Pith. sign in

REVIEW 3 major objections 5 minor 61 references

NAT: Neural Acoustic Transfer for Interactive Scenes in Real Time

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 A per-scene neural field predicts acoustic transfer maps for moving, resized, or re-materialized objects in a few milliseconds, making real-time dynamic sound rendering practical.

desk verdict A solid, genuinely useful method paper whose central claim survives scrutiny; the main weakness is an unproven singularity treatment in the data-generation pipeline, which is addressable and not fatal. read the letter →

arxiv 2506.06190 v1 pith:YOLGWEBD submitted 2025-06-06 cs.SD cs.GReess.AS

Xutong Jin , Bo Pang , Chenxi Xu , Xinyun Hou , Guoping Wang , Sheng Li This is my paper
classification cs.SDcs.GReess.AS
keywords NeuralacoustictransferBoundaryelementmethodMonteCarlointegrationModalsoundsynthesisReal-timerenderingfieldsFar-fieldmaps
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 argues that a neural network trained on boundary-element data can stand in for precomputed acoustic transfer throughout a continuous space of scene conditions. The target is real-time interactive audio: when an object moves, rotates, changes size, or changes material, the network re-derives the sound field in one to four milliseconds rather than re-solving a costly wave equation. The authors claim this matches or beats the leading prior neural method in accuracy while being orders of magnitude faster, and supports animated scenes with coupled moving sources and resonant cavities. This makes dynamic virtual reality and audio production a practical setting for instant, geometry- and material-responsive sound.

What carries the argument

The machinery is the neural FFAT map, a continuous function $\Phi(\theta, \phi, r, v, f) = \mathrm{MLP}(G(\theta, \phi, r), P(v), P(f))$, where $G$ is a multi-resolution hash grid in spherical coordinates and $P$ is sinusoidal positional encoding. The hash grid supplies high-frequency spatial detail, the positional encoding lets a single network span ranges of frequency and scene conditions, and the MLP combines them into transfer values. Training data come from two solvers: a CUDA-accelerated boundary element method (BEM) for high precision, and a Monte-Carlo BEM approximation (BEM-MC) that replaces mesh quadrature with Poisson-disk boundary sampling plus a local singularity treatment, used for smooth Neumann scenes.

What would settle it

Run BEM-MC and standard BEM on a sphere with a known analytic radiation solution and a deliberately curved, well-resolved mesh; if the FFAT-map error of BEM-MC grows with surface curvature or does not shrink as Poisson-disk sample count increases, the neglected gradient integral is biasing the solver, and NAT inherits that bias.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single neural network per scene can absorb the full variation of the acoustic transfer function over a continuous family of scene configurations. The network maps listener coordinates in a spherical hash grid plus condition parameters (position, rotation, size, material, shape proxy) and frequency through a small MLP to the far-field acoustic transfer map, so that at inference the expensive boundary-element solve is replaced by one forward pass. In validation, the predicted FFAT maps reach SNR 9.9–17.4 dB and SSIM 0.81–0.93 against BEM ground truth, while a 60-mode material and size edit takes about 2 ms and dynamic-scene transfer masks for 30 s of audio take 1–4 ms, several orders of magnitude faster than the prior neural method and with higher reported precision.

Load-bearing premise

The training-data pipeline rests on the assumption that a certain singular term in the boundary integral — the gradient of the Green's function integrated over a tiny disk around the evaluation point — is exactly zero because the radius vector is perpendicular to the surface normal; that is only exactly true at the center of a flat tangent disk, and the paper gives no error bound for curved surfaces.

Editorial extensions

If this is right

  • A trained NAT model makes material and size editing of modal sound objects interactive: new material and size settings produce FFAT maps for all modes in about 2 ms, so a designer can hear the change live.
  • Dynamic scenes with moving sources and resonating cavities, such as a phone moving in and out of a cup, can render a full acoustic transfer mask for 30 seconds of audio in 1–4 ms.
  • Because the neural fit averages over nearby condition samples, NAT can report higher SSIM than the Monte-Carlo solver that produced its training data, meaning network inference acts as a variance reducer.
  • The per-scene cost is roughly an hour of precomputation, about five minutes of training, and around one megabyte of model storage, so it is practical to replace stored transfer data with a small network.

