acceptodds
Under review as a conference paper at ICLR 2027

Beyond Point Samples: Acquisition Modeling as a Separate Axis for Implicit Neural Representations

Abstract

Implicit neural representations (INRs) provide continuous representations of signals, yet recovering a continuous field from discrete measurements depends on both the representation and the process by which measurements constrain it. This relationship is specified by the acquisition model, which maps the continuous field to discrete observations. However, INR reconstruction has largely treated this measurement process as fixed. Most INR work has focused on representation design, while the measurement constraint itself is usually kept fixed as pointwise supervision. Here we isolate the effect of acquisition choice by holding the representation fixed and varying only the acquisition model. We find that acquisition choice materially changes the recovered continuous field, with increasingly pronounced effects as reconstruction relies increasingly on interpolation between sparse observations. Spatial acquisition improves continuous-field recovery overall, but its effect is spatially nonuniform, with larger reductions around boundaries, thin structures, and regions of rapid spatial variation, while point acquisition can remain locally preferable. Acquisition choice also changes the spectral structure of reconstruction error, particularly at intermediate and higher spatial frequencies. These results motivate treating the acquisition model as an explicit part of INR reconstruction rather than a fixed supervision choice.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Implicit Neural Representations, image acquisition, Continuous Representation Learning, Image Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_bXkXOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper decouples the measurement operator / acquisition model from the representation design of INRs. Under fixed representation, fixed measurement, and fixed optimization steps, it compares point acquisition and Gaussian-pixel spatial acquisition. Using PSNR, spatial error maps and spectral error, the authors demonstrate that the acquisition model systematically alters continuous-field recovery. And they claim that acquisition modeling should be treated as a design axis on par with representation.

Critical strengths
  1. The problem studied in this paper is practically meaningful, and the chosen angle is worthwhile.
  2. The experiments yield reasonably consistent results.
Critical weaknesses
  1. The novelty is limited. The core findings lack clear distinctions from existing theories and methods. Finite-support acquisition, pixel integration, anti-aliasing, and measurement-aware reconstruction are not new concepts. Related ideas have already been developed in Thera, mip-NeRF, and sampling theory.
  2. The experiments suffer from acquisition-model mismatch, which weakens the interpretability of the main results. The main experiments use the Gaussian–pixel operator to generate observations. However, the Point baseline interprets these observations as point samples, whereas the Spatial method adopts the correct forward operator. Consequently, the performance gains may partly stem from properly modeling the physical observation process, rather than from a novel insight for INR reconstruction.
  3. The coverage of experimental tasks and acquisition models is limited, and stronger mechanistic evidence is lacking. Although four architectures are tested, all main experiments are restricted to natural image reconstruction and rely primarily on one synthetic Gaussian–pixel kernel.
  4. The computational overhead is substantial, and a sufficient efficiency–performance analysis is missing. Training time for spatial acquisition is roughly 9–24 times that of point acquisition, with peak GPU memory consumption increased by approximately 26–35 times.
  5. The writing and presentation are poor. Figure 1 appears too early in the manuscript, making it difficult to grasp its intended message. Additionally, the * symbol in Eq (4) is left undefined, with no specification of the operation it represents.
Questions for authors

Please refer to weaknesses.

4 comments

We thank the reviewer for recognizing the practical relevance of the problem and the consistency of the results. We address each concern below.

C1. Novelty relative to Thera, mip-NeRF and sampling theory. Thera and mip-NeRF incorporate the pixel footprint into their architectures, through the activation function in Thera and the input encoding in mip-NeRF. Our study examines a different question: how does acquisition supervision affect reconstruction when the representation and optimization remain fixed? (Sec. 2, L. 134–140) While sampling theory establishes the importance of prefiltering, it does not explain how different INR architectures respond to acquisition supervision across sampling densities. We investigate these differences through controlled experiments (Table 1; Tables R10 and R14). We further demonstrate the practical relevance of this analysis by improving existing INR pipelines without changing their representations, obtaining +0.81+0.81 dB for SuperF and +0.44+0.44 dB for FINER's NeRF pipeline (Table R8). These results establish the value of studying acquisition supervision independently of representation design, with four main findings summarized in our response to Reviewer dAYP (C3).

