A diffusion model learns by watching pictures get gradually buried in random noise, like a photo fading into television static, and then learning to reverse each step. To make something new, it starts from pure noise and cleans it up a little at a time until an image appears that matches the prompt.
Many AI image and video generators use diffusion or a close relative. A 2020 paper by Jonathan Ho and colleagues made the approach practical for high-quality images, and it now also appears in some world models and robot models.