acceptodds
Under review as a conference paper at ICLR 2027

CoRe-INR: Coordinate Rescaling for Spectral Expansion and Convergence Improvement for Implicit Neural Representations

15 views

Abstract

Implicit Neural Representations (INRs) create continuous representations for a broad range of signals by mapping the signal coordinates and corresponding values. However, INRs have limitations in learning high-frequency components due to the spectral bias. In this research, we explore the advantage of coordinate rescaling to mitigate spectral bias. We theoretically explain that scaling would improve the spectral representation characteristics using frequency atoms with a trade-off of introducing a spectral jitter. This spectral-aligned scaling improves convergence, explained through the Neural Tangent Kernel and the novel perspective of the spectral representation of tangent features. Specifically, we observe that some signals demonstrate a delayed learning behavior while scaling reduces this barrier to reach high-frequency content, explained through the tangent features. Further, through experimental results, we argue that there is a potential to outperform a deeper network with a shallow network, with the support of scaling. Moving further, we introduce CoRe-INR that demonstrates axis-based scaling could improve representation and inverse task modeling of high-dimensional signals.CoRe-INR outperforms the state-of-the-art algorithms, notably obtaining a +1.7 dB improvement in image representation. Our groundwork provides a new opening for future research in this domain.

Primary area:
learning theory
Keywords:
Implicit Neural Representation, Coordinate Scaling, Spectral Bias, Tangent Features, Multi Branch INR

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_UiE3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a new architecture for building INRs for image representation. The method consists of having multiple parallel MLPs that each get a differently scaled input, and who’s output gets merged at the end again. The authors provide an argument using NTK to justify why they’d expect their method to actually improve the representational capability of INRs. The authors claim it surpasses state-of-the-art in various tasks including just simple image representation, super resolution, image denoising, image inpainting etc. Some of the experiments are done on small datasets, others on single samples.

Critical strengths

The authors propose this multi-branch architecture with different coordinate scaling in each branch. This is a novel idea and based on the presented experiments it warrants some investigation. Figure 2 nicely explains the idea and gives a good overview on how it is designed.

Critical weaknesses

The experiments the authors provide can easily serve as a motivation for some further experiments, but are fundamentally insufficient to support the claim of “outperforming” the baselines. I’ll go into more detail why:

Experiment 1 (the main experiment) uses the kodak image dataset, consisting of 24 images. Experiments 2-6 were performed on a single sample, each of them with slightly different parameters. To actually have convincing results, each of those should ideally include larger datasets, maybe even from heterogeneous sources, to make sure the chosen hyperparameters are not just specific to this image, expecially considering the very thin improvements that are being reported. Also, the chosen settings might not be optimal for the baseline. For this we’d need to see a pipeline to carefully optimize of the hyperparameters of all of the baselines, and the authors did not report doing that. Furthermore, from what we can gather from the paper, the proposed method uses roughly 4x (in 2D) the amount of parameters compared to the baselines. This should more prominently be mentioned and taken into account with all the comparisons and ablations.

I’d like to add a few points that the authors could possibly address:

  • The NTK argument boils down to saying that for certain values of the scaling factor, the eigenvalues increases and the convergence therefore decreases. This argument appears quite post-hoc and provides no further intuition, I’d recommend shortening it.

  • An ablation is only provided in the appendix. One experiment I’m missing there is a fair (as in equal number of parameters with a suitable architecture, same amount of memory or FLOPs used or similar notion of “fair”) comparison with the baseline architectures. The authors provided the comparison with a wider SIREN network but this is on the one hand not the optimal way to scale SIRENs, but also omits the other baselines.

  • Please use larger datasets to evaluate your proposed method on and actually report some measure of spread, especially since the reported improvements are quite marginal.

  • Please first define the expression “frequency atom” before using it.

  • A few claims refer to figure 1, but without specifying to what subfigure. Similarly, figure 3 has three subfigures, but not all of them are actually relevant. E.g., 3a is not referenced anywhere and provides no additional value.

  • The resolution of the images in figure 6 are insufficient to display the changes well. Furthermore, the poses of the displayed mesh should be identical to provide a good comparison.

  • Equation 8 is referenced a lot, but it is only listed in the appendix. Consider moving it to the main paper.

Questions for authors

The experiments use mainly scaling factors around 0.5, however, the more “theoretical” parts investigate mostly scaling factors of 10, 20 or even larger. Of course I see that a scaling of e.g. (1, 0.5) or (2, 1) results in the same network (modulo learning rate), but a factor of 2 is still quite different than a factor of 10 -100. Can you comment on why there is this mismatch?

