acceptodds
Under review as a conference paper at ICLR 2027

Fast Multivariate Time Series Generation via Sampling from a Spherical Manifold

Abstract

We propose a multivariate time series generator that embeds the training data onto a uniformly populated spherical manifold, from which new points are sampled and decoded into high-fidelity, diverse samples. The approach uses no attention mechanisms, adversarial training, diffusion, flow matching, or vector quantization, yet achieves generation quality competitive with the state of the art while sampling one to four orders of magnitude faster. We identify which ingredients are essential to the success of the approach in time series generation, and which are not. A controlled ablation shows that replacing the Gaussian prior of a variational autoencoder with the bounded sphere is what makes the approach competitive, and we find that uniform coverage of the sphere is critical to generation quality, which we enforce with a KoLeo regularizer. Conversely, iterative refinement and consistency regularization prove unnecessary. The model is also readily made class-conditional, and its samples are effective for data augmentation in long-term (128-step) human activity recognition. Our results suggest that a bounded spherical manifold is a natural fit for time series generation, and they motivate further research in this direction.

Primary area:
learning on time series and dynamical systems
Keywords:
time series generation, autoencoders, multivariate time series, generative 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_ER47OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a fast multivariate time-series generator using a uniformly distributed spherical latent space with KoLeo regularization, enabling single-pass sampling without attention, adversarial training, or iterative methods.

Critical strengths
  1. Achieves competitive generation quality with a single forward pass, enabling 10–10,000× faster sampling.
  2. Uses a simple, reproducible architecture without attention, GANs, diffusion, or flow matching. 3, Introduces a uniformly distributed spherical latent space with KoLeo regularization instead of a Gaussian prior.
  3. Provides informative ablation studies identifying essential components and unnecessary complexities.
  4. Improves HAR classification accuracy from 0.841 to 0.906 when real training data is limited.
  5. Highlights limitations of standard evaluation protocols, particularly their inability to detect memorization.
Critical weaknesses
  1. The abstract should be revised to improve its clarity, conciseness, and overall readability, as the current version is somewhat confusing, unnecessary long sentences and lacks a clear presentation of the key points.
  2. The paper is not clearly written. There are several grammatical mistakes and sentence making issues. The description of the method should be improved.
  3. How do authors handle bad samples over the sphere? Nothing guarantees that D(M(Z_rand)) is distributed like the real data.
  4. The authors approved the evaluation scores generators against their own training data, so high fidelity could partly be memorization. Out-of-sample tests on Sines and MuJoCo help but don’t settle it.
  5. Competitiveness is measured against published numbers, and the authors note that their re-run baselines score worse. It makes the “state of the art” claim harder to verify. How can authors justify this?
  6. The use of MLP and LSTM backbones without explicit temporal structure in the latent space may limit the modeling of long-range dependencies and reduce generation quality for long sequences.
  7. Augmentation is evaluated only on HAR, using a small real dataset of 221 samples, and exhibits a performance gap of up to 3.2 percentage points compared with real data. The generalizability of these results to other tasks and domains remains unclear.
  8. The KoLeo regularizer pushes each point away from its nearest neighbour within a mini-batch. This is a local, batch-based criterion, so it does not guarantee uniform coverage of the sphere across the whole dataset, and its strength depends on the batch size and latent dimension. The loss weight may therefore need retuning when these change.
  9. Duplicate or near-duplicate points in a batch would make the logarithm term diverge, so numerical safeguards (e.g. a small epsilon) are needed.
  10. What is the motivation or reason for selecting these six datasets and using fixed training configurations (e.g., 2,000 epochs and a fixed batch size)? Are these choices justified by the characteristics of the datasets or supported by empirical evidence?
  11. The training cost and hyperparameter tuning effort are not sufficiently discussed or evaluated. How much training data the method needs?
  12. Ablation should be performed on more complex, high-dimensional or irregular real-world series.
  13. The evaluation would benefit from including recent baselines such as ImagenTime, COSCI-GAN, and TimeLDM, which have shown promising performance in time-series generation, particularly in latent-space modeling and inter-channel correlation preservation.
  14. Lack of novelty and limited technical contributions. The sphere latent, KoLeo regularizer and deterministic autoencoder with post-hoc sampling are all existing ideas. The contribution is mainly their combination.
  15. Code reproducibility is necessary to verify the results. I have some concerns regarding the outcomes. Authors must share the code or GitHub link (if any).
Questions for authors

Please see the section on weaknesses, e.g. W3, W5, W10, W11.

