Pith. sign in

REVIEW 3 major objections 5 minor 24 references

HuSc3D: Human Sculpture dataset for 3D object reconstruction

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read HuSc3D is a dataset of six white sculptures captured under realistic, messy conditions, and the paper claims it separates 3D reconstruction methods more sharply than conventional benchmarks.

desk verdict A genuinely useful small dataset for 3D reconstruction under amateur capture conditions, but the benchmark numbers for the two video scenes are compromised by a random rather than temporal test split. read the letter →

arxiv 2506.07628 v1 pith:AEK5OIMD submitted 2025-06-09 cs.CV

Weronika Smolak-Dyżewska , Dawid Malarz , Grzegorz Wilczyński , Rafa{l} Tobiasz , Joanna Waczyńska , Piotr Borycki , Przemys{l}aw Spurek This is my paper
classification cs.CV
keywords 3DreconstructiondatasetbenchmarkNeuralRadianceFieldsGaussianSplattingwhitebalancesparseviewsreal-worldcapture
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

HuSc3D is a new dataset for benchmarking 3D reconstruction from ordinary, unpolished captures: six white, highly perforated human sculptures photographed in a real gallery with dynamic backgrounds, automatic white-balance shifts, sparse views, and widely varying image counts. The paper claims that these conditions, which are typical when a casual user films with a smartphone, are missing from existing benchmarks that favour synthetic or professionally curated scenes, and that HuSc3D exposes them. To support this, the authors evaluate six popular reconstruction methods and report a clear performance spread, with Gaussian-splatting based methods ahead of NeRF-based ones on PSNR, SSIM, and LPIPS. If the claim holds, HuSc3D gives the field a testbed for robustness that clean datasets cannot provide.

What carries the argument

The load-bearing object is the dataset itself, defined by six scenes of white sculptures and three acquisition styles: sparse DSLR photo sets (Kacper, Paulina), high-resolution phone photos (Jarosław, Hanna), and frames extracted from phone video (Tosia, Wiktor). The argument is carried by the contrast between these designed challenge dimensions and the fixed evaluation protocol: the same six methods, each trained for 30,000 iterations, with 12.5% of each scene's images reserved as a test set. The dataset is meant to force methods to cope simultaneously with dynamic background, white-balance inconsistency, low-texture intricate geometry, and sparse or oversized inputs.

What would settle it

Re-run the Tosia and Wiktor scenes with a temporal holdout, training on one block of frames and testing on a later block, and compare PSNR and LPIPS to the paper's random-split numbers; if the metrics drop sharply or the method ranking changes, the claimed separation is an artifact of near-duplicate test frames rather than a robustness difference.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that a deliberately unpolished, small dataset is a better discriminator of reconstruction methods than the polished datasets currently in use. HuSc3D combines several real-world failure modes at once: low-texture white surfaces that give feature matching and color estimation little to work with, fine perforated geometry that demands high-frequency detail, photometric inconsistency from automatic white balance, moving people in the background, and some scenes with only 66 or 104 registered views. The reported experiments show Splatfacto-MCMC leading all metrics (average PSNR 32.45, SSIM 0.953, LPIPS 0.109), Nerfacto trailing (average PSNR 21.98), and Instant-NGP running out of memory on two of the six scenes. The authors read this spread as evidence that HuSc3D differentiates model performance along axes that conventional benchmarks mask.

Load-bearing premise

The evaluation assumes that a random 12.5% test split measures how well a method generalizes, but in the two video-derived scenes the test frames are extracted from the same continuous footage as the training frames, so they are near-duplicates and the reported metrics are likely inflated.

