes. 4.2. Inference Improvements Multiple o… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 14 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry.original add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 i...

  3. [3]

    : Neural point-based graphics

    Aliev K.-A., Sevastopolsky A., Kolos M., Ulyanov D., Lempitsky V. : Neural point-based graphics. In ECCV (2020), pp. 696--712. https://doi.org/10.1007/978-3-030-58542-6_42 doi:10.1007/978-3-030-58542-6_42

    work page doi:10.1007/978-3-030-58542-6_42 2020

  4. [4]

    Barron J. T. : A general and adaptive robust loss function. In CVPR (2019), pp. 4326--4334. https://doi.org/10.1109/CVPR.2019.00446 doi:10.1109/CVPR.2019.00446

    arXiv 2019

  5. [7]

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

    Barron J. T., Mildenhall B., Verbin D., Srinivasan P. P., Hedman P. : Zip-NeRF : Anti-aliased grid-based neural radiance fields. In ICCV (2023), pp. 19640--19648. https://doi.org/10.1109/ICCV51070.2023.01804 doi:10.1109/ICCV51070.2023.01804

    arXiv 2023

  6. [8]

    : Fast F ourier convolution

    Chi L., Jiang B., Mu Y. : Fast F ourier convolution. In NIPS (2020), vol. 33, pp. 4479--4488

    work page 2020

  7. [9]

    : TensoRF : Tensorial radiance fields

    Chen A., Xu Z., Geiger A., Yu J., Su H. : TensoRF : Tensorial radiance fields. In ECCV (2022), pp. 333--350. https://doi.org/10.1007/978-3-031-19824-3_20 doi:10.1007/978-3-031-19824-3_20

    doi:10.1007/978-3-031-19824-3_20 2022

  8. [10]

    Duckworth D., Hedman P., Reiser C., Zhizhin P., Thibert J.-F., Lu c i\' c M., Szeliski R., Barron J. T. : SMERF : Streamable memory efficient radiance fields for real-time large-scene exploration. ACM ToG 43, 4 (2024). https://doi.org/10.1145/3658193 doi:10.1145/3658193

    work page doi:10.1145/3658193 2024

Show all 42 references
  1. [11]

    : Modern hardware accelerated point based holography

    Fricke S., Caspary R., Castillo S., Eisemann M., Magnor M. : Modern hardware accelerated point based holography. Opt. Express 32, 15 (2024), 26994--27009. https://doi.org/10.1364/OE.523829 doi:10.1364/OE.523829

    2024 doi

  2. [12]

    : VR-Splatting : Foveated radiance field rendering via 3D Gaussian splatting and neural points

    Franke L., Fink L., Stamminger M. : VR-Splatting : Foveated radiance field rendering via 3D Gaussian splatting and neural points. PACMCGIT 8, 1 (2025). https://doi.org/10.1145/3728302 doi:10.1145/3728302

    2025 doi

  3. [13]

    : Plenoxels: Radiance fields without neural networks

    Fridovich-Keil S., Yu A., Tancik M., Chen Q., Recht B., Kanazawa A. : Plenoxels: Radiance fields without neural networks. In CVPR (2022), pp. 5491--5500. https://doi.org/10.1109/CVPR52688.2022.00542 doi:10.1109/CVPR52688.2022.00542

    2022

  4. [14]

    : VET : Visual error tomography for point cloud completion and high-quality neural rendering

    Franke L., R\" u ckert D., Fink L., Innmann M., Stamminger M. : VET : Visual error tomography for point cloud completion and high-quality neural rendering. In SIGGRAPH Asia (2023). https://doi.org/10.1145/3610548.3618212 doi:10.1145/3610548.3618212

    2023

  5. [15]

    : TRIPS : Trilinear point splatting for real-time radiance field rendering

    Franke L., R\" u ckert D., Fink L., Stamminger M. : TRIPS : Trilinear point splatting for real-time radiance field rendering. CGF 43, 2 (2024). https://doi.org/10.1111/cgf.15012 doi:10.1111/cgf.15012

    2024 doi

  6. [16]

    M., Tagliasacchi A

    Govindarajan S., Rebain D., Yi K. M., Tagliasacchi A. : Radiant Foam : Real-time differentiable ray tracing. arXiv:2502.01157 (2025)

    2025 arXiv

  7. [17]

    Halton J. H. : Algorithm 247: Radical-inverse quasi-random point sequence. Commun. ACM 7, 12 (1964), 701--702. https://doi.org/10.1145/355588.365104 doi:10.1145/355588.365104

    1964

  8. [18]

    : INPC : Implicit neural point clouds for radiance field rendering

    Hahlbohm F., Franke L., Kappel M., Castillo S., Eisemann M., Stamminger M., Magnor M. : INPC : Implicit neural point clouds for radiance field rendering. In 3DV (2025)

