acceptodds
Desk rejected at ICLR 2027

Teaching Diffusion to Understand Wireless Physics: Helmholtz-Guided Signal Generation

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

20 views

Abstract

Generative models have recently demonstrated strong capabilities in learning complex data distributions and synthesizing signals. However, for wireless signals with real electromagnetic propagation characteristics, purely data-driven approaches can learn statistical patterns from observations but often struggle to capture the underlying propagation structures behind signal variations. Meanwhile, physics-based methods provide propagation-consistent signal modeling but typically require accurate scene information, limiting their applicability in practical wireless environments. Therefore, how to incorporate propagation principles into generative models without scene priors and enable signal generation with both statistical and propagation characteristics remains a key challenge. In this work, we propose Helm-Diffusion, a novel diffusion process guided by the Helmholtz equation that incorporates signal propagation principles into the generation trajectory. Helm-Diffusion exploits the propagation consistency implied by the Helmholtz equation to construct approximate representations of propagation structures from finite time frequency observations, and further derives static consistency for characterizing shared cross-frequency propagation structures and dynamic consistency for describing the evolution patterns of different frequency responses under changing propagation environments. Based on these propagation representations, Helm-Diffusion reformulates the stochastic generation process of conventional DDPMs by embedding propagation-structure degradation and recovery mechanisms into the progressive diffusion trajectory. This enables the generation process to incorporate propagation structure information while learning the statistical characteristics of wireless signals, achieving a synergistic integration of physics guidance and data-driven generation. Experiments on multiple wireless sensing and communication datasets demonstrate that Helm-Diffusion effectively improves generation quality over existing generative approaches while enhancing the consistency between generated signals and propagation characteristics, highlighting the effectiveness of physical laws as structured priors for generative models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Wireless Signal Generation, Diffusion Models, Physics-Informed Generative Models, Physical Consistency, Wireless Propagation Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_Y9p9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Helm-Diffusion is a conditional diffusion framework for wireless signal generation. It constructs static cross-frequency cues using leave-one-frequency-out Loewner prediction and dynamic cues using normalized local temporal spectra, motivated by a Helmholtz-based propagation model and local multipath approximations. Frequency-domain Gaussian smoothing and Gaussian noise define the forward process; a denoising network and learned physics-guided refinement predict the clean response during reverse sampling. Training combines reconstruction with matching static and dynamic consistency scores to measured responses. The paper reports evaluations on seven datasets spanning wireless sensing and uplink-conditioned downlink channel generation.

Critical strengths
  1. A coherent physics-informed generative design for an important shared problem. Wireless sensing and communications both need synthetic signals that reproduce time-frequency statistics while retaining useful propagation structure. The paper addresses this recurring problem without requiring explicit scene geometry. Its complex-valued Transformer backbone, task conditioning, and attention-based refinement are well matched to phase-sensitive, structured RF responses. The contribution is the combination of observation-derived static/dynamic representations with a modified diffusion process (Section 3, pp. 4–6). Appendices A.3 and A.5 account for the structured Gaussian covariance introduced by frequency mixing when defining initialization and reverse sampling. This is a substantive domain-adapted learning design of interest to ICLR readers working on structured generation and scientific machine learning.

  2. Broad within-domain evidence with consistent improvements and several complementary ablations. For separately trained dataset-specific models, Table 1 (p. 8) reports lower WD and MMD on all five sensing datasets and higher SNR on both FDD datasets than the listed baselines. Relative to RF-Diffusion, the reductions in the five-dataset mean WD and MMD are approximately 34.6% and 28.0%; the SNR gains are 1.84 dB on Argos and 1.73 dB on COST2100. Appendix A.6 documents five independent training and sampling runs, with means and standard deviations in Tables 4–6 (pp. 27–28). Figure 5 compares degradation processes and backbones, and Figure 7 tests the static and dynamic branches separately. Together, these results support the usefulness of the complete method across multiple wireless tasks and give the contribution more weight than a single-dataset demonstration.

Critical weaknesses
  1. The physical interpretation is less firmly established than the empirical improvement. The architecture and component ablations support an effective combination of propagation-inspired features and generative modeling. They leave some uncertainty about how much improvement specifically comes from the proposed physical information. Figure 5(a) holds the architecture fixed, but No-Physics removes the entire learned refinement module (Section 4.3, p. 9), and Figure 7 does not fully specify the corresponding changes to trainable components and consistency losses. Table 2 (p. 17) demonstrates useful held-out-frequency prediction, while Equations (143)–(145) (p. 27) match observation-level consistency scores. These findings support structured inductive biases, but do not fully distinguish their propagation-specific value from generic frequency/temporal modeling or additional refinement capacity. This limits the strength of the mechanism attribution. A matched generic refinement control would strengthen that attribution; the existing results already provide evidence for the usefulness of the combined method.