Editorial extensions

If this is right

  • A benchmark that includes casual acquisition conditions can reveal robustness differences that clean datasets hide; methods that look similar on synthetic or curated scenes separate sharply on HuSc3D.
  • Gaussian-splatting methods, especially Splatfacto-MCMC, are more reliable than NeRF-based methods on low-texture white objects with sparse views and photometric inconsistency; however, they pay for this with longer training and lower frame rates.
  • High-resolution captures (Jarosław, Hanna) force 8x downsampling, and even the best method loses visible fine detail; methods that can process large images more efficiently would directly improve reconstruction fidelity on such scenes.
  • In the video-derived scenes, NeRF-based methods kept a more stable white balance while Gaussian-splatting methods better captured sudden white-balance changes, so method choice changes how the same photometric ambiguity is resolved.

Reading between the lines

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

  • A stricter evaluation of the two video scenes would use a temporal split, training on one contiguous block of frames and testing on a later block, rather than a random 12.5% split, because frames extracted every 20th frame from continuous footage are near-duplicates; without such a holdout, the reported video-scene scores likely overstate generalization to genuinely new viewpoints.
  • The dataset is well suited as a controlled testbed for preprocessing: one could quantify how much of the method gap disappears when automatic white balance is post-corrected or when dynamic background regions are masked out before training.
  • The low-texture, perforated white surface is close to an adversarial input for feature-based structure-from-motion; extending the dataset with more artists or materials could turn it into a broader robustness benchmark rather than a single-style stress test.
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 introduces HuSc3D, a new dataset of six real-world scenes showing white, intricately perforated sculptures by Paweł Althamer, captured under conditions that mimic casual acquisition by a non-expert user: sparse COLMAP-registered views (Kacper, Paulina), high-resolution phone stills requiring downsampling (Jarosław, Hanna), and phone-video frames with automatic white-balance shifts (Tosia, Wiktor). Six popular reconstruction methods—2DGS, Mip-Splatting, Nerfacto, Splatfacto, Splatfacto-MCMC, and Instant-NGP—are benchmarked with PSNR, SSIM, LPIPS, training time, and FPS. The central claim is that HuSc3D's combination of dynamic background, photometric inconsistency, sparse views, and low-texture objects 'effectively differentiat[es] model performance,' exposing limitations masked by conventional datasets.

Significance. If the quantitative claims are made secure, HuSc3D would be a useful benchmark: it addresses a real gap, uses real public sculptures, combines several realistic acquisition challenges that existing datasets treat only partially, and evaluates a reasonable set of open-source baselines under a stated protocol. The qualitative results support the rough ordering (Splatfacto-MCMC best, Nerfacto worst) and the dataset artifact itself, with code and data publicly announced, is a concrete community resource. However, the central 'distinctiveness' claim is currently supported mainly by a single quantitative table on HuSc3D alone, without statistical error bars, without a temporal holdout for video-derived scenes, and without same-protocol comparisons on existing datasets; these issues must be addressed before the benchmark's differentiating power is established.