C2. Acquisition-model mismatch. The acquisition-consistency control reveals an important interaction between acquisition choice, sampling density, and INR architecture. Under sparse sampling (×4\times4), spatial acquisition consistently outperforms genuine point acquisition across all four INR families (+0.90+0.90 to +1.29+1.29 dB; Appendix E), demonstrating that the advantage observed in our main experiments is not solely attributable to the mismatched point baseline. However, at denser sampling, the results become architecture-dependent, and point acquisition can outperform spatial acquisition. These findings show that the effectiveness of spatial acquisition depends on the sampling regime and representation, rather than being universally superior. This provides further evidence that acquisition modeling should be considered an independent design axis in INR reconstruction. (Sec. 4, L. 246–249; Sec. 5.1, L. 321–323, 349–350; Appendix E, L. 869–873, Table 8)

We further examine whether these improvements require precise knowledge of the acquisition operator. Across five acquisition kernels, spatial supervision using the corresponding operator achieves the highest mean PSNR, with gains of +0.40+0.40 to +1.39+1.39 dB over point supervision on all eight images per kernel. Even a simple box operator, without knowledge of the blur, improves reconstruction under all five kernels by +0.38+0.38 to +0.46+0.46 dB (Table R4). (Appendix I, L. 983–989, Table 14) Spatial supervision also maintains a positive mean gain when the assumed Gaussian width differs from the true width by up to 30% (Appendix B).

The practical importance of acquisition supervision is further demonstrated by improvements of +0.81+0.81 dB in SuperF and +0.44+0.44 dB in FINER's NeRF pipeline, without changing their representations (Table R8). Our contribution is therefore to establish acquisition supervision as an independent design axis for INRs and show how its effect depends on the representation and sampling density, rather than simply confirming the known importance of an accurate forward model. (Sec. 5.1, L. 358–361; Appendix J, L. 1057–1069, Table 19)

C3. Coverage and mechanistic evidence. We have extended the evaluation to ×8\times8 super-resolution, the WIRE SISR protocol, image denoising, and NeRF novel-view synthesis (Tables R1–R3). We also evaluate five acquisition kernels and an anisotropic Gaussian (Tables R4–R5), and expand the comparison to seven INR families and a classical B-spline representation (Tables R9–R10). (Sec. 5.1, L. 317–320, 352–361; Appendices H–J)

To better understand these improvements, we provide additional spectral and spatial analyses. The Fourier residual spectrum shows that most of the error reduction occurs within the measurement bandwidth and across all orientations (Figure 5, Tables R11–R12). The spatial analysis further shows that edges and textured regions account for the overall error reduction (Figure 4, Table R13). (Sec. 5.2, L. 397–407; Sec. 5.3, L. 411–421; Appendices G and K)

C4. Computational overhead. The computational cost reported in the submission is based on 16×1616\times16 quadrature. We further evaluate 4×44\times4 quadrature, which retains 88–97% of the reconstruction gain across all four INR families, while reducing training time to 2.5–4.9$\times$ that of point supervision, with no increase in peak GPU memory (Tables R6–R7). (Appendix A, L. 715–717, Table 2; Appendix D, L. 813–816, Table 6; Sec. 7, L. 514–515)

We also examine whether additional training or model capacity can close the performance gap. Even after training the point model for 54,142 steps (18$\times$ the original number) or increasing its parameters by 2$\times$ and 4$\times$, spatial supervision maintains a mean PSNR advantage of 1.53–1.83 dB, improving all 24 images (Table R15). The inference cost remains unchanged, as the acquisition operator is used only during training.

Table R15: Capacity and optimization controls (Appendix F of the paper): Spatial minus each enlarged Point model; FINER, Kodak-24, ×4\times4.

