acceptodds
Under review as a conference paper at ICLR 2027

SPAN-INR: SPATIAL ASSIGNMENT NETWORKS FOR DEEP SINUSOIDAL IMPLICIT REPRESENTATIONS

140 views

Abstract

Real signals are detailed in some places and smooth in others, so a sinusoidal implicit neural representation (INR) has to produce different harmonics at different locations. Existing sinusoidal INRs, however, spend their parameters on a separate matrix in every layer, which decides which channels can combine (*reach*), and do little to control what is produced where (*assignment*): the scale and offset of the argument entering each sine, which by the Jacobi–Anger expansion decide how high the harmonic orders go and the balance between odd and even harmonics, are the same at every location or are adjusted only by coarse masks on groups of features. In this paper, we first show that in our setting reach is cheap and assignment is what has to be learned. Analytically, which harmonic interactions are available depends on the layer matrix only through its connectivity, and the scale of each of its rows can be taken over by a gain; empirically, a single dense mixing matrix shared by all stages and frozen at its random initialization loses only 0.14 dB against a fully learned one, whereas restricting it to its diagonal loses 5.81 dB, and a model with an independent matrix per stage loses 3.48 dB at the same budget. Based on this observation, we propose a new sinusoidal INR, called Spatial Assignment Network (SPAN-INR), which spends its parameters on assignment rather than on reach: most of them go to coordinate controllers, one per stage, that predict a gain and a shift for every channel at every location, while the single shared mixing matrix that provides the reach is small by comparison. Since a new stage needs only a new controller and no new matrix, the same budget buys more stages, and every stage is one more place to assign harmonics locally. Experiments show that SPAN-INR achieves 38.90 dB PSNR on image fitting with 126K trainable parameters and the best results among the compared methods on super-resolution, denoising, 3D occupancy and novel-view synthesis. Analyses of the fitted models further confirm that the gain tracks the local spectral centroid in the later stages, the shift tracks the odd/even balance, and the empirical neural tangent kernel is 9.0 times narrower at edges than in smooth regions, against 4.8 times for SASNet.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
implicit neural representations, spatial modulation, harmonic composition, channel mixing

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_XfuZOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SPAN-INR, a sinusoidal INRs that combines a shared channel-mixing matrix with stage-specific coordinate networks producing per-channel gains and shifts. Its main contribution is to allocate model capacity primarily to spatial modulation, supported by theoretical analysis of how modulation and mixing affect harmonic composition. Experiments across image fitting, super-resolution, denoising, 3D occupancy reconstruction, and novel-view synthesis show improvements over parameter-matched baselines. Ablations and spectral analyses further support the roles of spatial modulation and dense channel mixing, including the finding that a frozen random mixing matrix performs close to a learned one.

Critical strengths
  1. Strong connection between theoretical analysis and empirical validation. The paper analyzes how spatial gains, shifts, and channel mixing influence harmonic composition, then investigates these roles through targeted ablations and spectral analyses. The finding that frozen random mixing performs nearly as well as learned mixing provides particularly compelling support for allocating capacity to spatial modulation.
  2. Broad and consistent experimental evidence. Evaluation spans image fitting, super-resolution, denoising, 3D occupancy reconstruction, and novel-view synthesis, with improvements over parameter-matched baselines. This breadth supports the practical value of the proposed architecture across different signals and reconstruction settings.
Critical weaknesses
  1. Training efficiency remains unclear. Figure 7b compares convergence by iteration count rather than wall-clock time. Matching parameter counts does not necessarily match computational cost, particularly given the repeated mixing operations and stage-specific controllers. Reporting training time and PSNR versus wall-clock time, especially against SASNet, would clarify the practical accuracy–cost trade-off.
Questions for authors
  1. Training cost. Figure 7b reports convergence against iteration count. What are the per-iteration and total training times of SPAN-INR compared with SASNet and FLAIR under the same hardware and batch settings? Does its accuracy advantage persist under a matched wall-clock budget?