Reading between the lines

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

  • The same conditioning formulation could be extended to listener position and global environment state, effectively yielding a parameterized room impulse response; the paper stops at object-conditioned transfer rather than full room acoustics.
  • Because the method trains a fresh network for each scene, it does not yet generalize to unseen scenes; a latent-space or multi-scene version would be the natural next step the paper leaves implicit.
  • The reported SNR range (roughly 10–17 dB) suggests the target is audible plausibility rather than exact wave physics, so a listening study would be the right test of whether the accuracy actually matters perceptually.
  • BEM-MC's sensitivity to the discarded gradient singularity could be probed directly by comparing against analytic radiation solutions on curved objects; if bias appears, NAT trained on BEM-MC inherits it.
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 proposes NAT, a per-scene implicit neural representation of acoustic transfer maps parameterized by listener position, frequency, and scene condition variables such as position, rotation, size, material, and shape. Training data are generated either by a CUDA-accelerated boundary element method (BEM) or by a proposed Monte-Carlo-based BEM approximation (BEM-MC) for smooth Neumann problems. The authors report 1-4 ms inference for 30 s audio, SNR 9.9-17.4 dB and SSIM 0.81-0.93 against BEM, and large speedups over NeuralSound, BEM, BEM-MC, and FDTD. The central claim is that, over a parameterized family of scene configurations, NAT predicts acoustic transfer in real time with accuracy close to BEM.

Significance. If the claims hold, the paper addresses a real bottleneck in interactive audio: making acoustic transfer editable and evaluable in real time for dynamically varying scenes. The empirical comparison is substantial: Tables 1, 2, and 4 include BEM, BEM-MC, NeuralSound, and NAT under controlled conditions, and Table 5 reports practical model sizes and inference times. The paper also honestly discloses that NAT can exceed its own training-data generator in SSIM, and attributes this to variance reduction, which is a reasonable and testable explanation. The experimental protocol is described in enough detail to be reproduced, and the speed-accuracy tradeoffs are reported clearly. The main weakness is theoretical: the BEM-MC singularity treatment in Eq. (13) is not derived or bounded, and since BEM-MC supplies most of the dynamic-scene training data, this gap propagates into the central claim.

major comments (3)
  1. [Sec. 4.2, Eq. (13)] The neglect of the Green's-function-gradient integral over the small disk is not justified for curved surfaces. The perpendicularity of r to n_y holds only at the disk center on a flat tangent plane; on a smooth curved surface r·n_y = O(ρ^2/R), so ∂r/∂n_y = O(ρ/R), and the omitted contribution is O(ε/R)p(y_i). The standard principal-value jump of the boundary integral does not vanish in general, and the manuscript supplies no bound in terms of ε, curvature, and wavenumber k. The pointer in Sec. 4.1.1 ('please refer to Section 6 for technical detail') is also unsatisfied, because Section 6 is the validation section rather than a derivation. Since BEM-MC is the training-data generator for the phone-cup scene and all extensive dynamic scenes (Tables 2, 4, 5; Figs. 3-6), any systematic bias in this term is inherited by NAT. Please provide the missing derivation and error bound, or empirically quantify the term by comparing BEM-MC solutions with and without the disk contribution on representative curved geometries.
  2. [Sec. 4.2, Eqs. (9)-(12)] The small-disk radius ε is a free parameter that is never specified or linked to the Poisson-disk sampling distance. The one-point estimate in Eq. (12) is dimensionally consistent but is a leading-order asymptotic approximation, not a standard N=1 Monte Carlo estimator of the disk integral (the area factor π ε^2 does not appear); the O(ε^2) and k-dependent error terms are not quantified. Because ε controls the magnitude of both the retained term in Eq. (12) and the neglected term in Eq. (13), a stated selection rule and a sensitivity study (e.g., varying ε at a fixed sample count) are needed to make the BEM-MC predictions reproducible and trustworthy.
  3. [Sec. 7, Table 5] For the six extensive scenes in Fig. 6, the paper reports only spectrograms and inference times; no SNR/SSIM or other accuracy metric against BEM or BEM-MC is given. These scenes include morphing geometry and multiple coupled sound sources, which are qualitatively more complex than the plate and phone-cup validations. The statement that NAT 'accurately captures' acoustic transfer across all these scenarios is therefore not directly supported by quantitative evidence. Adding at least one held-out quantitative validation for a morphing-shape or coupled-source scene would strengthen the generality of the central claim.
minor comments (5)
  1. [Sec. 6.3.1] The sentence 'The comparison of time cost and average precision ... is presented in Table 1' should refer to Table 4, which contains the 8-mode BEM-MC experiment; Table 1 reports the 60-mode BEM-based experiment.
  2. [Sec. 5.1 vs Sec. 7] The phone displacement range is stated as 0-0.3 m in Sec. 5.1 and as 0-0.2 m in Sec. 7; please reconcile the two descriptions.
  3. [Sec. 4.1.1, Eq. (7)] The symbol N is used both for the number of Monte Carlo samples in Eq. (7) and for the single-sample approximation in Eq. (12); using a distinct symbol (e.g., M_s) would avoid confusion.
  4. [Sec. 6.2] The statement that BEM-MC 'surpasses' CUDA-accelerated BEM is based on qualitative FFAT images in Fig. 8; please provide a quantitative metric for the analytical test cases to support this claim.
  5. [Sec. 8] The limitations paragraph mentions untuned hyperparameters and limited scene parameters, but does not mention the reliance on the unproven BEM-MC singularity treatment or the fact that only amplitude (not phase) is modeled; both are relevant caveats for the reported accuracy.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: NAT is trained on BEM-generated labels and independently validated against BEM ground truth on held-out cases.

