All language subtitles for 2. The Diffusion Model explained

I already told you.

We work with stable diffusion.

And stable diffusion is.

It's a diffusion model.

In this video we will take a look at what diffusion models are and what they do.

So I have found a really, really nice article from medium.

This article is relatively long but we won't make it that far.

All I need is this picture right here.

Let's assume we have a big, big computer and we train our computer on images on images like this.

So we give the computer images, for example, of this beach, and we describe it with a text.

We give the computer the image and we say maybe a beach with the blue ocean, blue sky, there's some

green on the mountains and so on.

We are really, really specific.

After that we add some noise to the picture, like you see here, but we still described what's on the

picture.

So a beach, blue ocean, blue sky and so on.

More noise, same text, more noise, same text, more noise, same text until you get only noise.

In this process, the computer learns how these pictures look like.

In this process, it simply understands that the words that you gave the computer yield to this picture.

So we can reverse this.

If we have only noise and we tell the computer a beach, blue sky, blue ocean, there is some green

on the mountains and so on.

The computer can reverse this and make out of the noise.

This picture, this is really, really a cool concept.

And of course we don't do this with just one picture.

We try to give the computer every picture that we can find.

And there are of course different diffusion models.

For example, there's also Adobe Firefly.

Adobe Firefly is trained on pictures of Adobe Stock Stable.

Diffusion is open source and it's free.

Everybody can use it.

And Stable Diffusion was trained on pictures from the internet.

And because of this, we also can create nearly everything that is on the internet.

We can create even celebrities.

We can create not safe for work stuff and so on.

Stable diffusion is not restricted.

Nearly everything that is in the internet we can create with stable diffusion if we give the right prompts.

The prompts are the descriptions that we give the computer to make our picture.

And for that instance, it's really, really important to make good prompts because we need good pictures.

If we are not specific, we can create a pictures that look like this.

If we simply tell maybe a beach, we will get a random beach.

If we tell him a beach, blue ocean, blue sky and so on, we will get exactly this picture.

So a quick illustration of this process because some people like this illustration, I use this a lot.

Just imagine you lay down on the ground and you look in the sky.

Besides, you is your girlfriend or your boyfriend or whoever you want and she tells to you, can you

see this cloud?

It looks a little bit like an apple, but you don't get it.

You don't see the apple.

But then she tells you, of course, just look, here is the apple.

And then you start to understand.

You see the cloud and now your eyes see an apple because your brain is trained on apples, your brain

most likely knows how a apple looks like.

And then you see the apple in the cloud, even if there is no apple there.

And if your girlfriend doesn't say it's maybe a green apple, maybe you think of a red apple.

And that's exactly why we need to use good prompt engineering.

Because if we don't are specific, we will get random pictures.

If you want to have a green apple, you need to tell the computer that you want to have a green apple,

just like your girlfriend.

Need to tell you that the apple in the clouds is green.

If she doesn't tell you that, maybe you think of a red apple, maybe of a green apple, maybe even

a yellow apple you doesn't know.

So you need to be specific.

So in this video we took a quick look at the diffusion.

Model.

The diffusion model works simple.

It's trained on pictures and on text.

Then noise gets added.

The computer learns in this process how this picture looks like, and if we give the computer text afterwards,

it can just create this pictures because it will randomly select the pixels that are right for our picture.

And I hope this makes sense for you.

And in the next video we will take an even closer look.

Because stable diffusion is a bit special.

We can use different checkpoints, Laura's seats, and so on.