3 comments
AuthorsOpenReview ↗

Response to Reviewer ER47

We thank the reviewer for the detailed review.

W1, W2: Abstract, writing and method description. We have revised the abstract. It now opens with the problem and motivation, uses shorter sentences, and states the key findings more directly. We found that our citation commands rendered parenthetical citations as in-text ones throughout, which made many sentences read as broken. This is fixed. We have also proofread the full paper. The Method section (Section 3) now opens with a notation paragraph and states the shape of every quantity. It removes overloaded symbols, and writes the projection, the generator, both loss terms and the latent noise as explicit equations.

W3: Bad samples; no guarantee that generated samples are distributed like the data. We do not filter or reject samples: every reported score is computed on all generated samples, as many as there are training samples. Instead, the model is designed so that bad samples are rare. The new Section 6.1 explains why we expect decoded draws from the prior to be distributed like the data (A). Bad samples arise as blur, where latents overlap, or from holes, where the prior has no training latents. The deterministic encoder is not penalized for keeping inputs apart. The projection places every latent on the support of the prior. KoLeo spreads the latents uniformly, which is equivalent to minimizing their divergence from the uniform prior. No generator offers a formal guarantee here, but the new ablation confirms this empirically. On every dataset, our model leaves fewer holes than the MMD-based variants. On most datasets, its prior draws lie closer to a training latent than the training latents lie to each other, at a low reconstruction error.

W4: Memorization. We agree, and highlighting this is part of our contribution. The standard suite scores generators against their own training data, and we show that it is largely blind to memorization (Appendix E.1). This applies to every method in the comparison, including ours. Sines and MuJoCo are the only benchmark datasets with a known generating process, so they are the only ones on which we can evaluate on fresh data at all. There, our scores barely change out of sample, which shows that tuning on the training set did not compromise generalization. As the reviewer notes, and as we state in the paper, this does not settle memorization. We name this as a limitation in the Conclusion.

W5: Comparison with published numbers; re-run baselines score worse. We apologize for the confusion. "Self-reported" means the scores that each method's authors published for their own method. We did not re-run any baseline. The re-run scores we mention come from other papers: PaD-TS and SMF re-ran Diffusion-TS and report worse scores than Diffusion-TS reports for itself, in most overlapping cells. This observation motivates our choice. Comparing against self-reported scores is the conservative option, because it compares our results with each baseline at its best. Re-running the baselines ourselves would likely make them look worse, so it would make our claim easier to support, not harder. We checked the public code of every baseline except PrismFlow, which has none, and confirmed that its self-reported scores were produced on the same datasets with a compatible evaluation protocol (Appendix D.3). A footnote in the Introduction now defines self-reported scores, and Section 4.3 states that we do not re-run baselines and why.

W6: MLP and LSTM backbones and long-range dependencies. We agree for the MLP backbones. Their latent has no explicit time axis, so temporal structure is encoded only implicitly. The LSTM backbones are different. Their latent keeps a compressed time axis, and the projection only rescales it. We use them for the 128-step HAR experiments, where a classifier trained only on our synthetic samples comes close to one trained on real data. That said, we see the competitiveness of the MLP models as a finding in itself. On four of the six benchmarks, a latent without any time axis is enough to be competitive. This suggests that these benchmarks do not require a latent that explicitly models temporal dynamics, and that more temporally demanding benchmarks are needed. Enforcing temporal structure in the latent more directly is part of the future work we name in the Conclusion.

AuthorsOpenReview ↗

Response to Reviewer ER47 (cont.)

