Architecture
Diffusion Model
The model learns backward: in training, noise is added to images step by step until nothing can be recognized — the model practices reversing each step. When generating, it starts with pure noise and denoises itself over dozens of steps into an image, guided by the text description. This is how most AI images are made, and increasingly video and audio too.
An example
A diffusion model begins with pure image noise — what a television shows without a signal — and removes a little noise from it in about fifty steps, guided by the text description. After the tenth step you sense a composition, after the thirtieth shapes are recognizable, after the fiftieth there is an image. It was trained by adding noise step by step to real images and letting the model learn the reverse path.
A closer look
In training, the model learns to remove one step of noise from a noisy image. When generating, it starts from pure noise and repeats this step fifty to a thousand times. The input text steers each individual step along the way.
Why it matters
Two properties matter in practice. The number of steps is a direct knob between cost and quality — fewer steps, a faster and cheaper image. And: diffusion models are not limited to images; the same idea now generates video, audio and molecular structures. Where they are used in communications, product photography or illustration, proof of provenance and labeling are the real open questions — not image quality.
Commonly misunderstood
More steps do not automatically mean a better image. Beyond a model-dependent limit, only the compute time rises.