full rationale

The central claim of the paper is that a per-scene neural network, NAT, can predict acoustic transfer maps under varying conditions (position, rotation, size, material, shape) in milliseconds. The training labels come from either CUDA-accelerated BEM or BEM-MC, and the evaluation is performed against BEM as an external ground truth on held-out cases (Tables 1, 2, and 4; Figs. 2-4 and 9). This is standard supervised function approximation, not a fitted parameter renamed as a prediction: the network is not optimized against the test-time BEM solutions it is compared with. The reported cases where NAT surpasses its own BEM-MC training data in SSIM are explicitly disclosed as variance reduction and are cross-checked against BEM in the same tables, so they do not constitute a circular student-beats-teacher argument. The only self-citation of note is the authors' prior work NeuralSound [11], which is used as a comparison baseline rather than as load-bearing justification for NAT's accuracy; the comparison is external and quantified. The paper does contain an omitted technical derivation in Section 4.1.1/4.2: Eq. (13) sets the singular gradient integral to zero using a perpendicularity argument whose validity is conditional, and the promised reference to 'Section 6 for technical detail' points to a validation section rather than a derivation. This is a correctness and numerical-accuracy risk inherited by BEM-MC training data, but it is not circularity: it is an unproven approximation, not an equation that reduces to its own input or a claim justified only by self-citation. Therefore, no load-bearing circularity is present, and the appropriate circularity score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on five hand-chosen numerical settings (disk radius, sampling counts, network shape, evaluation radius range, spatial sample density) and on three load-bearing assumptions: the disk singularity treatment of Eqs. (12)-(13), the smooth-Neumann applicability condition for BEM-MC, and the Helmholtz scaling used to parameterize material and size edits. No new physical entities are introduced; the 'neural FFAT map' is a computational surrogate for the standard sound-pressure field, not a new entity.

free parameters (5)
  • small-disk radius epsilon = not reported
    Chosen by hand for the singularity treatment in Section 4.2; its value never appears in the paper, and the BEM-MC results depend on it.
  • Poisson disk sample count M = 1000, 2000, 4000
    Sampling counts in Section 6.1 chosen to trade accuracy against runtime; reported accuracies (SNR 2-15 dB vs BEM) depend on these counts.
  • spatial sample points per condition = 10,000
    Section 3.4: for each sampled condition and frequency, 10,000 listener positions are generated; the training coverage density is chosen by hand.
  • network hyperparameters = 4x128 MLP; 4 hash levels at 8^3 to 64^3; feature length 4; 6 positional-encoding frequencies
    Section 3.4: chosen by hand and admitted by the authors not to be fine-tuned (Section 8, limitations). These affect the reported SNR and SSIM.
  • evaluation radius range = 1.5 to 3 times the scene bounding-box radius
    Section 3.4: the FFAT map training region is restricted to this range; accuracy outside the range is not validated.
assumptions (6)
  • standard math Standard boundary element and Burton-Miller boundary integral equation theory (Eq. 6)
    Used to formulate the exterior Neumann problem; standard results from [5, 55].
  • ad hoc to paper The integral of the Green's function gradient over the small disk around the collocation point vanishes (Eq. 13)
    Section 4.2: justified by 'the perpendicularity of vector r to the surface normal ny', which holds only at the disk center on a flat tangent disk; curvature terms are dropped without error bounds.
  • ad hoc to paper One-point Monte-Carlo estimate of the singular disk integral (Eq. 12)
    Section 4.2: the disk integral is replaced by the integrand at the center times the disk area; the induced error is uncontrolled, and the derivation cross-references a section that contains no derivation.
  • domain assumption Sufficiently smooth boundary for the local disk approximation
    Section 4.2 states 'Given the smoothness of the domain boundary, local approximations near yi are feasible and mathematically sound'; BEM-MC is restricted to smooth Neumann scenes (speaker Neumann 1, passive surfaces Neumann 0).
  • domain assumption Neural network variance reduction makes BEM-MC data adequate for training
    Sections 5.1 and 6.3.1: NAT is claimed to outperform its own noisy training data by fitting a smooth field; this is validated only empirically in the reported scenarios.
  • domain assumption Helmholtz scaling: acoustic transfer depends on frequency times size
    Section 6.3: frequency multiplied by size is used as the condition parameter for material and size editing; assumed valid for all 60 modes used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NAT: Neural Acoustic Transfer for Interactive Scenes in Real Time." pith.science (2026). https://pith.science/paper/YOLGWEBD

