Transcripts
1. Introduction to AI-based Digital Art: In the future, auditors will
have to incorporate AI into their workflow to
be able to keep up with the demand
that is put on. In this class, I want to
teach you how to make beautiful art pizza digital
art on your computer. My name is Jesper Dramsch and I'm a scientist
for machine learning. I have been working in
machine learning for a couple of years now and
I'm fascinated by it. I've talked to radiologists, geologists, and they all
ask the same questions. Artists will AI replacement
and AI will not replace. But artists with
AI will replace. So in this class, I wanted to teach you
the introduction to a fascinating new tool
where you can use sentences to generate
beautiful digital art that has not existed before
you came up with it. And I think this is
a fascinating time. The tool that I'm showing you
is two months old at the at the point that I'm
making this video came out in August 2022. And it is on the cutting
edge of our research. And our understanding has seen over 5 billion different images and understood the captions
of those and mentors. And it has cost
$600,000 to generate. And you have access
to this for free. So I hope to see you on the other side because
I will show you how to generate stunning imagery
with this amazing tool.
2. Class Project: Welcome. On the other side. I'm so happy you chose
to take this class. I had so much fun creating this. Let's talk about
the class project. This is a fairly easy one
because everything we do during this class is going to be generating really
interesting things. So if you want to
upload something cool that you made,
then please do. This can be your project,
which waving image. So take one of these images and if you feel
confident enough, also in the description, share their sentence that we
use to generate that image. Because in my experience, through collaboration and
learning from each other, we can really thrive and make possible to improve our craft and improve how we
understand these tools. If you want to post these to social media as
well, please tag me. My name is in the description
of this class as well. You can use Twitter, you can use Instagram or almost any other social
media that you like. And I love to see what you do. So make sure to post those. Creating projects is a great
way to show what you learn. In this entire thing. We're going to use a tool
called stable diffusion, which is the AI that the AI system that
we're working with. Throughout this class. I will slowly introduce you to more sophisticated
ways to interact with this and sprinkle
that in-between. Also to understand
how this works. Don't worry. It's an
introductory class. I want this to be focused on you on how
to use these tools. There's going to be no
math, no coding involved. That's going to be a different
class if you wanted. But for now, let's dive in. Let's start making our first
generative art with AI.
3. Turn Sentences into Digital Art with Prompt-based AI: So this is the first real class. And I want to generate
beautiful art pictures. All of these out of
simple language, out of just the text. And this should be
possible, right? Because we use language to describe so much
in our lives like we can describe everything
we see, hear, feel. But even the weird
dreams that have no real anchor in reality. And those funny sketches that
we think are up fantasy, we often have words to be
able to describe those. So we really want an AI to be able to take our
language as an input. The funny, the weird, all those sentences that we can generate and turn
those into images. And this year, we are at the cutting edge of machine
learning and AI right now. This year we got these
models where you can write a sentence
and get out. Amazing, really good
looking pictures. And I want to show this to you. However, I don't want to
overwhelm you with math. I don't want to
overwhelm you with code. So for this introductory class, we're going to work in a
website because this model that has been built by a huge international
collaboration, it costs $600,000 to create. It's available for free. And this is mind-boggling to me. You can just you can run
it on your computer. You can install it
and run it yourself. But this is not, this is not in the scope of this
class, maybe a future one. Let me know in the reviews or in your project if you would
be interested in that. But right now, we're just using a web interface which offers, that is incredible that
we can just use a website to use this machine
learning system, this AI to generate pictures. And yeah, I'll slowly
build this up. We'll start simple,
just go there, build something, and then we're getting better and better
at this in this class. So let's get on our computers. Remember, all the links are going to be in
the resource section. I'll show you how to
get those as well in the modern AI community. And I'm going to use AI
and machine learning here interchangeably because I don't really think this is an AI yet. This is a
machine-learning system. But I call it AI just to, just to conform to the
expectations, right? But basically we're
using a lot of Hugging Face and Hugging Face as a startup in this community, it is used by Google,
Microsoft, Facebook, Intel, all the big tech
companies that exists today. So yeah, but first things first, I don't want you
to have to pause this video all the
time and like click, Copy this and yeah, get typos. You can always go to the class. This is another class of mine. And you can go here to
Projects and Resources. And I will leave links in
here where you can find all of these websites that I'm
sending you in this class, instead of painstakingly typing
those into your browser. So yeah, you'll also be able to find
all the resources here. And when you have created
something awesome, you can share your
project right here. I always share my own project so people know what to expect. Yeah, I, I expect you to make something really
cool and this one. So let's, let's dive
right into this. We're using stable diffusion, which I'll explain later. Just take this as a, as the brand name, let's say. And what this is, it is prompt based
image generation. Now, a prompt is
just a sentence. Essentially, you can get
really creative here. The beautiful thing is you can come up with whatever you want, play around with this. This is free. There isn't a pop-up saying, oh, you used five of this. Now you have to pay. This is completely
for you to use. If you don't know
what to do yet. You can go to the
examples right here. And it's very diverse. So like a solar punk utopia
and the Amazon rain forest, Pikachu, fine dining in
view of the Eiffel Tower. All simply a cat lying on a
rug in front of a fireplace. So this should work. Then we put in generate image. We can see where in
the queue right here. So there's ten people
before us right now. It takes approximately
20 s to generate. We skip the queue.
We were lucky. And now we should be able
to see this quite soon. You can go to this website, play around with
whatever you like, get really creative and
try out different things. So we can see right
here this one, it's a fireplace, but I
would like to see more fire. It's a very Derby
cat, but I love it. Another count, the fireplace
is not on unfortunately, but it does look very cozy, although like it could
use some pupils. You'll, you'll see that these unprocessed outputs have
problems with two things. And that is faces. And also texts. I'll
show you in a minute. This cap looks well. It's definitely doing tattoo
yoga. I hope it's okay. I don't see a fireplace, but this one, I
think is very cute. The cat is cute. It looks fairly realistic. And the fireplace
is also very nice. It's a nice rug
going on right here. So if we wanted
to keep this one, we can save those as I'll open it in a new tab right now
to show you another thing. This is a 512 by 512 image, so they're not huge. But this is what we're working with in this free
online version. Now, normally, we also
have advanced options, but these are
temporarily unavailable. I am assuming this is due
to the popularity of this. I think this one is the most popular thing I've
seen on Twitter recently. This is what everyone
is using right now. So it might happen that there is a big red box popping
up right here. I'll show you a
screenshot and a second. We can also look at a
smaller version of this, also free cat sleeping on a
rug in front of a fireplace. Now this may take up to 2 min. We also see the little count on. This was the, this is basically the little sibling
of this diffusion demo, the stable diffusion demo. Dolly is another model that we'll get to know
a little bit later. And mini deli as just a way
for us to actually use it in the browser because these can
be quite chunky, quite big. But when we go back here, I promise you that
we will have a look at how it doesn't really
do great with texts. And we can simply test
this by saying, assign, saying, hello, Skillshare,
and generate image. So let's have a look. If crayon, crayon finished this, this a little bit
green, but very cute. You can see that the face
is a little bit mangled, but not the worst
nightmare material yet. Then let's have a look on
this one is quite cued. Again, face terrible. I don't know what's
going on here, but it might be a cat. Now, do we have a cute one? Not really. No. This is kinda cute. So at least is getting
the fluffiness right. But we can see that this one is performing a little bit
poorer than the big sibling. And if we wanted to
save one of these, this one was quite okay. We can also see that
these images are smaller. We have nine different ones. So it can be very nice to get a couple of different ideas. But yeah, I mean, it's a little bit freakish. I know. So just to be aware, when you're generating cats, the phases might not look
that great right now. But we'll have a look if
we can get better at this. So I'll just generate
a few other ones. This is also why you
should save these. They are generated out of
nothing, out of noise, out of the ether of
this, out over this AI. So when you rerun
the prompt here, we'll get something
completely different. Let's have a look at the signs. Well, she yellows Galileo. These show it. There is text, but it doesn't understand to put
texts into this. But I would say this is a great starting off
point for some Photoshop. So I mean, this is very
easy to fix, right? We just take this out,
take this one out. And we have a perfectly
good sign that we can use for a basis for adding
hello Skillshare to this, maybe the little logo. So, yeah, not terrible, especially if you
expect that you will have to do some editing, which most people do. Let's have a look. Do we have some cute cats going
on right here? This isn't too bad. I really liked this. I'm going to save this one. But yeah. Honestly, if you're already generating some very
cool ones right here, don't be afraid to
just head on over to the projects tab and add
it to your projects. This would be the easiest
way to complete the project. And of course, you can check
out my other courses here. But of course this doesn't
have to be your final output. But, you know, that's,
that's the first one. But if you're just
like your outputs, you can definitely
post them on Twitter. Tag me, on Twitter, on Instagram, where I'm
also just for drums. I have all of these
linked in here, but also in the main tab
of my course right here. So you can always, always find me if you want
to put them on LinkedIn, that be funny as well. But you obviously don't have
to go to LinkedIn for this. But yeah, this is your start and to prompt
based image generation, you can just go to a website, input, whatever
you come up with. And it can even be
branded things. Obviously, I don't want
to show them right here because I don't
want to get in trouble with Skillshare or in trouble with
whoever's brand it is. But play around with it. You will be surprised how well this does
on different things. Isn't this incredible? And this is two months old. This started existing
in August 2022. This is mind-boggling. Honestly, I work in this and this blows my mind that
we can do this now. And this is available
for free to play around. And yeah, we'll we'll have a
look at the license later. But this is available for
you to use commercially, non commercially, just
in an ethical way. And honestly, this is
incredible, isn't it? So we have some
very simple ones. And remember, play around
with this, figure, something really
funny, owl, weird out. Whenever I show this to friends, they can stop making it funny
Pokemon or imagine what, uh, one was really funny. Someone imagined
a cow writing and Monterey because we were just diving and we saw a mantra rays. So really play around with it. But also in the next class, we're going to
have a look at how to customize this a little
bit and write better prompts. So how can we change
these sentences to get different outputs and to fine-tune what we
get out of this AI.
4. Getting Creative with Watercolour Styles, van Gogh, and Hyperrealistic Art with AI: Welcome back. So in the last
class we had a look how we can basically form
basic sentences, how to use the website and just play around
with it a little bit. In this class, I
promised you watercolor. I promised you Van Gogh. I promised you hyper realistic paintings
or drawings rather. And how is this possible? Well, this
machine-learning system, I often don't call it a, it's called AI because that makes you click
on this class. But in my eyes, this
isn't sentient, right? So this is a system
that has seen, but this billion with a B, this has seen 5 billion images, or rather a subset
of those images that some people said were
aesthetically pleasing. And with those images, it also got a description of
what is seen in this image. So this machine-learning system is learning how
images look like. And it is learning how
these are described. So that way, it can learn how to make these images out
of our descriptions. And in those images, we have van Gogh, we have watercolor pictures. We have basically any form of photography that you can
think about and drawings. And so we can really fine tune the sentences and get a very different
outcome than before. So let's have a
look at how we can change our very simple
sentence and make it a little bit more interesting
and more on the style that we're really interested
to generate out. And now, after your
first interaction with this prompt based AI, you may think that there's no real art
behind this, right? So there's a lot of
criticism around submitting this to digital
art competitions, e.g. and like, there's
some merit in this, but we can definitely get very
creative with our prompts. So instead of just talking
about the New York skyline, which would give us a
fairly normal skyline. We can now start thinking
about how we can modify this prompt to look more
like what we actually want. So there's a lot of randomness. Of course, in this bud, we can do prompt engineering to make this look even
more realistic. So e.g. if we wanted this to
be a better picture, we can add picture for
k and also eight K. These are just little tricks that you learn when
working with this. And a way to find these out is by looking at what other
people used in the prompt. Now, looking at what other people did can
be a little bit tricky because not everyone is
producing safe for work content. So take this all with
a grain of salt. Especially when looking on Twitter for these
kinds of things, you'll probably, well, there is no safe search, so just be aware that that may not be the
first place to look. So we can see here that
this looks very much like those cheap images
that you can get on Amazon. But this one looks
like an HDR picture. This one has nice depth to it. So very clear foreground and
then the blue background. So definitely some ways to
be able to modify this, but let's, let's
think about this. We can also make this in 1920s gray-scale picture and
see what comes out of this. Now, while we wait for this, there is another nice website that you'll find
in the resources, which is lexica art. I already looked for the New
York skyline right here. And if we wanted
to find this one, we can see the
prompt right here, Bob Ross painting of
a New York cityscape. And see, see the different generated
pictures out of this, which is very nice. I think. I'm just happy
little accidents. And this is a collage of them. Now this one is quite
a bit different. You can add line brush, minimal paintings for this. I'm not sure if this is
overlaid Russian texts. I don't see it. See it. But yeah, you can change the time
to make it more wealth, more accustomed to if you
want the full moon, e.g. the reflections
here, quite nice. We could try that out. So this is very different. Just by adding 1920s
gray-scale picture. We have these very cool views of the New York City skyline. So let's see if we
can make it at night. Selecting in some liver. And that way you can, you can fine tune how you want the AI to generate
exactly what you want to, then be able to take it into Photoshop to further process it, to like clean up some of
the mess that you can see. This is interesting. I'm sure there was an
apple somewhere here. But yeah, you can see right here that this has
much more River in it. This one is quite nice, I think. Beautiful. So this is how you
can really get very, very custom with what you want. Yeah, get more into, into the style you want as well. So what happened if we
had the skyline at night? I don't want it to
gray scale anymore, but painted in watercolor. I think this is also very
good to get inspiration because we often have a very particular
image in our mind. And especially when we work
with clients or other people, then it can be
difficult to come up with a variety of
different ideas. But right here we can see four different
color palettes and also slightly different styles and how you would paint this. The paper has a very nice green. Here. We can see the
reflections as well. And how this is done. And this obviously
very different. I'm still very nice paper grain. We can see if we
can maybe change this on linen texture. Maybe that works. That way. I get like
more diverse idea before we come up with different
pictures that we present. Our clients are, well, if we're doing this
for ourselves, we can definitely get different ideas for
this and use this as an inspiration without
the fear of accidentally copying something that
we like a little bit too much, which it happens. But it's not that great. But yeah, I'm here. We can see a little bit more
of that texture as well. Yeah, this is nice. This is how you can get easy inspiration
for your watercolor. Now. Well, let's, let's move away from the New
York skyline because we've had a lot of this now. So how about we talked, we have a look at the catcher
in the Amazon forest. Now this has got an acute, we have different styles. Again. This looks like
something you could find on a travel
Instagram, to be honest. But what if we have this
in the style of Van Gogh? I think it is already pretty
amazing that we can get pretty good pictures of Pikachu and prolly other
Pokemon as well, or whichever. Like, funny thing you see. But we can also generate like Van Gogh like
images from this. Obviously it has
seen a starry night. Now this is slightly
terrifying to me on this, but sometimes we do fuel our nightmares
unfortunately with this. But yeah, you can generate these different styles maybe you're not
a fan of and go, which I can't fault you for. He's one of my favorites, but that's totally okay. Maybe Edward Munch
from the screen, I think that's what
it's called in English. And that way get very, very different outcomes from
basically the same prompt. I mean, this looks very close
to what I would expect. This is what we call
prompt engineering, or basically changing our
prompt to exactly see what, what we want out of this. Let's take another one where I promised you a
hyper-realistic art. So hyper-realistic art of a cyber punk bowl
with fruit in it. We'll see if this works. Hyper-realistic can be,
can be a nice modifier. I like to use it to get these very bright pictures
that look very nice for k. K is to get very realistic,
like photographs. And yeah, this, this looks
like something I would expect. This is very funny. And here we can
see that these are basically hyper-realistic
drawings, are, well, they are drafted up in this style of
hyper-realistic drawings. This is very cool. And that way you can change
something that you already have and put it into the
style of something you want. So if you go back to the cool pictures
that you had before, then you can essentially
now play around with them. But I want to show
you another tool that you can use to get
inspiration for your prompts. Let's have those slowed while Fraser is
another free tool, it tries to get you to the
dream studio better as well. Oh, this is adorable. I'm going to keep
this for later. Yeah. So this is what
happens if you have hyper-realistic instead
of like pictures. But anyways, Fraser, so this will point you to the
Dream Studio Beta, but you can get all
of these as well. So we'll have a look at
Fraser in our next lecture. Isn't it interesting how we
can change just a little bit about our prompt and we get such different
outcomes out of it. I love this, like this
is so fascinating. Yeah, I, the resources. I have a small book
that goes through some different styles that
you can have a look at. It's free. I made this for this class. So please download it and if you want to share it, go ahead. But like, be nice about it
please. This is for you. This is for you to just check out what different
styles you can do and play around with and eventually get
your project going. So really, this is
for you as a help, so you can use AI
to your advantage. And in the next class, I want to introduce
you to a tool or different tools that you can use to get some inspiration and
some better information. How to generate more diverse prompts or
different prompts. Or maybe it's something
you didn't think about. I was really
surprised at some of the things that work
really well for, for generating better,
better inputs to have nice images or images
closer to what I imagined. So, yeah, see you
in the next lesson.
5. Using Tools and Lookbooks for Prompt Inspiration: So I may be a little bit too
analytical for this one, but I really like to read what other people have
been doing this or see information how other people generate really pretty outputs. So I use other tools for this. E.g. in the resource section, I linked to a couple of
different look books. Essentially, like
there's the gallery, gallery where you can
basically have a look at a little book that has
different art styles in it. And I'm going to link to
a Google Doc as well that has some different styles and
with examples which I love, I also created a little e-book for you that you can download
in the resource section. Which is also a thing where you currently Skillshare
sometimes changes, but I think right now
that is where you would upload your project later. So keep that section in mind. It has a lot of value
packed into it for you. And yeah, now, let's have
a look at two websites. One way you can look at art that was generated
by other people. And you can actually
look at what prompts they use to
generate this out. And then we also
have a tool that actually can generate
these prompts for you and you just
click through it. And in the end, you get a nice succinct prompt that should generate
exactly what you want. And then you can delete stuff
and play around with it to really see and
fine-tune what you're getting out of this
machine-learning system. I'm always baffled by the
generosity of people, by collaboration, what
people can achieve. These documents and these
websites are just amazing for, for us to find out what we
can actually do with this ai. And I think for a
beginner like you, this is a fantastic way to
find out what you can actually achieve by using
stable diffusion when you're taking this class, which might be a couple
of months from now. That means there
might already be new tools right here that
you can play around with. Or there might be another free or pay for tool to
create prompts. This space is moving
extremely fast. But either way, what
you're learning here is applicable to
most of these models. So we select stable diffusion, which is what we're
working with right now. Then we can see are maybe we
do a 3D render right now. You can do it in
different languages, but we'll keep to
English right now because this class
is in English, it checks for you
if your prompt has enough of that information. Let's take the sleeping cat in front of a fireplace again. And it says, okay, this is good. These are similar
prompts that we had before in all these
different models. And let's click Next. What style do we want? Well, we can take
this Blender guru, Leonardo da Vinci,
Vincent van Gogh. You saw all those before. Let's, let's have
Picasso, give it a go. The coloring. Well, I think we should
have a nice orange maybe. And if there's a texture, we played around with this a
little bit before. I think. We can, we can
take any of these, maybe the night sky. Now, the resolution that
we're aiming for is medium. Right now, this doesn't change how large your
image actually is. It changes how it looks. And now we can also
give this a feeling. And we want this to be, maybe I want
inspiration as nice. And this one is contemporary. Now, this whole login thing. This stumped me the
first time as well. But we can just click on
go full page right here. And we can see orange
3D render made off night sky texture that
the data copy this over. I saved that cute cat. Oh no, it's busy. So we'll get back in a few. Okay, here we go again. It is rendering for us. Now this prompt is not natural. Natural, obviously. We're often talking about
natural language processing. So we can see right here
