All language subtitles for 1. Super Resolution using Latent Upscale and Generative Adverserial Networks

So let's take a look at how you can enlarge images inside of stable diffusion we're

Is it going to be looking at methods that can be used

Here in Stapleton 1.5 then what is it going to be mixing and matching

According to taste so we're going to be looking at the control net

An enlargement using a built-in method which is called latent upscale

And what's going to be looking at another way of enlarging using models so this is going to be

Feature packed video so you want to just

Grab a seat and hold on

So let me explain what we're looking at here first of all we've got the checkpoint coming in

That is the dream Point checkpoint

And that is taking us through to the

Load vae

And that's obviously the standard vae but we're now using a TI adapter why not

It's very quick

And we then have some text in almost that you seen from above forming an alien housing structure ocean

Okay so that's the description

No sailing ships we've got

An empty Layton just 512 by 512

Which is fairly standard for stable diffusion 1.5

Wake up the depth map of the depth control

Which I think we should really change to TI

And we'll call that

Apply

Number two

Adapter

So we've got the empty latent image we've got the adapter

Coming in and changing this to

Beautiful

Death mask

Is being used to create this image which is the 51255 1/2

And this image is again we're looking at an enormous that you've seen from about forming an alien housing structure

Ocean

Photographic we don't want text disgusting ugly Watermark or dark cuz I like everything's nice

We've got some ocean in the background it's looking good

When we zoom in we can see quite a little detail which is something I really like about this model and how it

And obviously stable diffusion itself

Now it might be a little bit surprising but we're looking at here we're seeing someone

Who

Obviously pretty alien

And I think it's the Alien Part in Alien housing structure

I want a futuristic

I wanted something futuristic I didn't woman who looks like an alien but I've got a woman who

Looks like an alien

And it's a beautiful structure she's got these beautiful eyebrows she's got

The things you know that

Her head definitely looks overall

The the color and shape

That we would expect from an alien

And aliens have done something beautiful with this with this little

It looks good it looks good I think we we humans will be proud to create something that looked as spectacular as

I really like that it's also respected the the idea that we have an enormous statue seen from above

It is enormous and we are actually seeing it from above it did not want to do that

With another of the images that I worked with

So we've got an image being seen above and it's fascinating to see how it's taking this image this

Deathmatch depth map and just really

Played with it to create something

Resembles the original image in terms of his overall composition

But which is really creative and very original

With an enlarged that image 1.5 times and you can see as we enlarge it we're getting

Detail more detail

We're getting we lose this with tentacles that he decided to create for a woman

She has something now that looks like a crown or something I don't know what's going on there

Are there these things coming out from her side it could be some kind of

I don't know I don't know you never know with these aliens they've got all sorts of things going on

And what we see now you can see

It's a huge structure you can see little bits and pieces of vegetation

Something that looks like maybe it could be a walkway there

And it looks to me like that they worked hard to build this this is this is a giant building then you'd see this

I'm a mile away

And then we have another one

Which is even larger one and a half times larger than this one so

I don't know you can figure out how long it is in Total 1 1/2 times 1 1/2 times

So it's now got a lot more detail it's almost as though we're looking at a whole kind of some kind of

Auditorium or some kind of

The people here will the aliens will be tiny if they were

In that area there

Please added more detail

We've got something that looks like some kind of vegetation like moss

Growing

Whole surface I don't know what that look like original it looks like it was planned here

It look like they they actually

Created some kind of

Effect where you have two colors and it looks really nice

Begins to break down here we begin to see all kinds of decay

And then

We begin to see like the you know

Nature is beginning to reclaim

This area

Her eyebrows have begun to really sort of Fall Apart

And what were these little things that were coming out from her sides these

I don't know what those are

One of them has turned into a kind of like decayed area where you can see sunlight coming through

Some kind of

I don't know deck

Dereliction I don't know what's going on there that they haven't maintained that part of the statue or the part of

Somewhat coherent but the original idea the original image

Is beginning to the original

Idea here this world built structure is we've lost it to some extent and we have something which is

