Architecture

Diffusion Model

A generative model that creates images (or other data) by learning to reverse a gradual noising process, producing high-quality outputs from random noise.

Explained at five levels

Level 1

The AI that makes pictures — it starts with TV static and slowly turns it into a beautiful image, like magic.

Level 2

The type of AI behind image generators like DALL-E and Midjourney. It learns by adding noise to images, then learns to remove it to create new ones.

Level 3

A generative model that creates images (or other data) by learning to reverse a gradual noising process, producing high-quality outputs from random noise.

Level 4

A generative architecture that learns a denoising process over a forward diffusion (noising) trajectory, enabling high-fidelity image, video, and audio synthesis via iterative refinement from Gaussian noise.

Level 5

A latent-variable model defining a forward Markov chain that gradually adds Gaussian noise and a learned reverse process that denoises — optimized via a variational bound equivalent to denoising score matching, with classifier-free guidance steering generation.

Definitions are educational summaries. Terminology can vary by source and context.

Sources