Demo: Clinically Diverse Chest X-ray Synthesis via Cross-Modal Conditioning

Latent diffusion models (LDMs) generate high-quality synthetic images from conditioning inputs but often face a trade-off between sample diversity and conditional fidelity, a tension that is acute in chest X-rays where subtle clinical cues must be preserved for diagnosis while maintaining variability for downstream tasks. We introduce a multi-conditional module that integrates multiple modalities into the conditioning signal and remains effective with any subset of available inputs, improving both diversity and fidelity. Notably, a classifier trained solely on our synthetic images matches the performance of a real-data baseline, indicating that the samples are both diverse and faithful to the conditioning. We further show that our method yields samples that better cover the real data distribution than strong baselines, and that combining our synthetic data with real images serves as an effective data augmentation strategy, improving both in-distribution and out-of-distribution generalization. These findings highlight the potential of our conditioning method as a data augmentation approach for enhancing model performance in other generative model applications, particularly in data-limited clinical settings.