acceptodds
Under review as a conference paper at ICLR 2027

Hybrid Diffusion Model for Sequential Data Generation

Abstract

Sequential data are ubiquitous in real-world applications, and generating realistic sequences is useful for simulation, data augmentation, and forecasting. Diffusion models have become a leading approach for this task, but standard time-domain diffusion treats the sequence as a whole and tends to capture global patterns more faithfully than local variations. Prior frequency-aware methods have added frequency information to time-domain models or moved diffusion entirely to the frequency domain; both approaches still use a single-domain generation process. We propose a Hybrid Diffusion Model that couples low-frequency generation in the time domain with high-frequency generation in the frequency domain. We decompose each sequence into low- and high-frequency components using the discrete Fourier transform and generate them with two diffusion processes that use different domains and schedules. The frequency-domain process runs on the full spectrum with a longer diffusion schedule. During joint sampling, the low-frequency component is generated by the shorter time-domain process, conditioned on the evolving high-frequency state from the frequency-domain process. On five real-world datasets, Hybrid Diffusion outperforms both single-domain diffusion baselines on performance metrics in both the time and frequency domains; on four additional datasets, it achieves the best mean performance in 10 of 12 dataset–metric comparisons against six time-series generative models.

Primary area:
generative models
Keywords:
Sequential Data Generation, Diffusion Models, Hybrid Time-Frequency 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 3 reviews, above 27% of the others.
Reviewer_r3PGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Hybrid Diffusion model, which decomposes time‑series data into low‑frequency and high‑frequency components via the discrete Fourier transform. Two diffusion branches with different schedules are adopted: the frequency‑domain branch uses the full total timestep T for high‑frequency component generation, while the time‑domain branch employs a shorter schedule t* to model low‑frequency components. During sampling, the frequency‑domain denoising runs independently first. After reaching timestep t*, the time‑domain denoiser is conditioned on the current high‑frequency state through cross‑attention, and the two branches perform joint iterations. Finally, the complete time series is reconstructed via inverse Fourier transform.

Critical strengths

S1. This method separately models low‑frequency trends and high‑frequency details. Cross‑attention is utilized for information interaction and constraints between high‑ and low‑frequency components, combined with an asymmetric diffusion schedule. It preserves the time‑domain capability for global trend modeling while leveraging frequency‑domain diffusion to recover fine‑grained high‑frequency details.

S2. Mechanism validation is first conducted on synthetic Gaussian‑process data. Two groups of experiments are carried out on real‑world datasets, with comparisons against single‑domain diffusion baselines and six mainstream time‑series generation models.

Critical weaknesses

W1. Ablation and sensitivity experiments show that model performance is highly dependent on the switching timestep t* (the breakpoint for standalone frequency‑domain denoising) and the frequency cutoff f_c. No automatic strategy is provided to determine these two hyper‑parameters, and grid‑search tuning is required. The optimal t* and f_c vary across datasets, increasing the tuning overhead for practical deployment.

W2. All experiments in this paper focus on unconditional generation. The effectiveness of this hybrid diffusion framework is not evaluated on conditional time‑series tasks such as time‑series forecasting and missing‑value imputation. A large number of real‑world time‑series applications require conditional modeling, yet the transferability of this architecture to conditional scenarios lacks experimental evidence.

W3. Built upon discrete Fourier transform (DFT) decomposition, the method assumes fixed‑length and uniformly‑sampled time series. It cannot be directly applied to variable‑length or non‑uniformly‑sampled time series, which necessitates preprocessing for length alignment. The paper does not discuss how to adapt the hybrid diffusion framework for non‑equispaced and variable‑length time‑series inputs.

Questions for authors

Q1. Since model performance is sensitive to the switching timestep t* and frequency cutoff f_c, could an adaptive mechanism be designed to automatically learn these hyper‑parameters and reduce dataset‑specific tuning efforts?

Q2. All experiments in this work target unconditional time‑series generation. What architectural modifications would be required if Hybrid Diffusion is extended to conditional time‑series tasks, e.g., time‑series forecasting and missing‑value imputation?