Point modelSpatial ahead by (dB)Images where Point wins
54,142 steps (18$\times$ original)1.531.530/24
2$\times$ parameters1.551.550/24
4$\times$ parameters1.831.830/24

C5. Presentation. We have defined ∗* as two-dimensional convolution in Eqs. (4)–(5), with ψ=kσ∗uh\psi=k_\sigma*u_h and (ψ∗f)(ci)(\psi*f)(\mathbf c_i), and use consistent notation for the observation and reconstruction operators throughout the paper. (∗* defined at L. 203)

Thanks for your response and new experiments. Most of my concerns are addressed. But I don't know why you only include image fitting in the main text. I think that adaptation to other kinds of INR tasks is important. Overall, I decide to raise my score from 1 to 2.

Thank you for reconsidering our submission and for the helpful suggestion. We agree that results on other INR tasks are important. In the paper, these results are discussed in Sec. 5.1 (L. 349–361), and the corresponding tables are provided in the appendix (Tables 14–19). Given the space budget, we kept the main text focused on the central question of the paper so that its argument remains clear. We appreciate your time and feedback throughout the discussion.

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

The paper studies how the measurement operator affects implicit neural representation reconstruction while holding the representation and training protocol fixed. It compares pointwise supervision against a Gaussian pixel spatial operator. The paper also examines spatial error patterns, power-spectrum discrepancies, acquisition consistency, kernel mismatch, and computational cost.

Critical strengths

One strong point is that the paper contains careful paired comparisons. Holding initialization, measurements, minibatch sequence, architecture, and optimization schedule fixed makes the main experiment informative about the effect of changing the reconstruction operator. I also note that the distinction between continuous queryability and faithful reconstruction is valuable. The paper acknowledges earlier spatial measurement models and mentions its substantial training overhead (while showing essentially unchanged inference cost).

Critical weaknesses

My assessment is that the contribution is incremental. The authors appropriately positions the paper as an empirical analysis, but the claim that earlier work does not isolate acquisition effects is a bit vague. In particular, SuperF (section 4.4, Table 3) explicitly ablates supersampling while holding other components fixed, including comparisons with the same direct transformation parameterization and fixed base frame. It's true that its multi-image setting differs from this paper, but this is closer precedent than this paper suggests. A similar issue exists with Efrat et al. as well.

Testing four INR families has marginal breadth. Also, the paper needs to identify what new, generalizable understanding emerges beyond the established importance of the forward model. A simple grid or spline reconstruction under the same operators would help establish whether the observed spatial and spectral trends reveal something distinctive about INRs.

Also, the spectral metric does not directly establish frequency-specific reconstruction accuracy. Appendix G measures the absolute difference between radially averaged log-power spectra. This discards Fourier phase and directional information. For example, a circularly shifted reconstruction can have exactly the same power spectrum as the reference while containing substantial reconstruction error. Please correct me if I am missing something.

Further, the spatial observations do not yet justify the proposed adaptive-acquisition interpretation. Figures 3, 6, and 7 provide useful qualitative evidence that improvements concentrate near structures and that point fitting sometimes wins locally. However, the paper does not quantify the relationship with gradient magnitude, edge regions, or texture.

Questions for authors

See above weaknesses.

3 comments

We thank the reviewer for the careful reading and the specific, actionable points. We address each below with new analyses.

C1. Relation to SuperF and Efrat et al. We acknowledge that the importance of accurate forward models has been established in classical reconstruction, particularly by Efrat et al. Our contribution is not to rediscover this principle, but to systematically investigate how acquisition supervision interacts with continuous neural representations under controlled reconstruction settings. Beyond demonstrating consistent improvements across INR families, our additional experiments reveal that the benefit depends strongly on representation architecture, sampling density, noise level, and acquisition-model accuracy. Notably, the acquisition-consistent control shows that spatial acquisition is advantageous under sparse sampling but can be inferior at denser sampling, while the noise experiments reveal a similar reversal under severe corruption. These findings provide a more nuanced characterization of when acquisition-aware INR reconstruction is beneficial. (Sec. 2, L. 134–140; Sec. 5.1, L. 321–323, 349–356)