  2. Matched-budget ablations. Please clarify the exact stage counts, feature widths, and controller architectures used for the shared-WW and untied-WW variants in Figure 7a. Which components were resized to match parameter counts? This is important for distinguishing the effect of weight sharing from that of reallocating capacity to the spatial controllers.

4 comments
AuthorsOpenReview ↗

Training efficiency and training cost

We thank the reviewer for their time and feedback. Table R8 reports computational cost on one RTX 3090. All the models under test use the full 768×512 image (393,216 coordinates) per iteration. SPAN-INR and SASNet have similar per-iteration and total training times, both substantially lower than FLAIR's. Repeated mixing operations use a small shared WW, and small stage-specific controllers read only the coordinate, keeping SPAN-INR's iteration comparable to SASNet.

Table R8: Computational cost.

ModelParamsFLOPs per queryTraining time per iteration (ms)Total training time (wall-clock, 2,000 iterations)Inference time per image (ms)
SIREN133,123265K41.082 s16.4
FLAIR128,706509K653.41,307 s139.9
SASNet126,211217K64.5129 s20.7
SPAN-INR125,871268K65.7131 s21.0

We further investigate the convergence behavior of the four tested models at different wall-clock training times. Table R9 reports the mean Kodak-24 PSNR at 30-second intervals, measured on the same hardware and under the same full-batch setting as Table R8.

Table R9: Mean Kodak-24 PSNR (dB) against wall-clock training time.

Model30 s60 s90 s120 s150 sEnd of training
SIREN27.9729.3629.4729.4729.4729.47 (82 s)
FLAIR22.5724.6125.2725.7226.1331.25 (1,307 s)
SASNet33.7235.5936.5836.9136.9536.95 (129 s)
SPAN-INR30.1133.5037.3938.8238.9038.90 (131 s)

SPAN-INR's per-iteration time is comparable to SASNet's (65.7 vs. 64.5 ms) and about 10× lower than FLAIR's (653.4 ms). Under a matched wall-clock budget, SASNet leads early and SPAN-INR leads from about 82 s onward (e.g., 38.82 vs. 36.91 dB at 120 s). Against FLAIR, SPAN-INR leads at every budget, and its final 38.90 dB at 131 s exceeds FLAIR's final 31.25 dB after 1,307 s. SPAN-INR reaches 37.39 dB at 90 s, exceeding SASNet's final 36.95 dB at 129 s (Table R9).

AuthorsOpenReview ↗

Matched-budget ablations.

Table R10 gives the configurations of the Untied WW variant of Fig. 7a and of SPAN-INR, and adds a Shared WW row at the same feature width as Untied WW. All three models have 10 layers and about 126K parameters, and are trained the same way on Kodak-24 at 768×512. Each layer has its own coordinate controller MkM_k: an MLP that reads only xx, with two hidden layers of width hh and sine scale ωctrl=30\omega_{\mathrm{ctrl}}=30, which predicts the channel-wise gain and shift fields gk(x)g_k(x) and sk(x)s_k(x). Untied WW uses a separate C×CC\times C mixing matrix per layer (ten matrices); Shared WW and SPAN-INR use one mixing matrix WW across all layers. The components resized to hold the budget are the feature width CC and the controller hidden width hh; depth and controller architecture are the same in every row. At the same feature width C=97C=97, sharing frees 84,681 parameters for wider controllers (hh from 14 to 47) and gains 2.70 dB, with higher PSNR on all 24 images. SPAN-INR moves still more of the budget to the controllers by narrowing CC to 34 and widening hh to 81.

Table R10: Untied WW, shared WW and SPAN-INR at the same depth and parameter budget (Kodak-24).

ModelDepthCChhMixing paramsController paramsTotal paramsPSNR (dB)
Untied WW10971494,09031,620126,00435.42
Shared WW1097479,409117,090126,79338.12
SPAN-INR1034811,156124,610125,87138.90

At the same depth and budget, SPAN-INR is 3.48 dB above Untied WW: weight sharing frees the capacity, and the gain comes from reallocating it from the mixing matrices to the coordinate controllers. We will include Table R10 and the exact configurations of all ablation variants in the revised manuscript.