Q3. Beyond existing baselines including Diffusion‑TS and TimeGAN, comparisons against additional state‑of‑the‑art baselines are expected.

4 comments
AuthorsOpenReview ↗

Response to Reviewer r3PG (Part 1/2)

Thank you for your time in reviewing our paper, and for your recognition of our design that separately models low-frequency trends and high-frequency details with cross-attention and an asymmetric schedule, as well as our mechanism validation on Gaussian processes and comparisons on real-world datasets. We now summarize the weaknesses and questions you raised and address them as follows:

Q1. Could an adaptive mechanism learn t∗t^* and fcf_c automatically to reduce dataset-specific tuning?

A1. Thank you for the suggestion. While one fixed setting already achieves strong performance across all nine datasets, learning t∗t^* and fcf_c adaptively is a promising direction for further study. Our current results provide a practical starting point:

  1. One setting is used for all nine datasets. All main results use the same t∗=T/2t^* = T/2 and fc=⌊N/4⌋+1f_c = \lfloor N/4 \rfloor + 1 (Tables 6 and 7), with no dataset-specific search over either value.
  2. Limited sensitivity to the cutoff at intermediate switching times (Appendix E). For t∗/T∈{0.25,0.5,0.75}t^*/T \in \{0.25, 0.5, 0.75\}, the cutoff has little effect: at t∗=0.5Tt^* = 0.5T, the spread across five cutoffs is 6% on NASDAQ-2019 and 3% on NASA-Discharge. Across the two datasets studied in Appendix E, the best t∗t^* among the evaluated settings is always 0.5T0.5T or 0.75T0.75T. The symmetric setting t∗=Tt^* = T, where the high-frequency condition is still pure noise when joint denoising starts, yields the highest SW distance for every cutoff on both datasets.
  3. Practical guideline. Our results suggest t∗=0.5Tt^* = 0.5T as a practical default, with 0.75T0.75T as another candidate to evaluate. A possible data-dependent extension is to choose fcf_c according to the spectral energy of the data, e.g. the frequency below which a fixed fraction of the training energy lies.

Developing an effective adaptive mechanism may require further study of how data-specific spectral structure should determine the cutoff and switching time. We view this as a promising extension of the current framework and leave it for future work.

Q2. What architectural modifications are needed to extend Hybrid Diffusion to conditional tasks such as forecasting and imputation?

A2. No architectural modification is required; only the sampling procedure changes. Forecasting and imputation both mean generating a sequence that agrees with some observed time steps. Diffusion Posterior Sampling (DPS) [1] does this with an unconditionally trained diffusion model: at each denoising step, it nudges the sample towards the observations using the gradient of the mismatch between them. For Hybrid Diffusion, this guidance can follow the two-stage sampler. In the frequency-only phase, it can use the clean-sequence estimate reconstructed from the full-spectrum frequency branch. During joint denoising, it can use the sum of the low-pass-filtered time-domain estimate and the high-frequency reconstruction from the frequency branch. This reconstruction and the observation mask are differentiable, so the guidance gradient can update both states during this phase. We will add this discussion to the revised version.

AuthorsOpenReview ↗

ponse to Reviewer r3PG (Part 2/2)

Q3. Comparisons with additional state-of-the-art baselines are expected.

A3. We have added PaCoDi [4], a recent frequency-domain diffusion model (KDD 2026). Hybrid Diffusion achieves lower mean scores on all three metrics on both ETTh and fMRI, using fewer parameters.