SuperF employs supersampling to improve sub-pixel alignment in satellite multi-image super-resolution, but does not systematically investigate how the choice of acquisition supervision affects INR reconstruction. Applying our analysis to SuperF, we identify a more suitable spatial supervision operator that improves reconstruction PSNR by +0.81+0.81 dB on 19/20 scenes at ×2\times2, without modifying its representation, alignment, or optimization (Table R8). This demonstrates the practical value of systematically selecting acquisition supervision rather than treating it as a fixed component of INR training. (Sec. 5.1, L. 358–361; Appendix J, L. 1066–1069, Table 19)

Table R8: Existing INR pipelines with an incomplete acquisition model: restoring the missing part of the operator, with nothing else changed. SuperF on SatSynthBurst, adding the Gaussian its own data generator applies; FINER NeRF trained on 2002200^2 box-averaged images, evaluated at 8002800^2.

PipelineChange to the lossSettingGain (dB)Improved
SuperFbox →\rightarrow Gaussian ++ box×2\times2, 20 scenes, 3 seeds+0.81+0.8119/20
FINER NeRFcentre ray →\rightarrow pixel box8 Blender scenes+0.44+0.448/8

C2. Breadth of INR families. We further extend the evaluation from four to seven INR families by including WIRE, COSMO-INR, and Fourier Features under the same protocol. Spatial supervision improves reconstruction by +0.83+0.83 dB for WIRE, +0.94+0.94 dB for COSMO-INR, and +0.92+0.92 dB for Fourier Features, as shown in Table R9. We also include a classical B-spline baseline, discussed in C4. (Sec. 5.1, L. 317–318; Appendix H, L. 952–955, Table 12)

Table R9: Additional INR families, Kodak-24, ×4\times4, three seeds. Gain: Spatial minus Point field PSNR.

RepresentationGain (dB)Images improved
WIRE+0.83+0.8324/24
COSMO-INR+0.94+0.9422/24
Fourier Features+0.92+0.9223/24

C3. Generalizable understanding. Our experiments establish four findings beyond the known importance of the forward model. First, the effect of acquisition supervision depends on INR architecture and sampling density. Spatial supervision improves all 24 main comparisons, and these gains persist even when point models are trained for 18$\times$ more steps or given 4$\times$ more parameters (Table R15). However, the acquisition-consistent control shows that spatial acquisition outperforms point acquisition across all four INRs at ×4\times4, while point acquisition performs better for three architectures at ×1\times1 (Table R14). Second, the benefit depends on measurement noise and operator accuracy, with severe noise or an unsuitable spatial operator reversing the advantage (Tables R3 and R4). Third, the reconstruction improvements are structured: most of the error reduction occurs within the observed frequency range, across all orientations (C5), while edges and textured regions account for most of the spatial improvement (C6). Finally, classical B-splines show similar gains under sparse sampling, closely tracking FINER across images (r=0.94r=0.94–0.960.96). This identifies an effect shared with classical reconstruction, while the differences across INR architectures and sampling densities characterize how this principle manifests in INRs (C4). (Sec. 5.1, L. 317–323, 349–361; Sec. 6.4, L. 493–497)

C4. Classical baselines. We also evaluate a classical cubic B-spline representation under point and spatial supervision on Kodak-24 and DIV2K-24 at all three sampling scales (Table R10). Spatial supervision improves all 24 images in each of the six dataset–scale comparisons, with gains increasing from +0.05+0.05/+0.09+0.09 dB at ×1\times1 to +1.29+1.29/+1.55+1.55 dB at ×4\times4 (Kodak/DIV2K). At ×4\times4, the per-image gains closely follow those of FINER (r=0.96r=0.96 on Kodak and r=0.94r=0.94 on DIV2K), showing that classical reconstruction also benefits from spatial supervision under sparse sampling. At denser sampling, the B-spline gains remain small due to its fixed control-point count, while INRs respond differently across architectures (Table 1; Table R14). These results distinguish the benefits shared with classical reconstruction from those that depend on the INR representation. (Sec. 5.1, L. 318–320; Appendix H, L. 964–971, Table 13; Sec. 6.4, L. 493–494)

Table R10: Classical reconstruction: Spatial minus Point (dB); every entry improves 24/24 images.

