From mess to picture

How does an AI that makes pictures actually draw?

Not line by line on white paper. It starts with nothing but mess.

The left box was still a mess

Back when you'd only nudged the slider.

The canvas still looked like nothing. But the guess on the right already had something flower-shaped in it.

It was guessing from mess alone.

Finding shapes in clouds

Ever looked at a cloud and thought "that one's a rabbit"?

That's what a picture-making AI does.

It looks at a mess and thinks "I reckon there's a rabbit in there." Then it takes away a little of what doesn't look like a rabbit.

Then it looks again. Takes a bit more away. Dozens of times over.

This method is called diffusion.

What it learned runs the other way

You pressed Backwards and the picture dissolved into mess.

That's what the machine did while it was learning.

Take a picture, stir in a measured amount of mess, and ask it: "what mess did I just stir in?"

Right answers stay. Wrong ones get nudged. Exactly what happened in 1.2.

Once it can spot the mess, it can take the mess out. Take it out enough times and a picture is left.

Now order one

The same order, a different picture every time

You pressed Exactly the same again and got something else.

The order was identical. What changed was the mess it started from.

Same as 4.2. An AI isn't fetching a stored answer — it's picking among the likely ones.

The same mess, different pictures

When you pressed Same mess, different order.

All four started from exactly the same mess. The only difference was the order.

Two people see different things in one cloud. Here the order decides what it sees.

That's why changing the prompt changes the picture. You saw it in words in 4.5; this is the same thing in pictures.

It doesn't create out of nothing

You pressed What the machine saw.

It looks like what the machine made. But not the same.

It didn't copy — it mixed. Colour from one, shape from another.

And here's the part that matters. Without those pictures it can make nothing at all.

Who drew those pictures?

The machine in this lab learned from 12,000 pictures we made with code. We made them, so there's nothing to worry about.

What did the real picture-making AIs learn from?

Things people drew and photographed. Hundreds of millions of them, from the internet.

So a share of what the AI makes belongs to the people who made those pictures.

Arguments about this are going on right now. Nobody has settled it.

Pictures, then sound

Could the same method make music?

That's the next chapter.