We trained PaCoDi with its official implementation, adjusting the model dimensions to obtain a size comparable to Hybrid Diffusion and retaining the default settings for the remaining hyperparameters. We used the same data as Table 3 (sequence length 24) and the same evaluation code. For the Table 3 configurations, PaCoDi has 0.54M parameters on ETTh and 2.15M on fMRI, while Hybrid Diffusion has 0.45M and 1.81M, respectively. The comparison is shown below:

DatasetMethodContext-FID ↓Correlational ↓Discriminative ↓
ETThPaCoDi [4]0.074 ± .0050.074 ± .0070.025 ± .015
Hybrid0.036 ± .0020.039 ± .0080.018 ± .009
fMRIPaCoDi [4]0.118 ± .0040.988 ± .0210.331 ± .005
Hybrid0.103 ± .0070.880 ± .0140.148 ± .013

Q4. The DFT-based method cannot directly handle variable-length or non-uniformly sampled time series.

A4. Both cases can be handled using existing data preprocessing techniques while retaining our two-branch diffusion framework. The benchmarks we follow use fixed-length, uniformly sampled sequences, as in Crabbé et al. [2], so we did not explicitly discuss these cases in the paper. For non-uniform sampling and variable-length inputs, practical extensions are as follows:

  • Non-uniform sampling: interpolate the observations onto a uniform grid so that the DFT can be applied as usual, and use a mask to down-weight the interpolated points in the time-domain loss.
  • Variable length: pad all sequences to a common length NN and mask the padded positions; the cutoff fc=⌊N/4⌋+1f_c = \lfloor N/4 \rfloor + 1 defines the frequency split relative to this common length.

We will revise the paper to discuss these approaches for handling non-uniformly sampled and variable-length inputs. Thank you for suggesting this useful addition to the discussion of our framework.

Thank you again for reading our response and for the constructive suggestions. We welcome further discussion.

References

[1] Chung et al. Diffusion Posterior Sampling for General Noisy Inverse Problems. ICLR, 2023.

[2] Crabbé, Huynh, Stanczuk, and van der Schaar. Time Series Diffusion in the Frequency Domain. ICML, 2024.

[3] Yuan and Qiao. Diffusion-TS: Interpretable Diffusion for General Time Series Generation. ICLR, 2024.

[4] Cai et al. Parallel Complex Diffusion for Scalable Time Series Generation. KDD, 2026.

Reviewer_r3PGOpenReview ↗

Reply to Authors

Thank you for your rebuttal. I will keep my current score.

Thanks for your time and support!

Reviewer_zM2ZOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
1(−1)
Read review
Paper summary

This paper introduces a hybrid diffusion model for sequential data generation through a two-stage sampling process in both the time and frequency domains, with different processing for low- and high-frequency components. The model is competitive in comparison to single-domain diffusion models and existing generative models.

Critical strengths
  • The model proposed in this paper is competitive with existing models. On 10 of the 12 metrics across the reported four datasets, Hybrid Diffusion is the best model.
Critical weaknesses
  • (1) Inconsistent use of datasets and benchmarks weakens the power of the experiments, and makes it difficult to assess the generality of the reported improvements.
  • (2) Code, details for reproduction are incomplete, and there are possible unfairness issues in experimental settings.
Questions for authors
  • (1) There has been research on jointly using time-domain and frequency-domain information for time-series data. For example, Jeong et al. (2025 CIKM doi: 10.1145/3746252.3761050) proposed a diffusion model with combined use of time and frequency domains. Although there are differences in the specific architectures, the conceptual overlap is substantial. Can the author provide more analysis and comparison with such similar models and clarify the novelty and contributions of this paper?

  • (2) There is no constraint on sθtimes_{\theta_{\text{time}}}, so unintended high-frequency components may be generated in this branch. Will unintended interference between the two denoisers be caused consequently?

  • (3) Could the author provide details on reproducibility to ensure that all models are fairly tuned, trained, and compared?

  • (4) A T-step hybrid sampler with t∗=T/2t^*=T/2 actually performs 0.5T+2×0.5T=1.5T0.5T+2\times0.5T=1.5T denoiser evaluations, 50% more than a single-domain sampler. The ablation baselines share the high-frequency denoiser, which may be unfair as the architectures, as well as the model capacity may differ. Some metrics are evaluated on training data, seriously affecting the validity of the results. Could the author give a detailed explanation of the fairness of the comparison?