major comments (3)
  1. [Section 4; Section 3 (Tosia and Wiktor)] The test split is not a valid generalization split for the video-derived scenes. Section 4 states that 'Following the protocol of 3DGS 12.5% of images from each scene are reserved as a test set,' while Section 3 reports that Tosia and Wiktor were obtained by extracting every 20th frame from continuous phone video. A random 12.5% split places each test frame within one or two extracted-frame steps of a training frame in capture time, so test views are near-duplicates of training views in viewpoint, lighting, and background. This likely inflates PSNR, SSIM, and LPIPS for those scenes and can change the method ranking reported in Table 2. Please rerun the benchmark with a temporal holdout (e.g., a contiguous block of frames or frames spaced by a sufficient capture-time interval) and report both sets of numbers, or justify why temporal leakage would not affect the conclusions.
  2. [Table 2; Section 4 (Quantitative results)] All quantitative results are single-run point estimates with no variance. The paper's central claim that HuSc3D 'effectively differentiat[es] model performance' is a statistical statement about rankings and margins, yet Table 2 reports no multiple seeds, confidence intervals, or significance tests. For example, the PSNR gap between Splatfacto and Mip-Splatting is about 0.7 dB on average, and per-scene differences vary in sign; without error bars it is unclear whether the ordering is stable. Please report results over at least three seeds (with different random initializations and test splits) and include standard deviations or confidence intervals, especially for the averaged metrics.
  3. [Section 4 (Comparative Analysis); Table 3] The 'distinctiveness' of HuSc3D is asserted from a qualitative feature-by-feature comparison in Table 3, not from a direct measurement. To support the statement that HuSc3D 'effectively differentiat[es] model performance' better than conventional datasets, the same set of methods should be evaluated under the same protocol on at least one or two representative existing datasets (e.g., LLFF or Mip-NeRF 360 subset), and the spread or ranking separability of metrics should be compared. Without such a head-to-head comparison, the claim that limitations are 'often masked by more conventional datasets' remains unsupported by the experiments presented.
minor comments (5)
  1. [Section 2] The paragraph describing Mip-NeRF 360 is duplicated verbatim; please remove one copy.
  2. [Throughout] There are several typos and grammatical slips, for example 'NICON D7000' should be 'Nikon D7000', 'are are particurarly' should be 'are particularly', 'sensivities' should be 'sensitivities', and 'more taggle' should be 'more tackle'; a thorough proofread is needed.
  3. [Table 2] Several time entries have missing separators, e.g., '7:328:20' for 2DGS on Tosia and '6:2110:45' for Splatfacto on Wiktor; these likely need spaces or a consistent format.
  4. [Table 1] The DeepBlending row uses a '‡' symbol with no accompanying footnote explaining the symbol, and the header text 'Min no of Max no of images per scene' is ambiguous; please clean up the table formatting.
  5. [Section 3 (Tosia and Wiktor scenes)] The sentence 'Tosia scene has 13.2% of scenes with misleading white balance and Wiktor 28.4%' should say 'frames' rather than 'scenes', and it would be helpful to state how this percentage was computed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: HuSc3D is a new artifact benchmarked with external methods; self-citations are only related-work examples.

full rationale

HuSc3D is a dataset contribution, not a derived theoretical claim. The central claim that the dataset separates reconstruction methods rests on measuring six external baseline implementations (2DGS, Mip-Splatting, Nerfacto, Splatfacto, Splatfacto-MCMC, Instant-NGP) with standard PSNR, SSIM, LPIPS, training time, and FPS metrics on held-out images. These methods come from public codebases with default hyperparameters and are not fit to the dataset's distinctiveness claim, so the observed performance spread is independent evidence rather than an input to the paper's reasoning. The only author self-citations are [12] and [14] in a related-work survey sentence; neither is load-bearing for the dataset's design, the evaluation protocol, or the conclusions. No equation, fitted parameter, or uniqueness theorem is invoked, and no prediction is defined in terms of the data it claims to explain. The random 12.5% test split for the video-derived Tosia and Wiktor scenes is a plausible temporal-leakage validity concern for the benchmark numbers, but that is an evaluation-design issue, not circularity of the kind defined in this analysis. Accordingly, the derivation chain is self-contained and the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The dataset paper introduces no new physical or theoretical entities. Its claims rest on experimental design choices (downsampling, AWB labeling) and domain assumptions about what constitutes a representative real-world capture.

free parameters (2)
  • AWB misclassification threshold = 13.2% (Tosia), 28.4% (Wiktor)
    The percentages of frames judged to have incorrect white balance are based on visual inspection with no stated objective criterion; they quantify the dataset's challenge level and are used to support the claim of photometric inconsistency.
  • Input downsampling factors = 4x (most scenes), 8x (Hanna, Jaroslaw)
    Chosen by hand to make training feasible on an RTX 4090; the high-resolution scenes are intentionally downsampled, which changes the measured difficulty and discards fine detail.
assumptions (3)
  • domain assumption COLMAP-registered subsets of 66 and 104 images represent a realistic sparse-view challenge for reconstruction
    Section 3 (Kacper and Paulina scenes) treats the low number of registered photos as a property of the capture conditions, not as an artifact of the chosen hyperparameters.
  • domain assumption White, low-texture, perforated surfaces are a representative hard case for contemporary 3D reconstruction
    Section 1 and Section 2 state this as the motivation for the dataset; it justifies the dataset's design but is not derived from data.
  • domain assumption Image-based metrics (PSNR/SSIM/LPIPS) capture the reconstruction quality relevant to fine geometry and color ambiguity
    The paper uses only render-quality metrics and does not evaluate geometric accuracy; it assumes these metrics reflect the challenges named.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HuSc3D: Human Sculpture dataset for 3D object reconstruction." pith.science (2026). https://pith.science/paper/AEK5OIMD

