Diffusion model
A model that starts from random noise and repeatedly cleans it up until a finished image, sound or block of text appears.
Diffusion is the technique behind most AI image generators. Training works backwards: you take real pictures, add noise until they are static, and teach the model to undo each step. Run it in reverse at the end and the model can start from pure noise and refine it, pass after pass, into something that looks like the training data. Each pass is called a denoising step.
The same idea now applies to text. Instead of writing one word after another, a text diffusion model takes a whole block of tokens and sharpens it over several passes, which means it can fix an early mistake before anything is final and can produce many tokens in parallel. The tradeoff is that each finished answer needs several passes through the network, so the method has to be made efficient to pay off. Google’s DiffusionGemma is a current example.