acceptodds
Withdrawn from ICLR 2027

Smooth-NGP: -smooth Differentiable Hash Grids

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Grid-based implicit methods like Instant-NGP greatly speed up neural field training, but their default multilinear interpolation yields only piecewise-smooth functions, which can destabilize higher-order derivatives and create grid-aligned artifacts. These issues limit performance in derivative-sensitive tasks, such as precise normal computation and PDE solving. We present *Smooth-NGP*, a -smooth variant of Instant-NGP that swaps multilinear interpolation for an infinitely differentiable kernel and extends the hash grid with a sliding-window design. The smooth kernel provides stable closed-form derivatives of any order, supporting derivative-driven losses while retaining the efficiency of grid encodings. However, because the kernel is flat at cell boundaries, its derivatives go to zero there, reducing gradient flow exactly where grid artifacts tend to appear. The sliding-window mechanism resolves this by ensuring each query point falls within the interior of several shifted grid windows, preserving useful gradients across the entire domain. On image fitting, signed distance field learning, and PINN-based PDE solving, Smooth-NGP reliably improves reconstruction quality, derivative fidelity, training stability, and surpasses leading grid-based baselines. We will release the source code and experimental scripts upon publication.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
$C^\infty$-smooth Hash grid, Image fitting, SDF, PDE

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qrVJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper targets the C0 limitation of Instant-NGP's multilinear hash-grid interpolation. First derivatives jump across cell boundaries, and pure second derivatives of the encoder vanish inside cells. This hurts derivative-driven objectives such as SDF normals and PINN residuals.

Smooth-NGP makes two changes: (1) It replaces the linear weights with a classical compactly supported C infinity partition-of-unity cutoff. All of its derivatives vanish at the vertices. (2) To counter the resulting vanishing gradients at cell boundaries, it uses N=4 diagonally shifted "sliding-window" grids. Each has its own feature tables and a shared decoder, and the four predictions are averaged.

Critical strengths
  1. The approach looks simple and effective. The analysis of why multilinear encoders hurt second-order objectives is correct and useful. Within a cell the encoder Hessian has zero diagonal, so the Laplacian of the encoder vanishes and any pure second derivative must come from the decoder Hessian.

  2. An efficient CUDA implementation that is much faster than the closest smooth-grid competitor. Building on tiny-cuda-nn with closed-form higher-order derivatives gives ~4 times lower runtime than ∞-Grid on SDFs at slightly better quality, and ~24 times lower on images.

Critical weaknesses
  1. The novelty is limited. Both components are classical: the cutoff function is adopted as-is, and averaging over shifted grids resembles known multi-offset grid ensembles. The contribution is thus mainly an engineering combination, so its value rests on careful comparison with simpler smooth alternatives (see W2).

  2. The most direct smooth-interpolation baselines are missing. tiny-cuda-nn already offers smoothstep interpolation, which is C1 and, like the proposed kernel, flat at the vertices; a smoothstep + sliding-window variant would show whether C infinity actually matters or whether C1/C2 is enough. Cubic B-spline interpolation is the standard way to obtain smooth (C2) derivatives, and because its derivatives do not vanish at the knots, it avoids the boundary-flat problem without needing multiple windows.

Questions for authors

Q1. How does Smooth-NGP compare with (a) tiny-cuda-nn's smoothstep interpolation, with and without your sliding windows, and (b) cubic B-spline grid interpolation, on the PINN and SDF tasks at matched parameters and runtime?

Q2. How does Smooth-NGP relate to prior work on smooth hash encodings for PDEs? Huang & Alkhalifah (2023, "Efficient PINN using hash encoding") study derivative computation for hash-encoded PINNs, and Hermite-NGP (He et al., 2026) stores derivatives at grid vertices, so the encoding is not flat there. A short discussion would help clarify the contribution. Given the timing, a direct comparison is not required, though it would be welcome if available.

