Source-linked AI summary
T2S: High-resolution Time Series Generation with Text-to-Series Diffusion Models
Yunfeng Ge, Jiawei Li, Yiji Zhao, Haomin Wen, Zhao Li, Meikang Qiu, Hongyan Li, Ming Jin, Shirui Pan
TL;DR
Text-to-time-series generation lacks general-purpose captions and flexible sequence-length handling. T2S introduces TSFragment-600K and a domain-agnostic diffusion framework combining length-adaptive encoding, flow matching, and a diffusion transformer. It achieves state-of-the-art performance across 13 datasets spanning 12 domains.
Problem
Text-to-time-series generation is limited by domain-specific caption datasets and models tied to predefined sequence lengths.
Method
T2S combines TSFragment-600K with a length-adaptive variational autoencoder, flow matching, diffusion transformer, and interleaved multi-length training.
Results
T2S achieves state-of-the-art performance across 13 datasets from 12 domains, outperforming diffusion-based and large-language-model approaches.
Takeaways & Limitations
T2S provides a domain-agnostic foundation for generating semantically aligned time series of arbitrary lengths.
Abstract
from arXiv · showhide
Text-to-Time Series generation holds significant potential to address challenges such as data sparsity, imbalance, and limited availability of multimodal time series datasets across domains. While diffusion models have achieved remarkable success in Text-to-X (e.g., vision and audio data) generation, their use in time series generation remains in its nascent stages. Existing approaches face two critical limitations: (1) the lack of systematic exploration of general-proposed time series captions, which are often domain-specific and struggle with generalization; and (2) the inability to generate time series of arbitrary lengths, limiting their applicability to real-world scenarios. In this work, we first categorize time series captions into three levels: point-level, fragment-level, and instance-level. Additionally, we introduce a new fragment-level dataset containing over 600,000 high-resolution time series-text pairs. Second, we propose Text-to-Series (T2S), a diffusion-based framework that bridges the gap between natural language and time series in a domain-agnostic manner. T2S employs a length-adaptive variational autoencoder to encode time series of varying lengths into consistent latent embeddings. On top of that, T2S effectively aligns textual representations with latent embeddings by utilizing Flow Matching and employing Diffusion Transformer as the denoiser. We train T2S in an interleaved paradigm across multiple lengths, allowing it to generate sequences of any desired length. Extensive evaluations demonstrate that T2S achieves state-of-the-art performance across 13 datasets spanning 12 domains.
1 Introduction
Text-to-time-series generation remains constrained by domain-specific captions and fixed sequence lengths, despite diffusion advances in related Text-to-X domains. T2S addresses these gaps with a new dataset and domain-agnostic, arbitrary-length generation, achieving state-of-the-art results across 13 datasets and 12 domains.
- Text-based conditioning is more flexible than label- and temporal-based conditioning, while diffusion models progressively refine noisy data into high-fidelity outputs.
- Existing methods lack general-purpose, high-resolution text–time-series captions and commonly require separate fixed-length training for different sequence lengths.These constraints limit generalization, scalability, and practicality for variable-length real-world series.
- T2S introduces TSFragment-600K, a fragment-level dataset containing over 600,000 high-resolution text–time-series pairs.
- The proposed domain-agnostic model combines a length-adaptive variational autoencoder, flow matching, and a diffusion transformer to generate semantically aligned series of arbitrary lengths.
- T2S sets a new state-of-the-art across 13 datasets from 12 domains, outperforming diffusion-based and large-language-model baselines.
2 Definition and Dataset
The paper defines text-to-series generation using captions at point, fragment, and instance granularity, then develops TSFragment-600K to expand fragment-level supervision. Its dataset pipeline produces natural-language descriptions focused on local temporal behavior and selects semantically coherent captions.
- 2.1 Problem Definition and Notation: Point-level captions annotate individual time points, fragment-level captions describe contiguous non-overlapping segments, and instance-level captions summarize the entire series.
- 2.1 Problem Definition and Notation: Text-to-series generation learns a model that generates time series conditioned on textual guidance and preserves semantic alignment with that guidance.
- 2.1 Problem Definition and Notation: Fragment-level descriptions balance fine-grained point annotations with holistic instance descriptions by capturing local trends while preserving contextual relationships.
- 2.2 TSFragment-600K: The captioning pipeline segments each series into contiguous fragments, uses expert-curated seed prompts, and applies a token limit to balance informativeness and brevity.
