All language subtitles for 2. SDXL ControlNet Canny

Now we've got here a workflow

For the control net

And for this one we've got

What's a candy which allows us to extract

Connect an image County image

And then we have

The outputs one output

Here is the

Unrefined and then we have refined one

So this is a fairly complex workflow

I should explain a little bit about before we explore some other details

But before we do that let's go and take a look at some of the images that I

So essentially what starting off with this image here

A woman sitting on the ground

Computer

On her

Making some sort of goofy sign

And we managed to get

For the

Up first render

One with the refined so you can see the refined is not actually adding very much at all

And then refined one

Then this one which is a tech girl

Supposed to be wearing cropped jeans

Anna is giving her a denim top which I suppose it's

Close enough to a crop Jean

And then another version with a different seed

From the

Tech n

Genre to an arm

Woman someone who is

Ready for battle

And then more ready for battle

Females but this one I think is actually a little bit more glamorous she's not only ready for battle

Who's looking glamorous with it

And I just did the candy a little bit and we got a few more

Including this one I think this one was

Pretty interesting because it was pretty much the same prompt as the other ones

And then I tried a different model

Which is different from the one that we looked at before

And sometimes it gave us black images

But we equally got something that

Pretty decent from the

Who won the we going to look at now

This is the original image and

File that we're working with

Actually a small version of the larger file the 5 GB

Are you blind

Then we're working with

So you may want to try some of these smaller ones it's actually experimental ones

And I want to just give you some kind of guidance as to where these ones are found because they're still experimental

And then not available as part of the mainstream

SD

What's 11.0 this is the one that we

Download it and we're going to use

What's the score of the most

But there are new ones

Came through and these ones are being

List separately

So I will give you a link to this page which is the page for the diffusers

And what they saying here is sdx air control and as we are training controllers for St for stable diffusion EXC

Small size control

Will experimenting with training spoiler controllers that can still retain the control capabilities

This is the

Would like you to give them feedback and the files are listed down here

Now the files are much much smaller so instead of 5 GB with 320

545

And we've got 320 and 545

So as you can see we've got both Kenny and

So you can try using

Play the depth map of

The depth map control

As well as the county for sdx 0.5

Nice by no means guarantee that they will work but the number that

Successful

Will increase over time

Not going back to where we were I want to just go through this particular workflow just explain

Play how everything is lined up

Broadly speaking the workflow runs from

Left to

So we start off by

Calling up the

What's safe tensor for sdx

And then we

Open up this image

You can

Choose your option here you can choose whatever image

This particular one will remember a list of images that you've chosen before so if you're working with several images it will

Give you an option to pick

I previous version previous image that you used before

We've got the candy here which is extracted very quickly and we can play around

With the threshold

For the Kenny low threshold high threshold

These are work differently from the ones that I gave you for the Laura file so this is

Slightly different setup

With different thresholds

A different maximum different minimum thresholds

I think with the Laura ones I think the maximum was $255 in the minimum was 0

Higher than 255

For the high threshold

The apply control net gives us a strength so we can basically

Increase that I think to about

Am I able to go higher can't quite remember

And then we have the

Very significant role as you can see

Particularly with the strength of the

Play control net

Just said to 9.75% feel free to experiment with that if you want to

I'm going for 10:24 by 10:24 and all of the images that I chose to use

What's 1024 * 1024

So I just added them inside of Photoshop just to get them to the to the same size

You can go above 1024 by 1024 but

At some point things Begin to Fall Apart

In sdx

I think

For maybe a widescreen maybe about

What's 2 megapixels

I think that's about the maximum that I've managed to get working

Without things falling apart

4 megapixel 248x 248

2048 by 2048

It's not going to work very well for you you kind of need to stick around that one megapixel

To two

Two megapixel maximum

Things to work well

Even before 2 megapixels

But what we have here is the size of the image

We've got

Clip text

Encoding the negative

Play

Positive

What is split into two so some of you might not have seen this before

But the eclipse texting code for sdx

Can you double

And

Quit texting code

It might not come up immediately so you might need to

Search for

And we've got one for

St XL the base one and one for the refiner which I'm using

So this is

I'm much more complex situation than we seen before

Leave there's a deaf

I wouldn't recommend going below 1024 by 1024

You can experiment with going up to say 2048 by 2048

Sometimes I've got good results that way but you can leave these at default if you want to

No these two sometimes people

There's no strong guidance as what to do with the the two settings here

This one here is the G setting for the positive prompt this one is the L setting for the positive prompt

