Diffusion Models
Generating an image from nothing in one shot is hard. Diffusion sidesteps the difficulty entirely: destroy an image with noise across many small steps, then train a network to undo just one of those small steps at a time — a thousand easy problems standing in for one hard one.