  2. Dataset-specific task definitions need clarification to assess the scope of the evaluation. Appendix A.6 (p. 22, Equation (112)) defines the localization condition as the mean receiver position. However, the cited espargos-0007 dataset has stationary transmitters and receivers and a moving passive target. The same appendix describes the HAR condition as a six-dimensional activity-related category, whereas the cited GaitSense study concerns gait-based identity recognition. The paper does not make clear how the position and category labels are mapped for these datasets. These could be terminology issues or valid subset/task choices, rather than implementation errors, but the distinction matters for interpreting the breadth and conditional validity of the results. The localization “Accuracy” endpoint in Figure 4(e) (p. 8) also needs an operational definition. The coherent design, consistent reported gains, and existing component studies support my current weak acceptance; clarification of these task definitions is important, and a substantive mismatch with the implemented experiments could change that recommendation.

Questions for authors
  1. Dataset-to-task mappings. For ESPARGOS, does the spatial condition represent the passive target position or the receiver position? Please identify the raw position field used and explain the receiver-position description accompanying Equation (112). For GaitSense, which labels and encoding define the six-dimensional categorical condition, and which participant or activity subset is used? Please provide a concise per-dataset label mapping and define the localization “Accuracy” endpoint in Figure 4(e). These answers would distinguish reporting issues from task-definition differences that could affect the recommendation.

  2. Attribution of the existing ablation results. For No-Physics and the static/dynamic branch ablations, which trainable components and loss terms are removed, and how are retraining and computational budgets matched? If a capacity-matched generic refinement control is available, how does it compare? These answers could strengthen the physical interpretation or indicate that the contribution should be framed more narrowly as an effective structured generative design.

2 comments

Weaknesses & Questions

W1 & Q2. Attribution of Physical Guidance and Ablation Studies

We thank the reviewer for the positive assessment of our overall design and experimental results, as well as for pointing out the need to further isolate the contribution of physical information. We agree that removing the Physics-Guided Module alone is insufficient to fully exclude the influence of additional neural refinement capacity. Therefore, we have added a parameter- and computational-complexity-matched generic refinement control in Appendix A.8, and clarified the component and loss modifications in each ablation.

Specifically:

  1. No-Physics: We remove only the Physics-Guided Module from the reverse generation process, while retaining the Denoise Module and the complete forward diffusion process.
  2. Static and dynamic branch ablations: We remove the static or dynamic branch together with its corresponding Static Loss or Dynamic Loss, respectively.
  3. Physical feature replacement: We replace the Loewner-based static features with conventional frequency interpolation features, and the dynamic physical features with local temporal features without physical normalization or cross-frequency alignment, together with their corresponding replacement losses. We also replace both physical branches with subcarrier amplitude differences and local temporal autocorrelation features.
  4. Generic refinement: We introduce a Generic Refinement module that retains general neural refinement capability without using physical features, while matching the trainable parameter count and computational complexity of the complete model.

All variants use identical dataset splits, training iterations, optimizers, random seeds, and diffusion steps, with matched trainable parameter counts. The results are presented in Table 1 (Sec. 4.4; Appendix A.8).

Table 1. Ablation results of the proposed physics-consistency designs and their generic alternatives.

VariantWD ↓MMD ↓σ_τ (ns)Env. Dev. ↓
Real Data––15.00.0
Helm-Diffusion0.2400.05513.03.8
w/o dynamic (Static Only)0.2600.09012.04.5
w/o static (Dynamic Only)0.2900.12017.59.4
No-Physics0.3700.18518.012.0
Static Physics + Generic Dynamic0.2520.07113.74.1
Generic Static + Dynamic Physics0.2570.08214.25.2
Generic Wireless Features0.2820.11016.26.9
Generic Refinement0.3380.15819.211.1

The results show that Generic Refinement, despite having matched parameter count and computational complexity, performs substantially worse than the complete model (WD: 0.338 vs. 0.240; MMD: 0.158 vs. 0.055). Generic wireless feature replacements also fail to match the performance of the complete model. These findings indicate that the improvements are not merely attributable to additional neural refinement capacity, but are closely associated with the proposed static and dynamic propagation features and their consistency constraints.

W2 & Q1. Dataset-to-Task Mapping and Conditional Labels

We thank the reviewer for identifying the inaccuracies and omissions in our original descriptions of dataset conditions and task definitions. We have re-examined the datasets and experimental settings and corrected the corresponding descriptions in Appendix A.7.

  1. ESPARGOS-0007: The original description incorrectly referred to the receiver position. We actually use the original Pos field associated with each CSI sample: c(i)=Pos(i)=(xi,yi,zi)\mathbf{c}^{(i)} = \mathrm{Pos}^{(i)} = (x_i, y_i, z_i). This field represents the passive target position, rather than the transmitter or receiver position. Both transmitters and receivers remain stationary in the experimental setup.

