diffusion model
/ dih-FYOO-zhun MOD-ul /
A diffusion model is the kind of AI behind most modern image generators, and its core idea is beautifully backwards: it learns to create pictures by learning to remove noise. During training the model is shown clean images that have been progressively corrupted with random static — a little, then more, then almost pure snow — and it practices reversing each step, predicting what the slightly cleaner version looked like. Do that enough and the model becomes an expert denoiser.
Generating a new image then runs that learned skill in reverse, from scratch. You start with a frame of pure random noise — TV static — and let the model denoise it step by step. At each step it nudges the static a little closer to a coherent image, guided (in a text-to-image system) by your prompt. After dozens of steps the noise has resolved into a picture that never existed before. It is a bit like a sculptor who sees a statue inside a rough block and removes everything that isn't it — except here the block is pure randomness.
Worth being plain about how this differs from a language model and what it cannot do. A diffusion model is not retrieving or collaging existing images; it is sampling a fresh one from patterns it learned. That also means it has no grasp of facts or logic — it produces what looks statistically right, which is why generated images can have six-fingered hands, garbled text, or physically impossible reflections. It is a powerful pattern sampler for the look of things, not an engine of understanding.
Start with a square of pure TV static. Step 1: the model nudges it and faint shapes emerge. Step 10: blurry outlines of a cat. Step 30: a sharp, detailed photo of a cat that has never existed. Each step removed a little noise, guided by what the model learned cats look like.
From random static to a coherent image, one denoising step at a time.
Diffusion models generate, they don't retrieve — the cat photo is sampled from learned patterns, not stitched from training images. That's also why they have no notion of truth or anatomy, and routinely produce extra fingers, fake text, or impossible physics that look right at a glance.