@misc{pith2026250606190,
  author       = {Pith},
  title        = {Pith review of: NAT: Neural Acoustic Transfer for Interactive Scenes in Real Time},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YOLGWEBD}},
  note         = {Machine review of arXiv:2506.06190}
}
read the original abstract

Previous acoustic transfer methods rely on extensive precomputation and storage of data to enable real-time interaction and auditory feedback. However, these methods struggle with complex scenes, especially when dynamic changes in object position, material, and size significantly alter sound effects. These continuous variations lead to fluctuating acoustic transfer distributions, making it challenging to represent with basic data structures and render efficiently in real time. To address this challenge, we present Neural Acoustic Transfer, a novel approach that utilizes an implicit neural representation to encode precomputed acoustic transfer and its variations, allowing for real-time prediction of sound fields under varying conditions. To efficiently generate the training data required for the neural acoustic field, we developed a fast Monte-Carlo-based boundary element method (BEM) approximation for general scenarios with smooth Neumann conditions. Additionally, we implemented a GPU-accelerated version of standard BEM for scenarios requiring higher precision. These methods provide the necessary training data, enabling our neural network to accurately model the sound radiation space. We demonstrate our method's numerical accuracy and runtime efficiency (within several milliseconds for 30s audio) through comprehensive validation and comparisons in diverse acoustic transfer scenarios. Our approach allows for efficient and accurate modeling of sound behavior in dynamically changing environments, which can benefit a wide range of interactive applications such as virtual reality, augmented reality, and advanced audio production.

Figures

Figures reproduced from arXiv: 2506.06190 by the authors.

Figure 1
Figure 1. Our Neural Acoustic Transfer implicitly models the acoustic transfer function under changing conditions, as well as a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Acoustic transfer tests on a scene involving material and size editing of a modal sound object with [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Comparisons of FFAT maps for different relative positions of a cup and a vibrating phone at frequencies of [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: A scenario where a vibrating phone, with only its bottom microphone position vibrating, is moved from inside a [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Visualization demonstrating how our NAT synthesizes sound effects for a dynamic scene. Initially, we set an audio [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Six extensive cases involve dynamic changes in acoustic transfer: (a) A phone playing audio moves in and out of a [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Acoustic transfer was tested on diverse objects with different materials. For each object shown in the leftmost [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 9
Figure 9. Figure 9: Both BEM-MC and NAT consistently exhibit closer [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 8
Figure 8. Figure 8: Analytical test case involves three dipole sound [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Acoustic transfer tests on a scene involving materials and size editing of a modal sound object with the first [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 51 canonical work pages

  1. [1]

    Scanning physical interaction behavior of 3d objects,

    D. K. Pai, K. v. d. Doel, D. L. James, J. Lang, J. E. Lloyd, J. L. Rich- mond, and S. H. Yau, “Scanning physical interaction behavior of 3d objects,” in Proceedings of the 28th annual conference on Computer graphics and interactive techniques , 2001, pp. 87–96

    work page 2001

  2. [2]

    Synthesizing sounds from rigid-body simulations,

    J. F. O’Brien, C. Shen, and C. M. Gatchalian, “Synthesizing sounds from rigid-body simulations,” in Proceedings of the 2002 ACM SIGGRAPH/Eurographics symposium on Computer animation , 2002, pp. 175–181

    work page 2002

  3. [3]

