(IJCAI'25) T2S: High-resolution Time Series Generation with Text-to-Series Diffusion Models

✅ T2S is the first domain-agnostic framework for text-to-time series generation.
📊 TSFragment-600K is the first well-aligned, fragment-level text–time series multimodal dataset across 6 classical domains.

🗞️ Updates / News

• 🚩 April 2025: T2S accepted by IJCAI 2025!
• 🚩 May 2025: TSFragment-600K is now available on 🤗 Hugging Face
• 🚩 May 2025: Pretrained models T2S-LA-VAE and T2S-DiT released

💫 Introduction

T2S is the first domain-agnostic model that enables text-to-time series generation. It allows users—both non-experts and professionals—to generate high-resolution, semantically aligned time series from natural language descriptions.

• Application Scenarios:

1. Inclusive Data Interaction
   Non-experts can describe temporal behaviors and generate synthetic data, democratizing access to data-driven tools, encouraging broader participation in time series data analysis.

2. Rapid Prototyping for Professionals
   Experts can use simple textual descriptions to quickly simulate system temporal dynamics. This capability supports rapid prototyping and analysis of system evolution under different conditions.

3. Stress Testing
   Simulate edge cases (e.g., "an extreme surge in demand") to evaluate system robustness—beyond what traditional diffusion models can do. Note that traditional models struggle to model these extreme cases because they rely on stationary source data distributions.

• Key Components

• T2S-DiT: A diffusion-based transformer tailored for conditional generation from natural language.

• LA-VAE: A pretrained Length-Adaptive Variational Autoencoder that supports generation of variable-length series.

• Dataset: TSFragment-600K: A large-scale multi-modal dataset with 600K fragment-level text-time series pairs annotated with fine-grained morphological captions.

📑 Datasets

• TSFragment-600K dataset is available on 🤗 Hugging Face. You can follow the usage example to call TSFragment-600K dataset:

from datasets import load_dataset
ds = load_dataset("WinfredGe/TSFragment-600K")

• You have access to download all well pre-processed [three levels datasets](include TSFragment-600K dataset), then place them under ./Data directory.

Note

We also open source the dataset construction and evaluation pipeline under ./Dataset_Construction_Pipeline/.

• Dataset Structure:

Data
├─ TSFragment-600K
│  ├─ embedding_cleaned_airquality_24.csv
│  ├─ embedding_cleaned_airquality_48.csv
│  ├─ embedding_cleaned_airquality_96.csv
│  ├─ embedding_cleaned_electricity_24.csv
│  ├─ embedding_cleaned_electricity_48.csv
│  ├─ embedding_cleaned_electricity_96.csv
│  │ ...
│  ├─ embedding_cleaned_traffic_24.csv
│  ├─ embedding_cleaned_traffic_48.csv
│  └─ embedding_cleaned_traffic_96.csv
├─ SUSHI
│  └─ embedding_cleaned_SUSHI.csv
└─ MMD
   ├─ embedding_cleaned_Agriculture_24.csv
   ├─ embedding_cleaned_Agriculture_48.csv
   ├─ embedding_cleaned_Agriculture_96.csv
   ├─ embedding_cleaned_Climate_24.csv
   ├─ embedding_cleaned_Climate_48.csv
   ├─ embedding_cleaned_Climate_96.csv
   │ ...
   ├─ embedding_cleaned_SocialGood_24.csv
   ├─ embedding_cleaned_SocialGood_48.csv
   └─ embedding_cleaned_SocialGood_96.csv

🚀 Get Started

Code Overview

The code structure is as follows:

T2S-main
├─ pretrained_lavae_unified.py
├─ train.py
├─ infer.py
├─ evaluation.py
├─ datafactory
│  ├─ dataloader.py
│  └─ dataset.py
├─ model
│  ├─ pretrained
│  │  ├─ core.py
│  │  └─ vqvae.py
│  ├─ denoiser
│  │  ├─ mlp.py
│  │  └─ transformer.py
│  └─ backbone
│     ├─ DDPM.py
│     └─ rectified_flow.py
└─ evaluate
   ├─ feature_based_measures.py
   ├─ ts2vec.py
   └─ utils.py

① Installation

• Install Python 3.10 from MiniConda, and then install the required dependencies:

pip install -r requirements.txt

Note: T2S requires torch==2.3.1 .

② Prepare Datasets

• You can access all well pre-processed three level datasets.
• You can also download our TSFragment-600K data only.

③ Pretrain LA-VAE

• You can access the well pretrained LA-VAE from T2S checkpoints in the folder ./results/saved_pretrained_models/
• Running the follow command to pretrain your own LA-VAE on different datasets. For example,

python pretrained_lavae_unified.py --dataset_name ETTh1 --save_path 'results/saved_pretrained_models/' --mix_train True

For the more detailed customize, please refer to the arg description of each hyperparameter in pretrained_lavae_unified.py.