  2. GaitSense: We use the original gait recordings from all participants, without selecting a participant subset. The six-dimensional condition consists of the original acquisition metadata: User ID, Room ID, Position ID, Orientation ID, Instance ID, and Receiver ID. These fields describe the acquisition conditions rather than six activity classes.

  3. Other datasets: WiDAR and FallDar use their original activity and event labels, respectively; DICHASUS uses its original spatial position annotations; and Argos and COST2100 use uplink CSI as the condition and downlink CSI as the generation target.

We have also clarified the Accuracy metrics in Figure 4(e): for FallDar, Accuracy represents fall-detection classification accuracy; for ESPARGOS, it represents the localization success rate, defined as the proportion of test samples whose Euclidean distance between predicted and ground-truth positions is no greater than 1.0 m.

We sincerely thank the reviewer for these detailed suggestions, which have helped improve the accuracy and reproducibility of our dataset descriptions and experimental evaluations.

Thank you for the detailed response. The clarifications are helpful, but they do not substantially change my overall assessment. I will therefore maintain my original rating.

Reviewer_z31eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Helm-Diffusion, a physics-informed diffusion model for wireless signal generation. It uses the Helmholtz equation and Loewner approximation to construct static cross-frequency consistency, and uses normalized temporal spectra to model dynamic consistency. These features are used in a physics-guided denoising module, together with frequency-domain degradation in the forward process. Experiments on seven datasets show improvements over RF-Diffusion and other generative baselines.

Critical strengths
  1. The paper addresses an important problem in wireless signal generation. The use of cross-frequency and temporal propagation consistency is interesting, especially as it does not require explicit scene geometry.
  2. The experiments cover seven datasets across sensing, localization, and communication. The improvements over RF-Diffusion are consistent, and the downstream localization results provide useful evidence beyond distribution matching.
Critical weaknesses
  1. The connection between the proposed consistency measures and physical validity is not fully convincing. A low Loewner prediction error may also indicate a smooth or predictable frequency response, rather than a physically valid channel. Similarly, the Gaussian blur used in the forward process does not directly depend on the Helmholtz-derived quantities. 2 . The ablation studies do not fully isolate the benefit of the physical guidance. Removing the Physics-Guided Module also changes the model capacity. A parameter-matched baseline without physical features would help show whether the improvement really comes from the proposed physics information.
Questions for authors
  1. Can the authors explain how the proposed consistency measures distinguish physically valid signals from statistically similar but physically inconsistent signals? A controlled experiment with known propagation parameters would be helpful.
  2. Have the authors compared the Physics-Guided Module with a parameter-matched refinement module that does not use physical features? This would help clarify the contribution of the physics guidance.
1 comment

Weaknesses & Questions

W1. Distinguishing Predictability from Physical Consistency

We thank the reviewer for raising this important concern. We acknowledge that the original description was not sufficiently precise and may have blurred the distinction between predictability and physical consistency. In fact, they are not equivalent. For example, inspired by blur-based diffusion models such as Cold Diffusion, progressively applying Gaussian blur makes an image smoother and easier to predict, but also removes its original edges, textures, and spatial structures. Similarly, frequency-domain Gaussian blurring progressively mixes neighboring frequency components, making the frequency response smoother and more predictable while destroying the multipath propagation structures induced by the underlying physical environment. Therefore, increased predictability does not necessarily imply improved physical consistency.

In our framework, the forward process progressively disrupts multipath structures that encode information about the shared propagation environment. The reverse process does not simply pursue predictable or smooth signal responses; instead, it aims to recover realistic multipath response distributions while preserving the cross-frequency associations induced by the same underlying physical environment. The Gaussian blur serves as a degradation operator, while the Helmholtz-derived propagation features provide physical guidance during reverse generation.

W2 & Q2. Isolating the Contribution of Physics Guidance

We thank the reviewer for pointing out this important issue, which helps us further demonstrate the role of the Physics-Guided Module in signal generation. We agree that simply removing the Physics-Guided Module also changes the model capacity, making it difficult to isolate the contribution of physical information.

To address this issue, we introduce a parameter- and computational-complexity-matched Generic Refinement baseline. This module retains the general neural refinement capability but does not incorporate any physical features or physical-consistency constraints. We further conduct static and dynamic branch removal experiments, as well as feature replacement experiments using generic wireless features, to distinguish the contribution of physics guidance from general feature extraction. All variants use identical data splits, training configurations, and diffusion steps.

The additional results are presented in Table 1 (Sec. 4.4; Appendix A.8).

Table 1. Ablation results of physical consistency designs and generic alternatives.