W7: Augmentation on HAR only; small real set; gap to real data. For the scope of the HAR experiment, please see (D). The small real set is deliberate. We test whether our generator can supplement training when real data is extremely scarce, which is where augmentation matters most. Both the generator and the classifier use the same fixed 221-sample subset, so no further real data is used. Adding only half as many generated samples as there are real ones clearly raises test accuracy.

On the gap to real data: a gap is expected when a classifier never sees a real sample, and it is small in this setting. More importantly, synthetic samples can replace real ones almost without loss. When part of the train set is replaced by synthetic samples, accuracy stays close to that of the full real train set throughout. When the same samples are instead removed, accuracy falls sharply once very little real data remains, and reaches chance level without real data (Fig. 4, left).

W8: KoLeo is local and batch-based; sensitivity to batch size and latent dimension. Locality is deliberate (A). KoLeo targets the same uniform distribution as a global criterion such as the MMD, but pushes hardest where latents crowd, and fills gaps between neighbors directly. Coverage over the whole dataset is not guaranteed by a batch-based loss, so we measure it. The new holes ratio is computed on the full training set, against prior draws of the same size. For our model, it shows that the latents cover the sphere more evenly than a random sample would, on four of the six datasets with latent noise and on all six without. On sensitivity: we keep the batch size fixed, while the latent dimension varies by two orders of magnitude across datasets. The ablation sweeps the KoLeo weight on all six datasets, and the best weight lies within a factor of 10 of our shared 1e-2 on every one.

W9: Numerical safeguards. We agree, and our implementation already includes them. It follows the KoLeo implementation of DINOv2: epsilon = 1e-8 is added to the nearest-neighbor distance before the logarithm, and is also used when normalizing and computing distances. Each term is therefore bounded by -log(epsilon) ≈ 18.4, even for exact duplicates. Each batch is also drawn without replacement. All of our runs, including the 1,040 models of the new ablation, trained without numerical problems. We omitted epsilon from Eq. 5 for readability; the code includes it (W15).

W10: Choice of datasets and fixed training configurations. For the datasets, please see (C). The fixed configuration is deliberate, and Section 4.2 now motivates it. The standard suite has no held-out validation split, so any hyperparameter tuned per dataset is tuned on the evaluation criterion itself. To keep this to a minimum, we tuned the full setup once, on Sines, and transferred it, sweeping only the latent size per dataset. Two further choices were introduced only where training otherwise failed: a cosine learning-rate schedule (MuJoCo) and LSTM backbones (Energy). A setup that transfers with this little tuning is also the stronger result. The new ablation supports the transfer: on every dataset, the best KoLeo weight lies within a factor of 10 of the shared weight (Appendix E.2). Out-of-sample scores on Sines and MuJoCo show that this tuning did not substantially affect generalization (Appendix E.1).

W11: Training cost, tuning effort and data requirements. We agree that these are important questions. To our knowledge, they are rarely studied in the time series generation literature, and characterizing how much data our approach requires is part of our future work. Within this paper, we address them where we can. Section 4.2 describes our deliberately minimal tuning protocol (see W10). On data requirements, the low-data HAR experiment trains both the generator and the classifier on only 221 real samples, and the generated samples still clearly raise accuracy.

W12: Ablation on more complex data. The new ablation now covers all six datasets, including the high-dimensional Energy (28 channels) and fMRI (50 channels) (A). Irregularly sampled series are outside the scope of our method and benchmark (see Uav3, W5).

AuthorsOpenReview ↗

Response to Reviewer ER47 (cont. b)

W13: Missing baselines (ImagenTime, COSCI-GAN, TimeLDM). These methods are relevant, but we cannot compare against every generator, so we selected baselines by explicit criteria (Appendix D.3). We aimed to cover each family with one well-established and one recent method: GANs (TimeGAN, GT-GAN), VAEs (KoVAE), diffusion (Diffusion-TS, PaD-TS), flow matching (SMF, PrismFlow) and discrete-token transformers (SDformer). The suggested methods belong to families that are already represented: COSCI-GAN is a GAN, and ImagenTime and TimeLDM are diffusion models. We did consider ImagenTime. We chose Diffusion-TS as the established diffusion baseline because it reports results on all six datasets and is the diffusion baseline used by the most recent method in our comparison, PrismFlow.