Reviewer_XfuZOpenReview ↗

Good rebuttal!

Thank you for the detailed response and additional experiments. All my concerns are addressed accordingly. I vote to accept the paper.

Dear Reviewer XfuZ,

We sincerely appreciate your careful consideration of our responses and additional experiments. We are grateful for your encouraging feedback. We will incorporate your feedback into the revised manuscript.

Reviewer_cHRnOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SPAN-INR proposes a parameter redistribution strategy for deep sinusoidal implicit neural representations (INRs). The authors argue that conventional periodic INRs spend significant capacity on layer-wise cross-channel mixing matrices (W) to establish "reach" (channel connectivity), leaving spatial "assignment" (local frequency scale and phase balance) insufficiently controlled. SPAN-INR decouples these roles by sharing a single dense mixing matrix W across stages, which can even be frozen at random initialization with minimal performance loss and allocating the parameter budget to stage-specific coordinate controllers (small MLPs) that predict channel-wise gain (g_k (x)) and shift (s_k (x)) fields. The approach is evaluated across image fitting, super-resolution, denoising, 3D occupancy, and novel-view synthesis.

Critical strengths
  1. Clean Theoretical Framing of Reach vs. Assignment: The paper offers a clear mathematical formulation separating channel connectivity from local harmonic modulation using Jacobi-Anger phase expansions (Propositions 1 and 2).

  2. Good Empirical Analysis of Frozen Connectivity: The empirical finding that a single frozen random dense matrix W preserves overall reconstruction quality compared to fully learned untied matrices (-0.14 dB PSNR drop) provides a strong argument for parameter efficiency in deep sinusoidal compositions.

  3. Parameter-Efficient Depth Scaling: Reusing the mixing matrix W across stages decouples the mixing parameter budget from depth, allowing parameter capacity to be shifted into light coordinate controllers to train deeper networks (K stages).

  4. Strong Benchmark Metrics: Achieves competitive performance across standard evaluation tasks (Kodak image fitting, DIV2K super-resolution/denoising, 3D occupancy, NeRF) against non-adaptive and group-masked baselines.

Critical weaknesses
  1. Unacknowledged Architectural Similarities to several Modulated INRs: The core mechanism of modulating sinusoidal layers via coordinate-conditioned MLPs to predict scale (gain) and shift parameters is structurally identical to existing dynamic modulation paradigms, most notably INCODE (INCODE: Implicit Neural Conditioning with Prior Knowledge Embeddings) and related FiLM-based coordinate modulators. The paper presents stage-wise coordinate controllers predicting gain and shift as a novel concept without explicitly contrasting its mathematical or structural formulation against InCode's harmonizer/conditioning MLPs.

  2. Omission of Key Baselines: The paper compares against standard non-adaptive periodic activations (SIREN, WIRE, FINER) and coarse spatial masking (SASNet, FLAIR), but entirely omits direct comparisons against prior conditioned coordinate shifting novelties like INCODE and Cosmo-INR. Because both utilize MLP-driven coordinate modulations for scaling/shifting sinusoidal arguments and report state-of-the-art PSNRs, their exclusion makes it impossible to verify whether SPAN-INR's shared-W design provides a true empirical advantage over prior coordinate-conditioned architectures under matched parameter budgets.

  3. Reproducibility Concerns & Missing Per-Test Training Configurations: While Appendix B lists task-level hyperparameter defaults (Table 3), individual training configurations, exact learning rate schedules, warm-up strategies, and seeds for specific test images/scenes are not specified. Given that performance varies significantly with the choice of sine scales ω_0 and ω_"ctrl" (Figure 7d), missing per-test configurations hinder full reproducibility prior to code release.

Questions for authors
  1. Mathematical and Empirical Comparison to key coordinate algorithms are encouraged: How does the proposed gain/shift coordinate controller mathematically and structurally differ from the coordinate-conditioned modulation networks in INCODE? Can the authors provide direct quantitative comparisons (PSNR, SSIM, parameter count, and FLOPs per query) against INCODE and Cosmo-INR under identical parameter budgets?

  2. FLOPs and Inference Speed: Since SPAN-INR requires running K separate MLP coordinate controllers M_k (x) at every queried coordinate x, what is the exact wall-clock training/inference time and total FLOP count compared to standard single-backbone INRs like SIREN and SASNet?

7 comments
AuthorsOpenReview ↗

Unacknowledged Architectural Similarities to several Modulated INRs

We thank the reviewer for their time and feedback. SPAN-INR's contribution is the reach/assignment analysis and the parameter allocation it leads to: one small shared dense WW provides channel reach, and nearly all parameters go to the coordinate controllers for local spectral assignment. Each layer's coordinate controller takes only the coordinate xx and predicts a gain gk(x)g_k(x) and a shift sk(x)s_k(x) that vary over coordinate, channel and layer; entering the next sine, the gain sets the local frequency scale and the shift the local phase. In contrast, INCODE predicts one set (a,b,c,d)(a,b,c,d) from an embedding of the target image, shared by all layers and constant over coordinates, and COSMO-INR predicts (T,ζ)(T,\zeta) per layer from the same kind of embedding; both keep a separate weight matrix per layer.

Table R4 gives the layer updates and architectural characteristics of the three models.

Table R4: Architectural characteristics of INCODE, COSMO-INR and SPAN-INR.

ModelConditioning inputLayer updateModulation granularityWeight WW
INCODEimage embeddingasin⁡(b,ω0z+c)+da\sin(b\\,\omega_0 z+c)+dglobalper layer
COSMO-INRimage embeddingrcT(z),e,2πi,ζz\mathrm{rc}_{T}(z)\\,e^{\\,2\pi i\\,\zeta z}per layerper layer
SPAN-INRcoordinate xxgk(x)⊙Wsin⁡(ω0uk−1)+sk(x)g_k(x)\odot W\sin(\omega_0 u_{k-1})+s_k(x)per layer, channel and coordinateshared

zz: the layer's pre-activation; uk−1u_{k-1}: the previous layer's output. rcT\mathrm{rc}_T: raised-cosine function with bandwidth TT; ζ\zeta: frequency shift. (a,b,c,d)(a,b,c,d) and (T,ζ)(T,\zeta) are predicted from the image embedding; gk(x),sk(x)g_k(x),\\ s_k(x) from coordinate xx.

SPAN-INR's 125,871-parameter image-fitting model allocates 99% to controllers for local spectral assignment and 0.9% to reach (channel connectivity) through shared WW.

AuthorsOpenReview ↗

Comparison with INCODE and COSMO-INR

Table R5 provides direct comparisons of SPAN-INR with INCODE and COSMO-INR under matched parameter budgets on Kodak-24 at 768×512. The SPAN-INR rows approximately match the baselines' Network and Total parameter counts.

Table R5: Kodak-24 PSNR under matched parameter budgets.

Params
ModelEncoderNetworkTotalPSNR (dB)
INCODE231,488205,287436,77535.36
COSMO-INR231,488205,419436,90736.98
SPAN-INR, matched network count–205,227205,22742.01
SPAN-INR, matched total count–436,931436,93146.99
INCODE231,488125,769357,25732.89
COSMO-INR231,488125,247356,73533.77
SPAN-INR, matched network count–125,871125,87138.90
SPAN-INR, matched total count–356,913356,91345.80

We used each method's official code and released optimiser settings: INCODE for 2,000 steps, and COSMO-INR for 2,000 epochs (12,000 optimiser steps, six times SPAN-INR's). With the official COSMO-INR code we could not reproduce the PSNR reported in its paper; Table R5 gives what the official code reaches: 36.98 dB at the paper's width 256 and three hidden layers. These results show an empirical advantage under matched parameter budgets for SPAN-INR's assignment-first design: 5.0–12.9 dB in mean PSNR across all eight comparisons. The gap is large because the baselines keep 231,488 parameters in an image encoder that outputs a few modulation values for the whole image, applied unchanged at every coordinate. SPAN-INR has no encoder: all of its parameters are in the per-coordinate network, where the controllers set the gain and shift at each coordinate. We will include both baselines in the revised manuscript.