that this is something, it's not a 3D render. But also like this prompt
seems a little bit overloaded. So let's take this night
sky texture right out. Let's change this render. Seems to be that that
time of the night where the application
is a little bit busy. As long as I don't get banned as a spammer because I'm having a lot
of fun with this. Everything's fine. Um, but yeah, you can see that the Fraser isn't always the
best for everything. But you can definitely get some inspiration
for your prompts and see what you come
up with right here. This is cute. I'm not sure if
it's definitely a Picasso, but it has orange in it. Obviously it's not a
picture because it's, it's a, it's supposed
to be a Picasso. It's sleeping cat in
front of the fireplace. And this way you can, you can really get, get different ideas
by using these tools. I'm also linking a phrase book
for the tool called Dolly. But just different ways
to get inspiration. And when you look through
this four different art, the lexicon art or these places, you can get a lot of
inspiration of what to do. This long exposure might
also work really well. Essentially they're
all, I'm explanations are no descriptions of pictures
that are on the Internet. So anything that someone
used to describe a picture that they
upload it somewhere on the Internet is probably
used in training, is used as something that
this AI has seen before. And that you can then replicate
to make something similar and just get these really
interesting images out of it. And yeah, yeah, tilt
shift is also a good one. I really liked tilt shift, but I think this is it
for Fraser at the moment. Just another tool. You'll definitely
find others that can help you generate
other prompts, better prompts, just
different prompts. And in the next class, we'll actually have a look at
how stable diffusion works. And don't worry, I don't
wanna get into code. I don't wanna get into math. But I want you to understand why I don t think this is an ai. And I think you should
understand your tools as well. So it's really great. It's an opportunity for me
to nerd out and for you to learn how this actually generates these
beautiful pictures.
6. What is the Stable Diffusion "AI"?: In this lesson, we'll have
a look at stable diffusion, the algorithm that
we're working with. So it's using a really, really neat trick that I think everyone that
read the paper was just baffled by how
simple that idea is and how well it works
and what's going on. So machine learning is this computer sciency
thing that is fairly new, where we're basically show a computer and an algorithm
running on that computer, a lot of data and
have it figured out relationships
within that data. In this case, we're
showing it images, billions of images
with descriptions. The really cool part
is those descriptions. We can turn those into numbers because we've been doing texts processing on
computers for so long. We are slowly figuring
out how to turn those into numbers that the computer
understands and images. Well, they've been digital
for a while as well. So we have those both components that a computer can understand. And now what we're doing and
this is the fascinating bit, is that we take the
image and we're slowly making the image worse. So basically we're
adding TV static to it. So we call it noise. And just a little bit, each round we're, we're putting a little bit
more noise on that image. And the fascinating bit is that we can train this
machine learning system, our AI, to recognize that noise. But we're actually doing
it the other way around. So we have our image and we have the image that
is a little bit worse. And we're teaching our AI, not the way that we're
making the image width. We're teaching our AI to create the original
image out of it. But trying to teach
the AI to get rid of the noise on our deteriorate
deteriorated image. And this way, we can apply this AI that is learning how
to reconstruct our images. And we can take
it and apply that to something that was noise from the beginning and
generate images out of it because it is slowly generating something
that it has seen before. So one of those beautiful images of a landscape or watercolor, a Van Gogh painting. It generates something
like that out of just TV stat, which
is fascinating. I'll show you right here. So we have a starting
image of just color noise, which is basically TV's
static but just colorful. And then we're
running our AI on at once and we see that
the noise changes, but there's nothing
really happening. And then suddenly it
flips and we're getting an actual image
out of it because so much noise has been
removed out of it. So we're generating a
image out of nothing, out of randomness, out of noise. And I think this is such
a fascinating idea. So you may be wondering, but how does it know
what to generate? Because right now it's
just doing something. Well. Because we have
this description. We can always tell the AI that, but it's reconstructing
right here from this noise. This image is this description. So basically we're nudging
it in a direction now. So when we give our AI a
sentence to work with, we are now notching this, this removal of noise into a direction that it knows
like a watercolor image. So it is now removing
the noise in a way that it knows how. Well, usually noise looks
like on watercolor image. And it knows how to remove
the noise on a Van Gogh. So really this very smart way of mixing these
media of text and images and using
them together to get your computer to do
basically magic. Of course, there's a lot
of math behind it and a lot of little
tricks that I use. But basically what's
happening is we're nudging this machine
learning system into a direction using text. And it knows how
to generate images by this reverse
process of removing noise from just
noise and making it into this image that it has
been nudge to do by the text. And that's it. So it's fascinatingly
simple to really do this. And I think this is
a beautiful concept. And this is also why I
don't think this is Ai. This is a very smart algorithm. This is fascinating an adult, things that were never
possible before. But it is not mentioned. There is no conscious
and there we are. Just showing it
lots of pictures. And we're getting it to
do exactly what we want, which is generate
pictures from sentences. And yeah, I think mats
beautiful in itself. It's a way to, well, to work with a
simple idea and make it work with billions
of pictures and your, your ideation and the n.
So with that in mind, I want to go to the
next class because I think this understanding
of what this is actually doing is enough for you to actually work with this
in a way where you're like, okay, I, I get what it's doing. It's just working with noise. And it's making this. It's not smart, it isn't
doing anything clever. All the cleverness was done by the people that created it, by the people that
infused it with this idea of making it smaller, this idea of removing the noise. But there is no inherent artificial intelligence
in this system. So, yeah, but still a
lot of fun to work with. And I think in our next class, well, in our next lesson, this is still the same class. We'll have a look
at how to create different styles of
photography with this AI.
7. Make Stunning Fake Photography with AI: In the very first lesson, I already showed you how to make a couple of nicer pictures. These pictures were in the
style of actual photography, but we can do a lot more. And in this lesson, I would like to
explore a couple of different ones just to
show you what we can do. But I don't want
to dive too deep because this is an introduction. We can go over hours of material of fine-tuning what
we want out of our pictures. But this should give you
an idea if you've ever tried photography on how to make different styles
and different kind of perspective on your picture that is generated out of
the sentences you're given. And let's go back
to the website. Let's go back to our
New York skyline. Now, when we start this, we get a fairly basic image. We can see, well, we already did part of this. We have all these
different ones, some of this obviously
like product photography. So the simplest way to get higher-quality images and to make sure it's a picture is putting in eight k for k. So this is because
pictures online and we're often tagged with those
high large images. So large images often mean
that it's a good camera. Like my camera can photograph in eight
K, I'm pretty sure. But yeah, this is
already much nicer. We can see a nice blue
hour here, I think. And yeah, Generally this
is one way to go about it. But we can always go through our photography vocabulary and see maybe we want a nice
tilt shift in here, the one that we saw
earlier in the lexica art. And that's just this one right now. I think that the one
before wasn't ideal. Maybe the mix out of 4k8k and tilt shift didn't quite work. I think this one
is getting closer. So we can see the tilt
shift right here. This is looking okay. I think we can try a couple
of different other ones. Maybe we try some long exposure. Might be worth it to add
some traffic in there maybe. Oh yeah, though we
can see this is already looking like it's
taken over a long time. This is obviously at night, but a very light night sky from taking long
exposures shots. And of course we can try HDR. This is a popular one. People like to take pictures, but also that is very colorful. Yeah, this is, well, this is how some HDR
photo does look like. I think this one is probably the best one out of the
out of the lot. It looks very nice. Very it's popping. Yeah, I like this. This is good. Some other things
that we can try is maybe depth of field. We can check if this looks nice. It does. So these are
ways how you can, how you can play around with actual technical terms
out of your field. Photographs. If photography isn't yours, but you rather work
with acrylics, you can, you can add terms
that are technical to describe styles into your, into your, your
prompt right here. So, yeah, experiment with it. If you have a couple
of nice ones, you can also create
a collage now to post as your project instead
of, Isn't that neat? I think it's fascinating
how we can take these technical terms
that make a lot of sense in your actual camera. But this ai is able to generate realistic representations
and perspectives of those inputs because it has seen so many of these images, um, during the training, during the conception of this
machine-learning system. And yeah, I think
it's fascinating. So now that we know how to make these realistic
looking things, I think we should have
a look at ethics and how to be responsible and
working with this type of AI.
8. Ethics of Generative Art and AI: When we work with
this type of system, we have an extremely
powerful tool. And that means basically, we have to go with
the Spiderman quote, with great power comes
great responsibility. I beg you to not
skip this class. Don't skip this lesson. Because I think this
is extremely important for artists and for people. Because there are
different aspects that we have to consider when we use these kinds of tools. One of them is how these
tools were trained. So basically, this machine-learning
system only knows what it has seen before. It can generate
creative new outputs. But we've seen it
a lot of times. Unfortunately, with
these systems that e.g. a. Machine-learning system was only trained to
recognize why people. And then it had huge trouble recognizing
people off color. So when we use these tools, we should have a look at how
these models were trained, how this AI actually
was created. And luckily for us, we have something
called modal cards. We can have a look
at that talks about biases in the data and
biases and training as well. So e.g. this has a strong
bias to a static pages. And we can go have a look at the model cop and actually
understand what it's doing. When we go to our
staple diffusion demo, we can scroll down. We see the license right here, which is important for you if you want to use
this commercially. Yeah, Check this out. But we want to look at
the model cart right now. Because yeah, just like
it says right here, And despite how impressive being able to turn
