Pith. sign in

REVIEW 3 major objections 9 minor 31 references

Level-Crossing Density as a Mesh-Free High-Frequency Auxiliary Loss for Implicit Neural Representations

T0 review · 3 major / 9 minor · reviewed 2026-07-08 · glm-5.2

Pith's one-line read Level-crossing density becomes a grid-free high-frequency loss for neural fields

desk verdict A clean, honest method paper that turns the Kac-Rice level-crossing density into a differentiable auxiliary loss for INRs. The math is correct, the experiments are disciplined, and the paper reports its own negative results. The central empirical claim is parity, not superiority, and the experimental scope is narrow—but the contribution is real. read the letter →

arxiv 2607.05815 v1 pith:FV6MNLEI submitted 2026-07-07 cs.LG

classification cs.LG
keywords eulertopologycharacteristiccostrepairsalonecomplexcosts
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 takes the Rice level-crossing density—a classical object from random-field theory that links how often a field crosses a threshold to its high-frequency energy—and turns it into a differentiable auxiliary training loss for implicit neural representations. The core mechanism is a smoothed Monte-Carlo estimator built on the co-area formula: at each sampled point, you evaluate the network's gradient magnitude weighted by a Gaussian bump centered at a target level. This estimator is unbiased for the epsilon-smoothed level-set density (crossings per unit length in 1D, level-set length per unit area in 2D, level-set area per unit volume in 3D), requires only automatic differentiation, and works at arbitrary scattered sample locations—no grid, no FFT, no mesh. The auxiliary loss matches the predicted crossing-density profile at 16 quantile levels to the ground-truth profile computed from the same batch points, alongside a standard reconstruction loss. The paper validates the estimator against exact crossing counts and the Rice formula, then evaluates the loss under strict information parity against the Focal Frequency Loss, normalized Sobolev supervision, and MSE-only training. The results are disciplined and mixed: where supervision is scarce and scattered, every auxiliary spectral loss helps substantially (+2.3–3.0 dB over MSE-only on PE-MLP, +1.4–1.8 dB on SIREN), and the crossing-density loss matches FFL within 0.2–0.5 dB on natural images. It overtakes FFL by 0.6 dB only on statistically homogeneous texture—the regime the Rice theory actually describes. On dense regular grids, auxiliary drives are unnecessary and gradient-based ones actively hurt. The loss's genuine distinctions are structural: it is the only method tested that needs neither a sampling grid nor trustworthy pointwise gradient targets, and an oracle-gradient diagnostic confirms it is uniquely insensitive to gradient-target quality because it consumes batch statistics rather than pointwise values.

What carries the argument

Smoothed Monte-Carlo level-crossing density estimator ĉε(u) = (1/N) Σ δε(fθ(xi)−u) ‖∇x fθ(xi)‖, derived from the co-area formula, used as an auxiliary loss matching predicted and target crossing-density profiles at L=16 quantile levels.

What would settle it

On a natural image with scattered supervision, the crossing-density loss would need to beat FFL (after grid resampling) by a clear margin in PSNR. It does not—it matches within 0.2–0.5 dB. The hypothesis that crossing statistics would broadly beat resampled frequency losses off-grid is not confirmed.

Watch

Extended reading notes

Core claim

The central discovery is that level-crossing density, estimated pointwise via the co-area formula as a Gaussian-smoothed average of gradient magnitudes at target levels, functions as a working mesh-free high-frequency auxiliary loss for implicit neural representations. The estimator is unbiased, differentiable through standard autodiff, and operates on scattered samples without any grid or FFT. Its effectiveness is content-dependent: it matches frequency-domain and gradient-matching losses on edge-dominated natural images but only surpasses them on statistically homogeneous signals, exactly where the underlying random-field theory applies. Its robustness to noisy gradient targets—demonstraed

Load-bearing premise

The loss constrains only L distributional numbers per batch (the crossing-density profile at 16 levels), which is a far weaker constraint than a full frequency spectrum or N pointwise gradient values. If a signal's high-frequency content cannot be adequately summarized by level-crossing statistics at those levels—as appears to be the case for edge-dominated, non-stationary natural images—the loss cannot outperform methods that constrain fuller spectral or gradient structure.