    Precomputed acoustic transfer: output-sensitive, accurate sound generation for geomet- rically complex vibration sources,

    D. L. James, J. Barbi ˇc, and D. K. Pai, “Precomputed acoustic transfer: output-sensitive, accurate sound generation for geomet- rically complex vibration sources,” ACM T ransactions on Graphics (Proceedings of SIGGRAPH 2006) , vol. 25, no. 3, pp. 987–995, 2006

    work page 2006

  4. [4]

    Harmonic shells: a practical nonlinear sound model for near-rigid thin shells,

    J. N. Chadwick, S. S. An, and D. L. James, “Harmonic shells: a practical nonlinear sound model for near-rigid thin shells,” ACM T ransactions on Graphics (Proceedings of SIGGRAPH 2009) , vol. 28, no. 5, pp. 1–10, 2009

    work page 2009

  5. [5]

    The boundary element method in acoustics: A survey,

    S. Kirkup, “The boundary element method in acoustics: A survey,” Applied Sciences, vol. 9, no. 8, p. 1642, 2019

    work page 2019

  6. [6]

    Interactive acoustic transfer approx- imation for modal sound,

    D. Li, Y. Fei, and C. Zheng, “Interactive acoustic transfer approx- imation for modal sound,” ACM T ransactions on Graphics (TOG) , vol. 35, no. 1, pp. 1–16, 2015

    work page 2015

  7. [8]

    Neural radiosity,

    S. Hadadan, S. Chen, and M. Zwicker, “Neural radiosity,” ACM T ransactions on Graphics (TOG), vol. 40, no. 6, pp. 1–11, 2021

    work page 2021

  8. [10]

    tiny-cuda-nn,

    T. M ¨uller, “tiny-cuda-nn,” 4 2021. [Online]. Available: https: //github.com/NVlabs/tiny-cuda-nn

    2021

Show all 61 references
  1. [11]

    Neuralsound: Learning-based modal sound synthesis with acoustic transfer,

    X. Jin, S. Li, G. Wang, and D. Manocha, “Neuralsound: Learning-based modal sound synthesis with acoustic transfer,” ACM T rans. Graph. , vol. 41, no. 4, jul 2022. [Online]. Available: https://doi.org/10.1145/3528223.3530184

    2022

  2. [12]

    Toward high-quality modal contact sound,

    C. Zheng and D. L. James, “Toward high-quality modal contact sound,” in ACM SIGGRAPH 2011 papers , 2011, pp. 1–12

    2011

  3. [13]

    Rigid-body fracture sound with precomputed sound- banks,

    ——, “Rigid-body fracture sound with precomputed sound- banks,” in ACM SIGGRAPH 2010 papers , 2010, pp. 1–13

    2010

  4. [14]

    Eigenmode compression for modal sound models,

    T. R. Langlois, S. S. An, K. K. Jin, and D. L. James, “Eigenmode compression for modal sound models,” ACM T ransactions on Graphics (Proceedings of SIGGRAPH 2014) , vol. 33, no. 4, Aug. 2014

    2014

  5. [15]

    Syncopation: Interactive synthesis-coupled sound propagation,

    A. Rungta, C. Schissler, R. Mehra, C. Malloy, M. Lin, and D. Manocha, “Syncopation: Interactive synthesis-coupled sound propagation,” IEEE transactions on visualization and computer graph- ics, vol. 22, no. 4, pp. 1346–1355, 2016

    2016

  6. [16]

    Kleinpat: Optimal mode conflation for time-domain precomputation of acoustic transfer,

    J.-H. Wang and D. L. James, “Kleinpat: Optimal mode conflation for time-domain precomputation of acoustic transfer,”ACM T rans- actions on Graphics (Proceedings of SIGGRAPH 2019) , vol. 38, no. 4, pp. 122:1–122:12, Jul. 2019

    2019

  7. [17]

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

    B. Mildenhall, P . P . Srinivasan, M. Tancik, J. T. Barron, R. Ra- mamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” Communications of the ACM , vol. 65, no. 1, pp. 99–106, 2021

    2021

  8. [18]

    Implicit neural representations with periodic activation func- tions,

    V . Sitzmann, J. Martel, A. Bergman, D. Lindell, and G. Wetzstein, “Implicit neural representations with periodic activation func- tions,” Advances in neural information processing systems , vol. 33, pp. 7462–7473, 2020

    2020

  9. [19]

    Fourier features let networks learn high frequency functions in low dimensional domains,

    M. Tancik, P . Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, and R. Ng, “Fourier features let networks learn high frequency functions in low dimensional domains,” Advances in Neural Information Process- ing Systems, vol. 33, ...

    2020

  10. [20]

    Efficient geometry-aware 3d generative adversarial networks,

    E. R. Chan, C. Z. Lin, M. A. Chan, K. Nagano, B. Pan, S. De Mello, O. Gallo, L. J. Guibas, J. Tremblay, S. Khamis et al. , “Efficient geometry-aware 3d generative adversarial networks,” in Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 202...

    2022

  11. [21]

    Tensorf: Tensorial ra- diance fields,

    A. Chen, Z. Xu, A. Geiger, J. Yu, and H. Su, “Tensorf: Tensorial ra- diance fields,” in European Conference on Computer Vision. Springer, 2022, pp. 333–350

    2022

  12. [22]

    Plenoxels: Radiance fields without neural net- works,

    S. Fridovich-Keil, A. Yu, M. Tancik, Q. Chen, B. Recht, and A. Kanazawa, “Plenoxels: Radiance fields without neural net- works,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 5501–5510

    2022

  13. [23]

    Acorn: Adaptive coordinate networks for neural scene representation,

    J. N. Martel, D. B. Lindell, C. Z. Lin, E. R. Chan, M. Monteiro, and G. Wetzstein, “Acorn: Adaptive coordinate networks for neural scene representation,” arXiv preprint arXiv:2105.02788 , 2021

    2021 arXiv

  14. [24]

    Mip-nerf: A multiscale represen- tation for anti-aliasing neural radiance fields,

    J. T. Barron, B. Mildenhall, M. Tancik, P . Hedman, R. Martin- Brualla, and P . P . Srinivasan, “Mip-nerf: A multiscale represen- tation for anti-aliasing neural radiance fields,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 5855–5864

    2021

  15. [25]

    Ref-nerf: Structured view-dependent appearance for neural radiance fields,

    D. Verbin, P . Hedman, B. Mildenhall, T. Zickler, J. T. Barron, and P . P . Srinivasan, “Ref-nerf: Structured view-dependent appearance for neural radiance fields,” in 2022 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) . IEEE, 2022, pp. 5481– 5490

    2022

  16. [26]

    Occupancy networks: Learning 3d reconstruction in 15 function space,

    L. Mescheder, M. Oechsle, M. Niemeyer, S. Nowozin, and A. Geiger, “Occupancy networks: Learning 3d reconstruction in 15 function space,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , 2019, pp. 4460–4470

    2019

  17. [27]

    Deepsdf: Learning continuous signed distance functions for shape representation,

    J. J. Park, P . Florence, J. Straub, R. Newcombe, and S. Love- grove, “Deepsdf: Learning continuous signed distance functions for shape representation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 165–174

    2019

  18. [28]

    Convolutional occupancy networks,

    S. Peng, M. Niemeyer, L. Mescheder, M. Pollefeys, and A. Geiger, “Convolutional occupancy networks,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16 . Springer, 2020, pp. 523–540

    2020

  19. [29]

    Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction,

    P . Wang, L. Liu, Y. Liu, C. Theobalt, T. Komura, and W. Wang, “Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction,” arXiv preprint arXiv:2106.10689 , 2021

    2021 arXiv

  20. [30]

    Learning gradient fields for shape generation,

    R. Cai, G. Yang, H. Averbuch-Elor, Z. Hao, S. Belongie, N. Snavely, and B. Hariharan, “Learning gradient fields for shape generation,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16 . Springer, 2020, pp. 364–381

    2020

  21. [31]

    Bsp-net: Generating compact meshes via binary space partitioning,

    Z. Chen, A. Tagliasacchi, and H. Zhang, “Bsp-net: Generating compact meshes via binary space partitioning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 45–54

    2020

  22. [32]

    Learning implicit fields for generative shape modeling,

    Z. Chen and H. Zhang, “Learning implicit fields for generative shape modeling,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2019, pp. 5939–5948

    2019

  23. [33]

    Pointflow: 3d point cloud generation with continuous normalizing flows,

    G. Yang, X. Huang, Z. Hao, M.-Y. Liu, S. Belongie, and B. Har- iharan, “Pointflow: 3d point cloud generation with continuous normalizing flows,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 4541–4550