3 comments

Thank you for the careful distinction between the multilinear encoder and the decoder contribution to pure second derivatives, and for recognizing the CUDA implementation. We agree that comparisons with simpler smooth interpolants are important for evaluating practical value.

Q1 Novelty and the role of shifted windows

Q. Both the cutoff and averaging over shifted grids are classical. Is the contribution mainly an engineering combination? [Weakness 1; review p. 5]

A. The representation contribution is their integration used for multiresolution hash encoding with a constructive guarantee of global C∞C^\infty continuity. Local smoothness of a kernel does not by itself ensure that a piecewise assembled grid field has this property. Our interpolation joins adjacent cells smoothly for arbitrary feature values, and the smooth decoder and prediction averaging preserve that continuity.

The role of shifted windows is more specific: they mitigate the endpoint derivative restrictions of the selected interpolation, rather than create infinite differentiability. Their empirical value is evaluated by the ablations. We agree that the structural guarantee should be accompanied by comparisons that establish the accuracy and cost of the complete design.

Q2 Smoothstep with and without shifted windows

Q. How does Smooth-NGP compare with tiny-cuda-nn smoothstep interpolation, with and without sliding windows, on PINNs and SDFs at matched parameters and runtime? [Question 1a; review p. 5]

A. We agree that this comparison would be valuable. We will add smoothstep baselines with and without shifted windows on the PINN and SDF tasks in a revised manuscript, with comparable parameter and training budgets, and report accuracy and runtime.

Q3 Cubic B spline comparison

Q. How does Smooth-NGP compare with cubic B-spline interpolation, which supplies C2C^2 continuity without the same boundary-flatness issue? [Weakness 2 and Question 1b; review p. 5]

A. We agree that cubic B-splines are a worthwhile method to compare with, and will add this comparison in a revised manuscript. Their overlapping bases offer C2C^2 continuity without forcing all derivatives to vanish at grid knots. However, a standard tensor-product cubic B-spline query accesses 4d4^d control coefficients per level, compared with 2d2^d corner features per window in our interpolation. This increases feature retrieval and aggregation work per interpolation query as dimension grows. Our four-window model also incurs 4×2d4\times2^d corner contributions and multiple decoder evaluations, so the overall runtime tradeoff must be measured under matched settings.

Q4 Relation to earlier derivative aware hash encodings

Q. How does Smooth-NGP relate to Huang and Alkhalifah 2023 and Hermite-NGP 2026? A short discussion would help, even without a direct comparison. [Question 2; review p. 5]

A. We will add these works and clarify the distinction. Our goal is a globally smooth NGP representation supporting arbitrary-order analytic derivatives; PDE solving is one application.

  • Huang and Alkhalifah, Efficient physics-informed neural networks using hash encoding (2023 preprint; JCP 2024): replaces automatic differentiation with finite differences to handle the derivative discontinuities of standard hash encoding, rather than resolving them through a globally smooth representation.
  • He et al., Hermite-NGP, ICML 2026: introduces derivative parameterization, learning mixed-derivative coefficients alongside value coefficients. This bypasses the limitations of the original value-only encoding by adding derivative degrees of freedom. The resulting Hermite field still supports analytic differentiation, with global C1C^1 continuity and cellwise analytic second derivatives.

Smooth-NGP instead modifies the interpolation and architecture so that arbitrary-order analytic derivatives follow directly from a globally C∞C^\infty field, without finite differences or additional learned vertex-derivative coefficients.

Thank you for the detailed response. I appreciate the clarifications and will maintain my original score (weak acceptance).

Your response to Q1 explains why the method works, but my concern about novelty remains. I still see it mainly as an engineering combination, and more comparisons are needed to demonstrate its practical value. For Q2/Q3, the actual comparison results are still missing. The discussion of prior work is helpful and addresses my concern.

Thank you for your follow-up feedback and for acknowledging our clarification of prior work. To address your requests for additional comparisons, we provide the following experimental results.

Additional PINN comparisons. To address the requests in Q2/Q3, we report the following comparisons on the six PINNacle cases. Entries are relative l2l_2 errors (lower is better), reported to three significant digits; the lowest error in each column is bold. INGP-Smoothstep uses smoothstep interpolation without shifted windows; Ours-Smoothstep retains our shifted-window architecture but replaces the C∞C^\infty interpolation kernel with smoothstep.

MethodBurg-1DPois-2DCPois-B2DNS-LDNS-BSWave-1D
INGP-Smoothstep3.33e-11.54e-17.70e-19.95e-13.51e-19.87e-1
B-spline1.17e-11.25e-12.62e-16.39e-13.44e-18.56e-1
Ours-Smoothstep1.46e-13.42e-32.23e-33.65e-12.00e-12.64e-1
Ours9.53e-21.06e-31.37e-31.93e-11.31e-16.10e-2

INGP-Smoothstep exhibits the same type of boundary derivative restriction as Ours w/o SW: the interpolation weights have zero first derivatives at cell endpoints, forcing the encoded field's derivative normal to a cell face to vanish there in the absence of shifted windows. This restriction can hinder convergence, consistent with its relatively high errors on several cases. Adding shifted windows reduces the smoothstep baseline's errors on all six cases.

B-spline shows a broadly similar pattern to ∂∞\partial^\infty-Grid: both attain relatively low errors on several cases, but their errors remain higher than those of Smooth-NGP across all six cases.

Ours-Smoothstep has higher relative errors than Ours across all six cases, despite retaining the shifted-window architecture. We attribute this difference to its limited higher-order smoothness: the smoothstep node basis is C2C^2 on the interior (−1,1)(-1,1) of its support, but not C3C^3, whereas our interpolation is C∞C^\infty. These results support the benefit of higher-order smoothness for fitting physical fields.

B-spline comparison on image fitting. We additionally report the image-fitting comparison with B-spline. All metrics are medians over the 100 DIV2K validation images.

MethodPSNR ↑SSIM ↑Training time (min) ↓
B-spline30.190.9336.65
Ours33.040.963.58

Smooth-NGP achieves higher PSNR (33.0433.04 vs. 30.1930.19) and SSIM (0.960.96 vs. 0.930.93) than B-spline, while reducing training time from 36.6536.65 to 3.583.58 minutes. This comparison demonstrates benefits in both image-fitting quality and training efficiency.

B-spline comparison on SDF reconstruction. The following results are medians over the 72 evaluated objects. CD and NE are scaled by 10−310^{-3} and 10−210^{-2}, respectively; runtime is in minutes.

MethodCD ↓NE ↓Total runtime (min) ↓
B-spline2.670.87276.4
Ours2.670.9035.6

Smooth-NGP matches B-spline in CD (2.672.67) and achieves a comparable NE (0.900.90 vs. 0.870.87), providing similar reconstruction quality while reducing total runtime from 276.4276.4 to 35.635.6 minutes. This substantial speed advantage mainly stems from our CUDA implementation and the smaller number of grid-feature parameters requiring differentiation per local interpolation in the NGP architecture compared with cubic B-spline interpolation.

Reviewer_f5kCOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Smooth-NGP, a C∞C^\infty-smooth variant of Instant-NGP (Instant Neural Graphics Primitives). Instant-NGP enables efficient neural field training by combining multiresolution hash grids with a small neural network. At each resolution, learnable feature vectors are stored at grid vertices and interpolated at arbitrary query locations using multilinear interpolation. However, this produces representations that are continuous but not continuously differentiable across grid-cell boundaries, limiting their applicability to tasks requiring accurate higher-order derivatives.

To address this limitation, Smooth-NGP replaces multilinear interpolation with compactly supported C^\{infty}-smooth kernels whose derivatives vanish at cell boundaries. To mitigate the resulting vanishing-gradient problem, the authors introduce multiple shifted hash grids with independent feature tables, ensuring that query points remain in the interior of at least some grids. The method is evaluated on image reconstruction, neural signed distance field reconstruction, and physics-informed neural network (PINN) approaches to PDE solving, demonstrating improved reconstruction quality and accuracy over existing hash-grid-based methods.