Editorial extensions

If this is right

  • On scattered or non-uniformly sampled domains where no regular grid exists, the crossing-density loss provides a high-frequency drive without the interpolation step that grid-based frequency losses require, avoiding smearing of exactly the high-frequency content being supervised.
  • The loss's insensitivity to gradient-target quality (shown via oracle diagnostics) means it can be used in settings where gradient estimation is genuinely hard—point clouds with noisy normals, adaptively sampled renderings—without the performance penalty that pointwise gradient-matching losses suffer.
  • The distributional character of the loss (constraining L batch-level statistics rather than N pointwise values) suggests a design principle: statistical summaries of geometric structure can serve as robust spectral surrogates where full spectral or pointwise information is unavailable, at the cost of constraining less information per batch.
  • The content-dependent advantage on homogeneous texture versus parity on edge-dominated images indicates that the method's applicability is governed by how well the signal's high-frequency content is summarized by level-crossing statistics—a testable criterion for when to prefer this loss over alternatives.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 9 minor

Summary. The paper proposes a mesh-free, FFT-free auxiliary loss for implicit neural representations (INRs) based on the Rice level-crossing density. The core methodological contribution is a smoothed Monte-Carlo estimator (Eq. 2) for the level-set density, derived via the co-area formula (Eq. 4), which is turned into a differentiable training objective (Eq. 5). The estimator is validated against exact crossing counts and the Rice formula. The loss is then evaluated under strict information parity against the Focal Frequency Loss (FFL), normalized Sobolev supervision, and MSE-only training on regular grids and scattered non-uniform samples. The results are reported with unusual honesty: the crossing-density loss matches but does not beat FFL on edge-dominated natural images, while overtaking FFL by 0.6 dB on statistically homogeneous texture. The paper also documents a negative result on dense regular grids, where finite-difference gradient targets actively hurt fidelity.

Significance. The paper makes a solid methodological contribution by instantiating the classical Kac–Rice machinery as a differentiable, grid-free training objective for INRs. The co-area derivation (Eq. 4) is parameter-free and standard, and the same-batch target design (§3.2) that ensures shared sampling noise cancellation under non-uniform sampling is a thoughtful practical contribution. The strict information-parity protocol—where all losses receive exactly the same scattered samples—is commendable and strengthens the comparative claims. The oracle-gradient diagnostic (Table 2), showing that Kac–Rice is uniquely insensitive to gradient-target quality while Sobolev benefits from exact gradients, provides clear mechanistic evidence for the distributional robustness claim. The paper ships reproducible code, raw result JSONs, estimator unit tests, and experiment scripts, which is a significant strength. The honest reporting of negative results (parity rather than dominance on natural images, active harm on dense grids) is a model for the field.