W14: Novelty. Please see (B).

W15: Code. We agree that code is necessary to verify the results. We have added an anonymized codebase as supplementary material. It contains the code to train every model in the paper and compute its reported scores, with one command per result, including all 1,040 models of the new ablation. We would be glad to address the reviewer's specific concerns about the outcomes if the reviewer could point us to them.

Reviewer_Uav3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a lightweight multivariate time series generation framework based on a non-variational spherical autoencoder. The method projects latent representations onto a hypersphere and employs KoLeo regularization to encourage uniform latent coverage, enabling generation through direct sampling and single-pass decoding. Experiments on six benchmark datasets demonstrate competitive generation quality and faster sampling compared with existing methods.

Critical strengths
  1. The proposed method achieves competitive generation quality using lightweight architectures without adversarial training or iterative sampling.

  2. The paper evaluates the method on six standard benchmarks using multiple metrics and further investigates conditional generation and data augmentation.

  3. The experiments investigate the roles of spherical latent representations, KoLeo regularization, and iterative refinement.

Critical weaknesses
  1. Limited methodological novelty. The core framework largely builds upon existing spherical autoencoders combined with established KoLeo regularization. The proposed framework is highly similar to Sphere Encoder [1], sharing the same core mechanisms of deterministic spherical encoding, uniform, spherical sampling, latent noise injection, and single-pass decoding.

[1] Image Generation with a Sphere Encoder.

  1. Insufficient evidence for the benefits of non-variational modeling. Although the proposed method outperforms the Gaussian VAE baseline, the comparison simultaneously changes the latent geometry, regularization objective, and noise mechanism. Therefore, the improvement cannot be attributed specifically to the non-variational formulation.

  2. Missing comparisons with closely related spherical generative models. Despite building upon existing spherical autoencoders, the paper does not directly compare against the original Spherical Auto-Encoder or hyperspherical VAEs. The comparison with Gaussian VAE alone cannot isolate the benefits of spherical geometry from those of non-variational modeling.

  3. Limited validation of KoLeo regularization. The KoLeo ablation is conducted only on Sines using an earlier model configuration rather than the final one. Its effectiveness across different datasets and final architectures remains insufficiently validated. More comprehensive ablations of KoLeo and latent noise are needed.

  4. Missing recent strong baseline. How does your method compare with Diff-MN [1]? Published results show that Diff-MN outperforms GT-GAN and KoVAE on the overlapping datasets.

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

  1. The proposed method shows limited overall improvements in generation quality and performs substantially worse than existing methods on several datasets and metrics, particularly on Energy and fMRI. These results raise concerns about its effectiveness and robustness across different datasets.

  2. Insufficient evidence for sampling efficiency advantages. The reported speedups are primarily measured against diffusion-based and other computationally expensive generators, while comparisons with VAE-based and other single-pass methods are missing. Since standard VAEs also generate samples through a single decoding pass, it remains unclear whether the proposed method offers any efficiency advantage over these simpler baselines.

  3. Insufficiently controlled efficiency comparisons. The reported sampling speedups rely on self-reported baseline results obtained using different hardware and sample counts. Comparisons under identical hardware, batch sizes, and generation workloads, including other single-pass generators, are needed to substantiate the efficiency claims.

  4. Limited evaluation of long-term generation. Most experiments focus on short sequences of length 24, while longer sequences are evaluated only on HAR.

  5. Insufficient evaluation of downstream data augmentation. Although the proposed method improves downstream classification accuracy, it lacks comparisons with other generative data augmentation methods, making it unclear whether it offers advantages over existing approaches. Moreover, the experiments are conducted on only one dataset, limiting the generalizability of the findings.

Questions for authors

Please refer to the Weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Uav3

We thank the reviewer for the detailed review.

