acceptodds
Under review as a conference paper at ICLR 2027

Text-Conditioned Time-Series Generation: Cross-Modal Alignment and Fine-Grained Control in Time-Frequency Image Space

Abstract

Text-conditioned time-series generation enables flexible and fine-grained control over generated time series, supporting a wide range of applications. Despite its potential, existing methods face two key challenges: (1) limited alignment between time series and text (TS-text alignment); and (2) limited adherence to fine-grained textual instructions. In this paper, we propose TextPlot, an image-based diffusion framework for text-conditioned time-series generation. To address the first challenge, we introduce the short-time Fourier transform (STFT) as an intermediate representation, mapping a one-dimensional time series to a two-dimensional time-frequency image with explicit temporal and frequency coordinates. The resulting time-frequency image provides a structured representation for relating textual descriptions to temporal dynamics. For the second challenge, TextPlot employs a dual-path text-conditioning mechanism. Global pooling provides sentence-level context, while token-level cross-attention preserves fine-grained textual information. These two conditioning paths jointly guide a 2D U-Net to generate time-frequency images through iterative denoising. Across one synthetic and four real-world datasets, TextPlot achieves superior TS-text alignment compared with the evaluated baselines. Further experiments show average relative improvements of 79.1% and 83.5% over the evaluated baselines in following original and edited textual requirements, respectively.

Primary area:
generative models
Keywords:
text-to-time-series generation, conditional generation, diffusion models, time-frequency representation

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_JpLzOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

For text-conditioned time-series generation, existing methods perform poorly in terms of text alignment and fine-grained control. The authors propose solving this problem by using STFT to organize time series into time-frequency images and adopting an image-based diffusion framework. The diffusion framework is a conditional 2D U-Net. Moreover, the authors design a dual-path text-conditioning mechanism, including global-level and token-level conditioning.

Critical strengths
  1. The motivation is clear, and all method designs, including image-TS mapping, architecture design, and the conditioning mechanism, have their own effects and work together to form a holistic framework that matches the motivation.

  2. This paper is well written and easy to understand.

Critical weaknesses
  1. All compared methods are from before 2026. As a submission to ICLR 2027, the authors should include some 2026 baselines.
  2. Although they are not targeted at text-conditioned time-series generation, several methods have already proposed mapping time series to images for time-series generation, such as ImagenTime and TimeOmni-VL. This greatly reduces the novelty of the first contribution "We introduce an STFT-based time–frequency image as an intermediate representation to bridge the modality gap between text and time series."
  3. The architecture design might have limited novelty. Token-level conditioning is standard cross-attention that has been widely used in image-generation models. Also, the 2D U-Net seems to have nothing special.
Questions for authors
  1. Why are only VerbalTS and T2S compared in Section 4.3? And why does it only report results on Synth-U and ETTm?
  2. Why does the naive retrieval baseline outperform all methods on AirQuality in terms of FID and J-FTSD?
  3. Comparing Table 1 and Table 3, delay embedding and Raw 1D already outperform the compared methods for CTTP on Synth-U and Traffic. The authors may add more analysis on this.
Reviewer_NTjXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

A paper proposes a diffusion U-Net-based model for generating TS from text. The use the STFT time series representation for time-series generation. The text descriptions are inserted are fused into features at every residual block as well as in cross-attention layer at the U-Net bottleneck. The model is tested on multiple real-world datasets + paired editing test (change requirement in text and see how model responds).

Critical strengths
  1. The representation ablation is useful. Comparing STFT with a 1D U-Net and a delay-embedding version under the same conditioning gives some real evidence that the input representation matters (at least in this case).
  2. Consistent results across datasets, easy to read, well documented implementation details.
Critical weaknesses
  1. Why STFT help with text alignment? It is clear to me from the results that it helps, but I don't get why exactly. Is it because it is easier for model to understand ts from spectral image? More theoretical grounding would be nice. Otherwise the novelty is limited as spectral representations of TS (including diffusion) are pretty well known.
Questions for authors

Why STFT help with text alignment? Why stft is better than other representations?

Reviewer_hTj6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces TextPlot, a framework meant to generate time series conditioned to a given text. The core is to transform the time series into a complex STFT spectrogram, use U-Net 2D for the diffusione and get back to time by means of iSTFT. This approach is integrated by a cross-attention mechanism on the text.

Critical strengths
  • The use of the STFT is sound to preserve the magnitude and to manage time series with local structure and frequencies
  • The use of a wide range of datasets for comparison
  • the paper is well written and the solutions is well described
Critical weaknesses
  • the novel content seems only incremental. The use of STFT and U-Net for time series generation is not novel. The paper claims to introduce a semantic bridge between text and data but this is not directly demonstrated. How would you address this point?
  • It is not clear to this reviewer the reason why you used "pointwise lower-median aggregation". Could you please clarify this point?
  • The cost of what proposed is not compared (in terms of latency, budget, stability) w.r.t. to 1D solution. Please address
Questions for authors

See previous box

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.