- 2.2 TSFragment-600K: Five candidate captions are generated per sample, and the caption with the highest average embedding similarity is selected for semantic coherence.
3 Methodology
T2S combines flow matching, a diffusion transformer, and a length-adaptive VAE to align captions with temporal latent representations and generate time series across lengths. Its architecture uses adaptive textual conditioning, latent-space reconstruction, and interleaved training for mixed-length datasets.
- T2S Diffusion Model: T2S-DiT uses flow matching and a diffusion transformer to align caption representations with temporal latent features during denoising.Textual conditioning enters through adaptive layer normalization, while the denoiser operates on patchified latent tokens with positional embeddings.
- Flow Matching Framework: Flow matching trains a denoiser to predict transition velocity from noisy latent states and samples by iteratively solving an ODE from pure noise.The method uses optimal-transport paths and classifier-free guidance for text-conditioned generation.
- Length-Adaptive VAE: The framework encodes variable-length time series into latent embeddings with LA-VAE, then decodes generated embeddings back to the original length.The latent representation is upsampled to a fixed-size embedding for diffusion and downsampled after generation; its loss combines time-series and latent-space reconstruction terms.
- Interleaved Training: Interleaved training mixes samples from datasets with different lengths within unified iterations, improving generalization across length settings.Samples are shuffled across datasets, grouped by length within each batch, and jointly used to update the LA-VAE and T2S-DiT.
4 Experiments
T2S is evaluated across 13 datasets spanning 12 domains using reconstruction and retrieval metrics, with comparisons across caption levels, lengths, components, inference settings, and data scales. It achieves strong benchmark performance while supporting cross-length training and showing sensitivity to component choices and inference parameters.
- 4.2 Performance Comparison on Fragment-Level Descriptions (RQ1): T2S secures 14 of 18 fragment-level MSE entries and achieves 0.039 average MSE on Exchange Rate, improving over DiffusionTS by 56.0% and TimeVAE by 68.9%.The comparison covers six datasets and evaluates WAPE, MSE, and MRR@10.
- 4.2 Performance Comparison on Fragment-Level Descriptions (RQ1): Interleaved training supports cross-length training within each dataset, removing the length-specific training required by baseline models and enabling evaluations at lengths 24, 48, and 96.This design is contrasted with fully trained models that require fixed-length inputs.
- 4.3 Performance Comparison on Point and Instance-Level Descriptions (RQ2): T2S achieves top performance across all metrics on seven point- and instance-level datasets, including 17 of 18 best MSE entries and 16 of 18 best WAPE entries.It also reaches the top MRR@10 of 0.314 on instance-level SUSHI, although it slightly underperforms DiffusionTS in some WAPE and MSE scenarios.
- 4.4 Ablation Study (RQ3): Replacing flow matching with DDPM increases average error by 311.00%, while replacing DiT with an MLP increases Exchange Rate error by 877.67%.Removing text guidance increases average error by 495.13%, 327.10%, and 205.23% at lengths 24, 48, and 96, respectively.
- 4.5 Parameter Sensitivity (RQ4): Figure 4 reports higher MRR@10 for CFG scales between 7 and 10 and generation time steps between 20 and 50, indicating sensitivity to inference-stage parameter selection.On Exchange Rate, performance improves with dataset size; using 50% of the data reaches 93.8% of full-data performance, while 48- and 96-length generations reach 92.4% and 91.3%.
5 Related Work
Text-time series datasets are organized by temporal alignment into point-, instance-, and fragment-level pairs, while generation methods divide into retrieval-oriented contrastive learning and generative modeling.
- Text-Time Series Datasets: Text-time series datasets are categorized as point-level, instance-level, or fragment-level according to caption–series temporal alignment.Point-level pairs describe individual events, while instance-level pairs align captions with whole series.
- Text-Time Series Generation: Contrastive text-to-time series approaches primarily support retrieval and cannot be directly applied to time series generation.
- Text-Time Series Generation: Generative approaches provide the alternative paradigm for mapping text to generated time series, alongside contrastive learning.
6 Conclusion
The paper introduces TSFragment-600K and T2S for domain-agnostic text-to-time series generation. T2S generates semantically aligned, high-fidelity series of arbitrary lengths and performs strongly across diverse domains.
- 6 Conclusion: T2S generates semantically aligned time series of arbitrary lengths with high fidelity using TSFragment-600K, LA-VAE, and T2S-DiT.
- 6 Conclusion: Comprehensive validation across 12 diverse domains demonstrates T2S’s superior performance and establishes a foundation for text-to-time series generation.