acceptodds
Under review as a conference paper at ICLR 2027

Wavelet Flow Matching for Time Series

Abstract

Synthetic time series are increasingly needed for data augmentation, privacy-preserving data sharing, and the evaluation of downstream learning systems, yet faithfully reproducing the multi-scale temporal structure and the cross-channel dependencies of real-world signals remains challenging. In this work we study generative modeling of multivariate time series by flow matching in the wavelet domain. Rather than learning the flow in the time domain, we learn it on the multilevel discrete wavelet coefficients of each window, a representation that places coarse trend and successively finer detail in separate levels and thus exposes the multi-scale structure directly to the model. The transform is applied to each channel independently, so it reorganizes temporal structure while leaving the dependence between channels entirely to the network. We therefore parametrize the velocity field with a transformer in which each token is a single channel, carrying its coefficients across all decomposition levels, so that attention acts directly across channels. We further derive a family of learnable orthonormal wavelets from the lattice factorization of paraunitary filter banks, with perfect reconstruction guaranteed at every training step, and find performance largely insensitive to the choice of filter, fixed or learned. We compare the resulting model against five recent generative models on seven benchmark datasets at four sequence lengths, using discriminative, predictive, correlational, and Context-FID scores. Our model is best or tied on a majority of the dataset–metric combinations at every window length, with the largest and most consistent gains on Context-FID and the discriminative score.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
time series generation, flow matching, wavelets, multivariate time series, generative modeling, multi-scale representations

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TAQrOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes unconditional multivariate time-series generation through flow matching in the wavelet domain. A channel-token transformer models the joint distribution of multilevel wavelet coefficients. The authors argue that differences in coefficient variance produce an implicit coarse-to-fine progression under a shared interpolation schedule. Experiments compare against five baselines across seven datasets and four sequence lengths, with particularly strong Context-FID and discriminative scores. Ablations examine channel identity, coefficient scaling, the wavelet representation, time sampling, and wavelet choice.

Critical strengths
  • At sequence length 24, the method achieves the lowest reported mean Context-FID on six of seven datasets and discriminative score on five. The evaluation covers multiple domains and sequence lengths, and the main comparison includes three independent training runs. This indicates strong empirical performance.

  • The paper includes a matched time-domain ablation and investigates coefficient standardization, channel identity, time sampling, and fixed versus learned wavelets. These experiments provide useful evidence about the design choices, including the finding that learning the wavelet transform does not systematically improve performance.

Critical weaknesses
  • Prior work has already explored diffusion-based time-series generation in transformed domains, including Fourier and wavelet representations. While the proposed combination with flow matching and channel-wise attention achieves strong empirical results, the additional insights into the benefits of transformed-domain generation are relatively limited. The SNR analysis provides useful motivation, but a clearer account of when and why the proposed approach is particularly beneficial would strengthen the paper’s contribution.

  • The experiments include four diffusion models and one flow-matching model, but no GAN or VAE baseline. GT-GAN [1] and KoVAE [2] would broaden this comparison. ImagenTime [3] is also particularly relevant because it applies diffusion to transformed time-series representations. Although all three are cited, their exclusion from the experiments is not explained, limiting the assessment of the method’s broader empirical advantage.

Furthermore, the comparisons at sequence lengths 32, 64, and 128 use only one generator training run; repeated metric evaluations do not capture training variability. These limitations leave some uncertainty about the breadth and robustness of the reported advantages.

[1] Jeon et al., GT-GAN: General Purpose Time Series Synthesis with Generative Adversarial Networks (2022)

[2] Naiman et al., Generative Modeling of Regular and Irregular Time Series Data via Koopman VAEs (2024)

[3] Naiman et al., Utilizing Image Transforms and Diffusion Models for Generative Modeling of Short and Long Time Series (2024)

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer TAQr - Part 1

We thank the reviewer for their positive feedback and suggestions. Below, we address each point individually.

Benefits of Wavelet-Domain Generation (Weakness 1)

We thank the reviewer for this observation and agree that a clearer understanding of when and why wavelet-domain generation is beneficial strengthens the contribution. We have therefore expanded both the analysis and discussion in the revised manuscript.

First, we clarify the role of the wavelet representation. Since the DWT is orthogonal, wavelet- and time-domain flow matching are mathematically equivalent under a change of coordinates (Remark 3.1). The wavelet transform does not introduce a different generative process, but explicitly separates temporal scales, which can facilitate optimization. This is particularly beneficial when fine-scale fluctuations are much smaller than coarse-scale trends. On Exchange, the finest wavelet level has a 70×70\times smaller standard deviation than the coarsest (Table 7). Stocks and Exchange show the largest improvements over our matched time-domain variant (2.8×2.8\times and 1.75×1.75\times in Context-FID, Table 15), and our model achieves the best Context-FID on both datasets across all evaluated sequence lengths.

Second, we directly investigated the learned sampling trajectories (new Appendix F.2, Figures 8–9). Rather than inferring coarse-to-fine generation solely from SNR crossing times, we track the model's running endpoint predictions along the sampling ODE. On six of seven datasets, coarse-scale structure emerges first, followed by progressively finer details, consistent with the predicted ordering. Equalizing the coefficient variances substantially collapses this progression, while the matched time-domain model exhibits similar trajectories. This confirms that the wavelet representation makes the progression explicit rather than creating it.