major comments (3)
  1. §5.3, Table 4: The content-dependence claim—that crossing statistics win on statistically homogeneous texture but only tie on edge-dominated natural images—rests on exactly one natural image (camera, Table 2) and one synthetic texture (Table 4). On the texture task, the +0.6 dB advantage of Kac–Rice over FFL (27.19±0.33 vs. 26.58±0.15) is approximately 1.8σ, which is suggestive but not conclusive. The paper acknowledges this in limitation (v), but the content-dependence conclusion is a load-bearing part of the paper's narrative (the 'parity with a content-dependent edge' framing in the abstract and §6). While the core methodological contribution stands independently of this scope issue, the generalizability of the content-dependence claim would be substantially strengthened by at least one additional image per category.
  2. §3.2, same-batch targets: The claim that estimator and target share sampling noise, which 'partially cancels in the difference,' is central to the robustness argument for non-uniform sampling. The paper states that 'the loss compares two estimates of the same functional under the same measure, so even under non-uniform sampling both sides are biased identically toward densely sampled regions.' This is a reasonable argument, but it is not formally proven or empirically isolated. A brief controlled experiment—e.g., comparing same-batch targets against independent-batch targets under non-uniform sampling—would directly validate this mechanism and strengthen the claim.
  3. §5.3, Table 2, oracle-gradient rows: The oracle-gradient diagnostic shows that Kac–Rice's performance slightly degrades with oracle gradients (21.14→20.86 dB on blobs), while Sobolev improves (21.59→21.80 dB). The paper interprets this as evidence for the distributional robustness mechanism. However, the degradation of Kac–Rice with sharper targets is somewhat counterintuitive—the explanation that 'sharper targets, if anything, ask for crossing mass the MSE anchor cannot yet place' is plausible but unverified. A brief ablation varying the noise level on gradient targets (rather than a binary oracle/non-oracle switch) would more directly demonstrate the insensitivity claim.
minor comments (9)
  1. Abstract: The phrase 'recovering +2.3–3.0 dB over MSE-only on PE-MLP and +1.4–1.8 dB on SIREN' could be misread as a claim of superiority over FFL. Consider clarifying that these gains are relative to MSE-only, not to the best auxiliary loss.
  2. Table 1: The FINER result on the 2D camera image (139.72±3.09 dB) is a numerically exact fit with very high variance on the 1D task (73.57±7.68 dB). A footnote explaining the source of this variance (seed sensitivity of the variable-periodic initialization) would help the reader contextualize this result.
  3. Figure 7 (ablations): The y-axis range for the bandwidth plot (19.5–21.0 dB) is narrow enough that the 'wide plateau' claim is visually understated. Consider widening the y-axis or adding a reference line for the MSE-only baseline to contextualize the plateau.
  4. §3.3, Eq. (5): The normalization denominator (c_gt(u_j) + c̄_gt)² is introduced to make the loss scale-free, but the choice of adding c̄_gt (the mean) rather than, say, max(c_gt(u_j), δ) is not motivated. A one-sentence justification would be helpful.
  5. Appendix A, Table 3: The learning rate for SIREN/FINER is listed as 5×10⁻⁴, but the schedule is 'cosine decay to 0.05×lr.' Clarify whether this means 0.05 × 5×10⁻⁴ = 2.5×10⁻⁵ or 0.05 × 10⁻³.
  6. §2 (Related Work): The paper mentions that 'the Kac–Rice machinery appears in machine learning mainly for counting critical points of loss landscapes.' A citation for this claim would strengthen the novelty positioning.
  7. Figure 8: The 1D scattered fitting result shows all methods tying within noise (28.1–28.6 dB), which is consistent with the paper's framing but is reported only in the appendix. A brief mention in the main text (§5.3) that 1D scattered results show no separation would make the narrative more complete.
  8. §5.4: The composability experiment with SIREN is reported only on the blobs task. A note explaining why the other sampling densities (ramp, uniform) were not tested would be appropriate.
  9. References: The citation for the co-area formula [5] (Evans & Gariepy 1992) is correct but could be supplemented with a more accessible reference for the machine-learning audience (e.g., a textbook treatment of the co-area formula in the context of geometric measure theory).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and generous reading. All three major comments are fair and actionable. We agree to revise the manuscript accordingly.

read point-by-point responses
  1. Referee: §5.3, Table 4: The content-dependence claim rests on one natural image and one synthetic texture. The +0.6 dB advantage is ~1.8σ. The claim is load-bearing in the abstract and §6. At least one additional image per category would strengthen generalizability.

    Authors: The referee is correct on all counts. The content-dependence claim is load-bearing in our framing, and one image per category is insufficient to support it. The +0.6 dB gap on texture (27.19±0.33 vs. 26.58±0.15) is approximately 1.8σ—suggestive, not conclusive—and we acknowledged this in limitation (v) but did not calibrate the narrative accordingly. We will address this in two ways. First, we will add at least two additional natural images (e.g., from the Kodak/CLIC set) and two additional synthetic textures with varying degrees of statistical homogeneity, run under the same blobs protocol at N=8,192. Second, we will soften the abstract and §6 language from 'overtakes FFL by 0.6 dB' to a hedged form that accurately reflects the evidence strength (e.g., 'shows a suggestive advantage on statistically homogeneous texture that does not reach significance at n=1'). If the additional runs replicate the effect, we will report the pooled estimate; if they do not, we will say so. The core methodological contribution—estimator derivation, validation, and the strict information-parity protocol—stands independently of this claim, but we agree the narrative must not lean on underpowered evidence. revision: yes

  2. Referee: §3.2, same-batch targets: The claim that estimator and target share sampling noise, which 'partially cancels in the difference,' is central to the robustness argument but is not formally proven or empirically isolated. A controlled experiment comparing same-batch vs. independent-batch targets under non-uniform sampling would validate the mechanism.