W1: Novelty, and similarity to the Sphere Encoder. Please see (B). We share the Sphere Encoder's deterministic spherical encoding, uniform sampling and single-pass decoding, and we credit it for them. What differs is the recipe that works for time series: uniform coverage must be enforced explicitly with KoLeo, while the Sphere Encoder's consistency losses and iterative refinement bring no benefit here (Appendix E.3).

W2: Non-variational benefit not isolated. We agree that the original comparison changed several factors at once. The new ablation changes one at a time (A). The step from the VAE to G-MMD replaces the per-sample KL with an aggregate-only MMD penalty, with the same Gaussian prior, backbone and training setup. Since G-MMD is trained both with and without latent noise, this step isolates the non-variational benefit from the noise mechanism. G-MMD clearly improves on the VAE on five of the six datasets.

W3: Original Spherical Auto-Encoder and hyperspherical VAEs. Setting aside backbone and training differences that follow from the data modality, Zhao et al.'s SAE is our model without latent regularization or latent noise, plus a centering step before projection. Our revised ablation includes KoLeo weight 0 on three datasets: with latent noise on Sines and Stocks, and without it on Energy. The last matches the SAE most closely. In every case it performs markedly worse than the regularized model. On Energy, its combined score lies far beyond the real-vs-noise ceiling (Appendix E.2, Fig. 6).

A hyperspherical VAE keeps the per-sample KL. On any prior, including the uniform sphere, this decomposes into the mutual information between input and latent plus the aggregate prior mismatch. It therefore retains the information penalty that our account identifies as the VAE's main limitation, and changes only the geometry. Zhao et al. also report that the S-VAE underperforms their deterministic SAE on MNIST and could not be trained on FFHQ. To isolate the geometry instead, our revised ablation compares the MMD-regularized model with Gaussian and spherical priors (G-MMD vs. S-MMD). The effect of the sphere alone is dataset-specific, while the sphere with KoLeo is at least as good as both on all six datasets.

W4: KoLeo and latent noise ablations. The new ablation sweeps the KoLeo weight on all six datasets, with the final configuration of each, and trains every model with and without latent noise (A; Fig. 6). Without KoLeo, draws from the prior land further from the nearest training latent than the training latents lie from each other, i.e. the prior has holes, and the combined score is clearly worse. Latent noise turns out to be optional for our model: without it, scores tie or improve slightly.

W5: Diff-MN. Diff-MN addresses a different task: generation from irregularly sampled data. All its results are obtained with a large share of observations removed, averaged over several sequence lengths, and it reports no results on fully observed data. Its advantage over GT-GAN and KoVAE holds in that irregular setting, against their irregular variants. Our method and benchmark assume regular sampling, so its published numbers cannot be compared with ours. Extending our approach to irregular observations is an interesting direction for future work.

W6: Limited gains, and weaker results on Energy and fMRI. We may not have made our aim clear enough, and have sharpened it in the introduction. We do not claim to outperform the state of the art. We claim generation quality competitive with it, reached with a mechanism widely considered non-competitive. We show that a plain autoencoder with MLP or LSTM backbones, given a suitably designed latent space, reaches the quality of diffusion and flow-matching models while sampling one to four orders of magnitude faster. Under the established metrics, the benchmark has no clear front-runner. Our method scores worst in fewer cells than Diffusion-TS, arguably the most widely used baseline. The one possible exception, SDformer, uses one to two orders of magnitude more parameters than the other methods.

On fMRI our method is in fact among the strongest. It has the second-best Discriminative score (after SDformer), ties on Predictive, and is mid-ranked on Context-FID and Correlational. Energy is where it is weakest. It scores worst on Context-FID and Correlational among the methods that report them, though still closer to the floor than to the ceiling on a log scale. Our floors and ceilings put such differences in perspective. On the other four datasets, most methods are already close to or below the real-vs-real floor, so further gains there are hard to interpret. Energy and fMRI are the two datasets with clear headroom. On them, no method except SDformer comes close to the floor on the Discriminative score.