Critical strengths
  1. The paper identifies provides a relatively simple and mathematically motivated solution to a fundamental limitation of Instant-NGP's multilinear interpolation, namely the discontinuity of its derivatives across grid-cell boundaries, which can hinder applications requiring accurate higher-order derivatives, through the combination of C∞C^{\infty}-smooth interpolation and shifted hash grids provides a relatively simple and mathematically motivated solution. Moreover, ablation studies demonstrate that both components are necessary to achieve the reported improvements. For instance, on the 2D Poisson equation, the relative error increases from 0.00106 to 0.429 when removing the shifted grids, and to 0.329 when replacing the smooth interpolation with linear interpolation.

  2. Smooth-NGP consistently outperforms the considered hash-grid baselines on image reconstruction, neural signed distance field reconstruction, and all six PDE benchmarks. The improvements are particularly substantial for PDEs involving second-order derivatives such as the 2D Poisson problem,and the 2D Poisson–Boltzmann problem. These results provide compelling evidence of the practical benefits of smooth interpolation for derivative-sensitive applications.

Critical weaknesses
  1. Limited novelty: The proposed method primarily combines existing techniques rather than introducing a fundamentally new representation or learning framework. C∞C^\infty-smooth interpolation kernel is directly adopted from existing numerical analysis literature, while infinitely differentiable neural grids have already been explored in prior work. The main contribution therefore lies in combining this interpolation scheme with multiple shifted hash grids to mitigate vanishing derivatives at cell boundaries. While the resulting architecture appears effective, the methodological advancement remains relatively incremental.

  2. While the authors compare Smooth-NGP against Instant-NGP and ∂∞\partial^{\infty}-Grid, the PDE experiments lack comparisons with other competitive physics-informed neural architectures, such as a Fourier-feature-based PINNs. This makes it difficult to assess the performance of Smooth-NGP relative to the broader state of the art in neural PDE solving. ). Moreover, the PDE experiments are restricted to six relatively simple 1D and 2D problems, without demonstrating applicability to more challenging high-dimensional PDEs or realistic engineering simulations.

Questions for authors
  1. Since the PDEs considered in the experiments involve derivatives of at most second order, could the authors compare their C∞C^\infty interpolation against simpler C2C^2- or C4C^4-continuous alternatives, such as B-splines? Would these alternatives achieve comparable accuracy/runtime without requiring the proposed sliding-window mechanism?

  2. Does Smooth-NGP improve the representation of physical fields even without physics-informed losses, or with a combination of physics informed losses and supervised learning from PDE solution data ?

  3. The paper would be strengthened by demonstrating that Smooth-NGP also improves the accuracy of spatial derivatives of learned PDE solutions. In many physical applications, the quantity of interest is derived from the solution field, such as the electric field computed from an electrostatic potential or the vorticity computed from a velocity field. Could the authors include an experiment demonstrating that Smooth-NGP provides more accurate estimates of such derived physical quantities compared to existing methods?

  4. Could the authors extend their PDE experiments to include comparisons with other state-of-the-art physics-informed neural architectures, such as a Fourier-feature-based PINN? Such comparisons would help establish whether Smooth-NGP offers meaningful improvements over the broader state of the art, beyond the considered hash-grid-based methods.

2 comments

Thank you for the assessment of the mathematical motivation and the evidence that both interpolation and shifted grids contribute to the reported results. We address the scope of the contribution and the requested comparisons separately.

Q1 Novelty of combining known components

Q. The kernel and shifted grids are existing ideas, so the methodological advancement appears incremental. [Weakness 1; review p. 4]

A. The cutoff is classical; our contribution is its integration into a multiresolution hash-grid architecture with global C∞C^\infty continuity for arbitrary learned feature values. A smooth kernel alone is insufficient: the interpolated features must also match smoothly across cell interfaces. Our corner-interpolation construction ensures this matching while retaining local corner queries. Its endpoint flatness, however, imposes derivative restrictions at grid boundaries. Shifted windows address this restriction by representing those locations in the interiors of other grids. The two components thus play complementary roles: smooth interpolation establishes global regularity, while shifted windows mitigate boundary-induced derivative degeneration.

