acceptodds
Under review as a conference paper at ICLR 2027

Dynamical Implicit Neural Representations

19 views

Abstract

Implicit Neural Representations (INRs) provide a powerful continuous framework for modeling complex visual and geometric signals, yet achieving high expressivity with compact architectures remains challenging. We introduce Dynamical Implicit Neural Representations (DINR), a model-agnostic framework that induces richer function classes from existing INRs by changing both what the network learns and how the signal is represented. Rather than directly parameterizing the target function with a static neural network, DINR learns a neural vector field that determines how a latent state evolves, while the target representation is given by the resulting evolution of that state. By augmenting static INRs with latent dynamics, this formulation allows compact neural parameterizations to induce substantially richer mappings without proportionally increasing the number of trainable parameters. We provide theoretical insights into the representational and optimization behavior of DINR through Rademacher complexity and Neural Tangent Kernel analyses. To balance representation capacity and generalization, we further introduce kinetic energy regularization that regulates the complexity of the learned dynamics. Across image representation, 3D field reconstruction, and scientific data compression, DINR consistently improves reconstruction fidelity, parameter efficiency, and robustness over conventional static INRs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Implicit Neural Representations, Latent Feature Evolution, Neural Tangent Kernel

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Ck87OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper replaces the static latent transformation in an INR with a time conditioned vector field. A kinetic-energy penalty regularizes the dynamics. The paper reports improved reconstruction and parameter efficiency across INR families and data modalities. The authors claim that these improvements come from increased expressivity and gradient diversity,

Critical strengths

It makes sense to reuse a compact transformation across integration steps to save storage I like the diverse experimental setting. The paper includes FLOP-matched and training-time-matched FFNet comparisons, recurrent-update ablations, and comparisons with weight decay.

Critical weaknesses

I feel that the paper over states its novelty. Parameterizing latent derivatives and obtaining outputs through ODE integration are already existing Neural ODE constructions. Same goes for Kinetic regularization of these dynamics. The connection is hidden in Appendix C. I agree that applying these ideas effectively to INR compression could still be a useful contribution. However, as I see, with fixed Euler steps, Eq. (5) is exactly a time-conditioned residual network with shared weights.

Some important hyperparameters seems to be missing. The paper does not provide enough information such as default integration steps, batch sizes, initialization settings, or validation procedure. These details are important.

I believe that the empirical NTK procedure is underspecified. Can you please define information such as effective rank, eigenvalue thresholds used for condition numbers, coordinate sample size etc.? Further, eigenvalue distributions alone also do not establish improved learning of spatial high-frequency modes without examining the corresponding eigenvectors or target alignment.

Questions for authors

See above weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer Ck87

I feel that the paper over states its novelty. Parameterizing latent derivatives and obtaining outputs through ODE integration are already existing Neural ODE constructions.

We acknowledge that parameterizing latent derivatives and obtaining outputs through ODE integration are established constructions in Neural ODEs. We therefore do not claim to introduce a fundamentally new architecture. Rather, our primary goal is to investigate the representational properties that continuous latent dynamics bring to implicit neural representations (INRs).

Specifically, our contribution is to establish latent dynamics as a source of representational capacity for compact INRs. By leveraging dynamical evolution, smaller network can achieve higher reconstruction fidelity than larger static counterparts, improving parameter efficiency and compression (Fig. 3), which are important objectives in implicit signal representation.

We further substantiate this contribution through extensive experiments, rigorous validation, and theoretical justification (Supported by a consensus of all three reviewers.). Rademacher complexity, which characterizes the expressivity and provides insights into generalization (Appendix A), and Neural Tangent Kernel (NTK) analysis, which examines gradient diversity and trainability (Appendix B). Both analyses are supported by our empirical results (Figs. 3 and 4).

In response to the reviewer's concern, we have revised the manuscript throughout to avoid overstating architectural novelty and clarify our intended contribution. Specifically, we have strengthened the positioning of DINR relative to Neural ODEs and related dynamical INR approaches also in the Introduction, and Method sections. We have also clarified that our contribution lies in investigating and establishing the representational role of dynamical evolution in compact implicit signal representations, rather than proposing a new dynamical architecture.

Same goes for Kinetic regularization (KE) of these dynamics.

