acceptodds
Withdrawn from ICLR 2027

RecFlowTime: Self-Conditioned Rectified Flow for Time-Series Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Flow matching offers an efficient alternative to diffusion models by learning continuous transport paths from noise to data that can be integrated in relatively few steps. We introduce RecFlowTime, a rectified-flow framework for time-series generation that learns velocities along linear data–noise interpolations. A bidirectional Transformer with rotary position embeddings represents temporal relationships, while self-conditioning progressively refines the estimated clean sequence. During training, minibatch optimal-transport coupling replaces independent pairing, reducing path crossings and producing straighter sampling trajectories. A floored minimum signal-to-noise ratio (min-SNR) weight preserves supervision near the low-noise endpoint and reduces sample roughness. Across four synthetic, financial, sensor, and biomedical benchmarks, RecFlowTime outperforms recent baselines on key measures of distributional fidelity and temporal realism. In addition, RecFlowTime uses only 20 function evaluations, compared with 200–500 for the diffusion baselines, resulting in more than an order-of-magnitude faster sampling and up to a 79% reduction in training time. Although trained unconditionally, RecFlowTime also supports conditional generation for imputation and forecasting without retraining or additional model evaluations.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series generation, flow matching, rectified flow, optimal transport, minibatch coupling, generative models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_iT9LOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes RecFlowTime, a rectified-flow framework for time-series generation using a bidirectional transformer with rotary position embeddings and self-conditioning. The results show that RecFlowTime outperforms multiple baseline methods across four synthetic, financial, sensor, and biomedical datasets and it also supports conditional generation for imputation and forecasting tasks.

Critical strengths

None

Critical weaknesses
  1. Limited Novelty: most of the components are proposed by previous work, and the remaining modification is minor. For example, the training and sampling framework of rectified flow, minibatch OT coupling, self-conditioning, and the min SNR weighting strategy are almost the same as the original work. In particular, the author doesn’t mention and cite the self-conditioning paper [1], which sounds like claiming to be your proposed method.
  2. Outdated Baseline Methods: two of the baselines, TimeGAN and TimeVAE, are old methods. The author should include more recent baselines such as SDformer [2] or the other Rectified Flow-based TS generation FlowTS [3].

[1] Ting Chen, et al. Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning. (ICLR 2023)
[2] Zhicheng Chen, et al. SDformer: Similarity-driven Discrete Transformer For Time Series Generation. (NeurIPS 2024)
[3] Yang Hu, et al. FlowTS: Time Series Generation via Rectified Flow.

Questions for authors

Clarification requests:

  1. The self-conditioning part is one of the main components for the proposed method but is poorly explained. It is hard to understand why we need this condition c as input in the first place and how it is actually used during the training.
  2. Several baseline results in Table 1 differ substantially from those reported in the original papers. For example, PaCoDi reports Context-FID 0.018 and discriminative score 0.010 on Sines at L = 64 in Table 8 in PaCoDi paper, whereas Table 1 in this paper reports 0.774 and 0.362. Could the author explain what is the reason for this large performance gap?

Suggestions for future work:

  1. The three additional metrics “spectral, increment and autocorrelation distances” are explained through definition and motivation only. The author should include more experiments to validate the reliability of these metrics, even using some synthetic data. Otherwise, it is unclear if we can trust these metrics.
  2. The conditional generation results are only compared to linear interpolation (imputation) and last value (forecasting) which may not be enough.
Reviewer_MiL7OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper applies rectified flow to multivariate time-series generation, combining a bidirectional Transformer with rotary position embeddings, self-conditioning, minibatch optimal-transport coupling, and a floored min-SNR loss weight. The experiments compare generation quality and computational efficiency with GAN, VAE, and diffusion baselines on four benchmarks and include component ablations. The authors report improvements on several generation metrics and faster sampling than the diffusion baselines using 20 Euler steps.

Critical strengths

None

Critical weaknesses
  1. The paper offers limited novelty over existing work. Rectified flow has already been used for time-series generation, for example in FlowTS, and several components adopted here are established techniques. The paper mainly adds these components to an existing generative framework. It does not clearly show what new method or insight goes beyond combining existing techniques.

  2. The motivation for adding these components is unclear. The introduction describes general requirements such as preserving local changes and long-range dependencies, but does not sufficiently explain where existing methods fail. It then introduces self-conditioning, minibatch OT coupling, and loss weighting without clearly connecting each design to a specific problem. The reader is told what components are used, but not which limitations require these additions or why this particular combination addresses them.

  3. The theory does not explain why self-conditioning should improve generation. Proposition 1 shows that converting a velocity prediction into a clean-sequence estimate does not amplify velocity errors at fixed inputs. However, it does not analyze what happens when this estimate is fed back into the network or show how the feedback improves the next prediction. The result therefore does not establish a beneficial property of the self-conditioning mechanism itself, such as reducing prediction errors or improving sampling stability. The theoretical explanation remains incomplete for a component presented as central to the method.

  4. The ablation experiments do not fully establish the benefit of the specific self-conditioning design. Removing self-conditioning provides evidence of its contribution within the current model, but does not show why feeding back a predicted clean sequence is beneficial compared with other feedback inputs. A comparison with previous sampling states or their history would help distinguish the value of this specific design from the general benefit of historical feedback. The additional training computation also needs to be considered. Furthermore, the reported standard deviations cover only evaluator seeds, so the consistency of the gains across independent generator training runs remains unclear.

  5. The experiments lack a basic rectified-flow baseline. The main comparisons include GAN, VAE, and diffusion models, but do not include a basic RF model with comparable network, training-budget, and sampling settings. Removing one component at a time does not replace this comparison. Without it, the results do not clearly separate the gains from using RF itself from those brought by the added components and their combination.

Questions for authors

Why use the predicted clean sequence for self-conditioning instead of the history of previous flow-sampling states, and what benefit does this particular design provide?

Reviewer_J2G1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes RecFlowTime, a rectified-flow model for unconditional time-series generation. A bidirectional Transformer predicts the velocity along straight paths between data and noise. On four datasets, the model beats TimeGAN, TimeVAE, Diffusion-TS and PaCoDi on most metrics and samples much faster than the diffusion baselines. The paper also shows imputation and forecasting through masked replacement without retraining.

Critical strengths
  1. The paper is clearly written and easy to follow. It reports mixed results too, and the code, datasets and exact baseline commits are provided.
  2. The results on the main metrics are strong and consistent, and the sampling speedup over diffusion baselines is large.
Critical weaknesses
  1. The novelty is limited. While the paper shows clear improvements, it is more about combination of existing techniques from different papers, closer to well-executed engineering, rather some novel approach.
  2. The evaluation setup makes it hard to place the results. The datasets are non-standard versions (a 5-feature subset of Energy, Stocks extended to 2022, a custom synthetic ECG) with longer sequences than the usual benchmark, so numbers cannot be compared with published ones.
Questions for authors
  1. Can you report results on the standard benchmark versions?
  2. Can you run several training seeds per method, not only two evaluator seeds?

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

Related papers

Discussion (0)

Sign in to comment.