VariantWD ↓MMD ↓στ (ns)Env. Dev. ↓
Helm-Diffusion0.2400.05513.03.8
w/o dynamic0.2600.09012.04.5
w/o static0.2900.12017.59.4
No-Physics0.3700.18518.012.0
Static Physics + Generic Dynamic0.2520.07113.74.1
Generic Static + Dynamic Physics0.2570.08214.25.2
Generic Wireless Features0.2820.11016.26.9
Generic Refinement0.3380.15819.211.1

The results show that even with matched parameter counts and computational complexity, the generic neural refinement module performs substantially worse than the complete Helm-Diffusion, confirming the contribution of the proposed physical guidance to signal generation.

Q1. Controlled Validation of Physical Consistency Measures

We thank the reviewer for this suggestion. We added a controlled multipath simulation with known propagation parameters in Appendix A.12 to test whether our measures distinguish signals with valid cross-frequency relationships from statistically similar but perturbed signals.

Using a model with a direct path and multiple reflections, we generated two signal sets:

  1. Ground-Truth Parameter Signals: All frequencies use predefined multipath parameters, preserving their cross-frequency correspondences.
  2. Cross-frequency Perturbed Signals: The marginal distributions of path gains and delays are preserved, but their cross-frequency correspondences are disrupted.

Table 1. Controlled multipath validation.

MetricGround-Truth Parameter SignalsCross-frequency Perturbed Signals
Static Consistency Error ↓0.094819.3025
Dynamic Consistency Error ↓0.013855.3212
Cross-band ToF Dispersion (ns) ↓3.9150.29
Environmental Deviation (ns) ↓2.0225.67

Although the marginal distributions are preserved, disrupting cross-frequency relationships markedly increases all four errors. This confirms that our measures capture cross-frequency propagation relationships, rather than marginal statistical similarity alone.

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

The paper proposes Helm-Diffusion, a physics-informed diffusion framework for wireless signal generation that extracts static cross-frequency and dynamic temporal propagation structures from finite observations using the Helmholtz equation and Loewner representations. These structures are incorporated into both the forward diffusion process and a physics-guided reverse projection module. Experiments across seven wireless datasets covering HAR, localization, and FDD channel generation show substantial improvements over RF-Diffusion and other generative baselines in WD/MMD/SNR, together with improved propagation-scale consistency and downstream localization robustness.

Critical strengths
  1. Interesting integration of physical structure with diffusion. The paper incorporates propagation structure into the diffusion trajectory itself, which is well aligned with the motivation of preserving physical structure throughout generation.
  2. Broad empirical evaluation. The evaluation spans seven datasets and multiple wireless applications, and the proposed method consistently outperforms the listed baselines on the reported generation metrics.
Critical weaknesses
  1. Insufficient evidence that the gains are specifically due to the proposed physics formulation. The method changes both the diffusion process and the network architecture/training objective, while the main comparison is against existing generative models. Although ablations are provided, a stronger controlled comparison against an equally capable RF-Diffusion style model would be important for isolating the contribution of the Helmholtz-guided components.
  2. Physical consistency evaluation is not sufficiently independent. The method is explicitly trained with static and dynamic propagation-consistency objectives, while the evaluation emphasizes closely related propagation quantities such as cross-band ToF dispersion and Environmental Deviation. This makes it harder to determine whether the method genuinely produces more physically realistic signals beyond the particular physical criteria encoded by the method.
Questions for authors
  1. Are RF-Diffusion and Helm-Diffusion matched in parameter count, training budget, diffusion steps, and conditioning information? This is particularly important given the large performance gap reported in Table 1.
  2. How much additional training and inference cost is introduced by the Loewner-based propagation processing and the Physics Guided Module, especially given that generation uses 300 reverse diffusion steps?
  3. Can the authors provide a physical-consistency evaluation that is not directly optimized by the Helm-Diffusion training objective, to better establish that the generated signals are physically realistic rather than simply optimized for the proposed consistency measures?
2 comments

Weaknesses & Questions

W1, Q1, and Q2

We thank the reviewers for pointing out that the original ablation studies did not sufficiently distinguish the contribution of physical information from that of additional network capacity, and for raising concerns regarding comparison fairness and computational overhead. To address these issues, we have conducted systematic ablation, feature replacement, and computational efficiency experiments.

First, to identify the individual contributions of the proposed physical-consistency designs, we introduce three groups of comparative experiments:

  1. Branch removal: We separately remove the static branch, the dynamic branch, and both branches, together with their corresponding consistency losses, to evaluate their individual contributions.
  2. Single-branch replacement: We replace the Loewner-based static features with conventional frequency interpolation features, and the dynamic physical features with local temporal features without physical normalization or cross-frequency alignment. The corresponding consistency losses are also replaced accordingly.
  3. Dual-branch replacement: We replace the complete physical design with generic wireless features and their corresponding consistency losses. Additionally, we introduce a generic neural refinement module with matched parameter count and computational complexity to exclude performance gains attributable to additional network capacity.

All variants use identical dataset splits, training iterations, optimizers, random seeds, and diffusion steps, with matched trainable parameter counts. The results are presented in Table 7 (Sec. 4.4; Appendix A.8).