So there

Play neural network

STX L uses

The G network is the larger one and is the dominant one

And L Network

What is the less dominant one is a smaller

Neural network and it's trained unless data

What I would do is

Probably

Put the most important

I promise in the G network in the top Network

And then

What I tend to do is to either put identical ones in the lower one

Or I will put

Things to do with style

Context

Not the most important

Did the most important

Always go in the G

Sometimes the important ones will go in the L

And what I tend to do is to

Put within the L style related so have put detailed textured on a studio lighting luxur

This one here is the inferior one it doesn't always influence the image

What's the sometimes it does

If you definitely want something to

Influence

The image you want to put them in

You wanted a very detailed image

Absolutely certain

You want to put that in both of them

When it comes to the

Now let's move on before coming to the other clip clip Network

We've got a case sampler

Standard K sampler where we ended 40

So that's a fairly high setting for

The

Control Netflix because

They tend to be very resource-heavy they will

Stop all your graphics processing

Play capacity

And 40 is fairly high but I think for this

The powerful control

What you really want to do is to use it to get the

What does professional looking

Images

So that probably means working with the large control net I mean I'm working here with a smaller control in it cuz I was testing it

But you probably want to work with the

A main control Nets the 5 GB

Well if you

Really run that one you want to run the smaller ones

But it may be better in many situations to actually go with the control laws because they do more or less the same thing

Now we have

The second part of the render

What comes on from the K sampler we connect to an advanced case sampler

And this one is running from 40 to 45

End

Here which is found in the advanced case sampler is 10000 which you're never going to reach

Determines where you end is the steps so we go from 404

At 2:45

That's where we end

We've got a different seeds which is I find important to have different seeds for the

The first one and the second case sampler

Add noises enabled which is important as well

As far as the

Checkpoint that I know is the sdx L1 from

Stability AI

I don't think it's perfect it's not

It's not brilliant but

Contributing very much to this image to this

Now

Quit texting code sdx

Refin

This is just a single box and what I've done here is to put the details from the L

Section from the lower section

Less important

And it doesn't contain details about what I want to see you just

It contains details about

The way I wanted to render so I don't want

I don't want

Refiner doing too much to create new detail I want to just refining the detail

That's why I've chosen

Particular approach

Sometimes it makes sense to actually put the entire

The entire

From both of these

Into the refiner

But for this one I found that the refiner was doing weird things which I didn't want

So I limited the number of steps

And I just gave it a little bit of a detail

Play instruction

You know it's just not

Elements and I want to see you but the style of the elements

Does less harm sometimes you can actually be

Can I see damage the image that

Who created with the

What is

So here I'm just giving it very limited instructions very limited activity

And also we've got some

Some of the same negative prompts

One of the things

You find here is I've changed the within the height numbers here to 4

Add 4096 I need to go up to 8 1 9

So what kind of working in multiples of 1024

Which is the

The size of the image

That seems to work out fine I don't think

Is a huge deal

Whether you change these numbers

I believe they need to go up in multiples of 64

For the numbers to be valid

496 and 496 it doesn't

Seems to

Any home and it might actually do some good

Good to have higher numbers certainly I found experimenting with these numbers here

Making those 81928192 or 2048

Connect to

Play produce a slightly different image and sometimes that might overcome

Little problems I found

In some of the images of work

Complex workflow

Which of course you're free to experiment with

Play the ace score is

Something which deals with Aesthetics and six seems to be a good number

But your fridge experiment obviously with this particular

Workflow

Then we have the output for the

Base layer and then the output for the refined layer

Now that's basically how it works and what I've done is to run a few images and you can

What kind of results that you can get

With a different prompts

We actually take a look at the prompts cuz I don't think you'll get a chance to see these again we can go to the history

And let's see

So for this one with Avengers

Is Vengeance Queen Royal clothing guns and ammunition conf

Attractive a photorealistic

And it's given us this this lady here

With the with a briefcase

It got very inventive with the with the laptop and made it different things in different images

Now this one was using one of the newer models so I went for

What are the

1.0 small diffusion

Control

So these are one of the ones I was telling you about the experimental ones

And it didn't work out too well and that might have been due to the

I think that was

It's probably due to the model because

Play everything just looks slightly weird here

I can't quite remember the prompt

But you get the idea The Prompt will influence the outcome if you keep the

The control next

Lights

That will give you a little bit more freedom to come up with a different outcomes

And sometimes I find there's a lot of impact by changing the

The sampler so can you get from Beulah

Sometimes to one of the other options has a big impact on the outcome

Feel free to experiment with this one

This one can produce very good outcomes

I think it's one way if you have a more powerful computer you can produce quite a lot of different variations

And see which one works best for what you want

What I'll do is

Give you

This particular one

With the settings that we just looked at or rather

Go to the final one

We've already looked at

Feel free to play around with it and experiment