A diffusion model learns to clean up pictures by seeing them buried in noise, at every level from none to all. Write blend(picture, noise, noise_weight): mix a 16-pixel picture with noise, putting the noise in at exactly noise_weight (from 0 to 1). How much of the picture to keep is up to you. The picture and the noise each have a spread of 1 (“How spread out?”) and have nothing to do with each other; the blend must keep a spread of 1 at every step, so the model always sees inputs of the same size.