Q2 Research objective and experimental scope

Q. Why use C∞C^\infty interpolation for second-order PDEs? What about data-only or combined supervision, direct evaluation of electric fields or vorticity, Fourier-feature PINN baselines, and higher-dimensional or engineering PDE benchmarks? [Review p. 4]

A. Our objective is a globally C∞C^\infty NGP neural architecture with continuous analytic derivatives of arbitrary order, rather than an optimized PDE solver (p. 2, lines 78–101). We use PDE experiments to test Smooth-NGP's higher-order differentiation capability and smoothness. PDE solving serves as an application for evaluating this architecture, rather than the motivation for proposing it.

Q3 Comparison with cubic B-splines

Q. Could B-splines achieve comparable accuracy and runtime without the sliding-window mechanism? [Question 1; review p. 4]

A. We agree that cubic B-splines are a worthwhile method to compare with, and will add this comparison in a revised manuscript. Their overlapping bases offer C2C^2 continuity without forcing all derivatives to vanish at grid knots. However, a standard tensor-product cubic B-spline query accesses 4d4^d control coefficients per level, compared with 2d2^d corner features per window in our interpolation. This increases feature retrieval and aggregation work per interpolation query as dimension grows. Our four-window model also incurs 4×2d4\times2^d corner contributions and multiple decoder evaluations, so the overall runtime tradeoff must be measured under matched settings.

Dear Reviewer f5kC (,

Thank you for your review of this submission. The authors have now provided a response that addresses the concerns raised in some detail, and I would appreciate hearing your thoughts on how well it addresses your points.

I realize the CVPR deadline and the ICLR discussion period deadlines have collided this year, so we are only a few days from the deadline. Please participate whenever you can, any time before the deadline works, and even a brief note on whether your concerns have been resolved would be very helpful.

Best regards, Your AC

Reviewer_nYrLOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents Smooth-NGP, a C∞-smooth variant of Instant-NGP that swaps multilinear interpolation for an infinitely differentiable kernel and extends the hash grid with a sliding-window design. On image fitting, signed distance field learning, and PINN-based PDE solving, SmoothNGP reliably improves reconstruction quality, derivative fidelity compared to leading grid-based baselines (DnfGrid and Instant-NGP).

Critical strengths
  • The paper addresses an important problem of scalability and accurate higher order derivatives needed for scientific signals. The derivative is not required for the representing visual signals, thus has been generally ignored in most of the prior literature. Thus, the problem is timely.

  • Smooth (C^inf) interpolation and sliding windows are both sound ideas to fix the derivative issue in the previous hybrid methods.

Critical weaknesses

Presentation

  • Presentation is quite bad. The paper is heavily written with AI with no explanation or intuition on why certain things should hold. Unless the paper is significantly rewritten, it should not be published in a top conference. See examples below
  • The paper claims at many places “Within each cell, the multilinear encoder Hessian has a zero diagonal, whereas its off-diagonal entries are not forced to vanish” Why should they vanish? If you are modelling f (x, y), a 2D to 1D signal then even mixed higher order derivatives should exist.
  • L046: this argument does not hold only for 3D, rather 3D is example. Does your method work only for 3D? If so, it should be clearly mentioned. On the other hand, experiments include 2D and PDE, why is the method description in 3D? L049: Does this PoU only work for trilinear?
  • L019: Why is the kernel chosen to be flat at cell boundaries? Need motivation
  • L170-179: What is the difference to these works? What are you solving that hasn’t been solved before?
  • L257-258: These sentences don’t make any sense, why complain about standard interpolation? L080: also, what does standard interpolation constraint even mean?
  • Figure 2 is very low quality, notations does not match with the method section (eg, r, W).
  • L038: argument is imprecise/incomplete; Instant-NGP also needs to be evaluate many times during both training and inference

Minor remarks

  • L104-106; Why are neural field methods not cited?
  • Typo in Equation 3, should be zero on left and same with Equation 4, or at least I don’t understand.
  • Citation format is incorrect everywhere; cited paper should be in parenthesis unless it directly part of the sentence.
  • L050: If you are saying later variants, they should be cited.

