Why do diffusion models work so well?

Google describes the operation of diffusion models. They operate by gradually contaminating the training data with Gaussian noise. This removes details in the data till it becomes pure noise. Then, it trains a neural network to reverse the corruption process.