A major barrier to progress in medical deep learning has been the limited accessibility of data driven by concerns of patient privacy. While various techniques have been proposed to mitigate privacy risks, such as federated learning and differential privacy, they often compromise model performance and/or are often susceptible to re-identification attacks. To address these issues, we present a large-scale deep learning model capable of generating synthetic medical images from 6 specialties and 10 image types, guided solely by user-provided text prompts.