The primary AI technology for image generation today. In the training phase, controlled amounts of Gaussian noise are added to images making them look like random static. The purpose of this image destruction is to learn how to put the image back together in a slightly different form, and the randomness is the key to creating an endless variety of images, not to restore the original. In the inference stage, the diffusion method repeatedly subtracts the noise to generate a unique result. The diffusion method is all about creativity. See
image model.