microsoft/TimeCraft

Implements a diffusion-based framework with three key innovations: semantic prototypes that encode domain-invariant temporal patterns for cross-domain generalization via lightweight few-shot adaptation, text-conditioned generation using multi-agent systems to pair time series with natural language descriptions, and influence-guided synthesis that optimizes generated samples toward downstream task performance rather than mere distributional fidelity. Integrates with foundation models through online augmentation (OATS), causal reasoning (CaTSG), and continuous-time modeling for irregular observations, supporting flexible conditioning across energy, finance, transportation, and healthcare domains.

1,060 stars.