    2019

  24. [34]

    Geometry processing with neural fields,

    G. Yang, S. Belongie, B. Hariharan, and V . Koltun, “Geometry processing with neural fields,” Advances in Neural Information Processing Systems, vol. 34, pp. 22 483–22 497, 2021

    2021

  25. [35]

    Generative neural articulated radiance fields,

    A. Bergman, P . Kellnhofer, W. Yifan, E. Chan, D. Lindell, and G. Wetzstein, “Generative neural articulated radiance fields,” Ad- vances in Neural Information Processing Systems , vol. 35, pp. 19 900– 19 916, 2022

    2022

  26. [36]

    Neural articulated ra- diance field,

    A. Noguchi, X. Sun, S. Lin, and T. Harada, “Neural articulated ra- diance field,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 5762–5772

    2021

  27. [37]

    A level set theory for neural implicit evolution under explicit flows,

    I. Mehta, M. Chandraker, and R. Ramamoorthi, “A level set theory for neural implicit evolution under explicit flows,” in European Conference on Computer Vision . Springer, 2022, pp. 711–729

    2022

  28. [38]

    Implicit neural spatial representations for time-dependent pdes,

    H. Chen, R. Wu, E. Grinspun, C. Zheng, and P . Y. Chen, “Implicit neural spatial representations for time-dependent pdes,” in Inter- national Conference on Machine Learning . PMLR, 2023, pp. 5162– 5177

    2023

  29. [39]

    Neural caches for monte carlo partial differential equation solvers,

    Z. Li, G. Yang, X. Deng, C. De Sa, B. Hariharan, and S. Marschner, “Neural caches for monte carlo partial differential equation solvers,” in SIGGRAPH Asia 2023 Conference Papers , 2023, pp. 1– 10

    2023

  30. [40]

    Synthesis of musical instrument sounds: Physics-based modeling or machine learning,

    S. H. Hawley, V . Chatziiannou, and A. Morrison, “Synthesis of musical instrument sounds: Physics-based modeling or machine learning,” Phys. T oday, vol. 16, no. 1, pp. 20–28, 2020

    2020

  31. [41]

    A survey on deep learning for symbolic music generation: Representations, algorithms, evaluations, and challenges,

    S. Ji, X. Yang, and J. Luo, “A survey on deep learning for symbolic music generation: Representations, algorithms, evaluations, and challenges,” ACM Computing Surveys , 2023

    2023

  32. [42]

    Deep-modal: real- time impact sound synthesis for arbitrary shapes,

    X. Jin, S. Li, T. Qu, D. Manocha, and G. Wang, “Deep-modal: real- time impact sound synthesis for arbitrary shapes,” in Proceedings of the 28th ACM International Conference on Multimedia , 2020, pp. 1171–1179

    2020

  33. [43]

    Fast acoustic scattering using convolutional neural networks,

    Z. Fan, V . Vineet, H. Gamper, and N. Raghuvanshi, “Fast acoustic scattering using convolutional neural networks,” in ICASSP 2020- 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 171–175

    2020

  34. [44]

    Machine-learning-based estimation and rendering of scattering in virtual reality,

    V . Pulkki and U. P . Svensson, “Machine-learning-based estimation and rendering of scattering in virtual reality,” The Journal of the Acoustical Society of America , vol. 145, no. 4, pp. 2664–2676, 2019

    2019

  35. [45]

    Ir-gan: Room impulse response generator for far-field speech recognition,

    A. Ratnarajah, Z. Tang, and D. Manocha, “Ir-gan: Room impulse response generator for far-field speech recognition,” arXiv preprint arXiv:2010.13219, 2020

    2010 arXiv

  36. [46]

    Learning acoustic scat- tering fields for dynamic interactive sound propagation,

    Z. Tang, H.-Y. Meng, and D. Manocha, “Learning acoustic scat- tering fields for dynamic interactive sound propagation,” in 2021 IEEE Virtual Reality and 3D User Interfaces (VR) . IEEE, 2021, pp. 835–844

    2021

  37. [47]

    Learning neural acoustic fields,

    A. Luo, Y. Du, M. Tarr, J. Tenenbaum, A. Torralba, and C. Gan, “Learning neural acoustic fields,” Advances in Neural Information Processing Systems, vol. 35, pp. 3165–3177, 2022

    2022

  38. [48]

    Acoustic classification and optimization for multi-modal rendering of real-world scenes,