DatasetRepresentation×1\times1×2\times2×4\times4rr with FINER (×4\times4)
Kodak-24Cubic B-spline+0.05+0.05+0.43+0.43+1.29+1.290.96
DIV2K-24Cubic B-spline+0.09+0.09+0.59+0.59+1.55+1.550.94

The fixed B-spline control-point count limits it at ×1/×2\times1/\times2. rr: per-image correlation of the ×4\times4 gains with FINER's.

C5. Spectral metric. We extend the spectral analysis by measuring reconstruction error directly in the Fourier domain using ∣F(f^−f)∣2|\mathcal{F}(\hat f-f)|^2. Unlike the original power-spectrum comparison, this measure accounts for phase differences and captures reconstruction errors caused by spatial shifts. Across all 24 comparisons, spatial supervision reduces the reconstruction error, with 81–100% of the reduction occurring within the measurement bandwidth (Figure 5, Table R11). We further analyze the error by orientation and find consistent improvements in all four directions (Table R12). (Sec. 5.3, L. 411–421, Figure 5; Appendix G, L. 916–930, Tables 10–11)

Table R11: Share of the residual-energy reduction by frequency band; range over four INRs and two datasets.

Band×1\times1×2\times2×4\times4
Low (ρ<1/6\rho\lt 1/6)10–45%34–65%91–104%
Intermediate51–82%33–66%−1-1–7%
High (ρ≥1/3\rho\geq1/3)1–10%−4-4–7%−2-2–4%
Below grid Nyquist100%87–99.4%81–95%

ρ\rho in cycles per pixel of the 5122512^2 field. Signed shares can exceed 100% when a small loss in one band is offset by a larger gain in another.

Table R12: Error reduction per orientation, 10log⁡10(EPoint/ESpatial)10\log_{10}(E_{\mathrm{Point}}/E_{\mathrm{Spatial}}) (dB), 0<ρ<0.50\lt \rho\lt 0.5; range over four INRs and two datasets, seed 0.

ScaleHorizontalDiagonal 45∘45^\circVerticalDiagonal 135∘135^\circImproved
×1\times1+0.20+0.20–1.781.78+0.13+0.13–1.321.32+0.20+0.20–2.042.04+0.13+0.13–1.381.38714/764
×2\times2+0.78+0.78–2.382.38+0.41+0.41–2.122.12+0.74+0.74–2.422.42+0.55+0.55–2.092.09765/768
×4\times4+0.84+0.84–1.581.58+0.98+0.98–1.981.98+0.63+0.63–1.671.67+1.10+1.10–1.911.91760/768

Last column: image–orientation pairs with lower error under spatial supervision (4 INRs ×\times 2 datasets ×\times 24 images ×\times 4 orientations; at ×1\times1 one reconstruction, SASNet on Kodak kodim01, is unavailable, giving 764 instead of 768).

C6. Quantifying the spatial pattern. We quantify the spatial error patterns by dividing each reference image into smooth, edge, and textured regions using the structure tensor. For each region, we measure the relative reduction in reconstruction error under spatial supervision and its contribution to the total error reduction (Table R13). Spatial supervision consistently reduces error in edge and textured regions, which account for most of the overall improvement. Point supervision can perform better in smooth regions, but these contribute relatively little to the total error change. These findings quantify the spatial differences observed in Figures 3, 7, and 8. (Sec. 5.2, L. 397–407, Figure 4; Appendix K, L. 1074–1079, Table 20) We have revised the adaptive-acquisition interpretation in the paper accordingly. (Sec. 6.2, L. 469–473)

Table R13: Reconstruction error by image structure. Relative regional MSE reduction (Red.) and contribution to total error reduction (Share), in percent. Ranges across four INRs and two datasets.

ScaleEdge Red.Edge ShareTexture Red.Texture ShareSmooth Red.Smooth Share
×1\times1+7+7 to +43+4365–97+1+1 to +21+214–36−20-20 to 00−5-5 to −1-1
×2\times2+21+21 to +49+4968–85+5+5 to +30+3016–33−43-43 to +1+1−3-3 to 00
×4\times4+24+24 to +39+3970–82+12+12 to +25+2524–30−100-100 to +11+11−11-11 to +2+2

Thank you again for your time and constructive feedback. As the discussion period is ending soon, we would like to kindly ask whether our response has addressed your concerns. If so, we would be grateful if you would consider updating your rating. We are happy to clarify any remaining questions.

