Starting from noise and denoising step by step
Diffusion models are trained using a clever trick: take a real image, gradually add random noise to it over many steps until it becomes pure static, and train a neural network to reverse each step, predicting the noise that was added so it can be subtracted back out. Do this over millions of images and the network learns a general skill: given a noisy image at any stage, estimate what noise needs to be removed to make it look more like a real photo.
Generating a brand new image reverses this process from scratch. You start with a canvas of pure random noise and run the trained network repeatedly, each pass nudging the noise slightly closer to a coherent image by predicting and removing a bit more noise. After enough steps, typically a few dozen, what began as static resolves into a recognizable picture. This is why diffusion models feel like they are 'dreaming' an image into existence rather than assembling it from parts.