Scaling Generative Foundation Models for Chest Radiography with Rectified Flow Transformers

Researchers have developed the first billion-parameter generative foundation model specifically designed for chest radiograph synthesis, trained on 1.2M radiographs. The model can generate synthetic chest X-rays with clinical-expert-level fidelity while supporting controllable generation across demographics, imaging views, and pathologies, addressing a critical need for diverse medical imaging datasets.

This advancement represents a significant milestone in medical AI by tackling a fundamental challenge in diagnostic model development: dataset diversity and generalization. Existing radiographic AI systems frequently fail across different patient populations, hospital settings, and imaging equipment, limiting their clinical adoption. By creating synthetic yet realistic chest radiographs, researchers enable the generation of balanced datasets that can evaluate model robustness without requiring additional patient data collection.

The technical achievement is substantial. The 1.3 billion-parameter model trained on 1.6 trillion tokens demonstrates how scaling foundation models—an approach proven effective in natural language processing and image generation—translates to specialized medical domains. The use of rectified flow transformers and expert-curated metadata ensures both fidelity and clinical validity, with synthetic images reportedly indistinguishable from real radiographs to clinical experts.

The practical implications extend across healthcare and AI development. Medical institutions can use synthetic radiographs to expand training datasets for underrepresented populations, potentially reducing algorithmic bias in diagnostic systems. Researchers can stress-test models against diverse pathologies and imaging conditions before deployment. Additionally, this work validates that foundation model scaling principles apply to specialized, regulated domains where data is scarce and quality is paramount.

Looking forward, similar generative models will likely emerge for other medical imaging modalities (CT, MRI, ultrasound), creating infrastructure for synthetic medical dataset generation. The approach may also influence how healthcare organizations validate AI systems for regulatory compliance, shifting evaluation from purely real-world data to hybrid synthetic-real methodologies.

• →First billion-parameter generative model for chest radiograph synthesis trained on 1.2M radiographs and 1.6 trillion tokens
• →Synthetic radiographs achieve clinical expert-level fidelity while supporting controllable generation across demographics and pathologies
• →Addresses critical healthcare AI challenge of dataset diversity and model generalization across patient populations and institutions
• →Foundation model scaling principles successfully transfer to specialized medical domains despite data scarcity constraints
• →Enables development of balanced training datasets and systematic evaluation of diagnostic model robustness without additional patient data