Hello my friends, welcome to
another AI for Architecture tutorial. Today we'll quickly cover Foucus,
which is an alternative to Automatic 11.11. My name's Matt. If you Google Fooocus with
three O's, you'll see this pop up. And I noticed that Olivio Sarracus popped
in as a recommendation in my video, so take that chance to say, definitely check
him out for AI generated tips and tricks. It's a great channel, taught me a lot. Thank you, Olivio. We're going to go to the main
GitHub page here for Fooocus. And if you scroll down
and install this version, it's great, but a
recommendation from Jens Anderson on my Facebook
page, he suggested that I check out the
fork, which is a separate offshoot of this repository. And this guy, Moonride303, has
created a better version, in my opinion. He's improved on it, a
couple features that, like control net, that we have
to use in architecture. So scroll down, go to
download, windows, and you're going to have a 7Z file, which is a compressed zip file. Can't open it, just download the
free version of Winrar, W-I-N-R-A-R. Download that, extract
it, put it into a directory. I keep all of my stable diffusion engines here under my C drive,
and so I have labeled this SDFoucusMRE. Go into the separate directory here, and
if you need to, you want to edit the paths. In this case, for Foucus,
you go to your paths example. Json file,
rename that to paths. Json, edit this with a text editor. In my case, I'm using, I have Visual Studio
installed, but you can just use Notepad. Go in here and change
these directories to where you can find each
type of LoRa, embedding, control net, etc. You'll notice the slash is the
opposite direction, so be mindful of that. These go opposite to what
you're normally used to in Windows. So save that, close it. That's only if you need to have
access to your other checkpoints. So paths.json. I'm going to go back to the
root and run the MRE.bat file. The first thing you'll
notice when you start. FoucusMRE in your
browser is how simple it is. The designer wanted
to have this mid-journey experience where it
was really nothing more than a simple prompt,
and you can get results right away by having a
simple prompt, something like a modern house in a
forest next to a lake with fog. I'm going to put fog again in brackets. The more brackets you have,
the more emphasis it is on that. Then I'm going to add photo detail. It'll generate, and the
default is to generate two images at once,
so you'll see that once we open the advanced tabs. We can reduce this number down to one,
and I'm going to do that for the testing. The additional prompt feature of Foucus
is this prompt expansion, Foucus version 2. That is evident when
you go to your back end. You can see that it's adding suffixes, and
I haven't figured out how it's figuring out what to add. In this tape, it's extremely
detailed octane render, 8K. When I'm rendering
something, I don't particularly want it to look like a render, so I'm going for photo realism, and
I'm not sure that that's something that we
particularly want in our prompting. You can uncheck that
and get similar results. Generally, right out the
gate, these look pretty good. We can generate again. I'll reduce this to one image and show you
the results without the prompt expansion. It doesn't change speed at all. These are typical prompts, especially when
we started using stable diffusion earlier this year. The addition of these
prompts were something people were experimenting
with that made a big difference, but I don't really
see that it adds much value now. We'll start with this image. It looks pretty cool. Just for curiosity, I ran the speed,
performance, and then custom. As you can see, they're all the exact same. It's not the compression on the MPEG
video that we're viewing on YouTube. There is absolutely no difference. The only way to get any
performance difference would be to use custom
and increase the steps. Then we have the switch to refiner. Go through some
testing, the refiner does almost nothing now with
these improved models. I'm using the base model. I believe that you might,
depending on when you're watching this
video, you may have the realistic stock photo as the default,
which is fantastic, so leave that as is. If not, you can go down
to Civit AI, download it, put it in the directory where the model checkpoints are. We can leave all this the same. That's fine. For the demonstration, I'll keep this
at 40, keep the resolution the same. We have our different styles
that we can add to our image here. This is also a prompt expansion. If we type in photo,
and I can do something like long exposure, and I'll add in nothing. I'll do photo long exposure. We have the fixed seed. That way we can see
each result from the image styles and what kind of
change they're bringing to our generated image. That's what the long exposure did. That is essentially
adding additional prompt. That's the idea there. Then we have the LORAs. You can add in your own
LORAs, which are basically like an add-on to the
base stable diffusion model that you're using here. In this case, this is a
contrast and detail sampler. You can experiment with
these values up to two. Next we have our sampling. CFG is how closely the image
generator sticks to the prompts. Higher the value, the more rigid the
encoder is to these words, these prompts. The lower the value, the more
flexibility it has with your prompts. I haven't really figured out what
base skip and refiner skip does. The examples on the
documentation really don't make it clear, so I would just leave those as blank. Your sampler can be... The
good default is the SD++. Keep that as default for your architecture. All right, the next big thing besides the
sampler, these will give subtle variations. However, the scheduler
between Keras and Exponential will give you completely different results. I really wouldn't be able to
explain the difference between them. It's just a matter of
trying and experimenting. This is the Keras, for example,
and this is the Exponential sampler. Within sampler, next we have FreeU. It's a really interesting add-on that
they've included in this version of Foucus. It is a separate paper, and I
can't understand at all how it works. However, through the
experimentation, if you increase these numbers a little bit up to, let's say, we'll just do 0.2, 1.2, you will
get... There's a little more drama and a little bit of interesting color
that seems to be added on. It's almost like you're passing it
through some type of filter in Photoshop. It's the best way I could explain it. You just have a little
more drama in the scene. It alters the image, but it also has this
incredible level of almost HDRI effect. You can play with that down here. Next thing that's important,
of course, is our control nets. They're only important if you're
going to be using an existing image. In this case, we're going
to take an old rendering. We're going to modify
it in stable diffusion. Here we load a image that doesn't
have any fog of, let's say, we'll use this. We're going to use this 1K image. This is a raw rendering. Since I already have the prompts that say
fog in brackets, purposes of visualization, I'm going to do 0.5,
which I would interpret as 50% of your prompts
and 50% of the base image. They go together and
it creates this noise matrix that they have
called a latent space. Then it's brought back and regenerated. Of course, right now, I didn't turn it on. You have to not only use it in the tab, but
you have to have image to image clicked. Let's do that again. Okay, so that's our input. Then you can see that
I've altered this image with adding, basically, adding some fog. However, we have changed a
little bit too much of the architecture. The higher we go in our noise value, the less of that base image
is going to be kept. In which case, especially for architecture,
we can't just randomly change the building. We need to hold on to the site
as the architect has designed it. That is not going to work for us. The way we get around
that is with Control Net. I go through a lot of Control Net and how
we use it in my tutorial series that I sell on my website, howlattvisual.com. If you want to see those results, this
is, I think we still have free U turned on. No, we don't. Okay, thanks so much for watching. If you want to like and
subscribe, that's cool. I don't really care. I'm not trying to be a YouTuber. I'm an architectural renderer and I also
provide some tutorials on my website. If you're interested
in learning more about stable diffusion in your
architecture work flow, you can check that out. The link's below, HallettVisual.com. You can follow me on
Instagram and Facebook. I'm usually on Facebook. That's where I post a lot of my work. Thanks for watching. Have a great day.
