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

A new framework called SRT reconstructs high-resolution signals from low-resolution time series inputs by separating trend and seasonal components before generating missing temporal detail, according to a paper submitted on 29 May 2026 [1]. The acquisition of fine-grained time series data is often constrained by cost and feasibility, limiting the accuracy of downstream analytics [1]. The proposed Super-Resolution for Time series (SRT) framework addresses this by using a disentangled rectified flow approach. SRT first decomposes an input into its trend and seasonal components, then aligns them to the target resolution through an implicit neural representation [1]. A cross-resolution attention mechanism subsequently guides the generation of high-resolution details that were lost in the low-resolution input [1]. The authors also introduce SRT-large, a scaled-up version of the model that undergoes extensive pre-training. This variant exhibits zero-shot super-resolution capability, meaning it can enhance temporal resolution on unseen data without additional fine-tuning [1]. Experiments across nine public datasets showed that both SRT and SRT-large consistently outperformed existing methods at multiple scale factors [1]. Super-resolution has been extensively studied in computer vision, but the paper notes that directly transferring image-based techniques to time series is not trivial [1]. The temporal dependencies and sequential nature of time series data require architectures that respect the ordering and periodic patterns inherent in such signals. The SRT framework’s component-based decomposition and attention mechanism are designed to meet these requirements [1]. While the paper does not specify the nine datasets used, the breadth of evaluation suggests the framework is tested across diverse domains where high-frequency sampling is expensive or impractical. The submission appears on arXiv, a preprint server that hosts research across physics, mathematics, and computer science, and does not represent peer-reviewed publication [1]. The absence of author-provided code or data links in the repository metadata leaves independent replication unverified at this stage [3][4][5].