Reviewer_2VkKOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper isolates acquisition modeling (the measurement constraint mapping continuous fields to discrete observations) as a distinct, critical design axis for Implicit Neural Representations (INRs), complementary to network representation capacity. By holding network architectures fixed and varying only the supervision/acquisition model (Point evaluation vs. Spatially integrated Gaussian-pixel footprint evaluated via fixed K×K quadrature), the authors evaluate four major INR families (FINER, MetricGrids, SIREN, SASNet) across Kodak-24 and DIV2K-24 at multiple sampling densities (×1,×2,×4).

Critical strengths
  1. Clear Conceptual Framing: Decoupling measurement constraints from representational capacity draws a strong, well-reasoned parallel to classical signal acquisition and sampling theory.
  2. Rigorous Controls: Thorough ablation studies in the appendix validate key variables including quadrature resolution (K=16), sensitivity to Gaussian width kernel mismatch, capacity and training epoch expansions (up to 4× parameters and 54"k" optimization steps), and measurement-reconstruction consistency.
  3. Multi-perspective Analysis: Combines standard aggregate metrics (PSNR) with spatial signed error maps (Δe(x)) and radially averaged spectral error curves to explain where and why performance differs. Global Reconstruction Gain: Spatial acquisition systematically improves continuous-field recovery (up to +2.17" dB" PSNR) across all architectures and scales.
Critical weaknesses
  1. Limited Evaluation across Standard INR Downstream Tasks: While the performance gains (PSNR hike) are clear on standard 2D image fitting, the paper lacks empirical verification on core INR downstream tasks such as super-resolution, image inpainting, image denoising, and 3D scene reconstruction / NeRF rendering. To prove that treating acquisition as an independent axis is universally beneficial, it needs validation on these standard benchmarks. It will also provide validity toward network’s continuity.
  2. Reliance on Gaussian Kernels vs. Prior Art: Gaussian-based blur and integration models during training/evaluation are common (though not identical) in INR-based super-resolution literature. Focusing heavily on an isotropic Gaussian pixel operator leaves the overarching claim somewhat vulnerable; ablating other, non-linear or complex physical acquisition operators would provide much stronger solidity for the paper's core proposal.
Questions for authors
  1. Validation on Downstream Tasks: Have you evaluated spatial acquisition on standard downstream INR applications like image super-resolution, inpainting, denoising, or 3D NeRF volume rendering?
  2. Non-linear & Complex Acquisition Models: Given that Gaussian integration formulations are already relatively popular in super-resolution contexts, have you considered testing or ablating non-linear, anisotropic acquisition operators to establish broader generality?
  3. Higher-Dimensional Scalability: How does spatial acquisition scale to higher-dimensional continuous representations (e.g., 3D occupancy fields or 5D Neural Radiance Fields), where k_d quadrature sampling could lead to severe memory bottlenecks?
4 comments

We thank the reviewer for the constructive suggestions.

C1. Downstream tasks. We have presented single-image super-resolution results at ×2\times2 and ×4\times4 in Table 1, where we reconstruct 512×512512\times512 high-resolution images from blurred measurements at 256×256256\times256 and 128×128128\times128, respectively. Spatial supervision improves all 16 comparisons by +0.68+0.68 to +2.17+2.17 dB. We further extend the evaluation to ×8\times8 super-resolution on Kodak-24, with results presented in Table R1 below. We also evaluate WIRE under its ×4\times4 SISR protocol, obtaining a gain of +0.83+0.83 dB on all 24 images. (Appendix J, L. 1017–1019, Table 16; Appendix H, L. 952–955)

We also evaluate NeRF novel-view synthesis using FINER, where the training images are downsampled from 800×800800\times800 to 200×200200\times200 through box averaging. By supervising the pixel footprint instead of a single centre ray, we improve the reconstruction PSNR by +0.44+0.44 dB across all eight Blender scenes, evaluated at the original 800×800800\times800 resolution. (Sec. 5.1, L. 358–361; Appendix J, L. 1057–1064)