Table 1. Ablation results of physical consistency designs and their generic alternatives.

VariantWD ↓MMD ↓στ (ns)Env. Dev. ↓
Real Data––15.00.0
Helm-Diffusion0.2400.05513.03.8
w/o dynamic0.2600.09012.04.5
w/o static0.2900.12017.59.4
No-Physics0.3700.18518.012.0
Static Physics + Generic Dynamic0.2520.07113.74.1
Generic Static + Dynamic Physics0.2570.08214.25.2
Generic Wireless Features0.2820.11016.26.9
Generic Refinement0.3380.15819.211.1

The results demonstrate that the performance improvements of Helm-Diffusion primarily originate from the proposed static and dynamic propagation-consistency designs, rather than additional network capacity or generic feature extraction (Sec. 4.4; Appendix A.8).

Second, regarding comparison fairness, we ensure that RF-Diffusion and Helm-Diffusion use identical dataset splits, training samples, conditional inputs, batch sizes, optimizers, learning rates, and training iterations. Both methods also employ 300 diffusion steps with the same Gaussian noise schedule. Their model sizes and inference efficiency are further compared in Table 10 (Appendix A.11).

Table 2. Comparison of model size and inference efficiency.

MetricRF-DiffusionHelm-Diffusion
Trainable Parameters (M)19.47419.542
FP32 Parameter Memory (MiB)74.28774.546
Single-Step Inference Time (ms)73.7499.55
Single-Step Throughput (samples/s)13.5610.05
Full Generation Time (s, 300 steps)22.09829.832
Full Generation Throughput (samples/s)0.04530.0335

Helm-Diffusion increases the trainable parameters by only 0.35% compared with RF-Diffusion, indicating comparable model sizes. Under identical testing conditions, its single-step inference takes 99.55 ms, while full generation with 300 steps takes 29.832 s, demonstrating improved generation quality and physical consistency with manageable computational overhead (Appendix A.11).

We sincerely thank the reviewers for raising these important questions. These additional experiments further strengthen our evaluation of the physical modeling contributions, comparison fairness, and computational efficiency of Helm-Diffusion.

W2 & Q3. Independent Evaluation of Physical Consistency

We thank the reviewer for this valuable suggestion. First, Helm-Diffusion not only improves physical consistency but also outperforms existing generative models in terms of WD and MMD, demonstrating improved overall statistical distribution matching.

Regarding the independent evaluation of physical consistency, we clarify that ToF Dispersion is neither directly involved in model training nor explicitly optimized during generation. Therefore, it provides an evaluation of physical consistency beyond the training objectives.

Meanwhile, we agree that ToF Dispersion alone is insufficient to fully evaluate the physical realism of generated signals. Accordingly, we further introduce four wireless channel metrics in Appendix A.10:

  1. Frequency Correlation Function (FCF): Measures the complex correlation between subcarriers within the same sample, reflecting cross-frequency channel relationships.
  2. Power Delay Profile (PDP): Compares normalized multipath delay-power distributions across subbands using Jensen-Shannon Divergence (JSD) and Wasserstein-1 distance (W1).
  3. RMS Delay Spread: Measures the temporal dispersion of multipath components to evaluate the consistency of delay-spread characteristics across frequencies.
  4. Group Delay IQR: Uses the interquartile range of group delays to evaluate cross-subcarrier phase variations and their associated delay characteristics.

Table 3. Comparison of physical propagation characteristics between real and generated wireless signals.

MetricReal TestHelm-DiffusionRF-Diffusion
FCF (0.625 MHz)0.99140.98610.9678
FCF (1.25 MHz)0.97350.95840.9371
FCF (2.5 MHz)0.92910.91480.8716
FCF (5 MHz)0.86250.84620.7723
FCF (8.125 MHz)0.83330.81750.7395
PDP-JSD0.00000.01370.0386
PDP-W1 (ns)0.009.7324.81
RMS Delay Spread (ns)181.52187.36197.91
Group Delay IQR (ns)14.9615.6116.88

None of these metrics is directly optimized during training. They provide independent evaluations of physical propagation characteristics in terms of cross-frequency correlation, multipath power distribution, delay spread, and phase variation. The results show that Helm-Diffusion consistently produces values closer to real measurements than RF-Diffusion across all evaluated metrics, further demonstrating its ability to preserve realistic propagation characteristics (Appendix A.10).

Reviewer_1tWGOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The main motivation of the paper is to build generative models for wireless signals with propagation principles embedded into them. They propose modified DDPM style Diffusion model where each diffusion step is modified to include EM propagation. This is called Helm-Diffusion.

The paper addresses two challenges (called theoretical and modelling - page 2): how to extract physically consistent structure from data and how to embed them in the generation.