Third, we added a multi-scale fidelity analysis (Section 5.1, Appendix F.5, Tables 13–14), comparing wavelet-level energy, spectral properties, and local high-frequency structure across seven generative methods. Our model achieves the best average rank across five statistics (2.11, compared with 3.00 for WaveletDiff and 3.36 for FourierDiffusion), ranking first or second on every statistic. In particular, time-domain models tend to misrepresent fine-scale energy, whereas FourierDiffusion matches the power spectrum but less faithfully reproduces the localization of fine-scale events.

Together, these findings clarify when and why wavelet-domain generation is beneficial: it provides an explicit multiscale parametrization that appears easier to optimize, particularly for signals with strongly unequal energy across scales, and yields more faithful multi-scale structure in the generated samples.


Additional Baselines: KoVAE, GT-GAN, and ImagenTime (Weakness 2)

We thank the reviewer for this suggestion. Our original evaluation focused on diffusion and flow-matching models, as these are most closely related to our proposed approach and have shown strong performance in recent time-series generation benchmarks. Nevertheless, we agree that including other generative paradigms and transformed-domain approaches would strengthen the comparison.

Regarding KoVAE, we have now included it as a VAE-based baseline across all seven datasets and six sequence lengths (Tables 1 and 8–12). KoVAE generally performs worse than the diffusion and flow-matching baselines, particularly on Context-FID and discriminative score. Our model consistently achieves substantially better Context-FID, reinforcing the conclusions of our original comparison.

We also attempted to reproduce GT-GAN but encountered substantial training instabilities and considerably longer training times than for the other baselines, preventing a reliable comparison within our computational budget. However, KoVAE already reports outperforming GT-GAN on all four datasets in both discriminative and predictive scores (Tables 1–2 of Naiman et al., 2024). We therefore prioritized KoVAE as a representative non-diffusion baseline, while acknowledging that this does not replace a direct comparison with GT-GAN. For the final version of the paper, after rebuttal, we aim to include GT-GAN as well, for completeness.

Finally, we agree that ImagenTime, which applies diffusion to transformed time-series representations, is a relevant baseline. We are currently evaluating it across all seven datasets and sequence lengths using the same protocol, expect to complete the experiments within the next few days, and will add the results to the revised manuscript.

References

Naiman, Ilan, et al. "Generative modeling of regular and irregular time series data via Koopman VAEs." International conference on learning representations. Vol. 2024. 2024.

AuthorsOpenReview ↗

Response to Reviewer TAQr - Part 2

Multiple Independent Training Runs at Longer Sequence Lengths (Weakness 3)

We thank the reviewer for highlighting the importance of accounting for training variability. We fully agree with this concern, which is why our original submission already included three independent training runs at (T=24), each evaluated five times.

While multiple independent generator training runs across all datasets and sequence lengths are not standard practice in diffusion-based time-series generation research, we agree that extending this evaluation strengthens the robustness of our conclusions. Such an extensive multi-seed evaluation is computationally demanding, particularly for an academic research group with limited resources.

Nevertheless, we have extended our experiments to three independent training runs at T∈{32,64,128}T \in \lbrace 32,64,128 \rbrace, in addition to T=24T=24. The revised results are reported in Tables 8–10, with the corresponding updates reflected in Figures 4 and 10. The results preserve our main findings: our method maintains the best mean rank in Context-FID and discriminative score across all six evaluated sequence lengths, including the newly added T∈{512,1024}T \in \lbrace 512, 1024 \rbrace.

Together with the additional experiments at T∈{512,1024}T \in \lbrace 512, 1024 \rbrace, this expands our original benchmark from 252 to 588 planned independent model trainings (2.33×2.33\times) across the original six methods. Furthermore, incorporating KoVAE and ImagenTime requires 98 additional trainings each, bringing the expanded benchmark to 784 trainings (3.11×3.11\times the original), excluding ablations and repeated evaluations. These extensions substantially strengthen the evidence for the robustness and breadth of our empirical comparison.

AuthorsOpenReview ↗

Follow-up: ImagenTime Baseline Results

Dear Reviewer,

As a brief update to our response, we have now included the ImagenTime results in the revised manuscript, as suggested. The experiments cover all seven datasets with three independent training runs for sequence lengths T≤128T \leq 128, following the same evaluation protocol as the other baselines.

We are also completing the ImagenTime evaluations for T∈{512,1024}T \in \lbrace 512, 1024 \rbrace and plan to include these results in the camera-ready version.

We thank you again for your feedback and hope that the expanded experimental comparison addresses your remaining concerns and strengthens your overall assessment of the paper.

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

Multivariate time-series generation needs to capture multi-scale temporal patterns and dependencies between variables. The authors propose Wavelet Flow Matching, which uses wavelet transforms to decompose time series into different scales and generates data in the wavelet domain. Differences in wavelet coefficient variances naturally lead to a coarse-to-fine generation process, while a Channel-token Transformer models dependencies between variables. The generated coefficients are then converted back into time series. Experiments on seven datasets with four sequence lengths show improvements in overall generation quality, especially Context-FID, but less consistent gains in cross-channel correlations and downstream prediction. Ablation studies further show the importance of preserving natural variance differences across wavelet scales.

