- 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
-
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.
-
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
-
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).
-
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 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 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 interpolation kernel with smoothstep.
| Method | Burg-1D | Pois-2DC | Pois-B2D | NS-LD | NS-BS | Wave-1D |
|---|---|---|---|---|---|---|
| INGP-Smoothstep | 3.33e-1 | 1.54e-1 | 7.70e-1 | 9.95e-1 | 3.51e-1 | 9.87e-1 |
| B-spline | 1.17e-1 | 1.25e-1 | 2.62e-1 | 6.39e-1 | 3.44e-1 | 8.56e-1 |
| Ours-Smoothstep | 1.46e-1 | 3.42e-3 | 2.23e-3 | 3.65e-1 | 2.00e-1 | 2.64e-1 |
| Ours | 9.53e-2 | 1.06e-3 | 1.37e-3 | 1.93e-1 | 1.31e-1 | 6.10e-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 -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 on the interior of its support, but not , whereas our interpolation is . 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.
| Method | PSNR ↑ | SSIM ↑ | Training time (min) ↓ |
|---|---|---|---|
| B-spline | 30.19 | 0.93 | 36.65 |
| Ours | 33.04 | 0.96 | 3.58 |
Smooth-NGP achieves higher PSNR ( vs. ) and SSIM ( vs. ) than B-spline, while reducing training time from to 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 and , respectively; runtime is in minutes.
| Method | CD ↓ | NE ↓ | Total runtime (min) ↓ |
|---|---|---|---|
| B-spline | 2.67 | 0.87 | 276.4 |
| Ours | 2.67 | 0.90 | 35.6 |
Smooth-NGP matches B-spline in CD () and achieves a comparable NE ( vs. ), providing similar reconstruction quality while reducing total runtime from to 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.
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∞ 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 C2 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 C2 continuity without forcing all derivatives to vanish at grid knots. However, a standard tensor-product cubic B-spline query accesses 4d control coefficients per level, compared with 2d 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 4×2d 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.
Smooth-NGP instead modifies the interpolation and architecture so that arbitrary-order analytic derivatives follow directly from a globally C∞ field, without finite differences or additional learned vertex-derivative coefficients.