We introduce ICI-Time, a novel framework that reframes time series forecasting as a visual inpainting task. By leveraging the generalisation power of large vision models (LVMs), the approach converts the conventional sequence prediction problem into image pattern completion.
The key idea is to map raw numerical series into structured area charts and embed known historical points into a grid‑structured prompt. A pre‑trained vision transformer can then fill the missing region within the prompt without any fine‑tuning or architectural changes, thereby producing forecasts for future values.
Temporal dependencies are encoded through spatial layout: the time axis aligns with the horizontal direction, while value magnitude maps to vertical height. This mapping is invertible, ensuring a one‑to‑one correspondence between the numeric and visual domains.
Extensive experiments were conducted on epidemiology, meteorology, and power‑system datasets. ICI-Time achieves performance comparable to state‑of‑the‑art deep‑learning baselines and shows superior adaptability in limited‑data regimes.
The work bridges the temporal and visual domains, offering a fresh cross‑modal perspective for forecasting tasks.
Blogger's Review: ICI-Time’s strength lies in treating time series as images and exploiting powerful visual models, eliminating the need for bespoke temporal architectures. Its promising results suggest broad applicability across diverse real‑world scenarios.