Critical strengths

The method is easy to understand.

Critical weaknesses

The proposed method appears to be a combination of several existing techniques, and its novelty is limited. The problem addressed in this paper is not new, as many existing studies have already explored multivariate and multi-scale time-series generation. Moreover, the paper does not provide a suitable application scenario that clearly demonstrates the advantages of the proposed method.

Questions for authors
  1. The selected baselines may not fully represent the latest state-of-the-art methods. Could the authors consider including more recent methods from 2026 that address similar time-series generation tasks to provide a more comprehensive and fair comparison?
  2. Many existing studies have already addressed multivariate and multi-scale time-series generation. Although the proposed method achieves good results on some generation metrics, the authors do not clearly explain the key limitations of existing methods. Could the authors provide a specific scenario where existing methods struggle but the proposed method performs better?
  3. Modeling cross-channel dependencies is one of the main goals of this paper. However, the proposed method achieves the best Correlational Score on only two out of seven datasets. How do the authors explain this result? Is the Channel-token Transformer really more effective than existing methods at modeling cross-channel dependencies?
  4. Could the authors visualize the generation trajectories of different wavelet scales during actual Flow Matching sampling to directly verify whether the model follows a coarse-to-fine generation process, rather than inferring it only from the SNR changes across scales?
  5. Table 2 shows that wavelet-domain and time-domain generation achieve similar performance on some datasets. For example, their Context-FID scores are 0.005 and 0.006 on ETTh1, while the time-domain model performs slightly better on MuJoCo. Given these limited improvements, have the authors evaluated the trade-off between the additional computational cost of wavelet transforms and the performance gains?
  6. The experiments only consider sequence lengths of 24, 32, 64, and 128, with 6–28 variables. Could the authors evaluate the method on longer sequences (e.g., 512 or 1024) and datasets with more variables to demonstrate its scalability?
  7. The authors claim that wavelet-domain generation better preserves multi-scale temporal structures, but the evaluation mainly uses overall generation metrics such as Context-FID and Discriminative Score. Could the authors compare the wavelet-scale energy distributions, spectral properties, and local high-frequency patterns of real and generated data to better support this claim?
4 comments
AuthorsOpenReview ↗

Response to Reviewer GQHL - Part 1

We thank the reviewer for their thoughtful feedback and suggestions. Below, we address each point individually.

Novelty and Practical Relevance (Critical Weakness)

Thank you for raising these concerns. We agree that multivariate and multi-scale time-series generation is an established research problem. However, we believe that the novelty of a contribution should be assessed not only by whether the underlying problem is new, but also by the methodological advances, insights, and empirical improvements it provides.

Our contribution goes beyond combining existing components. We investigate how the natural variance differences across wavelet scales interact with a shared flow-matching probability path, characterize the resulting scale-dependent SNR progression, and analyze its manifestation in learned sampling trajectories. We further establish the equivalence between wavelet- and time-domain formulations, clarifying that the benefit of wavelets lies in their explicit multiscale parametrization and its optimization advantages rather than a fundamentally different generative process. Combined with our channel-token architecture, this yields strong empirical performance across heterogeneous datasets and sequence lengths.

Regarding practical relevance, we have expanded Section 5.2 to identify concrete settings where our method is particularly effective, notably signals with strongly unequal energy across scales and systems with heterogeneous, coupled channels. We provide quantitative comparisons and discuss limitations in our response to Q2 below. The additional baselines (Q1), trajectory analysis (Q4), and multi-scale fidelity evaluation (Q7) further strengthen the evidence supporting these contributions.

We hope these clarifications and additional experiments address the concerns regarding novelty and practical relevance.


Recent State-of-the-Art Baselines (Q1)

Our evaluation already includes five recent and established baselines spanning different generative approaches and representations: Diffusion-TS (2024), FourierDiffusion (2024), SigDiffusion (2025), FlowTS (2025), and WaveletDiff (2026). In particular, WaveletDiff was published in TMLR in September 2026, making it a particularly recent and relevant baseline. Our method demonstrates strong overall performance against WaveletDiff across the evaluated datasets, sequence lengths, and metrics, with particularly consistent improvements in Context-FID and discriminative score.

To further strengthen the comparison, we have added KoVAE as an additional baseline and are exploring comparisons with ImagenTime, as also suggested by Reviewer TAQr. Combined with our extended evaluation across seven datasets, additional independent training runs, and longer sequence lengths, these additions provide a broader empirical comparison with recent time-series generation methods.


When Does Our Method Outperform Existing Approaches? (Q2)

Our results point to two concrete scenarios where existing methods struggle, and one where our method has no advantage.

