TimeCraft

Why we included this project

Teams that need realistic synthetic time series, because real data is scarce, costly, or too sensitive to share, will find TimeCraft useful in healthcare, finance, energy, and traffic monitoring. It is a diffusion-model framework that learns a shared set of semantic prototypes, so reusable patterns like trends and seasonality carry over to a new domain from just a handful of examples instead of a full retraining run. The standout part is controllability: instead of unconditional output, you can steer generation with free-form text prompts and domain characteristics to fit a specific scenario. It also checks whether the data actually helps, using influence functions to target downstream forecasting, classification, or anomaly-detection performance rather than just mimicking the training distribution. That makes it a solid reference if you are evaluating synthetic-data pipelines for privacy-preserving analytics, risk-free simulation, or data augmentation.