Simply, Diffusion models somewhat (and this is extremely hand wavy) work as the “reverse” of an image recognizer.
It is taught the concepts of images through image learning (build neural circuits) to detect image features. Then, you do the reverse, iterate on noise to generate features.
![[Rusty Creates] 'Artists'](https://lemmy.world/pictrs/image/ee6ccbcc-d001-4b28-a88f-d39c2b75f860.png){kind=link}
Simply, Diffusion models somewhat (and this is extremely hand wavy) work as the “reverse” of an image recognizer.
It is taught the concepts of images through image learning (build neural circuits) to detect image features. Then, you do the reverse, iterate on noise to generate features.