Much more decay

Much less impressive

And in many ways I don't know if it's almost like it's grown in size and also

It's aged as well

Whilst it was growing

What we're saying is the creation of detail was seeing

The creation of something that's meaningful within the context of

The thing that we're looking at it it is this weird alien structure

Scene from above with the ocean around it

Is coherent but it's not coherent across the scaling so we lose something

Play the original design

It creates detail but the detail takes us away from where we were originally

Now let's take a look at how we did this this affect here

The effect is achieved by what's known as upscale latent so

Upscale method

Nearest exact you can change this to different options

What else can you buy 1.5 as I mentioned before this method allows us to use a case sampler

Upscale

Moving from 20 to 40 in terms of the steps

So the original one is the first 20 steps

And then we moved from 20 to

240 in the second one and the third one I've moved from 40 to 60 so we're just building on what

What was the original which might explain this kind of progression that we see

The this image and that image

And also from this image to that image there's a progression

But there's something of the essence of the of the idea that's retained

It is in this in this building on steps

20 to 40 and 40 to 60 that's why we see that progression

A lot of what we're seeing here

In terms of the difference is is the desire for the software

To create detail and it's doing that creation of detail

Within latent space so we're upscaling violent

And that allows it to invent detail

Using its own knowledge

Of what's happening and what's happening

Explain to it in two different ways

We have one prompt which explains what we want to see which is the control net

And then we have enough 100% is 0.86

And then we have another prompt

The word

And it's using those two to create

This structure

At every stage and the reason why it's using both of them is because

On all of these K samples

I've connected the positive

Note to the prompt

To the control net

To the negative to the positive

And also to the positive here

We have

The negative

So we could have tried something like text disgusting ugly Watermark dark

And we could have tried

I could have put in decay

Or we could have put in

Plants to try to remove some of the vegetation here

In the negative that might have worked because the negative are all being taken into this case

I'll give you this particular workflow

And it will allow you to play around and that's really the same here there's so many ways of upscaling

That you really need to play around with them to see what they can do

And there's so many different outcomes that we can get

It's a good idea to play around play around with all of the the settings here not all of them but

Is there any play around

With the sampler name so change some of these

Very much play around with the

Scheduler so exponential Cara's normal

See which one produces the better

The better outcome

So for instance

What I could do is to change this to let's say SG

Uniform

And we could take this from simple to exponential

Then we accuse that and see what happens

Oh that's beautiful

Okay okay

And this one takes quite a long time this is the one where we're going from the

Enlarged one to the ultra enlarged one

You can see that

Play probably I don't know about you but I probably wouldn't use exponential again so that kind of experimentation can

Tell us a lot about what's working and what's not working

So what I'm going to do I'm going to reload

The one we just did will change this one maybe to normal again

Because that seemed to work a little bit more

A little bit better and maybe we'll put in

Deck

And we'll run the previous one again so

I'm going to run this one again to see what it does with the new

Wording and also with normal

Now you might be wondering

What happens

So I've been incremented the

Okay so it's changed a little bit in terms of the

The noise

We're getting more or less the same outcome

No I think it is actually just the changes in the in the word prompt that created this result

So that's something that actually work better without the

Without the changes to the to the words to the

The word prompt let's go back to where we were

And let's change this to

Normal

And what will do is take a look at another option

For enlarging and so what I'll do first of all is I'll explain how we get this this kind of enlargement

What we do is that we start off with the

Output

That we have here so we have the sample the

V a e d code

And then the the outcome the the image

And what we need to do is just basically grab the latent

Push it through to a latent upscale

We getting late in upscale by

And that allows us to choose the method I'm going to keep it at that

We can take the latent and bring it to a case sampler advanced

And that allows us to set these settings so here would go from 60 to 80 or whatever whatever we wanted

And then obviously we would do the latent

The to the vaed code

And then we'll create our image preview

Preview image

Like that

And that would allow us to preview the image which is what I've got here these are all preview images

That's a preview image

This particular situation up with the upscaling

And it's a very effective way of using the built-in processes to upscale

I'm going to now delete these cuz we don't really need another upscale do we and it was getting a little bit slow

So we don't want

Push it

When we can do now is to look at another technique for upscaling another

The thing that you should be aware of is if you just shift click

A bunch of nodes

Control C

And then Ctrl shift

It copies everything with all the links in place

And then what you would do in that situation is the decide

Basically where you going to get your positive from so you could say I'm going to produce a completely new