    Authors: The referee correctly identifies that our argument for same-batch noise cancellation is informal—we state the intuition but provide neither a formal bound nor an empirical isolation. We will add a controlled ablation on the blobs sampling configuration: same-batch targets (current design) versus independent-batch targets (estimator and target computed on disjoint random subsets of the batch at each iteration), under the same non-uniform sampling pattern. If the mechanism is as described, same-batch targets should outperform independent-batch targets, with the gap widening as sampling becomes more non-uniform (blobs > ramp > uniform). We will also add a brief formal statement: under non-uniform sampling with density p(x), both the estimator and target are biased by the same reweighting factor p(x)/p̄, so the leading-order sampling bias cancels in the difference; the residual is the Monte-Carlo variance of the difference, which is what the same-batch design minimizes. This is a standard variance-reduction argument (common random numbers) and we should have stated it as such rather than leaving it as prose. revision: yes

  3. Referee: §5.3, Table 2, oracle-gradient rows: Kac–Rice degrades with oracle gradients (21.14→20.86), which is counterintuitive. The explanation is plausible but unverified. A noise-level ablation on gradient targets (rather than binary oracle/non-oracle) would more directly demonstrate the insensitivity claim.

    Authors: We agree. The binary oracle/non-oracle switch shows that Kac–Rice is insensitive to the extreme of the gradient-quality axis, but it does not trace the response curve, and the slight degradation with oracle targets is indeed counterintuitive without further evidence. We will add an ablation that interpolates between estimated and oracle gradients by mixing them: target gradient g_target = (1−α)·g_estimated + α·g_oracle for α ∈ {0, 0.25, 0.5, 0.75, 1.0}, on the blobs configuration, for both Kac–Rice and Sobolev. The expected pattern, if our mechanism description is correct, is that Sobolev improves monotonically with α while Kac–Rice is approximately flat (or slightly decreasing, if the 'sharper targets ask for crossing mass the MSE anchor cannot place' explanation holds). We will also add a second ablation that injects controlled Gaussian noise of standard deviation σ_noise into oracle gradients, sweeping σ_noise ∈ {0, 0.1, 0.5, 1.0}×σ_gradient, to directly test robustness to gradient-target noise rather than to gradient-target sharpness. These two ablations together should either confirm or falsify our mechanistic explanation, and we will report the outcome honestly regardless of direction. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; derivation is self-contained against external benchmarks.

full rationale

The paper's derivation chain is clean and parameter-free. The co-area formula (Eq. 4) is a standard measure-theoretic result cited to Evans & Gariepy (1992), an external source. The Monte-Carlo estimator (Eq. 2) follows directly by substituting g = δε(·−u) into the co-area formula—this is a one-line derivation with no hidden assumptions. The Rice formula (Eq. 1) is used only for validation in Section 4, not as a premise for the loss; the paper explicitly states 'No probabilistic assumptions enter: (4) holds for the single deterministic field being fitted.' The loss (Eq. 5) compares the estimator applied to the prediction versus the same estimator applied to ground truth at the same batch points—the paper transparently notes this means shared sampling noise partially cancels, which is a design choice with stated trade-offs, not circularity. There are zero self-citations: the author is an independent researcher with no prior work referenced. All citations are to external, established sources (Rice 1944, Kac 1943, Evans & Gariepy 1992, Adler & Taylor 2007, Azaïs & Wschebor 2009). The claims are honestly scoped—the paper reports parity rather than dominance on natural images and explicitly acknowledges the 'no win' mechanism in Section 6. No step in the derivation reduces to its inputs by construction, no prediction is a renamed fit, and no load-bearing argument rests on unverified self-cited work.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new mathematical entities, particles, forces, or postulated objects. It applies existing mathematical machinery (co-area formula, Kac-Rice) to a new domain (INR training). The three free parameters are standard hyperparameters with ablation studies, not ad hoc constructs. No axioms are invented for the paper; all are standard results from measure theory and random field theory.