@misc{pith2026250607628,
  author       = {Pith},
  title        = {Pith review of: HuSc3D: Human Sculpture dataset for 3D object reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AEK5OIMD}},
  note         = {Machine review of arXiv:2506.07628}
}
read the original abstract

3D scene reconstruction from 2D images is one of the most important tasks in computer graphics. Unfortunately, existing datasets and benchmarks concentrate on idealized synthetic or meticulously captured realistic data. Such benchmarks fail to convey the inherent complexities encountered in newly acquired real-world scenes. In such scenes especially those acquired outside, the background is often dynamic, and by popular usage of cell phone cameras, there might be discrepancies in, e.g., white balance. To address this gap, we present HuSc3D, a novel dataset specifically designed for rigorous benchmarking of 3D reconstruction models under realistic acquisition challenges. Our dataset uniquely features six highly detailed, fully white sculptures characterized by intricate perforations and minimal textural and color variation. Furthermore, the number of images per scene varies significantly, introducing the additional challenge of limited training data for some instances alongside scenes with a standard number of views. By evaluating popular 3D reconstruction methods on this diverse dataset, we demonstrate the distinctiveness of HuSc3D in effectively differentiating model performance, particularly highlighting the sensitivity of methods to fine geometric details, color ambiguity, and varying data availability--limitations often masked by more conventional datasets.

Figures

Figures reproduced from arXiv: 2506.07628 by the authors.

Figure 1
Figure 1. The HuSc3D dataset is designed for 3D reconstruction from 2D images, simulating a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualisation of cameras in different scenes in HuSc3D. The visualisation was made in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Examples of Paulina and Kacper scenes’ images. Kacper and Paulina scenes To simulate chal￾lenges arising from data acquisition limitations, the Kacper and Paulina scenes introduce the diffi￾culty of sparse input views. These sets were made with photos from NICON D7000 and AF-S DX Nikkor 18-105mm f/3.5-5.6G ED VR lens using fixed ISO, aperture and shutter speed. The autofo￾cus was locked between sets of photos. The m… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The necessary 8x downsizing of input images led to substantial detail loss, preventing even [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Example of variety in auto￾matic white balance (AWB) in Tosia scene. Especially, when considering the wall in the background, a human eye can find incorrect AWB photos easily. Tosia and Wiktor scenes To simulate common acqui￾sition issues with current mobile devices an…
Figure 6
Figure 6. Figure 6: In scenes with low training samples (i.e. Paulina scene) methods based on NeRF struggle [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results on Jaroslaw, Kacper and Paulina scenes. The majority of methods strug￾gled on scenes with sparse number of training views (Kacper and Paulina). We have chosen the same frames for a fair comparison, however, Mip￾Splatting also showed artifacts (see …
Figure 8
Figure 8. Figure 8: Qualitative results on Hanna, Tosia and Wiktor scenes. Splatfacto-MCMC achieved results [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Examples of views from Paulina and Kacper scenes (sparse training views) in 2DGS, [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 14 canonical work pages

  1. [1]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. InECCV, 2020

    2020

  2. [2]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4), July 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4), July 2023

    2023

  3. [3]

    Pixelwise view selection for unstructured multi-view stereo

    Johannes Lutz Schönberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. InEuropean Conference on Computer Vision (ECCV), 2016

