Diffusion, in plain language
How machines paint with noise.
Generative image and video models learn to create by first learning, with enormous patience, to destroy.
It sounds like a paradox, and it is the whole trick. To teach a model to paint, you first teach it to ruin a painting, one careful step at a time, until nothing is left but static.
Learning by destroying
In training, the model takes a real image and adds a little noise, then a little more, then more again, until the picture has dissolved into pure random grain. At each step it learns one thing only: exactly which noise was just added.
Generation simply runs that knowledge backwards. Start from a field of pure noise and ask, step by step, “what noise should I remove here?” Peel it away, and an image the model has never seen emerges from the grain.
The creativity is a learned sense of what to take away.
Steering the static
Your prompt enters at every step, nudging each act of denoising toward meaning: this should be a face, that should be sky. Video adds one more demand: keep the picture coherent not just in space but across time, so that frame fifty still remembers frame one.