free parameters (3)
  • ε (bandwidth) = 0.15σ_y
    Gaussian kernel bandwidth for the smoothed Dirac δ_ε. Ablated in §5.5; wide plateau above 0.15σ_y. Tuned on camera image family.
  • L (number of levels) = 16
    Number of quantile levels at which crossing density is evaluated. Ablated; saturates by 16–32. Tuned on camera image family.
  • β (loss weight) = 0.05
    Weight of the Kac-Rice auxiliary loss in the total objective. Ablated; flat over 20× range [0.01, 0.2]. Tuned on camera image family.
assumptions (4)
  • standard math Co-area formula for Lipschitz functions (Eq. 4)
    Invoked in §3.1 to establish that the integral of g(f(x))‖∇f(x)‖ equals the integral of g(u)H^{d-1}(f^{-1}(u)∩Ω)du. Standard result from Evans & Gariepy (1992).
  • standard math Kac-Rice integrand (Eq. 3)
    Invoked in §3.1 to connect crossing counts to the expectation of |f'(x)| conditioned on f(x)=u. Standard result from Azaïs & Wschebor (2009).
  • domain assumption Spectral bias of coordinate MLPs
    Invoked in §1 to motivate the problem. Established by Rahaman et al. (2019) and Basri et al. (2020).
  • domain assumption Stationarity/Gaussianity for Rice formula (Eq. 1)
    Invoked in §1 and §3.1 for the spectral-moment interpretation. The paper correctly notes this is not required for the deterministic co-area estimator, only for the Rice formula used in validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Level-Crossing Density as a Mesh-Free High-Frequency Auxiliary Loss for Implicit Neural Representations." pith.science (2026). https://pith.science/paper/FV6MNLEI

@misc{pith2026260705815,
  author       = {Pith},
  title        = {Pith review of: Level-Crossing Density as a Mesh-Free High-Frequency Auxiliary Loss for Implicit Neural Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FV6MNLEI}},
  note         = {Machine review of arXiv:2607.05815}
}
abstract

The Minkowski functionals of a field's excursion sets -- area, boundary measure, and Euler characteristic -- describe its level-set morphology; the Euler characteristic is the cheapest handle on topology. We derive smooth Monte-Carlo estimators for all three of a continuous neural field, evaluated at scattered points via the co-area formula and Gauss-Bonnet, using only autodiff: no grid, no complex, no persistence. The estimator is accurate to 1-3% against exact topology in 2D and 3D, and costs about 3 ms per iteration where a persistent-homology (PH) loss on a cubical grid costs 650-1000 ms -- a 250x gap. We establish four design rules without which these losses silently fail: a dense level ladder (invariants are flat in the parameters away from transitions), a $C^2$ backbone (ReLU nets hide curvature in kinks), the full Minkowski vector (Euler characteristic alone is an alternating sum, gamed by debris-hole cancellation; pricing perimeter closes the channel), and sampling-scale coverage. In 2D the vector-valued cap is the only method in a controlled comparison that both repairs topology (3/3 seeds) and preserves fidelity -- uniform smoothing repairs at 11-17x the fidelity cost, and the Euler term alone repairs nothing. In 3D neural-SDF fitting, however, a failure mode we believe general to any sampled soft topology objective appears: gradient descent adversarially hides topological noise below the sampling density, where the estimator is blind -- spurious-feature counts are invariant to 4x more samples, and closing the window needs cubically many points, erasing the cost advantage. A grid-based PH baseline, whose complex is the evaluation resolution, solves the same benchmark ($4/9$ exact; median $b_1$ error 1 vs. ours above $10^4$). The 250x cost of persistence is, at present, the price of having no null space. We release estimators, receipts, and benchmarks.

Figures

Figures reproduced from arXiv: 2607.05815 by the authors.