Originality/Significance/Soundness

  • The proposed paper doesn’t properly position itself in the context of some of the prior literature. There have been hybrid neural fields (He et al. 2026, Chetan et al. 2025, Kairanda et al 2026, Wang et al. 2024) support higher order derivatives, and most of these are directly built on Instant-NGP. None of these are cited or compared to (except Kairanda). Thus, the proposed paper is misleading and suggests that hybrid smooth NGP hasn’t been attempted before.
  • Moreover, there is a smooth cubic interpolation version of Instant-NGP for SDF in the original paper. Its not mentioned anywhere.
  • How is your method different from and why is your result better than [Kairanda et al. 2026? This needs to clearly written.
  • I don’t see the relevance of experiments in Section 4.1 and Section 4.2. These are results of overfitting to signals, where the proposed method does not have any advanatage over DInf-Grid or Instant-NGP. Moreover, Figure 3, Table 1 is not believable – for InstantNGP, there is no reasons why image reconstruction is worse than Smooth-NGP, the advantage should be only with smother (otherwise you can attribute to some hyperparameter), For DInf-Grid, the table metnions 84 mins for image fitting, whereas the original papers reports 25 seconds for gradient reconstruction or 15 mins for Laplacian (which are much harder version of similar task).

References:

[He et al. 2026] Hermite-NGP: Gradient-Augmented Hash Encoding for Learning PDEs, ICML 2026

[Chetan et al., CVPR 2025] Accurate Differential Operators for Hybrid Neural Fields

[Wang et al., AAAI 2024] Neural Physical Simulation with Multi-Resolution Hash Grid Encoding

Questions for authors

Please see weaknesses

6 comments

Q1 Pure and mixed second derivatives

Q. Why should the off-diagonal Hessian entries vanish? For a 2D scalar signal, mixed higher-order derivatives should exist. [Presentation; review p. 2]

A. The manuscript explicitly distinguishes the encoder Hessian diagonal from mixed derivatives:

  • Section 3.1, p. 4, lines 211–213: the diagonal entries vanish within cells, whereas off-diagonal mixed derivatives “are not structurally constrained to vanish.”
  • Lines 214–215: “the Hessian is not identically zero but is structurally constrained.”
  • Section 3.4, p. 6, Eq. (7), lines 297–299: pure second derivatives of the decoded field can arise from the decoder-Hessian term.

For example, h(x,y)=a+bx+cy+dxyh(x,y)=a+bx+cy+dxy gives hxx=hyy=0h_{xx}=h_{yy}=0 but hxy=dh_{xy}=d. Our claim concerns the encoder's missing pure second derivatives, not vanishing mixed derivatives or a zero Hessian of the complete network.

Q2 Dimensionality and partition of unity

Q. Does the method work only in 3D? Does the partition-of-unity construction only apply to trilinear interpolation? [L046 and L049; review p. 2]

A. Partition of Unity (PoU) is a standard property of multilinear interpolation, not a new construction proposed by us. It preserves constant corner values and applies in any dimension:

=\prod_{j=1}^d[\phi_0(t_j)+\phi_1(t_j)]=1.$$ The manuscript defines a general $d$-dimensional domain (**p. 4, lines 185–187**) and tensor-product corner weights (**lines 195–201, Eq. (1)**). **P. 5, lines 221–223, Eq. (3)** state the inherited vertex-interpolation and PoU requirements; **line 235** retains the tensor-product structure. Three dimensions is therefore an example, not a restriction. **Smoothness is required by the intended higher-order differentiation, rather than by PoU or multilinear interpolation itself.** We change the interpolation to meet that derivative requirement while preserving the standard interpolation properties. ### Q3 Why the cutoff is flat at its endpoints **Q.** Why is the kernel chosen to be flat at cell boundaries? [L019 and L257–258; review p. 2] **A.** This is a mathematical necessity. The manuscript already states this property: **p. 5, lines 218–223, Eq. (3)** requires global $C^\infty$ smoothness and constant values outside $[0,1]$; **Section 3.3, lines 258–259** explicitly states that all positive-order endpoint derivatives vanish. Since the exterior is constant, all its derivatives are zero, and continuity of every derivative requires the interior derivatives to match these zeros at the endpoints. In our local two-endpoint interpolation, flatness also follows from combining **complementary PoU weights and global smoothness**. For the two cells adjacent to node $i$ on a unit-spaced grid, matching any derivative order $k\geq1$ requires $$\tilde\psi^{(k)}(1)(g_{i-1}-g_i) =\tilde\psi^{(k)}(0)(g_i-g_{i+1}).$$ For arbitrary independent features, this forces $\tilde\psi^{(k)}(0)=\tilde\psi^{(k)}(1)=0$. Shifted windows mitigate this restriction through different local coordinates. The argument depends on two-endpoint locality; PoU and smoothness alone do not force flatness for wider-support bases.

Q4 Interpolation conditions and Equations 3 and 4

Q. What do the standard interpolation constraints mean? Should the left side of Equation 3 be zero, and is there a similar typo in Equation 4? [L080 and minor remarks; review p. 2]

A. Eq. (3) states the vertex-interpolation and PoU conditions; Eq. (4) gives a classical function satisfying them with global C∞C^\infty smoothness (p. 5, lines 221–233). The interpolation within a single cell uses the local coordinate t∈[0,1]t\in[0,1]; Eq. (4) additionally specifies its smooth constant extension outside this interval.

Figure 1 joins the same grid node's contributions in its two adjacent cells to display them on [−1,1][-1,1]. Place that node at plotted coordinate s=0s=0. In the preceding cell it is the right corner, with s=t−1s=t-1; in the following cell it is the left corner, with s=ts=t. Its contributions are consequently

\text{Preceding cell:}\quad &\tilde\psi(-s)=\tilde\psi(1-t)=1-\tilde\psi(t),\quad s=t-1,\\ \text{Following cell:}\quad &\tilde\psi(s)=\tilde\psi(t),\quad s=t. \end{aligned}$$ Thus the plotted $[-1,1]$ interval is a concatenation of two single-cell contributions, each evaluated on $[0,1]$; the figure and formulas describe the same interpolation. ### Q5 Scope of the representation contribution **Q.** What does Smooth-NGP solve beyond earlier work on hybrid neural fields and higher-order derivatives? [L170–179 and originality; review p. 2] **A.** Our objective is to extend **Instant-NGP as a general neural representation with globally continuous analytic derivatives of arbitrary order**, rather than develop a particular PDE solver (**p. 2, lines 78–93; p. 4, lines 185–191**). Smooth corner interpolation and a smooth decoder establish global $C^\infty$ regularity; shifted windows mitigate endpoint derivative restrictions. Images, SDFs, and PDEs evaluate this representation in different settings. The cutoff is classical and explicitly attributed (**p. 5, lines 225–227**). Our contribution concerns its integration into the representation. As detailed in Q6, prior work includes numerical derivatives, learned derivative coefficients, and derivative correction; our distinction is the global arbitrary-order guarantee without separately learned vertex-derivative coefficients. ### Q6 Additional related work **Q.** The paper omits He et al. 2026, Chetan et al. 2025, and Wang et al. 2024, and does not adequately position itself relative to these works. [Originality; review pp. 2–3] **A.** Our primary motivation is to develop **an Instant-NGP architecture with global smoothness and continuous derivatives of arbitrary order**. PDE solving is one application used to evaluate this representation, alongside image fitting and SDF reconstruction; optimizing a particular PDE solver is not our primary objective (**p. 2, lines 78–101**). The relevant distinction is how earlier methods address the higher-order derivative limitation of standard hash grids: - **He et al., Hermite-NGP, ICML 2026:** addresses the limitation through **derivative parameterization**, learning additional mixed-derivative coefficients alongside value coefficients. It therefore bypasses the limitation of obtaining higher derivatives from the original value-only encoding by introducing derivative degrees of freedom. The resulting Hermite field still supports analytic differentiation, with global $C^1$ continuity and cellwise analytic second derivatives. - **Wang et al., Neural Physical Simulation with Multi-Resolution Hash Grid Encoding, AAAI 2024:** circumvents direct analytic higher-order differentiation through numerical finite differences and boundary-aware stencils (p. 5413). - **Chetan et al., Accurate Differential Operators for Hybrid Neural Fields, CVPR 2025:** estimates derivatives through local polynomial fitting and uses self-supervised fine-tuning to improve the field's autodifferentiated derivatives. Smooth-NGP changes the representation itself so that **arbitrary-order analytic derivatives follow directly from a globally $C^\infty$ field**, without additional learned vertex-derivative coefficients or numerical differencing. We will add these references and make this distinction explicit.