  • (5) Four different dataset suites are used for four groups of tests. Although the author follows the experimental setup of Diffusion-TS (Yuan & Qiao, 2024, arXiv:2403.01742) for comparison with other existing models using 4 different datasets, 2 simulated datasets in Yuan & Qiao's settings are omitted. Can the author comprehensively and consistently conduct the experiments and report the results, or provide a reasonable explanation for the current design?

5 comments
AuthorsOpenReview ↗

Response to Reviewer zM2Z (Part 1/2)

Thank you for your time in reviewing our paper, and for your recognition that Hybrid Diffusion is competitive with existing models, achieving the best result on 10 of the 12 dataset–metric comparisons. We now summarize the weaknesses and questions you raised and address them as follows:

Q1. Provide more analysis and comparison with similar models such as Jeong et al. (CIKM 2025), and clarify the novelty and contributions.

A1. The key difference is that Hybrid Diffusion runs the diffusion SDE itself in both the time and frequency domains, whereas existing methods all run the diffusion in a single domain. For example, Jeong et al. [1] condition a time-domain diffusion process on frequency features, so every frequency component is still generated in the time domain. Other frequency-aware methods are also single-domain: Diffusion-TS [2] uses a Fourier loss with time-domain diffusion; FIDE [3] amplifies high-frequency components and transforms the signal back to the time domain before diffusion; Crabbé et al. [4] perform diffusion entirely in the frequency domain.

In ours, the high-frequency components are diffused and generated by a frequency-domain SDE, and the low-frequency components by a time-domain SDE, so each part is learned in the domain where it is represented best (Figure 1, Appendix B). We then combine the high- and low-frequency components effectively through cross-attention conditioning and an asymmetric schedule (Table 8), using the complementary strengths of both domains.

Q2. Will unconstrained high-frequency components in the time-domain branch cause interference between the two denoisers?

A2. No, the two denoisers do not interfere, for two reasons. First, the frequency-domain denoiser is unconditional and never receives the time-domain output, so the time-domain branch cannot affect it. Second, even if the time-domain branch generates some high-frequency content, it is not combined into the final sample: we low-pass filter X0(l)X^{(l)}_0 before combining, so the high-frequency components come exclusively from the frequency-domain branch. We will add this step explicitly to Algorithm 1.

Q3. Could you provide details on reproducibility to ensure that all models are fairly tuned, trained and compared?

A3. Yes. Hybrid Diffusion uses the same datasets, preprocessing and evaluation code as Crabbé et al. [4] for Table 2 and Diffusion-TS [2] for Table 3, and its method-specific hyperparameters are not tuned per dataset: t∗=T/2t^* = T/2 and fc=⌊N/4⌋+1f_c = \lfloor N/4 \rfloor + 1 are fixed across all nine datasets. We tuned the batch size, the learning rate, dmodeld_{\rm model} and nheadn_{\rm head} to fit each dataset. The full configuration of the Hybrid Diffusion model is listed in Appendix C (Tables 6 and 7), and the code will be released soon. The baseline results in Tables 2 and 3 are taken from these two papers (Table 2 and Table 1, respectively), using the results reported by the original authors under their optimized configurations.

AuthorsOpenReview ↗

Response to Reviewer zM2Z (Part 2/2)

Q4. Is the comparison fair, given 50% more denoiser evaluations, the shared high-frequency denoiser in the ablation, and metrics evaluated on training data?

A4. We address computational efficiency, the controlled ablation design, and held-out performance in turn.