Figure 1
Figure 1. Estimator validation. Smoothed Monte-Carlo estimator cˆε(u) (red, N=4 × 105 points, ε=0.05), exact crossing counts from dense-grid sign changes (dots), and the Rice formula at empirical spectral moments (dashed), for an approximately Gaussian random field (60 random-phase sinusoids). The estimator tracks the exact counts across all levels, including the non-Gaussian wiggles that the Rice formula idealizes away. 1.00… view at source ↗
Figure 2
Figure 2. Left: a multisine signal, one level u, and its crossings; the crossing density at level u is the number of such intersections per unit length. Right: crossing-density profiles c(u) of the full signal versus its low-pass approximation—the profile collapses when high frequencies are missing, which is the signal the loss exploits. Tasks. Regular grids (sanity check): a five-tone multisine in 1D (frequencies 2, 5, 11, 2… view at source ↗
Figure 3
Figure 3. Regular-grid convergence. PSNR versus iteration (mean over 3 seeds; shaded ± s.d.). Left: 1D multisine—all PE-MLP losses converge to parity. Right: 1282 image—with dense supervision, MSE-only is best and the finite-difference gradient targets of Sobolev/Kac–Rice actively cost fidelity (see text). 5.2 Regular grids: parity, as expected 1D: parity. On the grid-sampled multisine, all four PE-MLP losses converge to stat… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Scattered-domain convergence (blobs sampling; mean over 3 seeds, shaded ± s.d.). Full-image PSNR (left) and high-frequency PSNR (right). MSE-only stalls immediately; all three auxiliary losses climb 2.3–2.7 dB above it. Kac–Rice leads on HF-PSNR early (its target does …
Figure 5
Figure 5. Figure 5: Per-band relative spectral error (log scale; lower is better; mean ± s.d. over 3 seeds). Left, regular grid: MSE-only is lowest in every band, mirroring [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results on scattered blobs sampling (seed 0): sample locations, reconstructions (top), high-pass residuals (bottom). MSE-only produces streaked, washed-out structure; every auxiliary loss restores coherent detail. Differences among the three auxiliary losse…
Figure 7
Figure 7. Figure 7: Ablations on camera/blobs (1,000 iterations, 2 seeds; error bars ± s.d.). Left to right: Dirac bandwidth scale ε/σy (wide plateau above 0.15), number of levels L (saturates by 16–32), loss weight β (flat over a 20× range, collapsing only at β=1), and sample budget N (o…
Figure 8
Figure 8. Figure 8: 1D scattered fitting (ramp density, N=384 samples, mean over 3 seeds). Left: reconstructions in the sparsely sampled region; all methods undershoot the highest tone where samples are scarce, with FFL producing the smoothest (most conservative) fit. Right: final PSNR—al…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

  1. [1]

    Adler and Jonathan E

    Robert J. Adler and Jonathan E. Taylor.Random Fields and Geometry. Springer, 2007

  2. [2]

    John Wiley & Sons, 2009

    Jean-Marc Aza¨ ıs and Mario Wschebor.Level Sets and Extrema of Random Processes and Fields. John Wiley & Sons, 2009

  3. [3]

    Frequency bias in neural networks for input of non-uniform density

    Ronen Basri, Meirav Galun, Amnon Geifman, David Jacobs, Yoni Kasten, and Shira Kritchman. Frequency bias in neural networks for input of non-uniform density. InInternational Conference on Machine Learning (ICML), 2020

  4. [4]

    Czarnecki, Simon Osindero, Max Jaderberg, Grzegorz ´Swirszcz, and Razvan Pascanu

    Wojciech M. Czarnecki, Simon Osindero, Max Jaderberg, Grzegorz ´Swirszcz, and Razvan Pascanu. Sobolev training for neural networks. InAdvances in Neural Information Processing Systems (NeurIPS), 2017

  5. [5]

    Evans and Ronald F

    Lawrence C. Evans and Ronald F. Gariepy.Measure Theory and Fine Properties of Functions. CRC Press, 1992

  6. [6]

    Zico Kolter

    Rizal Fathony, Anit Kumar Sahu, Devin Willmott, and J. Zico Kolter. Multiplicative filter networks. In International Conference on Learning Representations (ICLR), 2021

  7. [7]