AuthorsOpenReview ↗

Response to Reviewer Uav3 (cont.)

W7: No speed comparison against single-pass methods. We agree that VAEs and other single-pass generators sample about as fast as our method on similar hardware and with similar backbones. We do not claim an efficiency advantage over them, since our method is itself a single-pass generator. The point of the speed comparison is a different one. Iterative generators such as diffusion and flow-matching models currently set the standard for sample quality, at the cost of many network evaluations per sample. Our method reaches generation quality competitive with theirs while keeping the speed of a single-pass model. Against other single-pass models, the relevant comparison is quality rather than speed. There, the new ablation shows that a Gaussian VAE with the same backbone and training setup scores clearly worse than our model on every dataset (A).

W8: Speed comparisons not controlled. We agree that a fully controlled comparison would be more precise. However, the speed gap is large, and the known differences in setup all favor the baselines. Our times are measured on a laptop RTX 4050 GPU, which is less powerful than the hardware the baselines report. We also generate the entire training set, whereas the baselines that report their sample count generate far fewer samples (Section 5.1). Even so, our method is one to four orders of magnitude faster. Controlling for hardware and workload would therefore widen the gap rather than close it. On other single-pass generators, see W7.

W9: Few long-sequence evaluations. Please see (C) and (D).

W10: Augmentation: no other generative augmentation methods, one dataset. Please see (D).

Reviewer_Yg5jOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a simple and efficient approach to multivariate time series generation based on a spherical latent space. The method projects encoded time series onto a fixed-radius hypersphere and uses KoLeo regularization to encourage uniform coverage of the latent space. New samples are generated by uniformly sampling points on the sphere and decoding them in a single forward pass, without diffusion, flow matching, or adversarial training. Experiments on six standard benchmarks and a conditional human activity recognition task demonstrate competitive generation quality and substantially faster sampling compared with several existing methods.

Critical strengths

The main idea is with an interesting direction with useful empirical insights. I appreciate the authors' attempt to revisit simple autoencoder-based generation rather than relying on increasingly complex diffusion or flow matching models. The observation that uniform latent coverage is important, while iterative refinement and consistency regularization are largely unnecessary for time series, is quite interesting. The significant sampling speedup and the additional analysis of existing evaluation metrics also make this work practically relevant.

Critical weaknesses
  • The novelty and necessity of the spherical latent space remain unclear. The main technical ingredients, including spherical autoencoders and KoLeo regularization, have already been explored in prior work, particularly Zhao et al. (2019) and Yue et al. (2026). While applying these techniques to time series is interesting, the paper does not sufficiently explain why a spherical latent space should be particularly suitable for time series generation compared with other well-regularized latent distributions. I would encourage the authors to better clarify the methodological novelty and provide a stronger justification for the spherical design itself.

  • The experiments do not sufficiently isolate the contribution of the proposed design. The main comparison with Gaussian VAEs changes several factors simultaneously, including the latent geometry, the regularization objective (KoLeo versus KL), and the latent noise mechanism. Therefore, the observed improvements do not necessarily establish that the spherical geometry itself is responsible, rather than better latent distribution matching or a different training strategy. Comparisons with stronger non-variational autoencoder baselines using alternative priors or distribution-matching objectives would be particularly helpful. Moreover, the six main benchmarks consist of only 24-step sequences, while the longer conditional generation experiment lacks direct comparisons with competing generators, leaving the generality of the findings somewhat unclear.

Ref:

[1] Zhao, D., Zhu, J., & Zhang, B. (2019). Latent Variables on Spheres for Autoencoders in High Dimensions. arXiv preprint arXiv:1912.10233. https://arxiv.org/abs/1912.10233

[2] Yue, K., Jia, M., Hou, J., & Goldstein, T. (2026). Image Generation with a Sphere Encoder. arXiv preprint arXiv:2602.15030. https://arxiv.org/abs/2602.15030

Questions for authors

Could the authors explain why a uniformly populated sphere should be preferable to other latent distributions with appropriate prior matching, beyond the empirical comparison with a standard Gaussian VAE?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Yg5j

We thank the reviewer for the thoughtful review.