  1. Extra denoiser evaluations. With t∗=T/2t^* = T/2, Hybrid Diffusion uses 1.5T1.5T evaluations. For the Section 5.1 configuration, its two denoisers have 1.42M parameters in total, compared with 3.21M for either single-domain model, and achieve better generation results (Appendix C, Table 4; Section 5.1, Table 2). Running the two denoisers concurrently on two GPUs gives shorter training and sampling times than either single-domain baseline on one GPU: sampling takes 137.4 s vs. 143.8 s / 144.0 s, and training takes 26.6 min vs. 30.8 / 30.7 min.
  2. Shared high-frequency denoiser. The frequency-domain training objective is unchanged across A1–A4. Reusing the same denoiser holds the high-frequency model fixed while isolating the effects of cross-conditioning and scheduling. From A3 to A4, adding cross-attention lowers the frequency-domain SW distance by 14%, 63% and 26% on NASDAQ-2019, NASA-Charge and NASA-Discharge (Table 8). These improvements are achieved with an additional 0.11M parameters, an 8% increase in model size.
  3. Evaluation on training data. This is the protocol of both benchmarks (Crabbé et al. [4]; Diffusion-TS [2]) and is the same for all methods. On the held-out test set of NASA-Charge, Hybrid Diffusion remains better than both single-domain models (SW distance):
NASA-ChargeTest (freq.)Train (freq.)Test (time)Train (time)
Time-domain0.462 ± 0.0130.270 ± 0.0060.542 ± 0.0140.316 ± 0.008
Frequency-domain0.194 ± 0.0070.211 ± 0.0080.202 ± 0.0070.229 ± 0.008
Hybrid0.109 ± 0.0020.181 ± 0.0070.134 ± 0.0030.192 ± 0.007

Q5. Four dataset suites are used and the two simulated datasets of Diffusion-TS are omitted; conduct the experiments consistently or explain the current design.

A5. The experimental groups address complementary questions: spectral motivation, comparison with single-domain models, component contributions, and performance against established generators. We use Gaussian-process data to examine the spectral motivation and focus the evaluation of Hybrid Diffusion on real-world datasets with complex high-frequency structure. This choice accounts for the omission of Sines and MuJoCo from the Diffusion-TS benchmark. In total, Hybrid Diffusion is evaluated on nine real-world datasets from six domains.

To connect the two benchmark suites directly, we have additionally evaluated the time-only and frequency-only baselines on Stocks using the same three metrics as Table 3:

StocksContext-FID ↓Correlational ↓Discriminative ↓
Time-domain0.015 ± .0010.011 ± .0060.023 ± .009
Frequency-domain0.056 ± .0110.010 ± .0020.019 ± .026
Hybrid0.043 ± .0040.009 ± .0060.009 ± .007

Hybrid achieves the lowest mean Correlational and Discriminative Scores among the three variants, while time-domain diffusion achieves the lowest Context-FID. These results provide a shared-dataset comparison of the hybrid and single-domain architectures within the generative-model benchmark.

Thank you again for reading our response. We hope the additional controlled experiments address your concerns about fairness, and we welcome further discussion.

References

[1] Jeong, Sohn, Jeon, and Suk. Frequency-Conditioned Diffusion Models for Time Series Generation. CIKM, 2025.

[2] Yuan and Qiao. Diffusion-TS: Interpretable Diffusion for General Time Series Generation. ICLR, 2024.

[3] Galib, Tan, and Luo. FIDE: Frequency-Inflated Conditional Diffusion Model for Extreme-Aware Time Series Generation. NeurIPS, 2024.

[4] Crabbé, Huynh, Stanczuk, and van der Schaar. Time Series Diffusion in the Frequency Domain. ICML, 2024.

Reviewer_zM2ZOpenReview ↗

Reply to Rebuttal

Thank you for the detailed response.

My primary concern on the inconsistent use of datasets across experiments remains in your current submission. The ablation study (Figure 4) only covers 3 datasets, and the hyperparameter sensitivity analysis in Appendix E uses only 2 datasets. In my opinion, experiment in Appendix E is quite important for verifying the effectiveness of the frequency split and hybrid generation process, and I'm interested in such figure on all the datasets you've mentioned, as this can be a powerful proof of your contribution (both for frequency split and the two-stage hybrid process) with sensitivity analysis. Current inconsistent use of datasets in these two experiments makes the empirical support for the contributions less convincing.

Taken together with the limited novelty, I decide to maintain my rating.

