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

2026-05-29 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

SRT (Super-Resolution for Time series) is a novel framework designed to reconstruct high-resolution time series data from low-resolution inputs, addressing limitations in data acquisition cost and feasibility. Unlike direct transfers of image super-resolution techniques, SRT tackles this challenge by employing disentangled rectified flow. The framework operates by decomposing input time series into trend and seasonal components, aligning these components to the target resolution using an implicit neural representation, and guiding the generation of high-resolution details with a novel cross-resolution attention mechanism. A scaled-up version, SRT-large, benefits from extensive pre-training, enabling robust zero-shot super-resolution capabilities. Experiments across nine public datasets consistently demonstrate that both SRT and SRT-large surpass existing methods across multiple scale factors, validating their performance and architectural components.

Key takeaway

For Machine Learning Engineers and AI Scientists working with time series data, SRT provides a robust solution for generating high-resolution signals from low-resolution inputs. If your projects are constrained by the cost or feasibility of acquiring fine-grained temporal data, you should evaluate SRT or its pre-trained SRT-large variant. This framework can significantly enhance data quality for analytics, enabling more accurate models and insights without needing extensive high-resolution data acquisition.

Key insights

SRT reconstructs high-resolution time series from low-resolution inputs using disentangled rectified flow and component decomposition.

Principles

• Decompose time series into trend and seasonal components.
• Align components using implicit neural representation.
• Cross-resolution attention guides detail generation.

Method

SRT decomposes time series into trend and seasonal components, aligns them to target resolution via implicit neural representation, and generates high-resolution details using cross-resolution attention and disentangled rectified flow.

In practice

• Reconstruct fine-grained data from coarse sensor readings.
• Enhance financial market data resolution.
• Improve medical signal analysis from limited inputs.

Topics

• Time Series Super-Resolution
• Disentangled Rectified Flow
• Implicit Neural Representation
• Cross-Resolution Attention
• Zero-Shot Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Conformal Prediction with Time-Series Data via Sequential Conformalized Density Regions — SCDR offers doubly robust, asymptotically valid conformal prediction for time-series, generating informative, potentially disconnected prediction sets.
• Pretraining Recurrent Networks without Recurrence — SMT enables parallel RNN training and stable long-range dependency learning by decoupling memory updates from credit propagation.
• TeCoNeRV: Leveraging Temporal Coherence for Compressible Neural Representations for Videos — TeCoNeRV enhances video compression via INRs by leveraging temporal coherence and spatial-temporal decomposition.
• Multi-ResNets for Subspace Preconditioning in Constrained Optimization — MResOpt uses staged residual networks and prioritized constraint decomposition for efficient constrained optimization.
• Self-Supervised Learning via Flow-Guided Neural Operator on Time-Series Data — FGNO enhances self-supervised learning for time-series data by dynamically varying corruption levels and using clean inputs for inference.
• Sharpness Aware Surrogate Training for Spiking Neural Networks — SAST optimizes SNNs by applying SAM to a smooth surrogate forward model, reducing the surrogate-to-hard transfer gap.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.