AuthorsOpenReview ↗

Training configurations and reproducibility

We use one fixed configuration per task for all test images/scenes of that task, with no tuning of per-test configurations. All test images and scenes use the same seed (0); image fitting is full-batch, so there the seed sets only the initialization. All 24 Kodak images use the sine scales ω0=10\omega_0=10 and ωctrl=30\omega_{\mathrm{ctrl}}=30. Table 3 in Appendix B lists the task-level hyperparameter defaults: architecture (KK, CC, hh), sine scales ω0\omega_0 and ωctrl\omega_{\mathrm{ctrl}}, and trainable parameter count. Appendix B describes every task's training schedule; no task uses learning rate warm-up. Image fitting uses AdamW with an initial learning rate of 3×10−33\times10^{-3} and cosine decay to 1% of that rate over 2,000 full-batch iterations. We will further add implementation details to the revised Appendix B.

Reviewer_X8nxOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

SPAN-INR combines a shared dense mixing matrix with stage-specific, coordinate-conditioned gain and shift to control spatially varying harmonic contributions in sinusoidal implicit neural representations. The method is evaluated on five reconstruction tasks and achieves promising results.

Critical strengths
  1. The motivation is clear.
  2. Consistent reported improvements across five reconstruction tasks(especially image fitting).
  3. Interesting insight and analysis, e.g., i. Frozen random mixing nearly matches learned mixing, while diagonal mixing performs substantially worse, suggesting that dense connectivity can support effective spatial controllers. ii. Harmonic expansions explain how gain and shift influence sinusoidal composition
Critical weaknesses
  1. More comparison with prior work is needed: Coordinate-dependent affine modulation has been explored in Coordinate-Aware Modulation for Neural Fields (Lee et al., 2024), and weight-tied sinusoidal recurrence in Recurrent Sinusoidal INRs for Efficient High-Fidelity Representation (Cho et al., 2026), so direct comparisons are needed to establish the benefit of SPAN-INR’s particular combination.
  2. Ablations are not isolated: The shared-versus-untied comparison also changes architectural dimensions to match parameter budgets. Removing controllers eliminates both spatial adaptation and basic gain/shift flexibility. Additional controls are needed to distinguish these effects.
  3. Minor problem: The exact baseline architectures, hyperparameter-tuning procedures, and resized ablation configurations should be provided
Questions for authors

See above

6 comments
AuthorsOpenReview ↗

More comparison with prior work is needed

We thank the reviewer for their time and feedback. Table R1 shows direct comparisons with CAM and RS-INR at approximately matched Trainable params and identical Image size. We trained every model in Table R1 for 2,000 steps on Kodak-24, each with its official code and optimiser settings. The comparisons use two model sizes: the prior method's original size (CAM's default; RS-INR's paper size) and our image-fitting budget of about 126K parameters. At both budgets, we used native images (768×512) and each paper's image size: 512×512 centre crops for CAM, or these crops resized to 256×256 with RS-INR's resize.

Table R1: Comparisons with CAM and RS-INR.

CAM

ModelImage sizeTrainable paramsPSNR (dB)Inference TFLOPsTraining TFLOPsTrain time / step (ms)Inference time / image (ms)
CAM768×512269,82731.520.2071,24185.945.0
SPAN-INR768×512269,84743.890.2681,608115.134.8
SPAN-INR, frozen WW768×512261,56643.760.2681,477107.634.8
CAM768×512126,61930.700.09456664.135.1
SPAN-INR768×512125,87138.900.10563265.721.0
SPAN-INR, frozen WW768×512124,71538.800.10561462.821.0
CAM512×512269,82733.150.13882757.330.0
SPAN-INR512×512269,84746.930.1791,07276.422.7
SPAN-INR, frozen WW512×512261,56646.900.17998571.522.8
CAM512×512126,61932.100.06337842.623.3
SPAN-INR512×512125,87141.480.07042144.513.9
SPAN-INR, frozen WW512×512124,71541.350.07040942.513.9