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

This paper proposes Hybrid Diffusion, a time-frequency diffusion framework for sequential data generation. To address the limitations of time-domain diffusion in capturing high-frequency details, the method decomposes time series into low- and high-frequency components and models them using separate diffusion processes. A frequency-domain diffusion model learns the full spectrum, while a shorter time-domain diffusion process generates low-frequency components conditioned on the evolving high-frequency states through cross-attention. The two processes are coupled through an asymmetric diffusion schedule and a two-stage sampling procedure. Experiments on nine real-world datasets demonstrate improvements over selected single-domain diffusion models and existing time-series generative baselines.

Critical strengths
  1. Clear motivation. The paper identifies the limitations of time-domain diffusion in recovering high-frequency information and provides supporting analyses.

  2. Intuitive framework. The proposed approach integrates time-domain and frequency-domain diffusion through a relatively straightforward architecture.

  3. Promising empirical results. The method achieves competitive performance across nine datasets and multiple generation quality metrics.

Critical weaknesses
  1. Limited novelty. The idea of combining time-domain and frequency-domain information for diffusion-based generation has been explored in prior work. For example, RF-Diffusion and Frequency-Conditioned Diffusion Models for Time Series Generation. Asymmetric noise schedules and multi-stage sampling have also been extensively explored in diffusion models.

  2. Questionable direction of cross-frequency conditioning design. High-frequency components guide low-frequency generation. However, low-frequency components generally represent global trends, while high-frequency components capture local variations. A coarse-to-fine generation strategy, where low-frequency information guides high-frequency generation, appears more intuitive. The paper does not sufficiently justify the proposed direction.

  3. Insufficient evaluation metrics. Although SW Distance, Context-FID, Correlational Score, and Discriminative Score are reported, they do not fully assess the claimed improvement in high-frequency generation. In particular, the overall frequency-domain SW Distance cannot directly demonstrate better high-frequency reconstruction. Additional distributional and utility metrics, such as MDD, KL Divergence, Privacy Protection Metric, and downstream predictive performance as used by Diff-MN [1], would provide a more comprehensive evaluation.

[1] Diff-MN: Diffusion Parameterized MoE-NCDE for Continuous Time Series Generation with Irregular Observations.

  1. Insufficient justification of the asymmetric schedule. The ablation results show that introducing asymmetric scheduling alone provides little improvement. Although the complete model performs better, the independent contribution and necessity of the asymmetric schedule remain unclear. Further experiments are needed to isolate its benefits and explain its interaction with cross-attention conditioning. Comparisons with simpler conditioning mechanisms would also be useful.

  2. Limited Reproducibility and Unclear Baseline Reproduction and Experimental Fairness. The code is not available, making it difficult to reproduce the results and verify the experimental details. The paper does not clearly specify whether the baseline models were independently reproduced or whether their results were taken from prior publications. Evidence suggests that previously published results may have been reused without explicit clarification. The authors should clearly state the source of each baseline result and ensure consistent experimental settings, including data preprocessing, training configurations, and evaluation protocols.

  3. Incomplete Baseline Comparisons. Tables 2 and 3 use different datasets and baselines. Table 2 lacks comparisons with existing time-series generative methods, while Table 3 does not include time-only, frequency-only, or ablation results. This makes it difficult to determine whether the improvements over existing methods actually come from the proposed hybrid architecture. More comprehensive comparisons on shared datasets are needed.

Questions for authors

Please refer to the Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer MRVN (Part 1/2)

Thank you for your time in reviewing our paper, and for your recognition of the clear motivation of our work, the intuitive design of the framework, and the promising empirical results across nine datasets. We now summarize the weaknesses and questions you raised and address them as follows:

Q1. Clarify the novelty over prior work that combines time- and frequency-domain information (e.g. RF-Diffusion, Frequency-Conditioned Diffusion), asymmetric noise schedules and multi-stage sampling.