    2025

  9. [19]

    : Efficient perspective-correct 3D Gaussian splatting using hybrid transparency

    Hahlbohm F., Friederichs F., Weyrich T., Franke L., Kappel M., Castillo S., Stamminger M., Eisemann M., Magnor M. : Efficient perspective-correct 3D Gaussian splatting using hybrid transparency. CGF 44, 2 (2025). https://doi.org/10.1111/cgf.70014 doi:10.1111/cgf.70014

    2025 doi

  10. [20]

    : Gaussian error linear units ( GELU s)

    Hendrycks D., Gimpel K. : Gaussian error linear units ( GELU s). arXiv preprint arXiv:1606.08415 (2016)

    2016 arXiv

  11. [21]

    : PlenopticPoints : Rasterizing neural feature points for high-quality novel view synthesis

    Hahlbohm F., Kappel M., Tauscher J.-P., Eisemann M., Magnor M. : PlenopticPoints : Rasterizing neural feature points for high-quality novel view synthesis. In VMV (2023), pp. 53--61. https://doi.org/10.2312/vmv.20231226 doi:10.2312/vmv.20231226

    2023 doi

  12. [23]

    : 2D Gaussian splatting for geometrically accurate radiance fields

    Huang B., Yu Z., Chen A., Geiger A., Gao S. : 2D Gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH (2024). https://doi.org/10.1145/3641519.3657428 doi:10.1145/3641519.3657428

    2024

  13. [24]

    : 3D Gaussian splatting for real-time radiance field rendering

    Kerbl B., Kopanas G., Leimkuehler T., Drettakis G. : 3D Gaussian splatting for real-time radiance field rendering. ACM ToG 42, 4 (2023). https://doi.org/10.1145/3592433 doi:10.1145/3592433

    2023 doi

  14. [25]

    : A hierarchical 3D Gaussian representation for real-time rendering of very large datasets

    Kerbl B., Meuleman A., Kopanas G., Wimmer M., Lanvin A., Drettakis G. : A hierarchical 3D Gaussian representation for real-time rendering of very large datasets. ACM ToG 43, 4 (2024). https://doi.org/10.1145/3658160 doi:10.1145/3658160

    2024 doi

  15. [26]

    : Point-based neural rendering with per-view optimization

    Kopanas G., Philip J., Leimk\" u hler T., Drettakis G. : Point-based neural rendering with per-view optimization. CGF 40, 4 (2021), 29--43. https://doi.org/10.1111/cgf.14339 doi:10.1111/cgf.14339

    2021 doi

  16. [27]

    : Tanks and Temples : Benchmarking large-scale scene reconstruction

    Knapitsch A., Park J., Zhou Q.-Y., Koltun V. : Tanks and Temples : Benchmarking large-scale scene reconstruction. ACM ToG 36, 4 (2017). https://doi.org/10.1145/3072959.3073599 doi:10.1145/3072959.3073599

    2017

  17. [28]

    Kheradmand S., Rebain D., Sharma G., Sun W., Tseng Y.-C., Isack H., Kar A., Tagliasacchi A., Yi K. M. : 3D Gaussian splatting as Markov chain Monte Carlo . In NIPS (2024)

    2024

  18. [29]

    : Deformable beta splatting

    Liu R., Sun D., Chen M., Wang Y., Feng A. : Deformable beta splatting. In SIGGRAPH (2025). https://doi.org/10.1145/3721238.3730716 doi:10.1145/3721238.3730716

    2025

  19. [31]

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

    Mallick S. S., Goel R., Kerbl B., Steinberger M., Carrasco F. V., De La Torre F. : Taming 3DGS : High-quality radiance fields with limited resources. In SIGGRAPH Asia (2024). https://doi.org/10.1145/3680528.3687694 doi:10.1145/3680528.3687694

    2024

  20. [32]

    T., Zhang Y

    Mai A., Hedman P., Kopanas G., Verbin D., Futschik D., Xu Q., Kuester F., Barron J. T., Zhang Y. : EVER : Exact volumetric ellipsoid rendering for real-time view synthesis. arXiv:2410.01804 (2024)

    2024 arXiv

  21. [33]

    : 3D Gaussian Ray Tracing : Fast tracing of particle scenes

    Moenne-Loccoz N., Mirzaei A., Perel O., de Lutio R., Martinez Esturo J., State G., Fidler S., Sharp N., Gojcic Z. : 3D Gaussian Ray Tracing : Fast tracing of particle scenes. ACM ToG 43, 6 (2024). https://doi.org/10.1145/3687934 doi:10.1145/3687934

    2024 doi

  22. [34]