RS-INR

ModelImage sizeTrainable paramsPSNR (dB)Inference TFLOPsTraining TFLOPsTrain time / step (ms)Inference time / image (ms)
RS-INR768×512593,66438.272.32213,934419.5154.1
SPAN-INR768×512593,51148.130.5973,583225.665.7
SPAN-INR, frozen WW768×512574,46747.790.5973,284207.266.1
RS-INR768×512125,66429.150.4892,932127.444.6
SPAN-INR768×512125,87138.900.10563265.721.0
SPAN-INR, frozen WW768×512124,71538.800.10561462.821.0
RS-INR256×256593,66458.190.3872,32273.026.7
SPAN-INR256×256593,51165.000.10059738.911.1
SPAN-INR, frozen WW256×256574,46766.450.10054735.711.1
RS-INR256×256125,66438.520.08148922.27.5
SPAN-INR256×256125,87152.410.01810522.03.6
SPAN-INR, frozen WW256×256124,71552.450.01810223.63.7

All SPAN-INR models in Table R1 use K=10K=10 stages and one shared mixing matrix WW. Only the feature width CC and controller width hh change with the budget: (C,h)=(34,81)(C,h)=(34,81), (91,93)(91,93), and (138,137)(138,137) for 125,871, 269,847, and 593,511 parameters, respectively.

Measured on one RTX 3090. In the “SPAN-INR, frozen WW” rows, WW stays at its random initialization while the spatial controllers and output layer are trained.

  • For image fitting, CAM (Lee et al., 2024) uses coordinate-dependent affine modulation with a scalar scale and shift per location read from 32×32 grids. It shares them across channels and applies them to normalised features before a ReLU. SPAN-INR predicts channel-wise, stage-specific, coordinate-conditioned gain gk(x)g_k(x) and shift sk(x)s_k(x). At every stage but the last, they enter the next sine, whose harmonic coefficients depend on the gain as ∏mJκm!(ω0,gk,j(x),Wjm)\prod_m J_{\kappa_m}\\!\bigl(\omega_0\\,g_{k,j}(x)\\,W_{jm}\bigr) (Prop. 1).

  • RS-INR (Cho et al., 2026) uses weight-tied sinusoidal recurrence, with about 99% of its parameters in one shared matrix whose weights act at every coordinate. SPAN-INR's spatial controllers allocate spectral content across locations and channels (assignment) and hold 99% of the 125,871-parameter image-fitting model; shared WW supplies dense connectivity with 0.9%. Across all four groups, the “SPAN-INR, frozen WW” rows exceed RS-INR by 8.26–13.93 dB in PSNR with fewer Trainable params. At 768×512, they remain at most 0.34 dB below SPAN-INR with trained WW. At 256×256 and the larger budget, “SPAN-INR, frozen WW” has the higher mean PSNR. This model has about three parameters per image value, with enough capacity in the controllers and output layer to fit well with WW fixed. A fixed random WW retains more diverse channel mixing than the trained WW, and the controllers fit better with it within the training budget. Keeping WW fixed also lowers training cost (“Training TFLOPs”) in every Table R1 group.

RS-INR's preprint appeared on 23 Jul 2026; its code was released on 29 Sep 2026, after our submission. SPAN-INR's particular combination of shared WW and spatial controllers achieves higher mean PSNR than CAM and RS-INR across all tested budgets and image sizes. We will cite both papers and mention them in the revised manuscript for comparison.

AuthorsOpenReview ↗

Isolating weight sharing and spatial adaptation

Tables R2 and R3 add one control for each of the two effects. Table R2 isolates weight sharing: within each pair, Untied WW and Shared WW have the same depth, the same feature width CC and about 126K parameters (setup as in W1). Untied WW uses a separate C×CC\times C mixing matrix per layer, and Shared WW uses one. The only resized component within a pair is the controller hidden width hh: the parameters that sharing frees widen the coordinate controllers.

Table R2: Untied and shared mixing matrices at the same depth, feature width and parameter budget (Kodak-24).

