Diffusion Model

Take an image and add noise until nothing is left. That destruction is easy and needs no learning. Now train a model to undo one step of it. Run that model repeatedly from pure noise and an image appears — generation as the reverse of a process simple enough to write down.

The forward process is fixed and noise-only; the network predicts the noise added at a given step. Sampling is the expensive part, taking tens to hundreds of sequential passes, which puts diffusion on the opposite side of the latency trade from a single-pass generator. Conditioning on text is what turns it into an image generator people use.

Forward: q(x sub _t  | x sub _t−1 ) = N(√(1−β sub _t )·x sub _t−1 , β sub _t  I), which composes so x sub _t  = √ᾱ sub _t ·x sub ₀  + √(1−ᾱ sub _t )·ε in one step. The model predicts ε and the loss reduces to ‖ε − ε sub _θ (x sub _t , t)‖², a plain regression — the variational bound's KL terms are Gaussian and have closed forms, so what is left to optimise is a regression rather than a high-variance estimate of a divergence.