Q: Why should a uniformly populated sphere be preferable to other latent distributions with appropriate prior matching? This question shaped our revision most. Please see (A). In short, the sphere is preferable not for its own sake. With a uniform prior, the projection makes prior matching exact in magnitude, and KoLeo makes it local in direction, without penalizing the information the latent carries. The new ablation compares our model with WAEs using a Gaussian prior (G-MMD) and a uniform spherical prior (S-MMD).

W1: Novelty, and justification for the sphere. For novelty, please see (B). For the justification of the spherical design, please see (A).

W2: The VAE comparison is confounded; stronger non-variational baselines. We agree, and the new ablation addresses exactly this (A). It moves from the VAE to our model one change at a time: aggregate-only regularization (G-MMD), then the projection (S-MMD), then KoLeo (ours), each with and without latent noise. G-MMD and S-MMD are the non-variational baselines with alternative priors and distribution-matching objectives that the reviewer asks for. Most of the gain comes from regularizing only the aggregate posterior. KoLeo on the sphere makes it consistent across datasets.

W2: Only 24-step sequences. Please see (C).

W2: No competing generators for the long, conditional experiment. Please see (D).

Thanks for the rebuutal, it resolved most of my concerns and I'll increase the score

Comments on the paper

4 comments
AuthorsOpenReview ↗

General comment to all reviewers

We thank the reviewers for their effort with the feedback. Here we address shared concerns and answer individual concerns in each reviewer's thread. We have uploaded a revision to the paper with an anonymised version of our codebase.

(A) Why a uniformly populated sphere, and which ingredients matter

We agree that the submission lacked an account of why the method works. Its ablations, a comparison with a Gaussian VAE and a KoLeo sweep on Sines, could not isolate which factors matter. We have used the rebuttal period to articulate this account and run a 1,040-model ablation to support it. This replaces the previous VAE-ours ablation (6.1) and sphere uniformity (6.2) sections respectively. We have also moved the section on iterative refinement to the appendix to make room for this analysis.

In short, a decoder-based generator fails in two ways: overlapping latents give blurred samples, and regions of the prior without training latents ("holes") give unrealistic ones. The VAE's per-sample KL penalizes the information the latent carries, so it trades one failure for the other. Our model avoids this in three steps: it regularizes only the aggregate posterior, the projection fixes the magnitude of every latent, and KoLeo spreads the latents evenly over the sphere. The ablation moves from the VAE to our model one step at a time, on all six datasets. Most of the gain comes from the first step, and KoLeo on the sphere makes it consistent across datasets. See Section 6 of the revised paper for details.

(B) Novelty

We respectfully disagree that prior work diminishes the contribution. Deterministic spherical autoencoders have so far been demonstrated as generators only on images. Zhao et al. (2019) evaluate on MNIST, FFHQ and CelebA, and Yue et al. (2026) on CIFAR-10, ImageNet, Animal-Faces and Oxford-Flowers. Concurrent and later spherical-latent generators (Do et al., 2026; Mazumder et al., 2026; Yue et al., 2026b) are likewise evaluated on images, with one exception for discrete grammar expressions. We have added these works to the related work section. Whether a bounded, uniformly populated latent works for multivariate time series, and with which ingredients, was an open question.

Answering such a question is an established kind of contribution. Nearly every baseline in our benchmark builds on a generative family first demonstrated on another modality, mostly images: TimeGAN and GT-GAN on GANs, KoVAE on the VAE, Diffusion-TS and PaD-TS on denoising diffusion, PrismFlow on flow matching, and SDformer on discrete token modeling from NLP. TimeVQVAE is explicitly "motivated by the successes in the image generation domain", and ImagenTime turns series into images to "exploit advanced diffusion vision models". Nor is our adaptation mechanical. The recipe that works for time series differs from the image one (Section 2, "Our position"). Uniform coverage must be enforced explicitly with KoLeo, whereas Yue et al. find that latent noise alone gives near-uniform coverage. Conversely, the consistency losses and iterative refinement central to the Sphere Encoder bring no benefit here (Appendix E.3). The revised paper adds a further contribution: the account and ablation of (A), which identify which properties of prior matching matter.

