7How does it work?

It clears the mess bit by bit

This is the machine that drew for you in chapter 1. Now let's open it up.

Push the slider. The left end is the very beginning.

The left box stays a mess for a long time. But look at the right box.

There's nothing but mess, and something like Poco is already in there.

The machine looks at the mess and guesses: "I think this is in here." Then it takes away the parts that aren't. Take away enough, and a picture is left.

Press Zoom in. The mess turns into little squares of colour.

One of those tiny squares is called a . A picture is a huge pile of them.

What the machine fixes is their colour. One square at a time.

The picture wasn't hiding underneath. It was guessed into being, every time.

A little deeper

For grown-ups, and for children who want more.

The left box is what is on the canvas right now. The right box is the machine's estimate of what the finished picture will be. A diffusion model remakes that estimate at every step.

The right box sharpening before the left is the point of this chapter. Looking at nothing but mess, the machine already has a guess, and it walks towards it a little at a time. It goes in many small steps rather than one leap because re-guessing after each small step is far more accurate.

Each square you see under Zoom in is a pixel. This picture is 128×128, so it holds 16,384 of them, and what the machine adjusts are their numbers. Each colour is three numbers — red, green, blue — so really 49,152 numbers.

We chose the phrase about finding shapes in clouds for a reason. Saying the machine clears away the mess implies the picture was hidden underneath. It was not.