We agree with the reviewer that KE regularization is an established technique, and we adopt rather than propose it. We have revised the manuscript to avoid implying otherwise. Our contribution lies in demonstrating its benefits for implicit signal representation. While prior work has primarily employed KE regularization in continuous normalizing flows to reduce numerical integration costs, we investigate its role in controlling latent trajectory complexity and improving generalization in INRs. This benefit is demonstrated through reconstruction from noisy or limited observations (Appendix F) and further supported by our theoretical analysis (Appendix A).

The connection is hidden in Appendix C. I agree that applying these ideas effectively to INR compression could still be a useful contribution. However, as I see, with fixed Euler steps, Eq. (5) is exactly a time-conditioned residual network with shared weights.

We appreciate the reviewer's observation. We agree that the network formulation introduced in Eq. (5) corresponds to a standard Neural ODE with fixed-step Euler discretization. We have explicitly acknowledged this connection in the revised manuscript.

However, our contribution is not the underlying dynamical architecture itself. As the reviewer notes, our contribution lies in effectively applying this established formulation to INRs, where representational capacity and parameter efficiency are central objectives. To clarify this distinction and better contextualize our work, we have added a Related Work section (Section 2) to the main manuscript and revised the Introduction, Methodology, and Conclusion to explicitly establish the connection to Neural ODEs and articulate our distinct objectives.

We substantiate this through theoretical analyses, together with extensive empirical evaluations. Specifically, DINR consistently improves the fidelity-storage trade-off across four INR backbones (Fig. 3), while a FLOP-matched static INR does not close this performance gap (Fig. 7). Furthermore, time-conditioned updates significantly outperform time-independent updates (Table 3), highlighting the importance of explicitly modeling latent feature evolution.

We hope these revisions clarify that, while DINR builds on the established Neural ODE formulation, our work investigates a distinct application and representational objective, supported by theoretical analysis and empirical evidence. Our contribution is to establish and characterize the benefits of dynamical evolution for compact implicit signal representation and compression, rather than to propose a new dynamical architecture.

AuthorsOpenReview ↗

Response to Reviewer Ck87 (Continue)

Some important hyperparameters seems to be missing. The paper does not provide enough information such as default integration steps, batch sizes, initialization settings, or validation procedure. These details are important.

We thank the reviewer for raising this point.

  • Documentation: We have consolidated all training details into a dedicated Appendix E (complementing current Section 5) to enhance clarity and reproducibility.
  • Open Source: We have linked the turbulence experiment code here https://anonymous.4open.science/r/DINR_Code-3EB2 and are cleaning up the remaining code and data to fully open-source everything.

Can you please define information such as effective rank, eigenvalue thresholds used for condition numbers, coordinate sample size etc.?

Thank you for pointing this out. We have updated Appendix F of the revised manuscript with the implementation details for our empirical Neural Tangent Kernel (NTK) evaluation pipeline. Below is a high-level overview of the NTK definition and its practical implementation on the turbulence dataset:

Parameter / MetricImplementation / ValueDescription
Coordinate Sample (NN)N=1,024N = 1{,}024 (32×3232 \times 32 uniform subgrid)Sampled from the full 1024×10241024 \times 1024 field
Kernel SizeK∈R1024×1024K \in \mathbb{R}^{1024 \times 1024}K=JJ⊤K = J J^\top, symmetrized
NormalizationTrace NormalizationK←K/(tr⁡(K)+10−12)K \leftarrow K / (\operatorname{tr}(K) + 10^{-12})
Computation PrecisionJacobian: float32, Eigensolver: float64Eigensolver via torch.linalg.eigvalsh
Effective Rank FormulaParticipation Ratioreff=(∑λi)2/(∑λi2+10−12)r_{\mathrm{eff}} = (\sum \lambda_i)^2 / (\sum \lambda_i^2 + 10^{-12})
Condition Number Thresholdϵ=10−12\epsilon = 10^{-12}κ=λmax⁡/min⁡{λi>10−12}\kappa = \lambda_{\max} / \min\{\lambda_i > 10^{-12}\}

Further, eigenvalue distributions alone also do not establish improved learning of spatial high-frequency modes without examining the corresponding eigenvectors or target alignment.