    2016

  4. [4]

    Structure-from-motion revisited

    Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  5. [5]

    Srinivasan, Peter Hedman, Ricardo Martin-Brualla, and Jonathan T

    Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, Peter Hedman, Ricardo Martin-Brualla, and Jonathan T. Barron. MultiNeRF: A Code Release for Mip-NeRF 360, Ref-NeRF, and RawNeRF, 2022

    2022

  6. [6]

    Barron, and Pratul P

    Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields.CVPR, 2022

    work page 2022

  7. [7]

    D-NeRF: Neural Radiance Fields for Dynamic Scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020

    work page 2020

  8. [8]

    Headnerf: A real-time nerf-based parametric head model

    Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20374–20384, June 2022

    work page 2022

Show all 24 references
  1. [9]

    Depth-supervised nerf: Fewer views and faster training for free

    Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12882–12891, June 2022

    2022

  2. [10]

    4d gaussian splatting for real-time dynamic scene rendering

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023

    2023 arXiv

  3. [11]

    Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes.arXiv preprint arXiv:2312.14937, 2023

    Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes.arXiv preprint arXiv:2312.14937, 2023

    2023 arXiv

  4. [12]

    D- miso: Editing dynamic 3d scenes using multi-gaussians soup.arXiv preprint arXiv:2405.14276, 2024

    Joanna Waczy´nska, Piotr Borycki, Joanna Kaleta, Sławomir Tadeja, and Przemysław Spurek. D- miso: Editing dynamic 3d scenes using multi-gaussians soup.arXiv preprint arXiv:2405.14276, 2024

    2024 arXiv

  5. [13]

    2d gaussian splatting for geometrically accurate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. InSIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024

    2024

  6. [14]

    Gaussian splatting with nerf-based color and opacity.Computer Vision and Image Understanding, 251:104273, 2025

    Dawid Malarz, Weronika Smolak-Dy˙zewska, Jacek Tabor, Sławomir Tadeja, and Przemysław Spurek. Gaussian splatting with nerf-based color and opacity.Computer Vision and Image Understanding, 251:104273, 2025

    2025

  7. [15]

    Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018

    Blender Online Community.Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018

    2018

  8. [16]

    A photogrammetry-based framework to facilitate image-based modeling and automatic camera tracking.International Conference on Computer Graphics Theory and Applications, 2021

    Sebastian Bullinger, Christoph Bodensteiner, and Michael Arens. A photogrammetry-based framework to facilitate image-based modeling and automatic camera tracking.International Conference on Computer Graphics Theory and Applications, 2021

    2021

  9. [17]

    3dgs render blender addon by kiri engine

    Kiri-Innovation. 3dgs render blender addon by kiri engine. https://github.com/ Kiri-Innovation/3dgs-render-blender-addon, 2025

    2025

  10. [18]

    Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar

    Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines.ACM Transactions on Graphics (TOG), 2019. 10

    2019

  11. [19]

    Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction

    Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. InInternational Conference on Computer Vision, 2021

    2021

  12. [20]

    Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (SIGGRAPH Asia Conference Proceedings), 37(6), November 2018

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (SIGGRAPH Asia Conference Proceedings), 37(6), November 2018

    2018

  13. [21]

    Gaussian opacity fields: Efficient high-quality compact surface reconstruction in unbounded scenes.arXiv:2404.10772, 2024

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient high-quality compact surface reconstruction in unbounded scenes.arXiv:2404.10772, 2024

    2024 arXiv

  14. [22]

    Nerfstudio: A modular framework for neural radiance field development

    Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGG...

    2023

  15. [23]

    3d gaussian splatting as markov chain monte carlo

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Pre...

    2024

  16. [24]

    Instant neural graphics primitives with a multiresolution hash encoding.ACM Trans

    Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding.ACM Trans. Graph., 41(4):102:1–102:15, July 2022. 11

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.