I like that

Grab this and say quit texting code we can put in a completely new

Text we can put in a completely new prompt

Or we could connect it to the original prompto we can connect it to the

Control prompt

Go to the control conditioning

So control C control shift fee

Allows you to copy everything

With the links in place let's take a look at another technique which we could use

For enlarging and this one

Relies on the use of a model

Change this to simple I think it was actually a simple

Not normal

Let's go ahead and create this new upscaling method so we'll go to the original one cuz this one is a huge upscale

And we can actually grab

Let's see

We'll grab the image from

From the vaed code

And then add

Lotus no let me let me do the different way so we'll just

Go to a fresh ground here and double click

And we'll choose

Preview image so we put the image there

We need to now connect this to something that will create an image

And that's going to be using a model

So we'll go ahead and right click add node loaders

And we going to choose load upscale model

That is on upscale model we're going to use Let's see we can use

Will use Astro again

Now there are a lot of models that you can use for upscaling this one

You should have at least one model

In your system but you can find a lot of models online

This is an area where you really want to

Differentiate and look at how

What models work best with your own kind of workflow so for me I think we'll try

Maybe this one here

These models are quite often called Ganz or generative adversarial Networks

They're very fast so what we can do is to

This one is uploading all models so it's like one of the other uploaders and you

Have to have

The model stored away in a specific part

What's

Configure I folder so let me show you where that is

So what you do is you would

Basically

Google

For upscale models for

Stable diffusion

And then you would go and grab your models and place them inside models upscale models so I've got the three

I think one of them is being picked up from automatic 11:11 and

Once you've got the models and you probably want to pick up quite a lot because they behave differently and all of them

Most of them are quite valuable because they can produce different results in different situations

What we want to do is to

Wrap this guy here and then just bring out

Let's say

Upscale image upscale with model okay

So this is going to upscale the image using this model and then we can output

To the image here

Is a simple as that and then we need to actually grab the image and where where do we grab the image

Will we grab the image from the vaed code here so that completes the loop

We get our image from the original VA

Deck

Nope that's the wrong one let's go and grab the right one

Right at the beginning that the very first one here

So we'll grab that one and let's move this here

Then I'll preview image let's run this one and you can actually see how quick

Next day

So if you

See how quickly is running now and you'll be able to see

Bake the images I think

This one is four times

So that was super quick and you can see we have a an image which is significantly larger than that one

I don't think this is the best example what I want to do is just to run it again

Using

A slightly different

Front

And let's see if we can do a comparison of the speed

So we're now running the case sampler for the first image

Are we on and then for the second image

And then we should see how slow this one is compared

To the game so I think

Straight from here

Through to that

And this guy here is still running so you can see this one is actually quite a bit slower than the Gan than the

Ezra again

Times

This is created an image four times larger than the original

Didn't take any time at all and if we zoom in we can see it's done interesting stuff for the detail

And I'm actually going to start running another one just to see what happens this time around

But you can see that that is created in an image which I think looks better than this one

And with this new reminder you can see again we have a very different outcome here

I think

Should have more of a bit more of resembles resemblance to this one

So I'm kind of wondering what would happen if we instead of we

Took this one and

Increase it by 1.5% by say 1.2 and then 1.2

And then 1.2

Would we retain some of the original features some of the original logic

Switch off the original image overall I think the original image is look quite beautiful

But as we move down the scale here

Play move down

We begin to lose some of the coherence of the original image some of the

Beauty of the original image and it becomes a bit more like almost a software is trying to find anything

What they can use to fill the details

In the image

And it doesn't remain as coherent as it was before where is with this one with Ezra again it's actually

I think still it retains the original kind of sense of what we had there

Something which still looks similar to what we had originally

And it is so fast every time I run a new one it is just

Done very quickly

You've got a number of options there and maybe what we can do is to give this to you maybe

Play what I can do to give this

This work Flow To You

You can play around with the settings and see what the results you get

I certainly don't want to give the impression that the latent

Upscale is bad or something you shouldn't use

I definitely don't want to give that impression

Are there any situations where you definitely want to use it and it can give good results but I think that there's a little bit of food

I thought for you right there and maybe

You can experiment with many of there are literally dozens of

Models that we can use for upscaling and they some of them are fantastic and give really good results so

Let me try to experiment with the upscaling

And I think it's an important part of getting professional-looking results because obviously the 512 by 512