We thank the reviewer for the suggestion. We therefore provide visualizations of NTK eigenvectors in Figure 5 of the revised manuscript. Moreover, we independently measure reconstruction error in spatial-frequency bands using full-resolution predictions and targets. Specifically, we apply an orthonormal 2D FFT to the prediction residual and sum its squared amplitudes within each band, dividing by the total number of pixels. The bands are defined by radial frequency ρ\rho, measured in cycles per original image pixel: low (ρ≤0.05)(\rho\leq0.05), middle (0.05<ρ≤0.15)(0.05<\rho\leq0.15), and high (ρ>0.15)(\rho>0.15).

DatasetFrequency bandFFNet MSEDynamical FFNet MSEError reduction ↓
TurbulenceLow3.480924 × 10⁻⁴1.145660 × 10⁻⁴67.09%
TurbulenceMiddle4.405221 × 10⁻⁴1.687602 × 10⁻⁴61.69%
TurbulenceHigh2.461093 × 10⁻⁴2.308004 × 10⁻⁴6.22%
DIV2KLow7.998935 × 10⁻⁴1.306821 × 10⁻⁴83.66%
DIV2KMiddle5.548542 × 10⁻⁴6.120459 × 10⁻⁵88.97%
DIV2KHigh4.468282 × 10⁻³1.691470 × 10⁻³62.14%

The results show DINR reduces error in every band in these two examples, including the high-frequency band. In the revised manuscript, the above table is included in Table 4 and the accompanying training curves is provided in Figure 5 of the revised manuscript.

AuthorsOpenReview ↗

Review Reminder

Dear Reviewer Ck87,

Thank you for your constructive feedback and thoughtful questions. We have carefully incorporated your suggestions into our revised manuscript (with key updates highlighted in red) and addressed each of your points in detail above. We invite you to review the updated paper and our responses, and we hope you will consider re-evaluating our work in light of these revisions.

Thank you for your time and service to the community!

Best regards, The Authors

Reviewer_Uh7jOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces a plug-and-play framework for extending any INR results. This extension is done by augmenting any static INR via a recurrent formulation similar to Neural ODEs and RNNs, reaching richer and more representative complex functions with fewer parameters. This is done via a dynamic implicit neural representation that is basically a numerical ODE solver applied to the static INR; however, this is not naively applied, as it may generate irregular or oscillatory trajectories; hence, the authors propose a Kinetic Energy Regularization. Theoretical mathematical background is then used not only to enhance the architectural theory but also to show that the network is learning better mappings via NTK and Rademacher complexity. Their results show a +2 PSNR in image reconstruction for any backbone used, and also improve across dynamically complex scenes like turbulent flow and 3D fields.

Critical strengths

This paper is timely; it solves a deep problem in INR and may indicate a future path. The paper shows a clear and concise exposition of its theory, grounded in important literature (e.g. NTK and Radamacher). The results are sound and show significant improvement in highly challenging tasks. Not only that, but the paper introduces a plug-and-play framework for extending any INR results.

Additionally:

  • The paper shows good ablations
  • roughly +2 PSNR in image representation
Critical weaknesses
  • You should have said that DINR adds a +2 PSNR in all baselines in the image representation task earlier in the text, maybe even in the abstract. This is a good result, and it seems that it is underappreciated.

  • Since spectral biases are concerned. Take a look at works [1] and [2]. I found that these works improve spectral biases. Could you use them as your background networks for your tasks to see if DINR improves it even further?

SASNet is super important because it reaches 39 PSNR db on DIV2K. Which is the current SoTA, if I am not mistaken. The lack of a direct comparison does not hinder DINR because SASNet is rather new, but SASNet+DINR may actually achieve a new SoTA.

  • One important task, even shown in the original Siren paper, is surface reconstruction. Could you run experiments on that? You may use [1] or [2] for that. Since image reconstruction performs better, I assume 3d rec will also. (Be aware that I am asking for the mesh reconstruction, not the novel view synthesis, although some experiments on this would be nice too.) Maybe a plot like the one for the image reconstruction task would suffice my askings (Fig. 3)

  • Although Neural ODEs are cited, no comparison is made. Could you make such a comparison on similar tasks as yours? Maybe turbulence. If this is not possible, it's fine to deeply compare with them in the related works.

Text structure

  • I got a bit confused about Figure 3 after reading it; after some time, I understood it. However, could you improve it? I think the subtitle is not helping, and maybe a small arrow would improve it further. Maybe the title of it could be "DINR improves all INR on image reconstruction" or something similar.