    Implicit geometric regularization for learning shapes

    Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. InInternational Conference on Machine Learning (ICML), 2020

  8. [8]

    SAPE: Spatially- adaptive progressive encoding for neural optimization

    Amir Hertz, Or Perel, Raja Giryes, Olga Sorkine-Hornung, and Daniel Cohen-Or. SAPE: Spatially- adaptive progressive encoding for neural optimization. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

Show all 31 references
  1. [9]

    Focal frequency loss for image reconstruction and synthesis

    Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. InIEEE/CVF International Conference on Computer Vision (ICCV), 2021

  2. [10]

    On the average number of real roots of a random algebraic equation.Bulletin of the American Mathematical Society, 49(4):314–320, 1943

    Mark Kac. On the average number of real roots of a random algebraic equation.Bulletin of the American Mathematical Society, 49(4):314–320, 1943

  3. [11]

    Spectral analysis and discrimination by zero-crossings.Proceedings of the IEEE, 74 (11):1477–1493, 1986

    Benjamin Kedem. Spectral analysis and discrimination by zero-crossings.Proceedings of the IEEE, 74 (11):1477–1493, 1986

  4. [12]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInternational Conference on Learning Representations (ICLR), 2015

  5. [13]

    Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein

    David B. Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. BACON: Band-limited coordinate networks for multiscale scene representation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

  6. [14]

    FINER: Flexible spectral-bias tuning in implicit neural representation by variable-periodic activation functions

    Zhen Liu, Hao Zhu, Qi Zhang, Jingde Fu, Weibing Deng, Zhan Ma, Yanwen Guo, and Xun Cao. FINER: Flexible spectral-bias tuning in implicit neural representation by variable-periodic activation functions. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

  7. [15]

    Longuet-Higgins

    Michael S. Longuet-Higgins. The statistical analysis of a random, moving surface.Philosophical Transactions of the Royal Society of London. Series A, 249(966):321–387, 1957

  8. [16]

    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. InEuropean Conference on Computer Vision (ECCV), 2020

  9. [17]

    Instant neural graphics primitives with a multiresolution hash encoding.ACM Transactions on Graphics, 41(4):102:1–102:15, 2022

    Thomas M¨ uller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding.ACM Transactions on Graphics, 41(4):102:1–102:15, 2022. 11

  10. [18]

    DeepSDF: Learning continuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

  11. [19]

    PyTorch: An imperative style, high- performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high- performance deep learning library. InAdvances in Neural Information Processing Systems (...

  12. [20]

    On the spectral bias of neural networks

    Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. InInternational Conference on Machine Learning (ICML), 2019

  13. [21]

    Beyond periodicity: Towards a unifying framework for activations in coordinate-MLPs

    Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-MLPs. InEuropean Conference on Computer Vision (ECCV), 2022

  14. [22]

    Stephen O. Rice. Mathematical analysis of random noise.Bell System Technical Journal, 23(3):282–332,

  15. [23]

    and 24(1):46–156, 1945

  16. [24]

    Baraniuk

    Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G. Baraniuk. WIRE: Wavelet implicit neural representations. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

  17. [25]

    Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

  18. [26]

    Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T

    Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. InAdvances in Neural Informat...

  19. [27]

    Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, et al

    Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, et al. SciPy 1.0: Fundamental algorithms for scientific computing in python.Nature Methods, 17:261–272, 2020

  20. [28]

    Bovik, Hamid R

    Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Processing, 13(4):600–612, 2004

  21. [29]

    Neural fields in visual computing and beyond

    Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 41(2):641–676, 2022

  22. [30]

    FreeNeRF: Improving few-shot neural rendering with free frequency regularization

    Jiawei Yang, Marco Pavone, and Yue Wang. FreeNeRF: Improving few-shot neural rendering with free frequency regularization. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

  23. [31]

    Sobolev training for implicit neural representations with approximated image derivatives

    Wentao Yuan, Qingtian Zhu, Xiangyue Liu, Yikang Ding, Haotian Zhang, and Chi Zhang. Sobolev training for implicit neural representations with approximated image derivatives. InEuropean Conference on Computer Vision (ECCV), 2022. A Hyperparameters B Reproducibility All code, ra...

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.