First, signals whose fine-scale fluctuations are much smaller than their trend, such as exchange rates and stock prices. On Exchange, the finest wavelet level has a standard deviation 70×70 \times smaller than the coarsest (Table 7). A time-domain model must resolve these small fluctuations within coordinates dominated by the trend, whereas the wavelet basis places them in separate coordinates. This is where our method gains most: on Stocks and Exchange, its Context-FID is 0.0050.005 and 0.0040.004, against 0.0170.017 and 0.0090.009 for FlowTS and 0.1750.175 and 0.0470.047 for Diffusion-TS (Table 1), and the matched time-domain variant of our own model is 2.8×2.8 \times and 1.75×1.75 \times worse (Table 15), the largest gaps across all datasets. Our model also has the best Context-FID on both datasets at every window length (Tables 1 and 8–12).

Second, systems with strongly coupled, heterogeneous channels, such as the 14 joint coordinates of MuJoCo and the 28 sensor measurements of Energy. Unlike Diffusion-TS, FlowTS, and WaveletDiff, which combine channels within time- or wavelet-position tokens, our architecture represents each channel as a separate token and models their dependencies directly through attention. On MuJoCo, our method achieves the best Context-FID at every sequence length, including T=1024T=1024 (Tables 1 and 8–12, Figure 10). On Energy, our model achieves the best correlational score at every length up to T=128T=128, and again at T=1024T=1024. At T=24T=24, WaveletDiff, reaches 1.1801.180 on MuJoCo and 0.4820.482 on Energy, against 0.006 and 0.010 for ours, and removing the channel identity from our model degrades it by more than two orders of magnitude (1.3901.390 and 2.6172.617, Table 2).

Conversely, when the wavelet spectrum is nearly flat, as on EEG, there is no scale separation to exploit, and our method is not the best in Context-FID. We have added this discussion to Section 5.2 (“When is the method most effective?”).

AuthorsOpenReview ↗

Response to Reviewer GQHL - Part 2

Cross-channel Dependency Modeling and Correlational Score (Q3)

Our overarching goal is to generate realistic multivariate time series, capturing both multi-scale temporal structure and cross-channel dependencies. The channel-token transformer supports this goal rather than exclusively optimizing pairwise correlations. While the correlational score measures lag-zero pairwise dependencies, Context-FID and the discriminative score assess broader aspects of the joint multivariate distribution. Our model achieves the best mean rank on both metrics at every window length (Figure 4).

On the correlational score, our model is consistently among the best. At T=24T = 24, it ranks first on two datasets and second on four. Across T∈{24,32,64,128}T \in \lbrace 24, 32, 64, 128 \rbrace and seven datasets, our model achieves the best or tied-best score in 14 of 28 settings, compared with 8 for FlowTS and 6 for WaveletDiff. At T=32T = 32 and T=64T = 64, it is best on five of seven datasets. We acknowledge that the margins are smaller than for Context-FID and often within one standard deviation of FlowTS.

The channel-identity ablation further demonstrates the importance of modeling heterogeneous channels. Since the DWT acts independently on each channel, their dependencies must be learned by the velocity network. Removing the channel identity efe_f makes the generated distribution exchangeable, preventing channel-specific correlation patterns. This worsens the correlational score by 3.7×3.7\times to 146×146\times on six of seven datasets (Table 15, row (a)). The exception is EEG, where recording artifacts distort the metric, as discussed in detail in our response to Q1 of Reviewer 2DJx.

Overall, these results support effective cross-channel modeling, even though our largest gains are in broader measures of generation fidelity.


Visualization of Learned Coarse-to-Fine Sampling Trajectories (Q4)

We thank the reviewer for this suggestion, which we have implemented in the revised manuscript. The new Appendix F.2 ("Trajectory analysis of the learned flow", Figures 8–9) directly examines the evolution of different wavelet scales along actual sampling trajectories of the trained models (N=100N=100 Euler steps, 512 trajectories per dataset).

Rather than visualizing the raw ODE state ztz_t, which reflects the prescribed noise mixture, we track the running endpoint prediction z^1(t)=zt+(1−t),vθ(zt,t)\hat z_1(t)=z_t+(1-t),v_\theta(z_t,t), which reveals the model's current estimate of the final sample.

Figure 8 visualizes wavelet scalograms of the predicted samples across all seven datasets. The low-frequency structure emerges early, followed by progressively finer octave bands. Figure 9 quantifies this using the per-level resolved variance fraction rj(t)=Var[z^1(t)]j/Var[z1]jr_j(t)=\mathrm{Var}[\hat z_1(t)]_j/\mathrm{Var}[z_1]_j. The measured resolution times follow the predicted coarse-to-fine ordering on six of seven datasets, with perfect rank agreement (ρ=1\rho=1) on ETTh1, ETTh2, EEG, and Energy. On Stocks and Exchange, nearly tied predicted crossings correspond to nearly tied measured times. MuJoCo is the exception, where all scales resolve early and almost simultaneously.

Importantly, per-level standardization substantially collapses the measured resolution schedules, consistent with the role of level-wise SNR differences in the progression (Figure 9). The analysis thus provides direct empirical evidence that the learned flow exhibits the coarse-to-fine behavior anticipated by our theoretical SNR analysis.

AuthorsOpenReview ↗

Response to Reviewer GQHL - Part 3

Computational Overhead and Performance Gains (Q5)

We thank the reviewer for raising this question. We have added a detailed computational analysis to Appendix D.4 (new Table 4), covering FLOPs, parameter counts and preprocessing time.