  • Still on Figures. This is just a minor caveat, but I would gladly appreciate it if they had a higher DPI or if they were vector scalable (like a PDF or SVG). On my big monitor, all figures became pixelated and ugly.

I consider raising my score if these are introduced, especially the 3d reconsturction task.

  • [1] SASNet: Spatially-Adaptive Sinusoidal Networks for INRs, CVPR 2026
  • [2] Tuning the Frequencies: Robust Training for Sinusoidal Neural Networks, CVPR 2025
Questions for authors
  • Orthogonal, but have you tried implementing this for learning forecasting dynamics? Looks promising.
3 comments
AuthorsOpenReview ↗

Response to Reviewer Uh7j

Since spectral biases are concerned. Take a look at works [1] and [2]. I found that these works improve spectral biases. Could you use them as your background networks for your tasks to see if DINR improves it even further? SASNet is super important because it reaches 39 PSNR db on DIV2K. Which is the current SoTA, if I am not mistaken. The lack of a direct comparison does not hinder DINR because SASNet is rather new, but SASNet+DINR may actually achieve a new SoTA.

  • [1] SASNet: Spatially-Adaptive Sinusoidal Networks for INRs, CVPR 2026
  • [2] Tuning the Frequencies: Robust Training for Sinusoidal Neural Networks, CVPR 2025

Thank you for suggesting SASNet. We evaluated its static default configuration and our selected dynamical variant on 15 DIV2K images with three seeds per image (45 fits per model). Dynamical SASNet improves mean PSNR on all 15 images, achieving higher reconstruction fidelity with fewer parameters.

MetricSASNetDynamical SASNetImprovement
PSNR37.785 dB39.453 dB+1.668 dB
SSIM0.970250.97858+0.00833
RGB MSE(1.818\times10^{-4})(1.211\times10^{-4})33.39% lower
Total parameters596,127571,6074.11% fewer
Exported weights2.280 MiB2.186 MiB4.12% smaller

Both models were trained for 300 epochs. Results are averaged over the 45 fits, using unclipped RGB predictions, targets normalized to ([0,1]), and a PSNR data range of 1. These results demonstrate that DINR further improves the fidelity–storage trade-off of SASNet, extending its benefits to the reviewer’s suggested backbone.

One important task, even shown in the original Siren paper, is surface reconstruction. Could you run experiments on that? You may use [1] or [2] for that. Since image reconstruction performs better, I assume 3d rec will also. (Be aware that I am asking for the mesh reconstruction, not the novel view synthesis, although some experiments on this would be nice too.) Maybe a plot like the one for the image reconstruction task would suffice my askings (Fig. 3)

We thank the reviewer for this valuable suggestion. We conducted additional experiments on signed distance function (SDF) learning from 3D point clouds for surface reconstruction. We used SIREN and SASNet as the underlying backbones, following their official implementations. The experimental details and reconstructed surfaces have been added to Appendix F.8 of the revised manuscript.

Since the official implementation of SASNet does not provide code for SDF training, we followed the SDF training pipeline released with SIREN. Specifically, we retained the network size, batch size, number of training epochs, and learning rate reported in the SASNet paper, while adopting the SIREN training loss rather than the loss used in the original SASNet paper. We conducted the experiments on a surface reconstruction benchmark data.

The average reconstruction metrics across the benchmark are summarized below. We report Chamfer Distance (CD) and Hausdorff Distance (HD) for both the reconstructed surfaces (S) and point clouds (P). Detailed per-dataset results and visualizations are provided in Appendix F.8.

MethodS-CDS-HDP-CDP-HDParameters
SIREN0.04620.69340.03030.7207132,865
Dynamical SIREN0.02370.43310.00810.457295463
SASNet0.04290.46450.02680.4582136,637
Dynamical SASNet0.02880.28860.01080.2709127,693

The results demonstrate consistent improvements across all four evaluation metrics for both backbones. These findings further show that the benefits of dynamical evolution extend beyond 2D image representation to 3D surface reconstruction, improving reconstruction quality while reducing the number of network parameters.

We would like to emphasize that the primary objective of this experiment is to assess the benefits of dynamical evolution within each backbone, rather than to compare the absolute performance of SIREN and SASNet. Thus, the key comparison is between each static INR and its dynamical counterpart under the same experimental setting. We are also conducting further experiments on 3D surface reconstruction and will update the results as additional findings become available.

Thanks again for this valuable suggestion.