A1. To our knowledge, Hybrid Diffusion is the first method to run the diffusion SDE itself in both the time and frequency domains. Existing methods all run the diffusion in a single domain, e.g. RF-Diffusion [1], a single diffusion process whose forward step combines time-domain noise with a frequency-domain blur, and Frequency-Conditioned Diffusion [2], a time-domain process with frequency conditioning.

Our core idea is to diffuse each component in the domain where it is learned best: the frequency-domain model reduces high-frequency spectral error, while the time-domain model better matches temporal covariance (Appendix B). The asymmetric schedule and two-stage sampling coordinate these two diffusion processes and support effective cross-frequency conditioning.

We will add RF-Diffusion and this comparison to Section 2.

Q2. Why do high-frequency components guide low-frequency generation, rather than a coarse-to-fine strategy?

A2. We justify high-to-low conditioning from two perspectives: the complementary strengths of the two models and a controlled comparison of conditioning directions.

  1. Complementary modeling strengths. Appendix B shows that frequency-domain diffusion captures high-frequency structure well, while time-domain diffusion better matches temporal covariance. High-to-low conditioning uses these complementary strengths: the frequency-domain branch supplies high-frequency information to the time-domain branch as it generates the low-frequency component, helping coordinate the two components during generation.
  2. Empirical comparison. We implemented a low-to-high variant, where the frequency branch is conditioned on the low-frequency state. Under identical settings, both variants generate each component in the same domain, so the difference comes only from the direction. On NASDAQ-2019, reversing the direction increases the SW distance by about 20% in both domains:
NASDAQ-2019Frequency-domain SWTime-domain SW
High→Low (ours)41.341 ± 1.90539.606 ± 1.853
Low→High49.817 ± 2.32947.462 ± 2.257

Q3. The overall SW distance does not directly show better high-frequency generation; additional metrics as used by Diff-MN are suggested.

A3. We conduct additional experiments to directly evaluate high-frequency generation and assess distributional fidelity using the MDD and KL metrics from Diff-MN [3]. Hybrid is the best or on par in all of them.

  1. High-frequency generation. We split each sample at fcf_c and compute the SW distance on the high-frequency components alone. Hybrid recovers the high-frequency components better than both single-domain baselines, and much better than time-domain diffusion.
  2. Additional metrics. To further evaluate how well the generated distribution matches the real one, we compute MDD and KL divergence with the implementation of Diff-MN. Hybrid achieves the lowest KL divergence, with an MDD on par with the frequency-domain baseline.
NASDAQ-2019High-freq. SW (freq.)High-freq. SW (time)MDD ↓KL ↓
Time-domain [4]4.930 ± 0.0264.898 ± 0.0280.00670.0407
Frequency-domain [4]0.834 ± 0.0270.840 ± 0.0270.00160.0221
Hybrid0.718 ± 0.0250.722 ± 0.0260.00160.0067

We will further cite Diff-MN [3] and include the additional results in the revised version.

AuthorsOpenReview ↗

Response to Reviewer MRVN (Part 2/2)

Q4. Is the asymmetric schedule necessary, given that it alone brings little improvement in the ablation?

A4. The asymmetric schedule improves generation quality when combined with cross-conditioning. Table 8 shows consistent improvements from adding the schedule with conditioning enabled, whereas adding it without conditioning provides no consistent improvement.

Table 8 separates these two cases:

  1. Without cross-conditioning (A1 → A3). Adding the asymmetric schedule provides no consistent improvement: the frequency-domain SW distances change from 45.490 to 48.111 on NASDAQ-2019, 0.493 to 0.495 on NASA-Charge, and 2.636 to 2.633 on NASA-Discharge.
  2. With cross-conditioning (A2 → A4). Adding the asymmetric schedule lowers the frequency-domain SW distance on all three datasets: 64.870 → 41.341, 0.192 → 0.181, and 2.087 → 1.955, respectively. The 36% reduction on NASDAQ-2019 is particularly pronounced. This comparison isolates the schedule's contribution to the coupled generation process (Section 5.3, Table 8).

