VETime: Vision-Enhanced Zero-Shot Time Series Anomaly Detection

Interactive demo: thu-sail-lab/VETime Space
Source code: yyyangcoder/VETime
Evaluation guide: evaluation/README.md

This Hub repository distributes the source code and evaluation utilities. The pretrained weights are hosted separately in yyyang0/VETime-checkpoints. The linked static Space embeds the running personal demo, which loads those weights; adding the model link does not change its inference host.

News: Our work has been accepted by NeurIPS 2026.

This repository contains the official PyTorch implementation of VETime, a vision-enhanced framework for zero-shot time-series anomaly detection.

Overview

Time-series anomaly detection must identify both point anomalies and long-range contextual anomalies. Existing approaches typically trade off fine-grained temporal localization against global pattern understanding. VETime addresses this trade-off by aligning temporal and visual representations and dynamically fusing the complementary information from both modalities.

The framework contains four main components:

• Traceable Image Conversion: converts time series into visual representations while preserving temporal correspondence.
• Patch-Level Temporal Alignment: aligns visual patches with the original temporal timeline.
• Anomaly Window Contrastive Learning: improves the separation of normal and anomalous windows.
• Task-Adaptive Multi-Modal Fusion: adaptively combines temporal and visual features for anomaly localization.

VETime is designed for zero-shot evaluation and supports comprehensive evaluation on the TSB-AD benchmark.

Repository Structure

VETime-main/
├── config/                         # Training, testing, and baseline scripts
├── dataset/
│   ├── dataloader.py               # Dataset loading and batch collation
│   ├── pre_image.py                # Time-series-to-image conversion
│   ├── Datasets/                   # TSB-AD-U/M data and file lists
│   └── TSB-AD-main/                # Bundled TSB-AD evaluation code
├── model/
│   ├── VETime.py                   # Main VETime model
│   ├── VTS_module.py               # Visual-temporal alignment and fusion
│   ├── Vision_encoder/             # Vision encoder components
│   └── TS_encoder/                 # Time-series encoder components
├── loss/                           # Training losses
├── evaluation/                     # Metrics and evaluation utilities
├── train.py                        # Training entry point
├── Test_TSB.py                    # TSB-AD evaluation entry point
├── requirements.txt                # Project dependencies
└── LICENSE

Installation

Requirements

• Python 3.8 or newer; experiments were tested with Python 3.11.
• PyTorch 2.3.0 or newer.
• CUDA 12.1 is recommended for GPU execution.

Setup

git clone https://huggingface.co/thu-sail-lab/VETime
cd VETime

conda create -n VETime python=3.11
conda activate VETime

conda install pytorch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt

Optional baseline-specific dependencies are documented in dataset/TSB-AD-main/models/README.md.

Data and Checkpoints

VETime uses the TSB-AD benchmark. Download and place the datasets under dataset/Datasets/:

• TSB-AD-U: univariate time series.
• TSB-AD-M: multivariate time series.

The expected layout is:

dataset/Datasets/
├── TSB-AD-U/
├── TSB-AD-M/
└── File_List/

Download the pre-trained checkpoints from Hugging Face when required:

pip install huggingface_hub
hf download yyyang0/VETime-checkpoints VETime.pth weight_v/mae_visualize_base.pth \
    --revision 2aba27af38b6beb4174c2616e40a5db5e1c90198 \
    --local-dir ./checkpoints

See dataset/Datasets/README.md for dataset details and file naming conventions.

Running and evaluating VETime

For evaluation, follow the evaluation guide. It explains input data, checkpoints, the fixed prediction protocol, metric definitions, output files, and differences from the upstream optimized metric helpers.

Run from this repository root after installing the requirements and downloading the checkpoints:

python Test_TSB.py \
    --model_name VETime \
    --dataset_dir ./dataset/Datasets/TSB-AD-U \
    --save_dir ./output/metrics/uni/ \
    --vetime_path ./checkpoints/VETime.pth \
    --vision_name mae_visualize_base.pth \
    --device cuda:0 \
    --num_workers 2

Use a CUDA-capable environment for the upstream benchmark pipeline. The online Space is a lightweight demo; it does not execute a full TSB-AD benchmark. No new benchmark scores or paper-reproduction claims are included in this update.

Training and baseline scripts remain available under config/; inspect their paths before use. The benchmark dataset must be downloaded separately.

Citation

If you find VETime useful, please cite our work:

@article{yang2026vetime,
  title={VETime: Vision Enhanced Zero-Shot Time Series Anomaly Detection},
  author={Yingyuan Yang and Tian Lan and Yifei Gao and Yimeng Lu and Liming An and Wenjun He and Meng Wang and Chenghao Liu and Chen Zhang},
  journal={arXiv preprint arXiv:2602.16681},
  year={2026}
}

References

• TSB-AD: The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024).
• Time-RCD: Towards Foundation Models for Zero-Shot Time Series Anomaly Detection: Leveraging Synthetic Data and Relative Context Discrepancy (ICML 2026).

License

This project is released under the Apache 2.0 License.