🐏 TimeOmni-VL: Unified Models for Time Series Understanding and Generation

We present TimeOmni-VL, a unified multimodal model for time series understanding and generation. It first builds a fidelity-preserving bidirectional Time Series ⇔ Image interface for near-lossless conversion between numerical sequences and TS-images. Then, temporal reasoning from TS-image understanding is used as an explicit condition to guide generation. Experiments show that TimeOmni-VL achieves top-tier forecasting performance and state-of-the-art imputation results. Finally, these findings support an “understanding-guided generation” paradigm for future multimodal time series models.

🎨 Task Illustration

TimeOmni-VL is designed for two complementary task families:

• Time series understanding: answer questions about TS-images, including variable counting, variable localization, cycle localization, mean comparison, anomaly detection, and trend analysis.
• Time series generation: generate missing or future values for multivariate time series, including zero-shot forecasting and zero-shot imputation.

🧠 Method

TimeOmni-VL follows a unified understand-then-generate pipeline. The input time series is first transformed into a TS-image by the TS2I converter. For understanding tasks, the model directly produces a reasoning chain and the final textual answer from the TS-image. For generation tasks, the model first generates temporal reasoning as a condition, then uses it to guide the generation module to complete the target TS-image. The generated TS-image is finally decoded back into numerical time series by the I2TS converter.

📊 Benchmarks

* Note: For forecasting and imputation, lower nMASE is better. “–” indicates that the success rate is below 10%, so the result is not reported due to insufficient statistical reliability.

Table 1. Forecasting Performance on GIFT-Eval Subset

Method	Short-term nMASE↓	Medium-term nMASE↓	Long-term nMASE↓
LLMs
Gemini-2.5-Flash	1.295	1.201	1.279
Qwen2.5-Instruct-7B	1.445	–	–
Time Series-based Models
ChatTime	0.983	1.439	4.164
Time-R1	1.162	–	–
TimeOmni-1	1.298	–	–
Image-based Models
VisionTS++	0.915	0.682	0.690
VisionTS	1.263	0.763	0.794
Bagel	16.303	17.840	16.530
TimeOmni-VL	0.878	0.816	0.784

Table 2. Imputation Performance under Different Masking Ratios

Method	[0.1, 0.2) nMASE↓	[0.2, 0.3) nMASE↓	[0.3, 0.4) nMASE↓	[0.4, 0.5] nMASE↓
LLMs
Gemini-2.5-Flash	0.920	2.028	2.434	1.160
Qwen2.5-Instruct-7B	4.878	1.854	–	–
Statistics Baselines
Nearest	0.975	0.958	1.003	0.929
Linear	0.943	0.905	0.965	0.968
Time Series-based Models
Moment-large	1.220	1.400	1.630	2.100
Moment-base	1.510	1.600	1.700	2.130
Image-based Models
Bagel	17.411	12.239	11.849	11.032
TimeOmni-VL	0.713	0.757	0.842	0.927

Table 3. TS-image Understanding Performance

* Note: Scores are normalized to [0, 1], where higher is better. QA1–QA3 are layout-level tasks, and QA4–QA6 are signal-level tasks. Bold marks the best value in each column.

Method	Layout Tasks			Signal Tasks
	QA1↑	QA2↑	QA3↑	QA4↑	QA5↑	QA6↑
Proprietary VLMs
Gemini-2.5-Flash	0.540	0.640	0.004	0.535	0.000	0.342
Gemini-2.0-Flash	0.230	0.290	0.261	0.279	0.000	0.220
Base Model
Bagel	0.000	0.502	0.012	0.182	0.000	0.254
Ours
TimeOmni-VL	1.000	1.000	0.931	1.000	0.667	0.841

Table 4. Time Series Reasoning Performance

* Note: Task 1, Task 2, and Task 4 use accuracy, where higher is better. Task 3 uses MAE, where lower is better. Bold marks the best result in each column.

Method	Perception		Extrapolation	Decision Making
	Task1↑	Task2↑	Task3↓	Task4↑
LLMs
Gemini-2.5-Flash	77.5	25.9	170.78	36.6
Qwen2.5-Instruct-7B	42.8	26.3	146.12	24.9
TSLMs
Time-MQA-8B	25.1	31.2	-	11.6
ChatTS	39.2	18.6	-	11.1
ITFormer	47.5	14.6	230.04	41.7
Time-R1	34.0	31.4	160.47	32.2
TimeOmni-1	87.7	64.0	145.53	58.9
Ours
TimeOmni-VL	84.0	61.3	163.79	61.4

🚀 Usage

This repository hosts the model weights for TimeOmni-VL. For installation, inference scripts, TS2I/I2TS conversion utilities, and complete examples, please visit our GitHub repository.

License

TimeOmni-VL is released under the Apache 2.0 license.

✍️ Citation

@article{guan2026timeomni,
  title={TimeOmni-VL: Unified Models for Time Series Understanding and Generation},
  author={Guan, Tong and Pan, Sheng and Barthelemy, Johan and Li, Zhao and Cai, Yujun and Alippi, Cesare and Jin, Ming and Pan, Shirui},
  journal={arXiv preprint arXiv:2602.17149},
  year={2026}
}