We further evaluate spatial supervision for image denoising at different noise levels and sampling resolutions. At native resolution, spatial supervision improves all 24 Kodak images for SIREN, FINER, and MetricGrids, with gains of +0.15+0.15, +0.66+0.66, and +1.10+1.10 dB, respectively, at σn=0.02\sigma_n=0.02 (Table R2). At ×4\times4 downsampling, FINER improves by +1.44+1.44 dB at σn=0.001\sigma_n=0.001, whereas point supervision performs better at σn=0.02\sigma_n=0.02 (−2.50-2.50 dB; Table R3). These results show that spatial supervision remains effective under noise, while point supervision can be preferable when measurements are both sparse and highly noisy. (Sec. 5.1, L. 354–356; Appendix J, L. 1022–1036, Tables 17–18) Regarding inpainting, it changes which pixels are observed through a mask, whereas our study concerns what each observed pixel measures, the two are complementary.

Table R1: ×8\times8 super-resolution on Kodak-24 (three seeds). Gain: Spatial minus Point field PSNR.

RepresentationGain (dB)
SIREN+1.17+1.17
FINER+0.80+0.80
MetricGrids+1.18+1.18

Table R2: Denoising at ×1\times1, Kodak-24 (three seeds), at both ends of the noise range: Spatial minus Point field PSNR (dB).

Representationσn=0\sigma_n=0σn=0.02\sigma_n=0.02
SIREN+0.16+0.16+0.15+0.15
FINER+1.08+1.08+0.66+0.66
MetricGrids+1.64+1.64+1.10+1.10

Table R3: Denoising with FINER at ×4\times4, Kodak-24 (three seeds). Gain: Spatial minus Point field PSNR.

Noise σn\sigma_nGain (dB)
0.001+1.44+1.44
0.005+1.03+1.03
0.010−0.04-0.04
0.020−2.50-2.50

C2. Acquisition operators. We further evaluate five acquisition kernels to examine whether the results extend beyond the isotropic Gaussian model. Using FINER at ×4\times4, we test Gaussian–pixel, Airy–pixel, disk–pixel, anti-aliased bicubic, and box kernels. Spatial supervision with the matching operator improves PSNR by +1.39+1.39, +1.24+1.24, +1.12+1.12, +0.61+0.61, and +0.40+0.40 dB, respectively, over point supervision, with improvements on all eight images for each kernel (Table R4). Spatial supervision using the corresponding acquisition kernel achieves the highest mean PSNR in each case, with bicubic and box effectively tied for box-acquired measurements (0.001 dB difference). Even the box operator, which requires no knowledge of the blur, outperforms point supervision under all five kernels. (Sec. 5.1, L. 352–354; Appendix I, L. 983–989, Table 14)

We also evaluate anisotropic Gaussian–pixel measurements to examine the effect of directional information. Accounting for the anisotropic footprint instead of using an isotropic operator improves PSNR by +0.91+0.91 dB for FINER and +1.01+1.01 dB for MetricGrids, each on all 24 images (Table R5). (Appendix I, L. 1002–1005, Table 15; Sec. 3.2, L. 235–236; Sec. 7, L. 508–509) Furthermore, Appendix B shows that spatial supervision maintains a positive mean PSNR gain even when the assumed Gaussian width differs from the true width by up to ±30\pm30\\%.

Table R4: Spatial reconstruction minus Point (dB) for every combination of acquisition kernel (rows) and assumed operator (columns). FINER, 8 Kodak images, ×4\times4, seed 0; images improved out of 8 in parentheses. Bold: matching operator.