AuthorsOpenReview ↗

Response to Reviewer Uh7j (Continue)

Although Neural ODEs are cited, no comparison is made. Could you make such a comparison on similar tasks as yours? Maybe turbulence. If this is not possible, it's fine to deeply compare with them in the related works.

We have carefully revised Pages 1, 2, and 10 of the main paper and Pages 27–28 of Appendix C to clarify DINR’s relationship to Neural ODEs. Our existing Turbulence ablation (Table 3) includes (zk+1=zk+Δt,f(zk)z_{k+1}=z_k+\Delta t,f(z_k)), an Euler-discretized autonomous Neural ODE. With architecture, parameter count, function evaluations, and training protocol held fixed, this baseline compares time-independent dynamics with DINR’s explicitly time-conditioned formulation.

I got a bit confused about Figure 3 after reading it; after some time, I understood it. However, could you improve it? I think the subtitle is not helping, and maybe a small arrow would improve it further. Maybe the title of it could be "DINR improves all INR on image reconstruction" or something similar.

@Area Chair: We deeply appreciate the reviewer’s craftsmanship and integrity. In an era of LLM-assisted reviewing, their thoughtful attention to every detail deserves special recognition and has helped us refine our work. We have revised Figure 3’s caption and added an arrow to clarify the direction of improvement.

Still on Figures. This is just a minor caveat, but I would gladly appreciate it if they had a higher DPI or if they were vector scalable (like a PDF or SVG). On my big monitor, all figures became pixelated and ugly.

We apologize for the inconvenience caused by the low-resolution figures. We have replaced them with high-resolution versions, most exceeding 300 DPI, in the revised manuscript. Please let us know if any specific figure still appears blurry.

Orthogonal, but have you tried implementing this for learning forecasting dynamics? Looks promising.

Thank you for this promising suggestion. We have revised the conclusion to highlight it as a direction for future work.

First, I apologize for my late reply. I sincerely thank the authors for their thorough rebuttal and effort in my concerns. I appreciate the additional experiments with SASNet and 3D surface reconstruction, as both represent more challenging evaluations than standard image reconstruction. The former demonstrates improvements over an already state-of-the-art method, while the latter extends the evaluation to the more complex task of 3D surface reconstruction. The consistent improvements across these tasks provide stronger evidence of the effectiveness and generalizability of DINR.

I also appreciate the clarifications regarding Neural ODEs and the improvements to the figures. Since the authors have satisfactorily addressed essentially all my concerns, especially the challenging experiments I explicitly requested, I am increasing my rating.

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

The paper proposes DINR, a model-agnostic framework that replaces the static latent transformation of an INR with a learned, time-conditioned neural vector field: the latent state evolves as an ODE (discretized via Euler steps with shared weights), and the output is decoded from the terminal state. The authors motivate this via "dynamical lifting" of function classes, provide informal theory (higher Rademacher complexity; higher NTK rank for gradient diversity), and add a kinetic-energy regularizer to control trajectory complexity. Experiments across images (DIV2K, Kodak), scientific fields (turbulence, weather), and 3D volumes (Cryo-EM) with four backbones (FFNet, SIREN, WIRE, FINER) show improved PSNR–storage Pareto frontiers, supported by FLOP-matched controls, recurrent-design ablations, and multi-seed paired comparisons.

Critical strengths
  1. The paper anticipates obvious confounds: FLOP-matched and time-matched static baselines, a recurrent ablation isolating time-conditioning (Table 2), a weight-decay sweep to separate KE from generic regularization, and paired multi-seed/multi-image comparisons with confidence intervals. This is above-average experimental rigor.

  2. Consistent gains across backbones and modalities, supporting the model-agnostic claim; the spectral-bias analysis (power spectra, NTK effective rank/condition number) connects the theory to observable training behavior.

  3. Clear writing and motivation; the quadratic-ODE lifting example is a nice pedagogical device, and anonymized code is provided.

Critical weaknesses
  1. Limited conceptual novelty. The method is essentially a Neural ODE / weight-tied residual network inserted into INRs, and the KE regularizer is standard from the continuous normalizing flow literature (e.g., RNODE-style kinetic penalties). Related work is deferred entirely to the appendix, making it hard to assess positioning against prior ODE-based or recurrent/weight-shared INRs.

