Title: TimeOmni-VL: A Unified Framework for Time Series Comprehension and Synthesis

Abstract:

Current approaches to time series modeling are characterized by a stark dichotomy between generating numerical data and comprehending its semantic meaning. Existing literature indicates that while generative models frequently depend on surface-level pattern recognition, models designed for understanding often fail to produce high-fidelity numerical outputs. Although unified multimodal models (UMMs) have successfully closed this divide within the domain of computer vision, their application to time series data has yet to be fully explored.

To address this limitation, we introduce TimeOmni-VL, a pioneering vision-centric framework that integrates time series understanding and generation. This model relies on two primary innovations:

1. Fidelity-preserving bidirectional mapping (Bi-TSI): This mechanism enhances the conversion processes between time series and images (TS2I) and vice versa (I2TS), ensuring that transformations are nearly lossless.
2. Understanding-guided generation: We present TSUMM-Suite, a new dataset comprising two generation tasks and six understanding tasks derived from time series analytics. By employing a calibrated Chain-of-Thought, TimeOmni-VL represents the first system to utilize time series understanding as an explicit control signal to drive high-fidelity generation.

Experimental results demonstrate that this unified methodology substantially enhances both semantic comprehension and numerical accuracy, thereby opening a new frontier for multimodal time series modeling.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC