Hello everyone, Rodney here at Kleebz Tech.
Today, I'm going to attempt to show you how to convert a realistic image to anime, and anime
back to realistic. This can also work with other styles as well, but anime was a suggestion from
the comments that got me to make this video, so that's what I'm going to go with. I'll
be doing this in Fooocus, but some of these techniques will work with other interfaces.
If you haven't used Fooocus, I do recommend adding it to your AI tool chest. Just download,
extract, and run, no installing python or any of that sort of stuff. I do have many other videos
on using Fooocus that you can check out as well. Now, onto the topic of this video, realistic to
AI is where we will start. So, I'm going to be using the standard run.bat file with this, but
I've already downloaded the anime checkpoint, which we'll be using. You can use the checkpoint
that is downloaded when you launch the anime.bat file preset, or any other one that you prefer.
By using a checkpoint specifically trained for a certain style, like anime, it'll make conversion
much easier. For this, we'll be using several of the image prompt options as well, so you want
to open up the image prompt area of Fooocus. Now, the first thing we're going to do is go over
to the describe section. We're going to drop the image that we're going to use, this one here. So
this is the image we'll be using for this purpose, and I'm going to change this over to the art
anime because I find this works much better for this. Then, we're just going to describe
this prompt in this image into the prompt, and once that's done, we'll have the prompt
up here. One thing I do suggest doing is going over this, making sure it got it
correct. If there's anything like the hair colors off, or you want something
different, you can change that. Like, I'm just going to remove 'realistic' from there
for now because it's not what I'm going for, but you can leave that in there. And as you
will find, depending on how you do this, text prompt may not have as big of an impact
depending on how much you use the image prompts. Now that we've done that, I can go over to
now. When you do the describe tab, one thing you want to be aware of is, you need to go over
to the Styles. It's going to actually check off certain Styles. So in this one, I'm actually not
going to check off any Styles initially. Now, the other thing that we're going to want
to do as well is go over to the model, uncheck any LoRAs we have enabled,
and I'm going to select the anime pencil model. That's the one that comes
with the anime.bat preset for Fooocus, but, like I said, you can use any of the
ones that you prefer. That's up to you. Now, we're going to want to change the settings.
So, I'm going to leave it on the quality settings because that's what I prefer my images. I'm
going to change the resolution to match the one that I'm working with as the reference
image. Now that I have all that set up, I'm going to actually head over to my image
prompts down here and then I'm going to check off Advanced because we do need the advanced
settings. And I'll go back, find my picture, I'm going to drop one into each two
of these boxes. You can actually use the face swap when you're doing these as well,
although I find unless you're doing close-ups, it really doesn't make much of a difference. For
this one, I'm not going to be using the face swap. Here, we're ready to pretty much start converting.
Now, how you do this will depend on the results that you're looking to get. Like, so for this, the
first one I'm going to be using is just the image prompt here, and we're going to leave that as the
imageprompt for selection down below, and we'll be adjusting the stop at and weight as we see fit
to generate the image that we want. Now remember, the stop at setting determines when the image
prompt stops influencing the image generation. So, with it at a 0.5, at halfway through,
50% away through the image generation, this will stop influencing that. And think of the
weight is more as the volume, so that's how much it's going to influence. So your stop at is for
how long, and then your weight is for how much. It's the same for the Pyracanny, the CPDS, any
of those things. It all works the same way. Now, the other box that we're going to
be changing is the second one. Now, we have a choice between Pyracanny and CPDS. I'm
not going to dive into the details of those in this video; I have covered them in one of
my other videos a little bit more detail. But if you want to get an idea of how these
work, well, Pyracanny, the way that works, let's get this onto image number one. I'm going to
go into our developer debug settings, and in here, there is an option to debug the pre-processors.
Now, when you do that, all it's going to do when I hit this generate button is, it's actually going
to show you the results of the Pyracanny here. So, this is exactly what Pyracanny is going
to do; it's going to create all those outlines for the image, and that's going
to be used as a guide when we generate. Now, the CPDS, basically, the way that
works is it removes the color from the image for the most part, and the result
is pretty much a black and white image, and that's used to guide the image generation.
So, each one works differently, and, in all honesty, the results you're going
to get are pretty close in most of these, but you're going to want to try each one to
see which one looks better because I found, depending on which what type of what you're using
for an image, is going to determine which one works better, in all honesty. So, it's hard to
say one is better than the other at this point. Now that we've got all those, now those are set,
I'm not going to change those at the moment. So, let's go ahead and generate an image, and we'll
get an idea of what this does. I'm going to leave these right now on the default settings, the
image prompt, and now let's go with Pyracanny, and we'll go ahead and generate.
There we go, they're pretty close, and for the most part, that gets you started
with creating the anime from a photograph. Now, you might be happy with the results that
you get immediately, but probably not. So, using the settings on default here, I
find gives it a much more realistic look, especially the background. If we
turn on one of the anime Styles, we're actually going to see a little bit
different. So, let's go in here, and I'm going to actually just turn on this anime style,
and now I'm going to do the same thing. Now, I've actually left it on the same seed so we
can compare the results that we get from that. Okay, so on the left was with just the
default, no Styles. The one on the right was with Styles. At this point, by using a lot
of the default settings, you're going to get a mix of influence from the text prompt, the
image prompt, everything in the Styles. Now, if you're going, if you want to get really,
you want to get it as close as possible, the biggest thing you're going to want to do is
increase the weight and the stop ats on these. so I went ahead just to give you an idea actually so I
don't run through all these so I did five generations
pyracanny and also five with cpds the image prompt was on the full weight and Pyracanny and cpds were on full
weight which equal one and the stop at was for one on both of these pyrocanny and CPDS the image prompt was at full weight
and what I started at was 0.2 for the stop at and then I did .4 .6 .8 and then
a full 1.0 so as you can see in here you start off you have more of the influence
from now this was with the Styles attached to it as well so you had more of an
influence from the Styles and the text prompt than from the image you still
had that influence from the image but not as strong as you would with the
higher amounts so as we increase those stop ATS so that influence
carried on further and further into the image it got closer and
closer to the original image and looking at both of these both cpds
and pyracanny are going to get pretty close to getting the similar results and
I'm going to increase these all to one just to give you an idea of what
exactly this is going to come out we'll use cpds for this one and I'll generate
this image so that's with the settings all at a much higher amount and it
ends up coming out much closer to how the original image looked
even the dress the skirt is very close now obviously there's you know
we can do other stuff as well so I'm going to show you just a quick example
with um a different one using a well let's go ahead and use so we'll go
in I'll show you the process again so I'm going to use this image
here instead we are going to describe this into the prompt okay so
that looks good everything else I do need to go over to the Styles I'm going
to uncheck these I will check off the anime one we do have a couple different
anime for this one actually let me go with the MRE anime and let's bring
that image down here and I'm actually going to bring this in for the face
swap as well I'll increase that higher weights now these I'm actually going to
leave these all at a high weight I want these to be as close as possible so
I wanted to have I want to have all these weights up as much as possible for
what I'm going to do here but like I said that depends on what results your looking for. My goal is to get it to be anime style but
look as close to the original as possible. I did a pretty decent job, in my opinion.
And that's really all there is to doing any of these. Like I said, this can be applied to
other things as well. You can apply it to comics, you can apply it to any of these things. But,
depending on what checkpoint model you use, the advantage for us using the anime
here is it does make it a lot easier to convert to anime than if we were just
using a regular checkpoint and trying to convert it to anime. I did find you're going
to get much better results doing it this way. Now, for the more interesting part that I actually
found to be more challenging, and it is, is going from anime back to a realistic image. It can be
done, but the results are a little more difficult to achieve, and the results may not be as good
as you're hoping for. It also depends a lot on the source image, how extreme or how exaggerated
the anime is versus realistic. So, for example, let's start off with the one that we originally
just went with and try to convert it back to a regular image. The same idea: we drop them into
the image prompts like I did before. I'm going to take the image and drop it into the describe tab.
