Slide the noise up to turn a clean pattern into random static. Diffusion models learn to run this easy process backwards to create images.
Destroying an image is trivially easy: just add noise. Diffusion works by learning to run that process backwards.