Acquired with ↓\downarrow / assumed →\rightarrowBoxBicubicGaussian–pixelAiry–pixelDisk–pixel
Gaussian–pixel+0.38+0.38 (8)+0.44+0.44 (8)+1.39\mathbf{+1.39} (8)+1.21+1.21 (8)+1.14+1.14 (8)
Airy–pixel+0.45+0.45 (8)+0.55+0.55 (8)+1.13+1.13 (8)+1.24\mathbf{+1.24} (8)+1.20+1.20 (8)
Disk–pixel+0.46+0.46 (8)+0.58+0.58 (8)+0.81+0.81 (8)+1.02+1.02 (8)+1.12\mathbf{+1.12} (8)
Bicubic (anti-aliased)+0.43+0.43 (8)+0.61\mathbf{+0.61} (8)−2.49-2.49 (0)−1.08-1.08 (0)−0.65-0.65 (0)
Box+0.40\mathbf{+0.40} (8)+0.40+0.40 (8)−3.57-3.57 (0)−1.91-1.91 (0)−1.62-1.62 (0)

Table R5: Anisotropic Gaussian–pixel observations (σrow/h=0.75\sigma_{\mathrm{row}}/h=0.75, σcol/h=0.25\sigma_{\mathrm{col}}/h=0.25): mean field PSNR (dB), Kodak-24, ×4\times4, three seeds.

Assumed operatorFINERMetricGrids
Point24.1124.00
Isotropic Gaussian–pixel24.6224.37
Anisotropic Gaussian–pixel (matching)25.5225.37

C3. Scalability. We evaluate the effect of quadrature resolution on reconstruction accuracy and computational cost in Tables R6–R7. Reducing the quadrature from 16×1616\times16 to 4×44\times4 retains 88–97% of the spatial-supervision gain across all four INR families, while requiring 2.5–4.9$\times$ the training time of point supervision. At K=4K=4, peak GPU memory does not exceed that of point supervision for all four families. (Sec. 3.2, L. 233; Appendix A, L. 715–717, Table 2; Appendix D, L. 813–816, Table 6; Sec. 7, L. 514–515)

We also measure the computational cost for NeRF, where spatial supervision requires 6.2$\times$ the training time and 2.05$\times$ the peak GPU memory of point supervision. The integration is performed over the two-dimensional pixel footprint using K2K^2 sub-pixel rays, without changing the sampling along each ray. Thus, for image-supervised 3D fields, the quadrature scales with the pixel footprint rather than the dimensionality of the underlying representation. We will clarify this distinction and include the measured costs in the revision. (Appendix J, L. 1057–1064)

Table R6: Quadrature accuracy, Kodak-24, ×4\times4, three seeds. Gain: Spatial minus Point field PSNR; retained =ΔPSNRK=4/ΔPSNRK=16=\Delta\mathrm{PSNR}_{K=4}/\Delta\mathrm{PSNR}_{K=16}.

RepresentationK=4K=4 gain (dB)K=16K=16 gain (dB)Retained
SIREN+1.21+1.21 (24/24)+1.24+1.24 (24/24)97%
FINER+1.38+1.38 (24/24)+1.45+1.45 (24/24)95%
MetricGrids+1.24+1.24 (24/24)+1.39+1.39 (24/24)89%
SASNet+0.68+0.68 (23/24)+0.77+0.77 (24/24)88%

Table R7: Training time and peak memory relative to Point. One RTX 3090; CUDA-synchronized median over four images and three repetitions.

RepresentationTraining time / Point, K=4K=4Training time / Point, K=8K=8Training time / Point, K=16K=16Peak memory / Point, K=4K=4
SIREN2.6$\times$5.0$\times$18.9$\times$1.0$\times$
FINER3.9$\times$8.3$\times$31.3$\times$1.0$\times$
MetricGrids2.5$\times$2.9$\times$7.3$\times$1.0$\times$
SASNet4.9$\times$7.1$\times$25.8$\times$0.9$\times$
Reviewer_2VkKOpenReview ↗

Rebuttal Acknowledgment

Thank you to the authors for the detailed response and the additional experiments provided during the rebuttal period. The new results and clarifications address my technical questions and strengthen the manuscript. Please ensure these updates are cleanly integrated into the revised version. I am maintaining my positive rating in support of the paper.

Thank you for carefully considering our rebuttal and additional experiments. We appreciate your acknowledgment that the new results address the technical questions and strengthen the manuscript. We have incorporated these findings and clarifications into the revised submission, and we appreciate your continued support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.