Data Augmenter — SARTHAK

A data-augmentation tool that uses a ControlNet diffusion model to turn simulated biological patterns into realistic synthetic training data, basically closing the gap between what we can simulate and what we actually see in experiments.

This is a spinoff from a physics-constrained deep learning project in the You lab at Duke, where the hard problem is that real experimental data is scarce and expensive but simulated data is cheap and plentiful - they just don’t look quite like the real thing. So I built a data augmenter around a ControlNet diffusion model that takes simulated biological patterns and generates realistic, experiment-like versions of them, basically manufacturing the training data you wish you had. The idea is that if the synthetic samples are close enough to real microscopy, you can train much stronger pattern-prediction models without running a thousand more wet-lab experiments, which is the kind of leverage that makes deep learning actually usable in a biology setting.