The DWT adds negligible overhead. First, training windows are transformed only once before training, and generated samples are inverted once after sampling; all 100 network evaluations use the same T×FT\times F array shape as in the time domain. Second, the variants differ by only JdmJd_m embedding/head biases (768 parameters, or 0.008%0.008\%, at T=24T=24); FLOPs per network evaluation and NFE are identical. Third, a fast DWT costs O(qTF)O(qTF) for filter length qq: at T=24T=24, the transforms require about 2.4×1032.4\times10^3 multiply-adds versus 4.8×1094.8\times10^9 for sampling (ratio 5×10−75\times10^{-7}), remaining only 1.5×10−61.5\times10^{-6} at T=1024T=1024. Preprocessing all ETTh1 training windows takes just 0.030.03 s and 2.22.2 s on one CPU, respectively, incurred only once. Moreover, orthogonality and linear embedders/heads allow the DWT to be absorbed into the network, removing explicit transforms at inference (new Remark 3.1).

For performance, although ETTh1's gain is within one standard deviation and the time-domain variant is slightly better on MuJoCo (0.0050.005 vs. 0.0060.006), across 28 dataset–metric pairs at T=24T=24, the wavelet model wins in 22, ties in five, and loses in one (Table 15). It improves discriminative and correlational scores on all seven datasets (one-sided sign test, p≈0.008p\approx0.008 for each), and Context-FID on six of seven. Time-domain Context-FID is on average 1.44×1.44\times higher, reaching 2.8×2.8\times on Stocks and 1.75×1.75\times on Exchange. Thus, the wavelet parametrization provides consistent gains at negligible additional cost.


Scalability to Longer Sequences (Q6)

We substantially expanded the benchmark to T∈{512,1024}T \in \lbrace 512, 1024 \rbrace across all seven datasets (new Tables 11–12; updated Figures 4 and 10). Our method remains competitive across all four metrics, maintaining the best mean rank in Context-FID and discriminative score across all six lengths. At T=512T=512 and T=1024T=1024, it achieves the best Context-FID on 5/7 and 6/7 datasets, respectively, and the best discriminative score on all seven among available results. WaveletDiff's EEG result at T=1024T=1024 remains outstanding. The longer-sequence results currently use one independent training run with five evaluations per config. We plan to add more independent training runs to assess training variability.

Our datasets cover 6–28 variables. We acknowledge that scalability to substantially higher-dimensional data remains untested and have added this to the revised Limitations and Future Work section.


Multi-scale Fidelity of Generated Samples (Q7)

We implemented the requested analysis in Section 5.1 and new Appendix F.5 (Tables 13–14, Figures 13–14), comparing real and generated windows for all seven methods and datasets at T=24T=24. Five statistics cover all three requested aspects:

  • Wavelet-scale energy: Mean and median energy at each level across temporal scales.
  • Spectral properties: Log-spectral distance (dB) between mean Hann-tapered periodograms.
  • Local high-frequency structure: Kurtosis of finest-level details d1d_1 and correlation between ∣d1∣|d_1| and ∣d2∣|d_2|, for localized activity and cross-scale co-occurrence.

Statistics are averaged across channels and five random subsets, excluding boundary coefficients.

Our method achieves the best average rank over all five statistics (2.11 versus 3.00 for WaveletDiff, 3.36 for FourierDiffusion, 4.00 for FlowTS, and 4.13 for Diffusion-TS; Table 13). It ranks first or second on every statistic: first on both energy measures and finest-scale kurtosis, second on spectral distance and cross-scale correlation. It also reveals baseline limitations:

  • Fine-scale energy: Diffusion-TS and KoVAE generate a median of 6.6×6.6\times and 41×41\times too little finest-level energy, while FlowTS generates 2.7×2.7\times too much. Ours and FourierDiffusion achieve median generated-to-real energy ratios between 0.78 and 1.00 at every level (Figure 13).
  • Spectral versus local fidelity: FourierDiffusion matches the power spectrum best, but its finest-level coefficients are nearly Gaussian-like (kurtosis 2.40–2.57 on six datasets), and its cross-scale correlations are weaker than real data (−0.11-0.11 to 0.160.16 versus 0.200.20 to 0.550.55). Spectral agreement alone therefore does not guarantee faithful localization of fine-scale events.
  • Missing high frequencies: SigDiffusion generates no power above 1/81/8 cycles per step due to truncated Fourier reconstruction (Figure 14).

These statistics do not uniformly favor our model: FourierDiffusion wins on spectral distance and WaveletDiff on cross-scale correlation. Together, the new results support multiscale fidelity beyond aggregate generation scores.

Thank you for the detailed response and additional experiments. The revisions have addressed most of my concerns regarding empirical validation and practical relevance. However, my main concern about methodological novelty remains, particularly since the coarse-to-fine behavior is not unique to the proposed wavelet-domain formulation. The additional evidence nevertheless strengthens the paper's empirical contribution, so I raise my rating to 2.

Reviewer_2DJxOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose Wavelet Flow Matching, coupling a channel-token transformer with multi-resolution wavelet representations to generate multivariate time series. Crucially, the variance variations inherent to different wavelet levels naturally guide a coarse-to-fine generative process.

