SRT: Super-Resolution for Time Series via Disentangled Rectified Flow

In three linesSRT introduces a super-resolution framework for time series using disentangled rectified flow. The method decomposes input into trend and seasonal components, aligns them via implicit neural representation, and employs cross-resolution attention for high-resolution detail generation. SRT-large, a pre-trained scaled version, demonstrates zero-shot capabilities across 9 public datasets.

Summary generated by Claude — human-verified