Novel visually prompted diffusion model generates spatially controlled histopathology images, extends to unannotated TCGA, and produces synthetic training data that outperforms real data for segmentation models.