As usual with works on generative modeling for wireless, one contribution is the physically consistent propagation representations with lack of full scene information, and the other one is diffusion based generation.

The main technical details start from Section 3. Section 3.1 presents Helmholtz physical consistency term, divided into dynamic and static terms. The static consistency is enforced by adjacent-frequency construction and leave-one-frequency-out prediction. Eq 7,8 define cross frequency consistency. For dynamic case, multi-scale dynamic consistency is used based on the variations of propagation at different temporal resolutions.

Section 3.2 introduces the idea of diffusion in in propagation structure space: just as vanilla diffusion adds/removes noise gradually, the Helm-Diffusion degrades the propagation structure with Gaussian blurring in frequency domain, followed by Gaussian noise addition. Section 3.3 studies the reverse diffusion process (the key is physics guided module). The training objective is given in Section 3.4.

In section 4, numerical results are presented. Baselines are mainly other pure generative models (VAE, GAN, diffusion). Overall performance is reported using quality of generated signal (WD, MMD); 4.3 looks at physical consistency. The experiments uses various real and synthetic datasets (DICHASUS, etc.) and show performance gains using the metric as the generation quality and consistency with propagation.

Critical strengths
  • The physics based consistency measures are properly derived from the foundations and are interesting.
  • The modified Diffusion in the structure space is also new.
  • It is encouraging to see that the paper uses real measurements for benchmarking. This is very important for any paper on wireless simulation.
  • The experiment section and results are well done with many useful ablations with one important weakness mentioned below.
  • In particular, ToF experiment to measure physical consistency is very well thought of.
Critical weaknesses

To summarize, many comments I put below, my main concern is the place of this work within the larger neural wireless modelling literatures. There are many prior works on adding physical consistency, so the question is what makes the current approach better or worse.

  • The paper is closely related to RadioDiff-k2k^2 (Helmholtz Equation Informed Generative Diffusion Model for Multi-Path Aware Radio Map Construction). Its goal is radio map prediciton, different from the current paper, but it is important that authors clarify what is different in the way diffusion is modified.
  • The baselines in the experiment section are focused fully on the pure generative models. The win over those is kind of expected given the new components of the paper. It is more interesting to see the comparison with other neural propagation models (at least a few representative ones).
  • Some of the other neural models use other information for example scene information. So admittedly, this is not an apple to apple comparison, however, it is a guidance for practitioners to select a given model based on the task and available information.
  • The choice of generation quality as metric is fine to start, but not necessarily representative of downstream task. The main question is this: if we have collected some data, what is the gain of using generative model over the data itself? What is the impact/gap on downstream task? For instance, what is the gap between BLER predicted by the generative model and one reported by the data. Or, report estae^{sta} on generated vs. real samples.
  • The idea of cross-frequency consistency has been discussed in some other papers:
  • Overall, it is difficult to assemble all the information about the training and inference of the model. The information is spread across sections. Also, please give losses like Lsta\mathcal L_{sta} explicitly.
  • ToF experiment only captures one aspect of consistency. There are other aspects more important in MIMO and mmWave communication setup, and that's angular consistency and spatial correlation of paths. The paper does not seem to address this.
  • The paper overlooks many important attempts in the wireless field for channel modelling and wireless signal generation. The radical separation between physics-based models and data-driven models is not accurate. There are many statistical channel models with various degrees of physical consistency employed for decades in wireless research. Examples are 3GPP channel modesl (TDL, CDL with specific ray-cluster modeling of MIMO communication, HST, WINNER, etc).
  • Also the claim of the paper on data-driven methods that they primarily focus on distribution characteristics is wrong. Many works in the field already addressed physical consistency, see the following references (there are older works on this going back to 2023, 2024, one need to track the references inside these works)
    • RadioDiff-k2k^2 - Helmholtz Equation Informed Generative Diffusion Model for Multi-Path Aware Radio Map Construction (https://arxiv.org/abs/2504.15623)
    • NeWRF: A Deep Learning Framework for Wireless Radiation Field Reconstruction and Channel Prediction, ICML24
    • RF-3DGS: Wireless Channel Modeling with Radio Radiance Field and 3D Gaussian Splatting, 2025
  • Related work sections can be improved and is quite thing now. Many representative works on generative modelling for wireless are missing. In general, I feel it can be expanded further.
  • A comment on "As emerging wireless systems continue to evolve toward higher frequency bands, wider bandwidths, larger antenna arrays, and greater intelligence": this is debatable in our age given how mmWave deployment in 5G turned out. Discussions on the new generation of wireless standards shows the development is more toward integration with other technologies (AI, sensing and perception).
Questions for authors

Apart from the above comments, few more questions below.