I'm going to describe it into here the same thing. Now that I've done that, we are going to change
the styles. For these, I'm actually going to use these because, at this point, I want to get it
more towards the photograph as much as possible. So I'm actually going to be using the Styles,
and I'm going to go in and change the model. So, instead of using the anime model, obviously,
we now want to use an image model. I'm going to use just the standard Juggernaut. You could
use a photorealistic one, anything that handles photorealism. You're probably going to be better
off with, but you'll try different models to see which one gives different results. Juggernaut's
more of a middle ground, so that's why I'm using that one. It is a little more difficult,
but so we'll go with these as our defaults. I get the photograph in here; I put that into the
image prompt tab. Now, the difference in this is the opposite, in a sense, of what we were doing
before. Before we wanted to really increase the influence that these had on the resulting image,
I find when you're doing it the opposite direction and going back to a realistic image, you actually
want to decrease a lot of that because it's going to be a lot harder to get that image to go back
to a photorealistic one. I'll just give you an idea here of what would happen if we go in, we'll
go with the defaults on these. I think it's 0.5, and these are usually 0.5; the weight is 0.6. So
let's go ahead and see how this one comes out. Okay, so this actually worked on the default,
which I wasn't quite expecting, in all honesty. But it all depends upon the image itself on how
well these are going to work, is what I've found. So, usually, I find you're going to have to
turn these down considerably to get it to work, but this one actually worked pretty well. I would
actually probably lower the weights just slightly. I think it could be just a little better if we
were to decrease these weights a little bit. Overall, I'd say that one did pretty well,
actually. It did better than I thought it would, so I don't really even have to go
any further with that one. Let's go with this image here. So this one's
a lot more extreme. Let's go ahead and describe that into the prompt. I'm going
to go back to the image prompts. Now here, I do have to change the styles again because
it's going to change those on me. Sometimes you can actually just put 'photo' in the beginning,
although when you're applying the photo style, that shouldn't necessarily be necessary per
se. But if you're not using any of the Styles, then definitely you want to start adding something
in there to tell it that you actually want a photo of it. That would be the biggest thing. I'm
using those Styles in Fooocus to handle all that, which adds those to the prompts, but otherwise,
you're going to want to signify in this prompt that you want it to be a photograph so it
knows that. So it can try to convert it. So once this starts converting, it's going to
have a harder time because with these weights and everything else, it most likely won't
be able to accomplish this. But let's give it a shot and see what happens here. It'll
probably prove me wrong. So as we can see, ithad a hard time overcoming the extreme
details of the anime, the exaggerated looks versus trying to convert it back to realistic.
What I found here is you really need to, one, change the prompt. I brought it in using the
anime, which brings in more of an anime prompt, although it doesn't say it here. The problem
is if you go with the photographic described, it's only going to describe a girl sitting at a
desk, and that's pretty much it. So you may get better results in one aspect, but you're going
to lose a lot of those details if you do that. The biggest thing here, what you want to do
when you have that much of an issue fighting the original style, is you're going to want to lower
some of these to a lower amount and just keep trying until you get the results that you want.
A lot of times, you're going to end up with some that are going to be a mix, actually, of realistic
images like the background and even the shirt, the hands will look realistic, but the face
is still going to be more exaggerated because what happens is as we're going through, we're
going to notice the stop at when the influence stops. If the AI can then, from that point,
convert it to more of a realistic image, it's going to have a harder time with the bigger
eyes and everything else getting that down. Now this one's a little bit different; you
can actually see that the hands and things like that are starting to have more of a realistic
look. So it really comes down to, at this point, adjusting a lot of these things. Like I said, you
can adjust the text prompt. This one's actually a very difficult image, and it's actually the reason
I'm using this one because it's very hard to get this one exactly where I'd want it. So we'll lower
that a little bit more and give it one more try. So as you can see, this one's much closer to being
realistic. Obviously, it does need a little bit of fixing with those hands, but that was actually a
more difficult image, and I wanted to demonstrate that. So those, you do definitely need to lower
the weights, and you'll have to keep decreasing and playing with those until you can figure out
which one gives you the better results. On these, you definitely want to try pyracanny and cpds;
you're going to have a harder time getting that extreme of an image to look identical to the
original one. One that will work pretty well, for example, would be this one. As we can see
from this image, the features aren't as extreme, so this one should be a lot easier to bring over
and convert back to realistic. So what I'm going to do is I'll drop that, put that, describe it
into the prompt. We're going to remove that, make sure there's nothing in here to reference anime
because sometimes it will make some references in there, which will make the prompt want to do
anime more. So you want to check those out. These, I'm actually going to increase those. We have to
go back into the style because it's going to do the same thing again, and everything else is fine.
