Hey Ilan, you know what's wrong with AI generated images?

They're they're they're too perfect.

So what happens is that if you want something that's a bit more

like artistic or realistic or has some blur, has some imperfections to it.

those models, they don't want to do that.

They really don't want to.

And uh so

recently a new model came out called Krea 2 and not only is it lightning fast and can run
locally um but it's great at these more realistic kind of images.

It turns out that it's harder to be worse.

It is harder to be worse.

Okay, so take this image for example.

This was generated by Krea, and I did it locally on my machine.

And you'll see that uh there's like the the the Fisheye lens, so you see the distortion of
it all.

but it still looks pretty darn realistic,

so let's dive in to this model.

Okay, Krea 2 is made by a company called Krea out of uh California.

And they recently open sourced this.

so this is a model that any of us can just download and run locally on your machine as
long as you have a video card with at least 16GB of VRAM.

Got it, so still not on my MacBook Air.

So not your MacBook Air.

Yep.

Alright, first things first, hop into ComfyUI.

If you don't have it, download it, follow the tutorials you find elsewhere online on how
to do this.

Then come over here to templates and then uh just search for Krea.

And don't confuse it with Korea.

Yes, yeah.

If your muscle memory types in Korea by accident, you're not alone.

Um so make sure you select this one.

Uh there's another one floating around where you're using the uh API that uses the cloud
server.

Uh so don't do that.

Uh choose this option here.

And uh so after you select that template, you're gonna have these uh missing models
errors.

And so uh all that you gotta do is well download them.

So you can click on the download all or download them one at a time and pop them into the
correct folder.

so over here you'll see where they ought to belong over here, and there's links here as
well.

But of course, if you just click on download, you'll get it get it right from the correct
spot.

Right.

Now I got a

considerably easier since we first started showing ComfyUI.

Yeah.

It's like uh it's like they're improving their products or something.

Mm-hmm.

Alright, so once you've got your files downloaded and you have them selected either in
that error box or down here there's a few more setup steps that you might want to do.

So first, this workflow comes out of the box with a prompt enhancement uh segment.

so what it'll do is if you give it uh a prompt, maybe like a lazy prompt, it'll send it
over to an LLM to make it a lot nicer and then generate the image based off of that

prompt.

Is it the LLM it sends it to local?

I think it uses one of the local models here that's uh that is part of the files that you
gotta include.

I think it's this Qwen model.

Hey, that's super cool.

yeah, it's pretty neat.

but it does slow down the workflow.

And I want to show you how just how quick this is.

So I'm gonna turn that off.

The other thing to note here is that there is a resolution selector.

So um unlike other workflows where you need to just type in the exact pixels, you can
choose, hey, do you want it to be a portrait photo or do you want it to be portrait

widescreen and so on.

So let's let's choose portrait standard.

Yeah, I think it's great not having to know as the user how many pixels am I supposed to
make my picture to make it a widescreen portrait.

Yeah.

the other variable here is the number of megapixels.

and so if we we just start with the one, we'll end up with a s like a 720p um image and we
can go pretty high as well.

Well we can go up to like 4K.

Okay, so uh what I'm doing is I'm throwing in a prompt here to hey, I want to make a photo
of like a a sedan drifting.

All right, so let's go ahead and uh generate an image.

So it's gonna be an portrait and just one megapixel.

And we can already see just how fast this thing is.

Um you know, the uh the ability for this to generate images rapidly is on par with the the
the paid models.

Alright, so uh hey, we've got this this image generated pretty quickly, and I mean because
it's so fast, I can just keep rolling the dice until I get one that I want.

But I mean that's pretty that's pretty good, I would say, right?

we can see the imperfections show up here, right?

Let me open this image fully.

Right?

Like we can see some of the the fact that it's like not not everything is fully in focus.

Like you know, there's lots of blur around here.

Um so it's pretty good.

We we can crank up the uh the megapixels.

So four megapixels now.

And let's see how long this takes.

Longer.

We're gonna make the audience wait with us.

ha ha ha

I wonder if it's it would be four times longer.

What if it scales linearly like that?

I would guess no, I would guess that it's exponential.

Something tells me.

Alright.

So here we have a 4K image.

let's open this up.

There you go.

Lots of clarity there.

Um and again we're getting a pretty good image here, a believable image.

Let's let's try something else.

All right, let's try this with text.

So I'm telling it to uh make a whiteboard with some text on it and it's gonna be written
in cursive.

Did you add to your prompt please garble it?

Okay, so you can see here that it does lettering pretty well.

As you scroll down here, it's all written nicely.

Hm.

That's very cool.

Yeah.

So I think it was one of those things where uh it was the GPT image generator, that "Oh
my, they they they fixed the text issue." but with this model, again, locally, you can do

the same thing.

Okay, so in Krea 2, the workflow also has LoRas that you can add to them.

And so LoRas are things that will help the model become a lot more overindexed in a
particular direction for a particular image style.

And so here on the bottom part of this node, there is the enable LoRa toggle.

So we're gonna turn that on.

And the we have LoRas that we can select from here.

Uh we're gonna set the strength to 1.

And if you don't know where to get the LoRAs, well come over here to this node and just
come to this repository, download the corresponding LoRa you want, and make sure that you

have the trigger word.

Uh so for this example, we're going to choose the rainy window style LoRa and we're gonna
see what we get there.

So just add that trigger word to this bottom text box here.

So that'll be in addition to uh your text up here.

And let's go.

very cool.

Yeah.

So I mean look at how realistic uh you know the effects are, the the the light being bent
with the water.

Uh if we didn't have this LoRA, let's ch turn that off and see the difference uh that the
LoRA has made.

Okay, so here's what it looks like without the LoRA.

And you can see, alright, you still have a little bit of the water effects, but it's just
not the same.

here it is uh without the LoRa here it is with the LoRA.

So a bit more artistic and again a bit more of the power of uh Krea 2.

Very cool.

Well David, thank you so much for showing that to us.

I think this is it's cool that this model exists, it's cool that this power is coming to
the hands of users in their local environments.

And we'd love to hear what you all are doing with Krea 2.

Let us know in the comments.

All right.

See you next time.

See you next time.

many times in getting ready for this episode I I typed Korea.

it's just like muscle memory.

my brain loads those characters and my muscles are like, I think you mean with an O.

That's funny.

Yep.

It turns out that it's harder to be worse.

It is harder to be worse.

All you need to know to validate that is look at our subscriber count.

Ha ha ha.

Exactly.

We we intentionally have fewer than one thousand subscribers.

That's that's a feature, not a bug.

harder to be worse.