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

This paper observes and theoretically explains that global/axial scaling of input coordinates for SIREN expands the set of available frequency atoms, raises NTK-related eigenvalues, and reduces the delay in high-frequency learning. It then proposes CoRe-INR: a model with four branches (1,1), (s_x,1), (1,s_y), (s_x,s_y) fused via spatial confidence softmax. Improvements are reported on images, audio, 3D occupancy, and several inverse problems.

Critical strengths
  1. The problem studied in this paper is practical and valuable. The spectral bias of INRs, slow convergence for high-frequency components, and insufficient frequency expansion in shallow networks are substantial issues. Coordinate scaling is a minimal intervention.
  2. The proposed method is simple and straightforward yet clear: coordinate rescaling for spectral expansion, alongside the CoRe-INR multi-branch axial scaling with confidence fusion.
  3. This paper covers a broad range of tasks, with experiments conducted on image representation, denoising, super-resolution, inpainting, 3D occupancy, and audio.
Critical weaknesses
  1. Coordinate scaling is potentially nothing more than reparameterization of the first-layer weights. More critically, there exist closely related prior works [R1].
  2. The theoretical analysis is insufficient to constitute an independent theoretical contribution. The paper claims to explain the benefits of coordinate scaling via frequency atoms and NTK. However, Appendix A.1 largely adopts the structured dictionary analysis from Yüce et al. (CVPR 2022). This prior work has already investigated INR frequency atoms and the relationship between NTK eigenvalues and signal recovery.
  3. The method exhibits poor scalability to high-dimensional problems. CoRe-INR uses 2N2^N branches for NN-dimensional coordinates.
  4. The experimental evidence is weak for most tasks. For denoising, super-resolution, inpainting, audio, and 3D occupancy, only a single sample is tested for each setting, with limited diversity in the evaluated images, audio clips and meshes.
  5. The writing quality is poor. There are missing spaces in multiple locations (e.g., Lines 25, 34), and two consecutive periods on Line 156. Section 3.2 cites Equation (8) many times, though the equation resides in the appendix. If this equation is critical, it ought to appear in the main body. Furthermore, inconsistent formatting exists for figure and equation references (mix of abbreviations and full names), and uniform styling is needed.

[R1] Fast Training of Sinusoidal Neural Fields via Scaling Initialization. ICLR 25

Questions for authors

Please refer to weaknesses.

Reviewer_NW32OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the effect of coordinate scaling on the spectral representation and convergence properties of sinusoidal implicit neural representations (INRs). The authors analyze this effect through frequency atoms, neural tangent kernels, and tangent-feature spectra. Based on these observations, they propose CoRe-INR, a multi-branch architecture combining differently scaled coordinate networks through learned confidence-based fusion. Experiments on image and audio representation, 3D occupancy, and inverse imaging tasks report improvements over existing baselines.

Critical strengths
  • Simplicity and empirical performance. The proposed approach is straightforward and demonstrates promising results in certain domains, particularly for image representation. The ablations provide some evidence for the benefits of coordinate scaling beyond increasing model capacity.

  • Analysis of optimization behavior. The investigation of spectral characteristics, tangent features, and delayed convergence provides an interesting perspective on the optimization of sinusoidal INRs.

Critical weaknesses
  • Limited novelty. The central observations overlap considerably with prior work, particularly Yeom et al. [A], which already analyzes how scaling sinusoidal network weights affects spectral characteristics and convergence, and FreSh [B], which investigates spectral alignment between INR initialization and target signals. While the proposed architecture is different, its additional benefits over these approaches are not sufficiently established.

  • Limited practical relevance. Most experiments involve fitting individual signals or small-scale image benchmarks, with marginal improvements in some settings. The paper does not convincingly demonstrate practical advantages for more realistic INR applications or establish whether the improvements justify the additional computational cost.

  • Limited theoretical insights. While the provided observations are interesting, the theory remains largely descriptive and does not provide principled guidance for selecting scaling factors or predicting convergence improvements.

[A] Yeom, T., Lee, S., and Lee, J. Fast Training of Sinusoidal Neural Fields via Scaling Initialization. ICLR, 2025.

[B] Kania, A., Mihajlovic, M., Prokudin, S., Tabor, J., and Spurek, P. FreSh: Frequency Shifting for Accelerated Neural Representation Learning. ICLR, 2025.

Questions for authors
  • How does CoRe-INR compare against appropriately tuned SIREN models, particularly those using FreSh [B] or scaling initialization [A] under matched computational budgets?

  • Could the authors provide an evaluation in a more practical INR application, such as neural scene representation or surface fitting to point clouds, and demonstrate whether the proposed approach offers a meaningful accuracy-efficiency advantage in these settings?

  • Can the theoretical analysis be used to predict suitable scaling factors or convergence improvements rather than explaining them retrospectively?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.