text into image is, be aware of the fact that the model may
output content that reinforces and exacerbates
societal biases. Yeah, so this model
card right here, I need to login for this. There we go. I was wondering why this was so light that we can see right here what this model
essentially is trained on, what the dataset is and
what thoughts go into this. So here you can see what the intended use
of this model is. How misuse and malicious use and out of scope obviously are. Basically, don't use it too. You do demeaning
or dehumanizing or otherwise harmful
representations of people, cultures,
religions, etc. Don't intentionally promote or propagate
discriminatory content. Because of course,
this type of AI, especially without a filter, is, I'm capable of doing
this, unfortunately. So we have to be ethical
users of this power for this powerful tool. Yeah, so these are
malicious use and misuse. But obviously we
have to have a look at the limitations and biases. Because I told you before
this cannot work with text. It is not perfectly
photo-realistic. I think we saw that
when we were looking at the making stunning
photography class before, there were some errors
and some of it, it does not work perfectly
on more difficult task. We saw that when I was
mixing tilt shift or some other keywords right here, faces and people in
generally are not great. So a lot of people then
go into Photoshop to actually put in
realistic pictures. Then the model was mostly
trained on English captions, which is something that
you'll have to adapt to. And this is a
limitation of this. Then we have a
lossy auto encoding that as a technical
term. But this was. Trained on lay on five B, which contains adult material. So you can also get some spicy images out of
this n here the biases, well, it consists mostly
out of Western images. So things that we're used to, they often have a certain bias. So from communities that are largely affluent
on the Internet, have access to two cameras. So we're talking about a lot western white
cultures in this, which will affect how, how this generates
images and perspectives. So be aware of this. Because we, well,
we don't want to make things worse just
because we're using a tool that makes life a little
bit easier for us. So just, you can
read this when you, when you click here and go to the bottom
to the model card and you can read
about the dataset it was generated from as well. This is the lay on
five for 5 billion. Yeah, just consider
this as as users, we have to be ethical
about the use of this. And of course, there are other considerations that
you should not generate images from in the style of living artists and
yeah, all of this. So really try to be a
good person about this. And yeah, let's go on. In addition to the model card, we are also subject to the
license of this model. This model is licensed to you and can be
used commercially, non commercially free of charge. But they do ask you to not
use it in a harmful way. Now, there are
unfortunately people that use these systems
to create fake news or recreate images of people that are in
the public eye, e.g. and yeah, that's
that's not great. Other uses that aren't exactly
totally kosher is creating art in this style of
artists that are still alive because they're
making a living with this and just taking
their style and saying, Oh yeah, I create this app
and maybe even selling that art is a bit
problematic, right? I think we can all
agree on this. So consider the impact
of what you're doing. Of course, if you're
playing around with it and you want it to create some art that you just
want to use for yourself. People have always
imitated artists. This tool just makes
it even easier to do. Consider the impact
of what you're doing. And be aware of how
this model was made up. And be aware of the impact that you're going to
have with your outputs. So these original models that were created by
OpenAI and Google before, they actually have
filters on them against hateful speech,
hateful symbology. And while they're there, this is an understandable
way to go. Bad actors will always find
a way to circumvent this and create harmful
imagery with other means. Just create, finding ways
around these filters, e.g. so this one is
unfiltered obviously. But yeah, consider the
impact that this can have and work directly with
these kinds of models. Because this should be something that can celebrate the
beauty of fantasy, of your imagination
and shouldn't be used to make the world a worst place because we can all use
a little spark and use a little imagination and fantasy
and beauty in our lives. And we don't have to spell a good thing just because
we're using it in a bad way. This is a problem that we call dual technology
and machine learning. This can be used for good, this can be used for bad. People will use a nefarious, but I think don't tell anyone. But this is a little bit of a side quest for me
in teaching this. By understanding how easy
it is to use these and how, well, how you can get
very creative with this. I think it's also more and more important for people
to understand how these models work and how these could be
used and if erroneously, to generate fake
news, fake images. And yeah, I think you're
in a better place by just understanding how
powerful this is and how you can use this to
make beautiful things. And yeah, this may be a
little bit Pollyanna. But I believe we can all have a little bit of a spark in our
lives right now. In our next class, I want, since this is a vibe. So in the next class, after this heavy topic, I wanted to talk about
how to get the vibe, the feeling of the image
right, that we're generating. So join me in that next lesson. Right now.
