SSDA adapts vision models for time series, outperforming LLM baselines

A team of researchers from multiple Chinese universities has introduced SSDA (Spectral-Structural Dual Adaptation), a method that unlocks the full potential of large vision models (LVMs) for time series forecasting. The key insight: rendering temporal data as images for LVM use suffers from two fundamental gaps. Spectrally, the power spectrum of such rendered images is much shallower than the natural images LVMs were pretrained on. Structurally, reshaping 1D sequences into 2D grids creates spurious spatial adjacencies and breaks genuine temporal continuity, misleading the model's inductive biases.

SSDA bridges both gaps with a dual-branch architecture. At the data level, a Spectral Magnitude Aligner (SMA) applies 2D FFT to selectively enhance the magnitude spectrum toward natural-image statistics while preserving phase information. At the model level, Structural-Guided Low-Rank Adaptation (SG-LoRA) injects position-aware temporal encodings into patch embeddings and adapts attention via low-rank updates. The two branches are then adaptively fused. Experiments across seven real-world benchmarks demonstrate that SSDA consistently outperforms strong LVM- and LLM-based baselines under both full-shot and few-shot settings. The code has been open-sourced on GitHub.