Q7 Difference from partial infinity Grid

Q. How is Smooth-NGP different from Kairanda et al. 2026, and why are its results better? [Originality; review p. 2]

A. The distinction is cell-corner interpolation versus distance-based RBF interpolation. Instant-NGP combines the 2d2^d corners of the containing cell through tensor-product linear weights. Smooth-NGP retains that organization but substitutes smooth interpolation functions, producing exact compact support and matching derivatives across cell interfaces (our p. 5, lines 235–254). A smooth decoder preserves global C∞C^\infty regularity (p. 6, lines 288–290).

∂∞\partial^\infty-Grid instead aggregates neighboring features using normalized Gaussian functions of their distances to the query (its Section 3.2, pp. 5–6, Eqs. (6)–(7)). Because Gaussians have infinite support, its efficient implementation restricts aggregation to a finite neighborhood (p. 6, “Efficient RBF Interpolation”).

This hard truncation cuts off small but nonzero values, which can cause jumps in field values and mismatches in derivatives where neighborhoods switch, affecting both continuity of the field and its smoothness. Increasing the neighborhood reduces truncation effects but does not guarantee exact continuity or smoothness. Our compact basis instead vanishes smoothly at its support boundary, avoiding this truncation problem. This establishes the structural regularity distinction; task accuracy and runtime are evaluated experimentally.

Q8 Smoothstep in Instant NGP

Q. The original Instant-NGP already has a smooth cubic interpolation option for SDFs. Why is this not discussed? [Originality; review p. 2]

A. We will discuss this option. Cubic smoothstep s(t)=3t2−2t3s(t)=3t^2-2t^3 has zero endpoint first derivatives, but s′′(0)=6s''(0)=6 and s′′(1)=−6s''(1)=-6. With shared corner features, it guarantees C1C^1 but generally not C2C^2 across cells. Our cutoff matches derivatives of every order. We agree that a comparison of accuracy and runtime under matched settings would be valuable and plan to include it in a revised manuscript.

Q9 Relevance of image fitting and SDF reconstruction

Q. Why are Sections 4.1 and 4.2 relevant? These tasks fit individual signals, and Smooth-NGP should only improve smoothness rather than image reconstruction quality. [Originality; review p. 3]

A. Our primary motivation is to develop an NGP-based neural representation with global smoothness and continuous derivatives of arbitrary order, rather than to optimize a PDE solver (p. 2, lines 78–101). Image fitting and SDF reconstruction are standard NGP applications. We include these experiments to assess whether the smooth representation can also improve reconstruction quality in conventional tasks.

Images evaluate function-value reconstruction, while SDFs additionally assess derivative-related geometry. Changing the interpolation and window structure changes expressivity and optimization, so the benefits need not be confined to tasks with explicit derivative supervision.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.