    C. Schissler, C. Loftin, and D. Manocha, “Acoustic classification and optimization for multi-modal rendering of real-world scenes,” IEEE transactions on visualization and computer graphics , vol. 24, no. 3, pp. 1246–1259, 2017

    2017

  39. [49]

    Scene-aware audio rendering via deep acoustic analysis,

    Z. Tang, N. J. Bryan, D. Li, T. R. Langlois, and D. Manocha, “Scene-aware audio rendering via deep acoustic analysis,” IEEE transactions on visualization and computer graphics , vol. 26, no. 5, pp. 1991–2001, 2020

    1991

  40. [50]

    Some continuous monte carlo methods for the dirichlet problem,

    M. E. Muller, “Some continuous monte carlo methods for the dirichlet problem,” The Annals of Mathematical Statistics , pp. 569– 589, 1956

    1956

  41. [51]

    Monte carlo geometry processing: A grid-free approach to pde-based methods on volumetric do- mains,

    R. Sawhney and K. Crane, “Monte carlo geometry processing: A grid-free approach to pde-based methods on volumetric do- mains,” ACM T ransactions on Graphics, vol. 39, no. 4, 2020

    2020

  42. [52]

    Pharr, W

    M. Pharr, W. Jakob, and G. Humphreys, Physically based rendering: From theory to implementation . MIT Press, 2023

    2023

  43. [53]

    Adam: A method for stochastic optimiza- tion,

    D. P . Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,” in ICLR (Poster), 2015

    2015

  44. [54]

    Automatic differ- entiation in pytorch,

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer, “Automatic differ- entiation in pytorch,” in NIPS-W, 2017

    2017

  45. [55]

    Bempp-cl: A fast python based just-in-time compiling boundary element library

    T. Betcke and M. W. Scroggs, “Bempp-cl: A fast python based just-in-time compiling boundary element library.” Journal of Open Source Software, vol. 6, no. 59, p. 2879, 2021

    2021

  46. [56]

    A prac- tical walk-on-boundary method for boundary value problems,

    R. Sugimoto, T. Chen, Y. Jiang, C. Batty, and T. Hachisuka, “A prac- tical walk-on-boundary method for boundary value problems,” ACM T rans. Graph., vol. 42, no. 4, jul 2023

    2023

  47. [57]

    Walk on stars: A grid-free monte carlo method for pdes with neumann boundary conditions,

    R. Sawhney, B. Miller, I. Gkioulekas, and K. Crane, “Walk on stars: A grid-free monte carlo method for pdes with neumann boundary conditions,” ACM T rans. Graph., vol. 42, no. 4, jul 2023. [Online]. Available: https://doi.org/10.1145/3592398

    2023 doi

  48. [58]

    Parallel poisson disk sampling with spectrum analysis on surfaces,

    J. Bowers, R. Wang, L.-Y. Wei, and D. Maletz, “Parallel poisson disk sampling with spectrum analysis on surfaces,” ACM T ransactions on Graphics (TOG) , vol. 29, no. 6, pp. 1–10, 2010

    2010

  49. [59]

    Toward wave- based sound synthesis for computer animation

    J.-H. Wang, A. Qu, T. R. Langlois, and D. L. James, “Toward wave- based sound synthesis for computer animation.” ACM T rans. Graph., vol. 37, no. 4, pp. 109–1, 2018

    2018

  50. [60]

    Improved water sound synthesis using coupled bubbles,

    K. Xue, R. M. Aronson, J.-H. Wang, T. R. Langlois, and D. L. James, “Improved water sound synthesis using coupled bubbles,” ACM T ransactions on Graphics (TOG), vol. 42, no. 4, pp. 1–13, 2023

    2023

  51. [61]

    Finite difference schemes and simulation in musical acoustics,

    S. Bilbao, “Finite difference schemes and simulation in musical acoustics,” 2009

    2009

  52. [62]

    The perfectly matched layer for acoustic waves in absorptive media,

    Q.-H. Liu and J. Tao, “The perfectly matched layer for acoustic waves in absorptive media,” The Journal of the Acoustical Society of America, vol. 102, no. 4, pp. 2072–2082, 1997

    1997

  53. [63]

    Physically based sound for computer animation and virtual environments,

    D. L. James, “Physically based sound for computer animation and virtual environments,” in ACM SIGGRAPH 2016 Courses , ser. SIGGRAPH ’16. New York, NY, USA: Association for Computing Machinery, 2016

    2016

Pith tools

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