All language subtitles for 3. Stable Diffusion Models, Checkpoints and Loras

We have learned that stable diffusion is a diffusion model.

In this video we will take a closer look and we need to see what Laura's are and what checkpoints are.

These two things are really, really important.

So we make this like really really simple in paint this right here.

Just think of this as stable diffusion.

So this is the big big computer.

This is just a program stable diffusion.

This is trained of course on a lot of pictures.

But this enormous diffusion model has of course more versions, more checkpoints.

And there are different checkpoints for different stuff that we want to make.

There is, for example, stable diffusion, Excel.

That's basically just a model that makes really, really good pictures.

Maybe from faces may be really, really good realistic ones.

We will use such a checkpoint.

So we have also stable diffusion 1.5.

And we have of course stable diffusion 2.0 and so on.

Just think about this.

There are a lot of different stable diffusion models stable diffusion checkpoints.

And they are always trained on more data.

So stable diffusion is the big thing.

And then we have this checkpoints stable diffusion.

Excel is for example something that makes really really good realistic pictures.

And it depends what you want to make.

We have also stuff that makes really, really good anime pictures and so on.

The important stuff is that the training of this pictures takes a lot of time and is not that specific.

Here is a big article.

We just take a quick look at this.

They want to explain to you that there are a lot of different diffusion models on stable diffusion.

So different checkpoints, stable diffusion 1.41.5 realistic version, dream shaper stable diffusion,

Excel anything.

Version three deliberate.

Version two.

This deliberate, for example, is perfect for Animas.

Stable diffusion is perfect for realistic pictures.

Also, this realistic version is good for realistic pictures.

The concept here is just that every different checkpoint is trained on a bit of different pictures.

So here we have a good example.

We have the realistic version, we have the entry and we have the dream bar.

You can see the realistic version makes of course realistic pictures.

The anything version is a bit more like anime and the Dream Sharper is also in this direction and it's

realistic painting style, so every different checkpoint is simply a different style.

And that's basically a checkpoint in stable diffusion.

The model version 1.4 is good in such pictures, and another model is better in other pictures.

Generally speaking, we need to use realistic versions and they are simply trained on realistic pictures

to get realistic outputs.

I hope all of this make sense.

Then we have this dream sharper that makes good animals and so on.

The important stuff is that all of these models so even the dream sharper and so on.

These are all ways, really, really big models.

And we only have a handful, but we have a really, really cool solution.

We have also Laura's.

There is the possibility to use Laura's and to fine tune these models even further.

Think about it this way.

Let's say we use stable diffusion, but we want a really, really specific output.

Maybe we want a picture that looks exactly the same in almost every picture.

We can fine tune this stable diffusion excel with a Laura and this works really nice.

So we work on this checkpoint and then we take a little piece of this checkpoint, this piece we call

Laura and this Laura.

We train on additional pictures and it's irrelevant what pictures you use.

I trained myself, for example, a Laura on this pictures, so it's simply my face in different poses.

And if we train a model on all of these pictures, we can create output that looks similar.

But I didn't use stable diffusion Excel.

I used a model that makes like animal pictures.

This is, for example, a picture that I have made with this model.

So first we train our model on a lot of pictures of me the.

Model or the checkpoint was stable.

Diffusion at Ram, sharper model and the low rise then are fine tuned.

Lora of myself.

That's also a picture of the model.

And this like you can see you can make different kinds of output.

And the cool stuff is we don't have to train our lowers ourself.

The training of the Lora takes a bit of time, but we can search for Laura's that we like because there

are millions of Laura's out there.

And if we take the right Laura and we take it every time, we will get consistent characters.

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

You have learned that stable diffusion is the big diffusion model.

Then there are checkpoints.

The checkpoints are something like the versions of stable diffusion.

They are of course trained on different styles.

You have a realistic one.

So stable diffusion Excel you have the dream sharper for animals and so on.

And they make really, really good pictures.

But they are really, really big.

It takes months of training for such models, for such checkpoints.

And then we have Laura's.

Laura's are smaller checkpoints that we can plug in in our big checkpoints.

And with Laura's we can fine tune our models, our checkpoints, we can fine tune our Laura's ourself.

This takes a lot of time.

Google Colab and so on.

And we won't do that in this course.

But we can also take Laura's that other people have trained and we will search.

Perfect.

Laura's.

That's basically the concept of stable diffusion.

In the next video, we take a quick look at the seat and then we are ready to go.