From Images to Signals: Are Large Vision Models Useful for Time Series Analysis?

ICLR 2026 Conference Submission14518 Authors

Keywords: Time Series Analysis, Large Vision Models

Abstract: Large Vision Models (LVMs) are emerging tools for transferring cross-modal knowledge to time series, but this potential is not well understood. This work addresses the gap by investigating LVMs for both high-level (classification) and low-level (forecasting) tasks. Our aim is to not only assess whether LVMs can succeed, but also reveal why they succeed or fall short. Through a comparative benchmark covering 4 LVMs, 8 imaging methods, 18 datasets, and 26 baselines, we identify the strengths and limitations of LVMs, as well as strategies for adapting them to time series modeling. Our findings indicate while LVMs are effective for time series classification, they face notable challenges in forecasting - the best LVM forecaster is limited to specific model types and imaging methods, exhibit biases toward forecasting periods, and struggle to leverage long look-back windows. We hope our findings can serve as both a cornerstone and a practical guide for advancing LVM- and multimodal-based solutions to different time series tasks.

Primary Area: learning on time series and dynamical systems

Submission Number: 14518