  2. Computation costs are under-reported in the main text. Training time, inference latency, and memory overhead of N-step integration vs. static baselines deserve a main-paper table, since "offset by reduced network size" is asserted more than demonstrated.

  3. Missing strong baselines. No comparison to grid/hash-based representations (e.g., Instant-NGP) or other state-of-the-art compression-oriented INRs, which are the practical competitors for the fidelity–storage claims.

Questions for authors

refer to weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer nXCb

Limited conceptual novelty. The method is essentially a Neural ODE / weight-tied residual network inserted into INRs, and the KE regularizer is standard from the continuous normalizing flow literature (e.g., RNODE-style kinetic penalties).

We acknowledge that Neural ODEs provide the mathematical machinery for continuous-depth transformations, and that DINR shares the mathematical formulation of a non-autonomous Neural ODE. We therefore do not claim to introduce a fundamentally new network architecture or dynamical operator.

Our contribution is to establish dynamical evolution as a source of representational capacity for compact, coordinate-based implicit representations, which are key objectives in implicit signal representation. Specifically, we investigate whether a compact latent vector field can enrich an INR’s representation and improve data compression. We substantiate this approach through extensive experiments, rigorous validation, and theoretical justification (Supported by a consensus of all three reviewers.).

Theoretically, we analyze DINR from two complementary angles: Rademacher complexity, to bound the capacity of the DINR hypothesis class and connect kinetic-energy regularization to generalization (App. A), and the NTK, to show how latent evolution diversifies parameter gradients relative to static INRs (App. B).

We agree that KE regularization is an established technique and adopt rather than propose it. We have revised the manuscript to clarify this distinction. Our contribution lies in demonstrating its benefits for implicit signal representation. Unlike prior work, which primarily uses KE regularization in continuous normalizing flows to reduce numerical integration costs, we investigate its role in controlling latent trajectory complexity and improving generalization in INRs. These benefits are demonstrated through reconstruction from noisy or limited observations (Appendix F) and supported by theoretical analysis (Appendix A).

Thus, DINR’s novelty lies in the representational role of dynamics and the resulting compression framework: Neural ODEs provide the mathematical machinery, while our work establishes how this machinery can be effectively and rigorously leveraged for implicit signal representation and compression. We have clarified throughout the revised manuscript that our contribution is not to propose a new model, but to establish dynamical evolution as a source of representational capacity for compact, coordinate-based implicit representations.

Related work is deferred entirely to the appendix, making it hard to assess positioning against prior ODE-based or recurrent/weight-shared INRs.

In the revised manuscript, we have included a high-level overview of related work in Section 2 that clarifies DINR's positioning, alongside additional discussion on ODE integration in the Appendix C.

While DINR adopts the established Neural ODE formulation, it differs in its application and systematic investigation for implicit signal representation and compression, where representational capacity and parameter efficiency are central objectives. Compared to the typical focus of Neural ODE research, DINR addresses a different fundamental problem and assigns a distinct representational role to dynamical evolution as follows:

PerspectiveGeneral Neural ODEDINR
Core problemContinuous-depth modeling and dynamical transformationLimited representational capacity of static INRs
Role of dynamicsTransformation of latent statesExpansion of coordinate-to-signal mapping expressivity
Primary interestDynamics, solver, stability, memory efficiencyFidelity-storage trade-off, spectral bias, trainability
Evaluation criteriaTask performance, dynamical modelingFidelity, stored size, frequency recovery, NTK geometry

In particular, unlike conventional dynamical modeling, where temporal evolution describes changes in an observed or latent system state, DINR uses an internal evolution variable to construct a coordinate-to-signal mapping. The dynamics are not intended to model the temporal evolution of the signal; rather, they serve as a representational mechanism that progressively transforms coordinate embeddings into expressive signal representations.

We hope these revisions clarify that, while DINR builds on the established Neural ODE formulation, our work investigates a distinct application and representational objective, supported by theoretical analysis and empirical evidence.

AuthorsOpenReview ↗

Response to Reviewer nXCb (Continue)

Computation costs are under-reported in the main text. Training time, inference latency, and memory overhead of N-step integration vs. static baselines deserve a main-paper table, since "offset by reduced network size" is asserted more than demonstrated.

We appreciate the reviewer’s suggestion. In the revised manuscript, we have leveraged the additional page to have a detailed computational cost table into the main text (Table 2).