ModelDepthCChhMixing paramsController paramsTotal paramsPSNR (dB)
Untied WW61386114,26411,952126,63332.35
Shared WW61385319,044107,550127,01137.16
Untied WW81195113,28811,784125,43233.19
Shared WW81194714,161110,568125,08937.64
Untied WW10971494,09031,620126,00435.42
Shared WW1097479,409117,090126,79338.12

Shared WW has higher PSNR than Untied WW on all 24 images in every pair: by 4.81, 4.45 and 2.70 dB at depths 6, 8 and 10. In each pair, sharing moves 85K–99K parameters from the mixing matrices to the coordinate controllers. One shared dense WW provides channel reach, even when frozen at random initialization (Table R1).

Table R3 isolates spatial adaptation from basic gain/shift flexibility at the paper's depth (10), feature width (C=34C=34) and one shared WW, with the paper's input. Only the gain and shift change between rows: each is a learned constant per layer and channel (basic) or predicted from the coordinate by the coordinate controller (spatial).

Table R3: Basic, spatial and removed gain and shift at depth 10, C=34C=34 and one shared WW (Kodak-24).

GainShiftPSNR (dB)
BasicBasic14.83
SpatialBasic37.25
BasicSpatial36.05
SpatialNone36.68
NoneSpatial24.37
SpatialSpatial (SPAN-INR)38.90

Over basic gain and shift, making only the gain spatial adds 22.42 dB, only the shift 21.22 dB, and both 24.07 dB, on all 24 images. Removing either field entirely (None) lowers PSNR on all 24 images.

At the same depth and feature width, sharing WW raises PSNR by 2.70–4.81 dB on all 24 images; spatial adaptation through gk(x),sk(x)g_k(x),\\ s_k(x) adds 24.07 dB over basic gain and shift.

AuthorsOpenReview ↗

Baseline architectures, hyperparameter tuning and ablation configurations

Hyperparameter-tuning procedures: no hyperparameter is tuned per image or per scene; every method uses one configuration per task, and SPAN-INR uses the task-level defaults in Table 3.

Baseline architectures: the baselines use the architectures and settings of their official code or papers, at the task's parameter budget.

Resized ablation configurations: every variant in Fig. 7a has 10 layers and about 126K parameters, and uses ω0=10\omega_0=10, ωctrl=30\omega_{\mathrm{ctrl}}=30 and 2,000 iterations on Kodak-24. The table below gives the feature width CC and the controller hidden width hh of each variant; we will add these settings to Appendix B.3.

Configurations of the variants in Fig. 7a (10 layers each, Kodak-24).

VariantCChhTrainable paramsPSNR (dB)
SPAN-INR3481125,87138.90
No gain field3494125,68124.82
No shift field3494125,68137.98
No controllers353–125,67114.67
Untied WW9714126,00435.42
Frozen random WW3482127,06538.76
Zero-diagonal WW3481125,83738.89
Diagonal WW3481124,74933.09

In the revised manuscript, we will add a new appendix (Baseline Configurations) with the exact baseline architectures, hyperparameter-tuning procedures, and resized ablation configurations.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We appreciate the time and effort the reviewers put into reviewing our work. We are glad that the reviewers found the work interesting, and we hope it will be useful for follow-up work on sinusoidal INRs.

We responded to every point in the individual responses, and all three reviewers replied that their concerns are addressed. The main additions are summarised below.

Summary of responses

  • Comparison with prior work: CAM, RS-INR, INCODE and COSMO-INR, each run with its official code, at approximately matched parameter budgets (Tables R1, R5, R6), and a structural comparison with INCODE and COSMO-INR (Table R4).
  • Ablations: untied and shared WW at the same depth and parameter budget (Tables R2, R10); basic and spatial gain and shift at fixed depth and feature width (Table R3).
  • Computational cost: FLOPs and wall-clock training and inference times against SIREN, SASNet and FLAIR (Tables R7, R8), and PSNR against wall-clock training time (Table R9).
  • Configurations: baseline architectures and hyperparameter tuning, the configuration of every ablation variant in Fig. 7a, and the training configuration of each task.

We will include the additional experiments and implementation details in the revised manuscript and cite CAM, RS-INR, INCODE, COSMO-INR and FiLM.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.