Note

LA-VAE use mix_train to convert arbitrary length data into the unified representation.

④ Train and Inference

• We provide some train and inference experiment pipeline in ./script.sh.
• [Example] Running the following command to train and inference on ETTh1.

python train.py --dataset_name 'ETTh1'

python infer.py --dataset_name 'ETTh1_24' --cfg_scale 9.0 --total_step 10
python infer.py --dataset_name 'ETTh1_48' --cfg_scale 9.0 --total_step 10
python infer.py --dataset_name 'ETTh1_96' --cfg_scale 9.0 --total_step 10

Note

You can tune the hyperparameters to suit your needs, such as cfg_scale and total_step. Please refer to train.py and infer.py for more detailed description of customized hyperparameter settings.

⑤ Evaluate

• You can evaluate the model using ./scripts_validation_only.sh directly.
• According to the configuration of inferce.py, set the corresponding hyperparameters of evaluation.
• [Example] Running the following evaluation command to evaluate on ETTh1.

python evaluation.py --dataset_name 'ETTh1_24' --cfg_scale 9.0 --total_step 10

Note

If you want to evaluate on MRR metric, please set --run_multi True in inferce.py.

📈 Quick Reproduce

1. Install Python 3.10, and then install the dependencies in requirements.txt.
2. Download the TSFragment-600K data and T2S checkpoints to ./
3. Evaluate the model using ./scripts_validation_only.sh directly.

📚Further Reading

1, Position Paper: What Can Large Language Models Tell Us about Time Series Analysis, in ICML 2024.

Authors: Ming Jin, Yifan Zhang, Wei Chen, Kexin Zhang, Yuxuan Liang, Bin Yang, Jindong Wang, Shirui Pan, Qingsong Wen

@inproceedings{jin2024position,
   title={Position Paper: What Can Large Language Models Tell Us about Time Series Analysis}, 
   author={Ming Jin and Yifan Zhang and Wei Chen and Kexin Zhang and Yuxuan Liang and Bin Yang and Jindong Wang and Shirui Pan and Qingsong Wen},
  booktitle={International Conference on Machine Learning (ICML 2024)},
  year={2024}
}

2, A Survey on Diffusion Models for Time Series and Spatio-Temporal Data, in arXiv 2024. [GitHub Repo]

Authors: Yiyuan Yang, Ming Jin, Haomin Wen, Chaoli Zhang, Yuxuan Liang, Lintao Ma, Yi Wang, Chenghao Liu, Bin Yang, Zenglin Xu, Jiang Bian, Shirui Pan, Qingsong Wen

@article{yang2024survey,
  title={A survey on diffusion models for time series and spatio-temporal data},
  author={Yang, Yiyuan and Jin, Ming and Wen, Haomin and Zhang, Chaoli and Liang, Yuxuan and Ma, Lintao and Wang, Yi and Liu, Chenghao and Yang, Bin and Xu, Zenglin and others},
  journal={arXiv preprint arXiv:2404.18886},
  year={2024}
}

3, Foundation Models for Spatio-Temporal Data Science: A Tutorial and Survey, in arXiv 2025.

Authors: Yuxuan Liang, Haomin Wen, Yutong Xia, Ming Jin, Bin Yang, Flora Salim, Qingsong Wen, Shirui Pan, Gao Cong

@article{liang2025foundation,
  title={Foundation Models for Spatio-Temporal Data Science: A Tutorial and Survey},
  author={Liang, Yuxuan and Wen, Haomin and Xia, Yutong and Jin, Ming and Yang, Bin and Salim, Flora and Wen, Qingsong and Pan, Shirui and Cong, Gao},
  journal={arXiv preprint arXiv:2503.13502},
  year={2025}
}

🙋 Citation

Please let us know if you find out a mistake or have any suggestions! If you find this resource helpful, please consider to star this repository and cite our research:

@inproceedings{ge2025t2s,
  title={{T2S}: High-resolution Time Series Generation with Text-to-Series Diffusion Models},
  author={Ge, Yunfeng and Li, Jiawei and Zhao, Yiji and Wen, Haomin and Li, Zhao and Qiu, Meikang and Li, Hongyan and Jin, Ming and Pan, Shirui},
  booktitle={International Joint Conference on Artificial Intelligence (IJCAI)},
  year={2025}
}

🌟 Acknowledgement

Our implementation adapts Time-Series-Library, TSGBench, TOTEM and Meta (Scalable Diffusion Models with Transformers) as the code base and have extensively modified it to our purposes. We thank the authors for sharing their implementations and related resources.

🔰 License

This project is licensed under the Apache-2.0 License.