    P., Tancik M., Barron J

    Mildenhall B., Srinivasan P. P., Tancik M., Barron J. T., Ramamoorthi R., Ng R. : NeRF : Representing scenes as neural radiance fields for view synthesis. In ECCV (2020), pp. 405--421. https://doi.org/10.1007/978-3-030-58452-8_24 doi:10.1007/978-3-030-58452-8_24

    2020 doi

  23. [35]

    : U-Net : Convolutional networks for biomedical image segmentation

    Ronneberger O., Fischer P., Brox T. : U-Net : Convolutional networks for biomedical image segmentation. In MICCAI (2015), pp. 234--241. https://doi.org/10.1007/978-3-319-24574-4_28 doi:10.1007/978-3-319-24574-4_28

    2015 doi

  24. [36]

    : ADOP : Approximate differentiable one-pixel point rendering

    R\" u ckert D., Franke L., Stamminger M. : ADOP : Approximate differentiable one-pixel point rendering. ACM ToG 41, 4 (2022). https://doi.org/10.1145/3528223.3530122 doi:10.1145/3528223.3530122

    2022

  25. [37]

    : Common Objects in 3D : Large-scale learning and evaluation of real-life 3D category reconstruction

    Reizenstein J., Shapovalov R., Henzler P., Sbordone L., Labatut P., Novotny D. : Common Objects in 3D : Large-scale learning and evaluation of real-life 3D category reconstruction. In ICCV (2021). https://doi.org/10.1109/ICCV48922.2021.01072 doi:10.1109/ICCV48922.2021.01072

    2021

  26. [38]

    : StopThePop : Sorted Gaussian splatting for view-consistent real-time rendering

    Radl L., Steiner M., Parger M., Weinrauch A., Kerbl B., Steinberger M. : StopThePop : Sorted Gaussian splatting for view-consistent real-time rendering. ACM ToG 43, 4 (2024). https://doi.org/10.1145/3658187 doi:10.1145/3658187

    2024 doi

  27. [40]

    L., Frahm J.-M

    Sch\" o nberger J. L., Frahm J.-M. : Structure-from-motion revisited. In CVPR (2016), pp. 4104--4113. https://doi.org/10.1109/CVPR.2016.445 doi:10.1109/CVPR.2016.445

    2016 doi

  28. [41]

    : Splatshop: Efficiently editing large Gaussian splat models

    Schütz M., Peters C., Hahlbohm F., Eisemann E., Magnor M., Wimmer M. : Splatshop: Efficiently editing large Gaussian splat models. CGF (2025). https://doi.org/10.1111/cgf.70214 doi:10.1111/cgf.70214

    2025 doi

  29. [42]

    : LinPrim : Linear primitives for differentiable volumetric rendering

    von L \"u tzow N., Nie ner M. : LinPrim : Linear primitives for differentiable volumetric rendering. arXiv:2501.16312 (2025)

    2025

  30. [43]

    M., Mirzaei A., Moenne-Loccoz N., Gojcic Z

    Wu Q., Esturo J. M., Mirzaei A., Moenne-Loccoz N., Gojcic Z. : 3DGUT : Enabling distorted cameras and secondary rays in Gaussian splatting. In CVPR (2025), pp. 26036--26046. https://doi.org/10.1109/CVPR52734.2025.02425 doi:10.1109/CVPR52734.2025.02425

    2025

  31. [44]

    : Mip-Splatting : Alias-free 3D Gaussian splatting

    Yu Z., Chen A., Huang B., Sattler T., Geiger A. : Mip-Splatting : Alias-free 3D Gaussian splatting. In CVPR (2024), pp. 19447--19456. https://doi.org/10.1109/CVPR52733.2024.01839 doi:10.1109/CVPR52733.2024.01839

    2024

  32. [45]

    : Gaussian Opacity Fields : Efficient adaptive surface reconstruction in unbounded scenes

    Yu Z., Sattler T., Geiger A. : Gaussian Opacity Fields : Efficient adaptive surface reconstruction in unbounded scenes. ACM ToG 43, 6 (2024). https://doi.org/10.1145/3687937 doi:10.1145/3687937

    2024 doi

  33. [46]

    A., Shechtman E., Wang O

    Zhang R., Isola P., Efros A. A., Shechtman E., Wang O. : The unreasonable effectiveness of deep features as a perceptual metric. In CVPR (2018), pp. 586--595. https://doi.org/10.1109/CVPR.2018.00068 doi:10.1109/CVPR.2018.00068

    2018

  34. [47]

    : EWA volume splatting

    Zwicker M., Pfister H., van Baar J., Gross M. : EWA volume splatting. In Proc. of the Conference on Visualization (2001), pp. 29--36. https://doi.org/10.5555/601671.601674 doi:10.5555/601671.601674

    2001

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.