Missing strong baselines. No comparison to grid/hash-based representations (e.g., Instant-NGP) or other state-of-the-art compression-oriented INRs, which are the practical competitors for the fidelity–storage claims.

We thank the reviewer for suggesting grid/hash-based and compression-oriented representations. We added turbulence-field experiments with Instant-NGP-style hash grids [1] and COIN++-style modulation architectures [2]. DINR improves reconstruction accuracy while reducing parameter count and checkpoint size in both cases.

RepresentationParametersCheckpoint sizePSNR
Instant-NGP120,183477.514 KiB38.8916 dB
Dynamical Instant-NGP111,581445.447 KiB39.1365 dB
COIN++119,809476.699 KiB21.8452 dB
Dynamical COIN++76,241304.893 KiB28.0616 dB

Both variants replace the feedforward decoder with repeated latent-state updates:
hk+1=hk+ΔτFθ(hk,τk;c),u^=Rϕ(hK;c),h_{k+1}=h_k+\Delta\tau F_\theta(h_k,\tau_k;c), \hat u=R_\phi(h_K;c), where τk\tau_k is internal feature-evolution time and cc is optional conditioning. Sharing transition weights across steps increases computational depth without adding a separate network at each step. For Dynamical Instant-NGP, we retain the multi-resolution hash encoding Hψ(x)H_\psi(x), initialize h0=Pα(Hψ(x))h_0=P_\alpha(H_\psi(x)), and evolve the latent state before readout. The hash grid preserves spatial detail while shared dynamics refine its features. For Dynamical COIN++, we retain patch-specific codes and sinusoidal modulation. Each code zpz_p conditions latent initialization, the shared transition, and readout. Decoder and transition weights are shared across patches. Training retains code adaptation followed by code fitting with the decoder frozen.

Together with the original backbone experiments, these results demonstrate the applicability of dynamical evolution across diverse representation families. More importantly, they support our central claim that dynamical structures can improve representational capacity while enhancing parameter and storage efficiency, which are fundamental objectives in implicit neural representations. This highlights the value of dynamical evolution as an effective means of improving the fidelity–storage trade-off in compact INRs.

AuthorsOpenReview ↗

Review Reminder

Dear Reviewer nXCb,

Thank you for your constructive feedback and thoughtful questions. We have systematically integrated your suggestions into our revised manuscript (with key updates highlighted in red) and provided detailed responses to each of your points above. We kindly invite you to review the updated paper and our rebuttal, and we would greatly appreciate it if you could reconsider your evaluation in light of these revisions.

Thank you for your time and service to the community.

Best regards, The Authors

Dear authors,

Thanks for your effort. I tend to keep my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Reviewers and AC,

Thank you all for your time and constructive feedback. We believe that we have addressed all the concerns raised during the review process and revised the manuscript accordingly, with the main changes highlighted in red.

Importantly, the feedback helped us further clarify and strengthen the presentation of our central contribution. In particular, the revisions focus on more clearly positioning our contribution with respect to existing literature, providing additional empirical validation, and elaborating on technical details. The main changes are summarized below:

  • We have revised the manuscript throughout to sharpen the central message of our work: establishing latent dynamics as a source of representational capacity for compact INRs, supported by theoretical analysis and extensive empirical results. We have also added a new Section 2 (Related Work) to better contextualize our approach within the existing literature and clarify its connection to Neural ODEs.

  • We have evaluated DINR against additional strong baselines, including SASNet, Instant-NGP-style hash grids, and COIN++-style modulation architectures. DINR consistently achieves better performance with fewer parameters, further supporting our main claims.

  • We have conducted experiments with SIREN and SASNet on 3D surface reconstruction (Appendix F.8). The results are again consistent with our main findings: DINR improves all four evaluation metrics while using fewer parameters.

  • We have added Fig. 4 to provide NTK eigenmode visualizations and a frequency-band error analysis. The results show that DINR reduces reconstruction error across all frequency bands (low, mid, and high) on both Turbulence and DIV2K.

  • We have added Table 2 to provide a clearer characterization of the trade-off between parameter efficiency and runtime.

  • We have added implementation details (Appendix E) and definitions of the empirical NTK (Appendix F.1). We have also replaced all figures with high-resolution versions.

We hope that these revisions address the reviewers' concerns and further clarify the contributions and empirical strengths of DINR. We sincerely appreciate your time and consideration, and would be grateful for your re-evaluation.

Best, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.