The sensitivity grid provides further evidence for this role (Appendix E). With conditioning enabled, t∗=Tt^* = T yields the highest SW distance for every cutoff on both evaluated datasets. Moving to t∗=0.5Tt^* = 0.5T lowers the frequency-domain SW distance by 36–43% across the five cutoffs on NASDAQ-2019. These results support the asymmetric schedule as an important part of the coupled generation process.

We also compared cross-attention with adaptive layer normalization (AdaLN) on NASDAQ-2019 to evaluate a simpler conditioning mechanism. A1 and A3 use no conditioning and are shared by both comparisons. With AdaLN, adding the asymmetric schedule (A2 → A4) also lowers SW distance in both domains:

NASDAQ-2019Freq. SW ↓ (A2)Freq. SW ↓ (A4)Time SW ↓ (A2)Time SW ↓ (A4)
Cross-attention64.870 ± 3.00441.341 ± 1.90562.025 ± 2.91839.606 ± 1.853
AdaLN61.530 ± 2.84556.863 ± 2.64258.898 ± 2.76054.296 ± 2.567

The schedule improves both conditioning mechanisms, and cross-attention achieves the lowest SW distance in both domains under the asymmetric schedule. These results support both the asymmetric schedule and our choice of cross-attention in the full model.

Q5. State the source of each baseline result, ensure consistent experimental settings, and release the code.

A5. The baseline results in Tables 2 and 3 are taken from Crabbé et al. [4] (Table 2) and Diffusion-TS [5] (Table 1), respectively, and are compared under consistent settings: Hybrid Diffusion uses the same datasets, preprocessing and evaluation code as the source paper of each table. We use the baseline results reported by the original authors under their optimized configurations and will state their provenance in the captions. The configuration of the Hybrid Diffusion model is listed in Appendix C, and the code will be released soon.

Q6. Tables 2 and 3 use different datasets and baselines; comparisons on shared datasets are needed.

A6. We conduct additional experiments to compare Hybrid Diffusion with the time-only and frequency-only baselines on Stocks, using the same three metrics as Table 3:

StocksContext-FID ↓Correlational ↓Discriminative ↓
Time-domain0.015 ± .0010.011 ± .0060.023 ± .009
Frequency-domain0.056 ± .0110.010 ± .0020.019 ± .026
Hybrid0.043 ± .0040.009 ± .0060.009 ± .007

Hybrid achieves the lowest mean Correlational and Discriminative Scores among the three variants, while time-domain diffusion achieves the lowest Context-FID. This shared-dataset comparison connects the single-domain analysis to the generative-model benchmark in Table 3.

The original tables address complementary questions. Table 2 follows the benchmark of Crabbé et al. [4] to compare hybrid diffusion with diffusion in either single domain. Table 3 follows the Diffusion-TS [5] benchmark to compare Hybrid with established time-series generators. The additional Stocks results bring the single-domain baselines into the latter benchmark.

Thank you again for reading our response. We hope these clarifications and additional results address your concerns, and we welcome further discussion.

References

[1] Chi et al. RF-Diffusion: Radio Signal Generation via Time-Frequency Diffusion. MobiCom, 2024.

[2] Jeong, Sohn, Jeon, and Suk. Frequency-Conditioned Diffusion Models for Time Series Generation. CIKM, 2025.

[3] Zhang et al. Diff-MN: Diffusion Parameterized MoE-NCDE for Continuous Time Series Generation with Irregular Observations. ICML, 2026.

[4] Crabbé, Huynh, Stanczuk, and van der Schaar. Time Series Diffusion in the Frequency Domain. ICML, 2024.

[5] Yuan and Qiao. Diffusion-TS: Interpretable Diffusion for General Time Series Generation. ICLR, 2024.

Dear Reviewer,

We would like to kindly follow up on our rebuttal and additional experimental results. As the discussion period is approaching its end (in around 3 days), we would greatly appreciate any further feedback and would be happy to address any remaining concerns or questions.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.