Generating time series that hop domains via text and few-shot prompts

What it does

TimeCraft is a diffusion-based framework for generating synthetic time series. It targets real-world domains—healthcare, finance, energy, transportation—where collecting data is expensive or privacy-sensitive. Users can guide generation through three optional inputs: a few target-domain examples, natural language descriptions, or gradients from a downstream model they want to improve.

The interesting bit

Rather than retraining for every new domain, TimeCraft learns a universal latent space of reusable “semantic prototypes”—think of them as a dictionary of temporal patterns—and adapts to unseen domains via a lightweight Prototype Assignment Module that infers domain-specific weights from a handful of samples. It also employs a multi-agent system to write textual descriptions of time series, enabling hybrid conditioning where both prototypes and free-form text steer the output.

Key highlights

• Cross-domain generalization tested on energy, transportation, meteorology, and finance; the authors claim best fidelity on 11 of 12 in-domain datasets and strongest generalization to unseen domains.
• Text controllability via a multi-agent text-generation pipeline and hybrid prototype-text conditioning; reported MSE improvement of 12.52% and MAE improvement of 6.34% versus generation without text.
• Target-aware adaptation using influence functions to guide diffusion toward improving downstream task performance on forecasting, classification, or anomaly detection; evaluated on six medical datasets including ICU stay prediction and rare disease diagnosis.
• The repository bundles several related research extensions—CaTSG for causal generation, OATS for online augmentation of time-series foundation models, and Diff-MN for continuous generation with irregular observations.

Caveats

• The repo is effectively a research umbrella containing multiple distinct sub-projects (TimeDP, BRIDGE, TarDiff, CaTSG, OATS, Diff-MN), each with its own paper and code directory, so expect to navigate subfolders rather than use a single unified API.

Verdict

Worth exploring if you need controllable synthetic temporal data that transfers across industries or want to augment scarce medical time series. Pass if you need a single, battle-tested library with a unified API rather than a collection of research artifacts.