  • It is probably useful to comment why such signal generation matters as such? Where are these models used in practice?
  • For the modified forward and reverse diffusion trajectory with structure degradation and recovery components, what is the impact on the stationary distribution the diffusion process samples from? The SDE perspective of diffusion comes with guarantees that Langevin sampling samples from true distribution given the proper score function. What is the guarantee here? The two step degradation is quite different.
  • Is the reverse diffusion process the analytically correct process of the forward diffusion defined in Section 3.2? If physics guided module is learned separately, why the forward diffusion is needed at all? If only for learning the statistical part (score function), why the same score can be applied?
5 comments

Weaknesses

W1. Distinction from RadioDiff in Diffusion Design

We thank the reviewer for pointing out this related work. As noted, RadioDiff-k2k^2 focuses on radio map construction and relies on known environmental geometry and base station locations. In contrast, our work targets wireless signal generation from limited signal observations, without requiring complete spatial environmental parameters.

The main differences lie in how the Helmholtz equation is used and how the diffusion process is designed. RadioDiff adopts a conventional diffusion process and uses the Helmholtz equation under known environmental conditions to guide radio map generation. In contrast, Helm-Diffusion uses the propagation Green's function derived from the Helmholtz equation to characterize cross-frequency consistency in channel responses. It further incorporates physical properties into the diffusion process: the forward process progressively disrupts multipath structures and temporal evolution through frequency-domain blurring and time-domain noise, while the reverse process uses physical consistency constraints to guide signal recovery at each step. Thus, the two works differ in their task objectives, input requirements, and integration of physical principles into diffusion generation. We appreciate the reviewer for suggesting this valuable comparison.

W2. Comparison with Neural Propagation Models

We thank the reviewer for this valuable suggestion. We agree that comparisons with neural propagation models can provide a more comprehensive evaluation of our method. Therefore, we have added two representative methods, NeRF² and Deep UL2DL (FIRE), to the comparison in the 5G FDD downlink channel estimation task. The results are presented below.

Table 1. SNR comparison among different methods (dB).

MethodSNR (dB)
NeRF²20.07
Deep UL2DL15.22
RF-Diffusion26.97
Helm-Diffusion (Ours)28.81

The results show that Helm-Diffusion achieves the highest SNR among the compared methods, further demonstrating its effectiveness in wireless signal reconstruction.

We sincerely thank the reviewer for this constructive suggestion.

W3. Comparison of Different Wireless Generation Models

We thank the reviewer for highlighting the differences in environmental information requirements among different methods. A key contribution of our work is that it enables high-quality wireless signal generation using only limited wireless signal observations, without requiring prior environmental information such as scene geometry, obstacle distributions, or material properties. This makes our approach particularly suitable for wireless sensing scenarios where environmental information is difficult to acquire accurately, sensing targets change dynamically, or real-world measurements are limited.

We agree that these methods rely on different input information and therefore do not constitute a strictly apples-to-apples comparison. Nevertheless, the reviewer's suggestion is valuable, as it helps clarify the applicability of different wireless generation models and provides guidance for practitioners to select appropriate methods based on available information and task requirements. We thank the reviewer for this valuable suggestion. First, we would like to clarify that our work focuses on CSI signal generation rather than payload generation. Therefore, BLER is not a direct metric for evaluating our generation task. Nevertheless, the reviewer's suggestion provides a new perspective for application-level validation. We agree that generated CSI should have the potential to support practical communication links. Accordingly, we have supplemented BLER comparisons under different noise levels and modulation schemes in Appendix A.9 to validate the applicability of generated CSI in communication links. The results are presented in Table 8. Furthermore, for the wireless sensing scenarios that are the primary focus of this work, we have already presented downstream task evaluations on localization and HAR in Section 4.3, Figure 4(e), further demonstrating the practical value of our model.

Table 2. Comparison of uncoded BLER using real FDD CSI and generated CSI.

SNR (dB)16-QAM RealRF-Diff.Helm-Diff.QPSK RealRF-Diff.Helm-Diff.
599.85%100.00%100.00%89.07%90.93%89.00%
1099.35%99.50%99.35%13.42%24.50%16.28%
1549.57%67.85%52.35%0.07%0.28%0.10%
203.35%6.37%4.25%0.00%0.00%0.00%

Therefore, these two additional experiments, motivated by the reviewer's suggestion, further demonstrate that generated data can benefit downstream sensing tasks while maintaining communication-link reliability comparable to that of real signals.

We sincerely thank the reviewer for this valuable suggestion, which further strengthens our work and demonstrates the practical value of generated signals for both communication and sensing applications.

W4. Related Work on Wireless Frequency Relationships

We thank the reviewer for providing these references. We note that both papers consider frequency relationships in wireless signals, but differ substantially from our work. As the reviewer pointed out, we model the Physical Consistency of the underlying propagation environment governed by the Helmholtz equation, whereas these two papers primarily learn the Statistical Correlation of wireless signals through different approaches.

Therefore, although both works consider structural relationships in wireless signals, they do not directly address the question studied in our work: whether generated signals at different frequencies can maintain consistency with a shared propagation environment.

