A latent diffusion model pre-trained on pairs of natural images and text descriptors can be adapted to generate realistic chest radiographs that are controlled by free-form medical text prompts.