I have it on the model that I want. So these, I've got the weights higher, so on this one, the
first try came out and made it into a realistic image. It may not be exactly as close to
the original that we'd want, and for that, I would go through and just keep tweaking some of
these things. We could increase the weight. Let's just try it with CPDS to see how that one does, a
little bit different, and we'll give this another shot. So that one came out a little bit closer
because we increased the 'stop at' and weight, and that had a bigger influence. So, when
you're going back in the other direction, you're really going to want to do a lot of
tweaking with those to find the level that you want. Because to get closer to the original
image but have it still be a realistic image and not the anime, you're going to want to
keep increasing these up. And, you know, for when you first start and then adjust
them until you get to a certain point, you'll start seeing that it starts bringing in
the anime, then you can reduce those a little bit until you can find that exact match of what
you're looking for. So, that is nowhere near as easy to go from a photo to the anime than it
is, or any other style really, than it is to go from those styles back to the realistic. But
it can be done, and they do come out pretty well. Now, there is another method that somebody
actually recommended when they were commenting. They talked about trying this differently, so I
gave it a shot, and it actually does kind of work, although it does depend on the image
itself. I didn't find that this method worked well for the more extreme ones. So, for
something like this image here, I didn't find this really worked at all. But when it came to
simpler images, it definitely would work. So, this one, the idea was so I'm going to stick
with the same image, so the prompt is already in there. So we do the same thing: you put
it in the describe tab prompt, and then the suggestion was to put that in the main tab. So
I'm going to take that image, the original image, now they said to basically mask off the whole
image, so we'll go ahead and completely mask that. Then the suggestion was to turn this to
'improve detail,' leave this prompt empty because it's actually going to use the prompt that's up
here; that'll actually work, it doesn't matter if you take that and put it down here or leave it up
here, you're going to get the exact same results, so that doesn't really make a difference. But
they did say to go to 'improve detail,' we'll go ahead and give that a shot. It didn't really
do it exactly, you know, didn't change anything, but I did play around with this a little more. I
don't know, there are a couple of images I tried, and this actually worked just by default without
adjusting this at all. But for a lot of them, I found I had to go in and actually, and now you
can do this, I did experiment with just a regular in painting, but I did find this works better, but
then to change the denoising strength because if you increase this a little bit, it's going to use
less of the original image when it regenerates. So we'll go ahead and generate that now, and we
should see a little bit better result out of that. So, as we can see, that actually did pretty good
there, compared to just the default. So you can try this method. I've, you know, it really comes
down to each image seems to be different on how which method works better. So a lot of times, you
try I find the image prompt, the first method I showed, works the best overall in general, but
this one actually does work in some respects better for certain images, especially if they're
not as extreme images. This one does pretty well, but you I did find most of the time I did have
to go in and increase that denoising strength. If I didn't, then a lot of times it didn't
necessarily get the results that I wanted. Let's look at that and compare that; that was the
one that we generated using the image prompts, and in my opinion, I actually think that's the
better one here. But like I said, each image is different. Those are basically the couple of
different methods that you can experiment with when it comes to bringing from the anime back to
the photo-realistic style. I would do the image prompts first, adjust those levels if you don't
get what you want, then try the in-painting method that I just showed; that can work as well. But
don't forget, if you're going to use this method, you probably want to go into the developer/debug
mode and then increase that de-noise strength a little bit up; you might not have to go too high,
it really depends upon the original image. The more extreme it is, the more exaggerated, then
the harder it is, and you're going to have to set that Denoising strength higher. It's almost
like turning the weight and everything down on the image prompts is the same idea as increasing
the denoising strength. That should give you a couple of different tools if you're looking to
do that. Like I said, it's a lot easier to go from realistic to anime than it is to go back the
other direction. If anyone has any suggestions for other ways of doing it or ways of improving
the results, please leave a comment because I actually do learn quite a bit from the comments.
If you found this video helpful, please consider hitting the like button because it definitely
does help. I'd also like to thank anyone who's bought me a coffee; it's really appreciated.
And that's all for now, and have a great day.