(C) The benchmark: six datasets at 24 steps

We chose the six datasets and the 24-step window because they are the standard benchmark for unconditional multivariate time series generation. The protocol was introduced by TimeGAN (NeurIPS 2019) and extended to its current six datasets by Diffusion-TS (ICLR 2024). It is used by, among others, GT-GAN (NeurIPS 2022), DiffTime (NeurIPS 2023), KoVAE (ICLR 2024), SDformer (NeurIPS 2024), PaD-TS (AAAI 2025) and SMF (NeurIPS 2025). Evaluating on it is what lets us compare against these methods' self-reported results. Three of the four time series methods suggested in the reviews use the same datasets at length 24 as well: ImagenTime (NeurIPS 2024), TimeLDM, which calls 24-step generation "the most common comparison in existing works", and Diff-MN (ICML 2026), in the benchmark's irregular-sampling variant.

We agree that the benchmark has limitations, and examining them is one of our contributions. We add empirical floors and ceilings for every metric. We show that the protocol is largely blind to memorization, and that the Predictive score is uninformative on two of the six datasets. Whatever its limitations, this benchmark is currently the best basis for a claim of competitiveness, because it is where the established methods report their own results. Beyond the benchmark, we evaluate long-term (128-step) generation on HAR (see (D)).

AuthorsOpenReview ↗

General comment to all reviewers (cont.)

(D) The HAR experiment

The HAR experiment is a capability demonstration, not our main comparison. Our main claim rests on the six standard benchmarks. HAR shows that the same model extends to class-conditional, 128-step, multivariate generation, and that its samples are useful downstream.

We agree that a comparison with other generators would strengthen it, but we could not find a basis for one. Conditional, long-term generation is not the setting in which time series generators are usually compared. Among our baselines and the related work we surveyed, we found no method that self-reports long-term, class-conditional generation evaluated through downstream classification. The conditional experiments of our baselines are imputation and forecasting. The closest exception, TimeVQVAE (Lee et al., 2023), evaluates class-conditional generation on the univariate datasets of the UCR archive, many with small training sets. That is a different setting from multivariate HAR (9 channels, 7,352 training windows). Since we compare only against self-reported results throughout (see our response to ER47, W5), we evaluate against the real data itself instead: the full real train set, reduced real train sets, and chance level (Section 5.2).

For the same reason, we do not claim generality from one dataset. A broader study of generative augmentation across tasks and domains is an interesting direction for future work.

AuthorsOpenReview ↗

Revised paper now highlights changes

Hi reviewers! Note that I just re-uploaded the revised paper with changes highlighted in blue. It is exactly the same as the previous (and only) revised version. It just highlights the changes relative to the original paper for your convenience.

AuthorsOpenReview ↗

Concluding remarks

We have not yet received responses to our rebuttal. To make the reviewer–AC discussion easier, here is a short map of how the revision addresses the main concerns. Details are in our individual responses and the revised PDF.

Why the method works. Section 6 is rewritten (and Appendix E.2 added) with a direct account of this, supported by new ablations (Figure 5 and 6). These show that dropping the per-sample posterior regularizer produces most of the gain, and that adding a hard projection onto the sphere together with KoLeo regularization makes the gain consistent across datasets.

Novelty. To our knowledge, a deterministic spherical prior has not been used for autoencoder-based time series generation, and the role of uniformity on the sphere (here via KoLeo regularization) has not been studied in that setting. This is now stated more prominently, and related concurrent work (outside time series) is listed in Section 2.

Evaluation suite. We clarified that we expanded the evaluation with floor, ceiling and out-of-sample comparisons (Figure 2, Table 3), and discuss the limitations of the existing evaluation suite in Sections 4.2 and 5 and Appendices D.2 and E.1. These points form part of our contribution over prior work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.