Critical strengths
  1. Combining wavelet representations with flow matching is intuitive, and the proposed architecture is simple and effective.

  2. The paper includes extensive comparisons, useful ablation studies, and a careful investigation of wavelet properties.

Critical weaknesses
  1. I'm wondering that for sequence lengths above 24, the experiments use only one training run. The reported standard deviations therefore do not capture training variability.

  2. Since the DWT is orthonormal, the population flow-matching problem is equivalent to a time-domain formulation under a change of coordinates. Different SNR crossing times alone do not establish that the learned generative process follows a coarse-to-fine progression.

Questions for authors
  1. On EEG at (T=128), the correlational score is 3.348, compared with 0.871 for FlowTS. What explains this substantial gap? Have the authors examined lagged cross-correlations or cross-spectral dependencies?

  2. The datasets use sliding windows with stride one. How exactly are training and evaluation windows separated? Are overlapping windows excluded across the train/test boundary to avoid potentially inflated fidelity scores?

  3. The experiments only consider relatively short sequences (T ≤ 128). How does the method scale to longer sequences (e.g., T=512 or 1024)? Since the level-specific embeddings depend on sequence length, can a pretrained model generalize to unseen lengths without retraining?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2DJx - Part 1

We thank the reviewer for their positive feedback and suggestions. Below, we address each point individually.

Multiple Independent Training Runs at Longer Sequence Lengths (Weakness 1)

We thank the reviewer for highlighting the importance of accounting for training variability. Our original submission already included three independent training runs at T=24T = 24, each evaluated five times, to capture variability beyond sampling and evaluation alone.

We would like to contextualize the computational requirements of extending this protocol to additional sequence lengths. Such exhaustive multi-seed evaluations are not standard practice in diffusion-based time-series generation, and none of our five original baselines explicitly reports them. Extending our benchmark to T∈{512,1024}T \in \lbrace 512, 1024 \rbrace while adding independent training runs expands the evaluation from 252 to 588 planned model trainings (2.33×2.33\times) across six methods and seven datasets, excluding ablations, repeated evaluations, and the additional KoVAE baseline. Longer sequences further increase the cost per run.

Nevertheless, we have extended our evaluation to three independent training runs for T∈{32,64,128}T\in \lbrace 32,64,128 \rbrace, in addition to T=24T=24. The revised results (Tables 8–10) closely match the originally reported means and preserve our main findings: our method achieves the best Context-FID on 6/7, 6/7, and 5/7 datasets at these lengths, respectively. We have also updated Figures 4 and 10 to reflect the expanded benchmark, including KoVAE and the new sequence lengths T∈{512,1024}T\in \lbrace 512,1024 \rbrace (Tables 11–12). Across all six lengths, our method maintains the best mean rank in both Context-FID and discriminative score. For T∈{512,1024}T\in \lbrace 512,1024 \rbrace, we currently report one training run with five evaluations per configuration, with additional independent runs planned to assess training variability.

These experiments substantially strengthen the assessment of robustness across sequence lengths while accounting for the computational demands of the expanded benchmark. We hope this sets new standards in the work to be published in this domain.


Coarse-to-fine progression of the learned flow (W2)

We thank the reviewer for this precise observation, with which we agree. Since W\mathbf{W} is orthogonal and the source isotropic, the linear path commutes with W\mathbf{W}: the coefficient- and time-domain flow-matching problems are conjugate and generate the same process on signals (new Remark 3.1). Our hypothesis class is also equivalent under this transformation, since the linear embedders and heads can absorb W\mathbf{W} and W⊤\mathbf{W}^\top. Thus, the coarse-to-fine progression is a property of the probability path and data statistics, not induced by the DWT. The wavelet basis instead makes scale explicit and changes optimization and initialization (AdamW preconditioning and per-level fan-in). This matters empirically: the wavelet model outperforms its matched time-domain variant in 22/28 dataset–metric pairs, ties in 5 and loses in 1 (Table 15).

We also agree that SNR crossings alone do not establish the behavior of the learned flow. We therefore directly instrumented the sampling ODE (new Appendix F.2, Figures 8–9), tracking the running endpoint prediction z^1(t)=zt+(1−t)vθ(zt,t)\hat z_1(t)=z_t+(1-t)v_\theta(z_t,t) and its per-level resolved variance fraction rj(t)=Var[z^1(t)]j/Var[z1]jr_j(t)=\mathrm{Var}[\hat z_1(t)]_j/\mathrm{Var}[z_1]_j. For independent Gaussian coefficients under the exact flow, rj(t)=SNRj(t)/(1+SNRj(t))r_j(t)=\mathrm{SNR}_j(t)/(1+\mathrm{SNR}_j(t)), crossing 1/21/2 at the predicted tj=1/(1+σj)t_j=1/(1+\sigma_j). Our findings are:

  1. The learned flow follows the predicted order on six of seven datasets. Rank agreement is perfect on ETTh1, ETTh2, EEG and Energy (ρ=1\rho=1). On ETTh1, measured half-resolution times are 0.08/0.15/0.29/0.640.08/0.15/0.29/0.64 for a3/d3/d2/d1a_3/d_3/d_2/d_1. Exchange and Stocks exhibit near-ties consistent with their predicted schedules, while nearly deterministic MuJoCo trajectories resolve all levels almost simultaneously. Resolution generally occurs earlier than predicted, reflecting learned cross-coefficient dependencies. Scalograms (Fig. 8) directly visualize the low-frequency skeleton emerging first, followed by finer bands.
  2. The progression responds to intervention. Per-level standardization collapses the ETTh1 resolution-time spread from 0.560.56 to 0.120.12 and degrades generation quality (ablation (b), Table 2, Fig. 6).
  3. The progression is representation-independent. Projecting the matched time-domain model's trajectories onto the wavelet basis yields nearly identical resolution times (0.08/0.16/0.29/0.640.08/0.16/0.29/0.64 on ETTh1), confirming the predicted equivalence. The representation affects learning quality, not the underlying progression (ablation (c)).

Accordingly, we have replaced "induces" with "exposes" in Section 3.2, clarified the equivalence in Remark 3.1, and added the learned-trajectory analysis in Appendix F.2.

AuthorsOpenReview ↗

Response to Reviewer 2DJx - Part 2

EEG Correlational Score and Recording Artifacts (Q1)

Thank you for this question, which led us to re-examine the EEG data more closely.

We found that the correlational score on EEG is dominated by four extreme recording glitches. Indeed, the EEG recording contains four time steps (out of 14,980) where some channels jump to raw values of up to 715,897, whereas the regular signal lies around 4,000–4,600 (channel medians), with per-channel interquartile ranges of 15–32. After z-scoring, these points reach ∣z∣|z| up to 122 and carry a median 88% of each channel's variance. As in Diffusion-TS, the correlational score is a lag-0 correlation pooled over all windows and time steps, i.e., an average of products zizjz_i z_j: a single glitch contributes about 1222/15,000≈1122^2 / 15{,}000 \approx 1 to it, against about 1/15,0001/15{,}000 for a regular point. The four glitches therefore outweigh the rest of the recording.

To quantify this, we scored modified versions of the real data with the same protocol as the models (1,0001{,}000 windows per draw, T=128T = 128):

Data scored against real windowsCorrelational score
Real windows (noise level of the metric)0.56±0.250.56 \pm 0.25
Glitches preserved, all other values shuffled in time independently per channel0.70±0.170.70 \pm 0.17
FlowTS (Table 10)1.12±0.361.12 \pm 0.36
Ours (Table 10)3.33±0.553.33 \pm 0.55
Real windows without the glitches4.37±0.124.37 \pm 0.12

Data with no cross-channel dependence outside the four glitches scores better than every method, while the real signal without the glitches scores worse than our model. On EEG, a low correlational score thus requires reproducing these four glitches, and says little about the dependence between channels in the rest of the recording. This suggests FlowTS reproduces these glitches more closely than our model, while our model, in turn, captures the regular signal well: its discriminative score on EEG is the best or tied of all methods at every window length except T=64, where it is second (0.007 vs. 0.006 for WaveletDiff). This issue also explains the high variability of the correlational score at (T=24), where only 0.6% of windows contain a glitch. Two independent draws of real windows yield a correlational score of (2.50\pm1.00), reaching 4.40 across 50 draws, encompassing the scores of FlowTS (1.89), WaveletDiff (2.27), and ours (3.87).

We did not evaluate lagged cross-correlations or cross-spectral dependencies. On EEG, they would be dominated by the same glitches, since they are also averages of products pooled over windows. A meaningful comparison would first require removing the glitches, i.e., modifying the benchmark dataset itself. This would depart from the standard preprocessing used by prior work, make our numbers incomparable with published results, and could be seen as tailoring the data to our method; we therefore keep the dataset unchanged.

We have added a description of these glitches to the EEG dataset paragraph and a note on their effect to the definition of the correlational score (Appendix E.1).

Finally, the values cited by the reviewer were updated with three independent training runs (Table 10; see our response to W1); the conclusions above are unchanged.


Training–Evaluation Protocol and Overlapping Windows (Q2)

We have made the protocol explicit in Section 4. For unconditional generation, we do not use a train/test split for the generator. Indeed, for each dataset and window length, the model is trained on all stride-one windows, and the four metrics compare an equally sized set of generated windows with these training windows. Since there is no held-out set, there is no train/test boundary across which overlapping windows could leak. All methods are trained and evaluated on exactly the same windows through the same pipeline, so the comparison is consistent across methods.

This is the standard protocol for unconditional time-series generation, used by TimeGAN, Diffusion-TS, FourierDiffusion, FlowTS and WaveletDiff for example, where the task is to learn the window distribution of a given dataset, and the metrics measure how closely the generated windows match it. Following it keeps our numbers directly comparable to published results.

A held-out evaluation would address a different question: generalization to unseen temporal segments rather than fidelity to the observed distribution. Removing the overlap requires holding out contiguous blocks of the series, separated from the training data by gaps of T−1T-1 steps. Since most of these series are non-stationary (e.g., Exchange spans 1990–2016 and Stocks 2004–2019), such blocks need not follow the training distribution, and metrics on them would measure distribution shift as much as generation quality: even real samples could obtain poor scores.