Nevertheless, we sincerely thank the reviewer for these valuable references. They help us further clarify the connections and distinctions among data-driven cross-frequency correlation learning, cross-modal association modeling, and explicit cross-frequency propagation-consistency constraints, allowing us to better articulate the motivation and novelty of our work. We have added discussions of both papers to the Related Work section of the revised manuscript.

W5. Clarification of Training, Inference, and Loss Functions

We thank the reviewer for this valuable suggestion, which helps us present the model's operation more clearly and systematically. Accordingly, we have added two algorithm boxes in Appendix A.3: Algorithm 1 summarizes the complete training procedure, while Algorithm 2 presents the inference and sampling procedure. In addition, we explicitly present the overall training objective and loss functions in Section 3.4, with their detailed formulations provided in Appendix A.6. We sincerely thank the reviewer for this suggestion, which helps improve the clarity and reproducibility of our method.

W6. Cross-Subcarrier AOA Consistency Evaluation

We thank the reviewer for this suggestion, which provides a new perspective for further demonstrating the advantages of our method in physical consistency. Accordingly, we further supplement the evaluation of cross-subcarrier angle-of-arrival (AOA) consistency from three aspects: angular dispersion, error relative to the real AOA, and angular variation trends.

Table 3. Comparison of cross-subcarrier AOA consistency.

MetricReal CSIRF-DiffusionHelm-Diffusion
AOA circular standard deviation (°)11.37814.28711.175
Mean median absolute error of AOA (°)—1.5110.619
AOA RMSE (°)—10.3407.044
AOA curve correlation coefficient—0.90180.9505

W7. Positioning Relative to Classical Statistical Channel Models

We thank the reviewer for pointing out this important issue. Classical models such as 3GPP TDL/CDL, HST, and WINNER characterize wireless channel properties through parameters such as multipath delays, propagation angles, and Doppler shifts. However, these models typically rely on predefined propagation parameters and their statistical distributions. While they effectively characterize typical wireless communication scenarios, they may struggle to directly reproduce complex propagation characteristics in diverse and dynamic real-world environments without scenario-specific parameter calibration.

This limitation makes it challenging for conventional statistical channel models to fully preserve the details of evolving propagation environments in dynamic and diverse wireless sensing tasks, such as human activity recognition and indoor localization. In contrast, our work focuses on generating multi-frequency channels governed by a consistent physical propagation environment for dynamic and diverse wireless sensing scenarios, without requiring scenario-specific parameter calibration.

Therefore, statistical channel modeling methods and our work differ in their target scenarios, input requirements, and the problems they address. We thank the reviewer for this suggestion and will incorporate discussions of these classical channel modeling approaches into the revised manuscript to more accurately position our research motivation and contributions.

W8. Positioning Relative to Existing Physics-Informed Wireless Modeling

We thank the reviewer for this important clarification. We note that RadioDiff-k2k^2, primarily focuses on radio map generation and applies the Helmholtz equation to model spatial signal-strength distributions, rather than the wireless signal generation task considered in our work. Meanwhile, NeWRF and RF-3DGS are both ray-tracing-inspired wireless radiance field modeling approaches that learn spatial propagation characteristics of specific environments for channel reconstruction and prediction. In contrast, our work investigates signal modeling inspired by electromagnetic field theory.

Although these works all incorporate physical principles, our work differs significantly in that it targets signal generation without prior environmental information. Specifically, we exploit the property that frequency responses are jointly governed by the same physical propagation environment and incorporate this property into the diffusion generation process. This aims to address the inconsistency in the physical representation of the propagation environment across generated frequencies, particularly for downstream wireless sensing tasks such as localization.

We thank the reviewer for providing these representative works. We have added discussions of these studies and revised the corresponding descriptions in the Related Work section to more accurately clarify the connections and distinctions between our work and existing approaches.

W9. Expansion of the Related Work Section

We thank the reviewer for this valuable comment. We acknowledge that the current Related Work section does not sufficiently cover representative studies on wireless signal generation, neural propagation modeling, and physics-guided wireless modeling. We have systematically revised this section in the updated manuscript by incorporating more representative studies and further clarifying the comparisons and positioning of our work relative to existing approaches.

W10. Clarification of Wireless Technology Development Trends

We thank the reviewer for this suggestion. We acknowledge that our original characterization of wireless system development trends was not sufficiently accurate. We have revised the statement to better reflect the growing integration of wireless communication with AI, sensing, and other emerging technologies.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for the review. The authors have responded to the points you raised, and the response period closes on 18 November (AoE). I would be grateful if you could post a brief reply on whether your concerns are resolved, partly resolved, or still open, so that the authors have a fair chance to address them. If anything still needs clarifying, please ask now while they can still respond.

Regards,

AC

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

Related papers

Discussion (0)

Sign in to comment.