9. Getting the Vibe of Pictures Right: Now that we know how we can
work with this responsibly, let's think about the vibe
that we want to portray. And I know this may sound
a little bit TikTok. But in the end, a lot of art, or a lot of what we do about
our art is about the vibe, the feeling that
we want to convey. There's a reason
that as musicians, we choose a minor
scale for most samba, or maybe even sad
piece of music. And there's a reason why we choose more muted
colors because, well, it's not a summer day. Maybe it's rainy. Maybe
we're processing our pain. On the other side. Maybe we're using neon colors, using something
sparkling to have joy, to have a party, a celebration. So we want to get
the vibe right. And our descriptions
often have emotions. So when we talk about art, when we describe odd, There's often emotion somehow
woven into that kept. And our AI learned
those emotions. So we can actually use
that to our advantage to tweak the vibe that we
actually get for our image. So let's have a look
how we can convey emotion with these
generative art pieces. When we're trying to
get the vibe right. We already have a couple
of ideas right here, where we have this
nice gray scale or some golden hours
and photography. But of course we can convey
emotions like gloomy, for example, and we can see
how that affects the output. Because often when we
describe that as gloomy, well, often when we
describe images online, we do assign them
emotions as well. So right here we can see that
this has way more clouds. We still see these
images that we've seen the entire
time with the with the Empire State Building
featured prominently and just but yeah,
very gray tone. We can tell it to have
muted colors, for example. To also go more on
the side of yeah, of some bonus, right? To not have it pop. But we can go the completely
opposite direction and see how this looks
in neon Cyberpunk. Wow, I love this image actually, I'm going to save
that for later. These are just ways for you
to play around with this. But also if we group
Friends hyper-realistic, think that's written without hyper-realistic and
then say happy. So they're all happy at the faces don't even look
terrible in some places. Thad is another signifiers. So we can really
modify what we're doing by assigning
it emotions as well. It's duration. This almost looks like friends. So this is how you, how you can use the same
style and the same motive, but change the entire
vibe of the picture. So in an alley way, we can make this fork a picture. Give it an eight k2 because
that's just how this works. And we can make this upbeat. So those are very
light pictures. Very nice. But we can change this because Gotham in Batman usually
has a dark vibe. It's still looks
fair, fairly nice. But yeah, experiment with
feelings and different vibes. This is much more
likelihood so we can see some dark parts and in the back, we can add gravity to it. Although graffiti
sometimes doesn't work. So you have to say spray
paint on the wall. Yeah. I love this. Okay. This is interesting because
it doesn't have a cap, but this is looking like a cat. But yeah, play around with this. Change the vibe of
a picture just by adding little modifiers
like this into it. And isn't it
interesting how the AI interprets our prompt
and changes how a, an image feels to us
using the same image. I love this. I think this is fascinating. And yeah, with that, I just want to finish on
our penultimate class where I will talk about a little bit what this machine learning
system can also do. What other things you
could be learning. And to be honest, this is a little
bit of a chance for me to also prompt
you for feedback. Because I would love to know from you what you
would like to learn next. I had so much fun
making this class. And I would like to do a
follow up if you want that. So see you in the next lesson.
10. What to Learn Next: This isn't quite the conclusion, but this is about this. Understanding the
power of the system. Because stable diffusion can generate these images
out of sentences. But it can do more, much more. But for this, we would have to install it ourselves or use online tools to customize
it a bit to our needs. And I wanted to keep
this introduction as friendly to anyone that
is interested as possible. So if you're
interested to see how to set this up and use this on other online tools where
you can do more with it or create more granularity about what you're doing with the ai. Let me know, write
it in the review, write it in your project. I would love that feedback
if you're interested. I might be able to
make that class because that is my expertise. And the other things that
you can do with this. Incredible, because
you can't, well, you don't only have to create
images out of nothing. You can also take an image and transform it into something new. With the text prompt. You can use this to do
something called out painting, where you have
your image and you then create what is
around the image. So the AI comes up with a
completion of the outsides. And we can do the
same thing, reverse. We can basically
take damaged images, whether we did
that intentionally or if it happened over time. And we can have the AI dream up what this
image should have been, where it has been damaged. So we get interesting new ways that this image
is then repaired. And I think you can see
how powerful this can be. And how you as an artist
can use this to match up, well, extensions
to your own work or variations on your own work. If a client doesn't quite like what you did there
in this image, you can use AI to get
new ideas, new inputs. And I think you
don't always have to use this to give this
rod to a client, right? You have your own style, you have your own ideas. But sometimes we're
stuck in our own ways. So we can use this
as inspiration. We can use this as a
schematic and it can help us, well, help us get more diverse outputs
and more diverse ideas. And that's what I love this for. What I think is the power
of having this type of tool in addition to
your own imagination. So you're now
combining what you do, your expertise and
your creativity with the power of 5 billion images that are baked into this tool. And yeah, that is
what you can do next. So you can take this and not just use the website
that has been built for you, but go to the code. Now, the next step
isn't coding yourself. You don't have to
know programming. The next step is to use code. Others have written and
do slight tweaks to it. So you can use a full powered
machine learning system, not with a nice website
in front of it, but with a little bit of tweaking and generate all
these fantastic things. And you can learn how to
do all of those tasks instead of just generating something from your imagination. Now, if you're
interested in any of those, please let me know. I would love to
create those as well. And with that, I'm going to let you go on to the last
lesson for our conclusion.
11. Conclusion: Congratulations, You
made it to the end. I know. I think this is a fun class, but still you sat through
a lot of material. I know that I tried
to sprinkle in. Well, making more
interesting problems and more interesting art. But we had some heavy
topics in between. You'll learn how
stable diffusion works with the noise
and everything. And then we even
talked about AI ethics and how to use this
to responsibly. So we had a thorough curriculum right here for an
introduction class for sure. So pat yourself on the
back and after that, please show me your favorite
image in the projects. I can't wait to see
what you come up with. And if you are so inclined, also post your promise so
others can learn from it. And yeah, maybe see variations
on what you did and play around with your ideas and
bounce off each other. So yeah, thank you for making
it all the way to the end. Make sure to also
leave a review. This helps others
find this class. This helps me make
future classes better. And also, if you find any topic that you would
like me to go more into, especially if they're out of the last lesson where we
talked about future learnings, then let me know. You can write to me. You can leave those
in the review, you can leave those
in the projects. I read all of them. And with that, thank you
so much for being here. This was incredibly fun to make. I hope you enjoyed
it just as much
Jesper Dramsch, PhD, Scientist for Machine Learning
