Diffusion Model

No jargon, no math

Training runs in two directions. Forward: take a real image and add a little random noise, over and over, until nothing but static remains. The model sees thousands of these stages and learns, at each noise level, to predict what the cleaner version looked like. That is the whole trick: a noise-remover trained on millions of before-and-after pairs. Generation runs the reverse. You give a text prompt, the model encodes it as guidance, starts from random noise and applies the remover repeatedly. Each pass removes a bit of noise and adds a bit of prompt-matching structure: blobs become shapes, shapes get edges, edges get textures. After tens to hundreds of passes, a coherent image remains. The boundary: it follows the statistics of the prompt, not your intent, and it cannot draw facts, so any text inside the image usually comes out garbled.