Notes on generative AI, applied data science, and research.
TSTR (train on synthetic, test on real) measures whether synthetic data is useful for training. TRTR (train real, test real) is the baseline you judge it against. The closer TSTR sits to TRTR, the more task-relevant signal the synthetic data kept, which is a different thing from how realistic it looks.
A classifier trained only on synthetic brain MRI reached an AUC of 0.754 on real scans, against 0.810 for the same model trained on real data. The synthetic scans carried genuine diagnostic signal, with a 0.056 gap that has to be reported honestly, and usefulness for training is a separate question from patient privacy.
A latent diffusion model compresses an image into a latent space and learns to denoise inside it, which is what makes high-resolution medical image synthesis affordable. This guide covers where LDMs earn their place, how to evaluate them with downstream utility instead of fidelity alone, and the privacy and fidelity trade-offs that decide whether the output is safe to use.