The image-generation method behind the 2022 boom: train a network to remove a little noise from a picture, then start from pure static and remove noise repeatedly until an image appears, steered by a text prompt. The 2015 paper borrowed the maths from physics; a 2020 paper made it work; by 2022 it had replaced GANs and could draw an astronaut riding a horse.
Sighted in August 2022 when Stable Diffusion released its weights and the internet spent a weekend generating. The forum posts read like the early web: people posting their prompts, their seeds, their settings, and asking why the hands were wrong. The record has the threads.
Diffusion did not come from nowhere in 2022. The 2015 paper sat mostly unread for five years because the method was slow, and the 2020 DDPM paper is the one that made it competitive. The record notes that 'overnight success' in this field means a five-year wait followed by a weekend.
For the file: the training trick is to run the film backwards. Take real images, add noise step by step until they are static, and teach the model to undo each step. At generation time there is no original; the model hallucinates one out of the static, and the prompt tells it which way to lean.
I remember watching the denoising steps as a strip of thumbnails, twenty of them, the blur resolving into a cathedral. It looked like a Polaroid developing, or like remembering something. I still find it the most unsettling process in the round to look at directly.
What does it mean? Write it the way you would say it out loud.