We agree that this protocol does not detect memorization. This limitation is shared by all standard time-series generation metrics, and we acknowledged it in our Limitations section.

AuthorsOpenReview ↗

Response to Reviewer 2DJx - Part 3

Scalability to Longer Sequences (Q3a)

We thank the reviewer for this suggestion and agree that evaluating longer sequences is important. Our original evaluation already matches or exceeds the lengths considered by several established baselines: WaveletDiff evaluates T∈{24,32,64,128}T\in \lbrace 24,32,64,128 \rbrace, while Diffusion-TS and FlowTS evaluate up to T=256T=256. Although SigDiffusion considers T=1000T=1000, systematic evaluations across multiple longer lengths remain uncommon.

Nevertheless, we have extended our evaluation to T∈{512,1024}T\in \lbrace 512,1024 \rbrace across all seven datasets, substantially expanding the benchmark (Tables 11–12, Figures 4 and 10). At T=512T=512, our model achieves the best Context-FID on five of seven datasets and the best discriminative score on all seven. At T=1024T=1024, it achieves the best Context-FID on six of seven datasets and the best discriminative score on all seven among the available results. The updated Figure 4 further shows that our method maintains the best mean rank in both metrics across all six sequence lengths. The WaveletDiff results on EEG at T=1024T=1024 remain outstanding. Currently, each longer-sequence configuration uses one independent training seed with five evaluations; additional independent runs are planned to assess training variability.


Generalization Across Sequence Lengths (Q3b)

A trained model does not directly generalize to unseen lengths: it is trained for a fixed window length TT, as are most baselines in our comparison. However, the dependence on TT is confined to a very small part of the model. Only the level embedders and heads,

\qquad \mathbf{H}_j \in \mathbb{R}^{L_j \times d_m},$$ with $L_j = T/2^j$, depend on $T$; they account for 0.14% of the parameters at $T=24$ (Table 3) and under 1% at $T=128$. The transformer blocks, time conditioning, and channel identities are independent of $T$. Moreover, for $T \leq 128$, where $J=3$, each level has the same meaning at every length: the detail level $d_j$ captures fluctuations at a scale of about $2^j$ time steps and $a_3$ the trend over blocks of 8 steps, so a longer window only has more coefficients per level, each describing a different position in time. Longer windows also add coarser levels ($J=6$ at $T=1024$), while $d_1$, $d_2$ and $d_3$ keep their meaning. Consequently, a model pretrained at one length could initialize training at another with the same token width $d_m$ by reusing the shared transformer and conditioning parameters, while reinitializing the length-dependent embedders and heads. We have added this point to the Limitations and Future Work section.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear reviewers,

We thank you all for your thoughtful feedback and constructive suggestions. We have carefully addressed the concerns and substantially revised the manuscript, with changes highlighted in blue. We believe that quality and clarity have improved.

The main additions and revisions are:

  • Expanded benchmark and additional baselines: We added KoVAE across all seven datasets and six sequence lengths (Table 1, Tables 8–12) and updated Figures 4 and 10 to summarize the expanded comparison. Our method maintains the best mean rank in Context-FID and discriminative score across all six lengths. Experiments with ImagenTime are ongoing.
  • Independent training runs: We extended the evaluation to three independent training seeds for T∈{32,64,128}T \in \lbrace 32, 64, 128 \rbrace, in addition to T=24T = 24, updating Tables 8–10 to account for training variability.
  • Longer sequences: We extended the benchmark to T∈{512,1024}T \in \lbrace 512, 1024 \rbrace across all seven datasets and the original five baselines (new Tables 11–12), with nearly all configurations completed.
  • Multi-scale fidelity: We added a detailed comparison of wavelet-level energy, spectral properties, and local high-frequency structure across generative methods (new Tables 13–14, Figures 13–14).
  • Mechanistic and computational analyses: We clarified the equivalence of wavelet- and time-domain flow matching, directly analyzed learned sampling trajectories (new Figures 8–9), and quantified the computational overhead of the wavelet transform (new Table 4).

We also expanded the discussion of when our method is most effective, clarified the evaluation protocol, and updated the limitations and future-work section.

We believe these additions substantially strengthen the empirical evidence and clarify the contributions of our work. We provide detailed explanations in our point-by-point responses below. We welcome any further feedback and would appreciate your consideration of whether the revisions address your concerns and warrant an updated assessment.

AuthorsOpenReview ↗

Follow-up on Revisions and Additional Baseline

Dear Reviewers,

We would like to briefly follow up on our detailed responses and revised manuscript, in which we have carefully addressed your concerns and incorporated substantial additional experiments and analyses.

As a further update, we have now added ImagenTime as an additional baseline across all seven datasets for sequence lengths T∈{24,32,64,128}T \in \lbrace 24, 32, 64, 128 \rbrace, using three independent training runs per configuration. The corresponding results are included in the updated manuscript. We are also preparing results for T∈{512,1024}T \in \lbrace 512, 1024 \rbrace, which we plan to include in the camera-ready version.

We would be grateful if you could take a look at our responses and revisions when you have the opportunity. We hope that the additional evidence addresses your concerns and would appreciate any further feedback or reconsideration of your assessment.

Thank you again for your time and constructive suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.