A new AI model synthesizes chest X-rays realistic enough that trained clinicians can't tell them from the real thing.
Researchers have released what they describe as the largest specialist generative model for chest radiography to date, weighing in at 1.3 billion parameters. It was trained on 1.2 million radiographs and clinical metadata totaling 1.6 trillion tokens. The model uses a rectified flow transformer architecture and supports controllable image generation across demographic subgroups, multiple acquisition angles, and roughly a dozen distinct pathologies. The team trained it from scratch rather than adapting a general-purpose foundation model.
The stated goal is dataset diversification. Existing diagnostic AI struggles when deployed outside the institutions where it was trained — different patient populations, different scanner settings, different acquisition protocols all degrade performance. Synthetic data, generated on demand and tuned to fill demographic or pathological gaps, is one way to patch those blind spots without waiting years to collect real cases.
The catch worth watching: a model this capable of producing indistinguishable fake X-rays is also a capable source of medical misinformation. The researchers frame synthetic data as a tool for improving robustness testing, which is legitimate — but the same capability that helps a researcher stress-test a diagnostic model could complicate how institutions validate AI-generated images in clinical workflows.