Intro to AI Art — Lesson 1: An Introduction to AI Art
Speaker 1 As we enter the second quarter of the twenty first century, human civilization is experiencing an accelerated pace of evolution. A technological and socioeconomic singularity is ahead of us, and it's driven by artificial intelligence and blockchain technology. NPEAK equips entrepreneurs and business professionals with practical knowledge to keep them ahead of the curve in the exponential age. Our live mentoring sessions, on demand training, and exclusive networking opportunities keep you at the cutting edge of Web three, NFTs, the metaverse, decentralized finance, automation, and so much more. Meet top industry leaders during our live mentoring sessions to ask your questions directly or simply follow the recorded sessions in your own time. In PEEP, inclusive, inspired, in the know, in this together.
Speaker 2 Hi. Hello, everyone. Welcome back. Hi, Benjamin, Ben. We're so excited to have you here today with us. This is a hot topic. Anything related to AI, we're all down for it. So for those of you who do not know Ben, he is part of our Genesys members, but he also has an impressive CV behind him. So he has over a decade of experience working with blockchain technology. That's probably longer than what I've known what blockchain is. He has been a featured speaker on blockchain technology and a guest speaker on podcast and Twitter spaces. He has also advised multiple projects in the web three space. As an artist, Ben has over two decades of experience with music and photography, mostly as a hobby, and he has been working with AI art tools for about nine months and has sold multiples NFTs.
In his Web two career, he has over two decades of experience in sales and management and has successfully started companies and held executive and managerial positions at multiple companies. He currently works in sales and consulting and lives in Phoenix, Arizona, while he also sits on the board of directors for a search fund and a charity. Ben, we're excited to have you here. The floor is yours. Everyone, head on over to the poll sections.
Ben Well, we're in web three, so GM GM, let me pull up my presentation here. And let's see. There we go. Alright. So if everybody can see that, I am gonna flip around screen so that I can see my notes, and then we will get rocking and rolling. I have to admit, I didn't know what introduction I was gonna get because I always hop in, like, five minutes late. I'm really bad about it. And so I actually did a slide on who I am, which I don't need to use anymore. So instead, I'll let everybody sit and look at this new AI photo that I saw the other day. This is called astronauts from Earth zero three four two. It's being displayed in LA, I think, this week coming up. And I was gonna talk about myself while I showed it to you.
I don't have to know since now, so that's great, but still an impressive piece of AI art. So we'll jump right into it. Why am I here? Why are we talking about AI art, and what are we talking about? So the goals today this is, titled lesson one, intro to AI art, and we're gonna go over the history of AI art from its earliest beginnings, all the way up to present day, which present day changes. Tomorrow, present day is gonna be different. We're gonna discuss different types and genres of AI art, and we're gonna define some terms so that when you hear a term like a GAN or a diffusion model, you have some understanding around that. We're gonna review some significant art and artists in the space. Obviously, that's really subjective. So I'm just gonna mention a few people that I think are important.
We'll look at a few that museums have said are important or that famous collectors are are saying are important in the space, and then some that, of course, have historical significance because of things they did in the last decade, two decades. And then we're gonna discuss some of the current controversies around the use of AI and art and kinda look at it from a couple different sides, kind of a a pro and a con. A couple notes for the purposes of the lessons that I'm doing right now. We're gonna focus on two d nonmoving images. So we're not gonna talk about three d modeling and sculpture and video and music. We may talk about some of those down the road in future lessons. But for right now, I wanna stick on kinda two d images, and then we'll go from there because I think that'll be plenty to deal with for the next several weeks.
So let's see. Next slide. What's next? So I like to tell people, like, where are we? Where are we going? So today's intro to AIR. I don't think I have to explain much more. We just did. Lesson two is gonna be basic tools for creating AIR. So we're gonna look at sort of the three major tools, DALI, mid journey, and stable diffusion next time in a couple weeks and really understand what they do, how they do it, how they're different, how they're similar. The next lesson after that, we're gonna look at tools and workflows for AI part part one. So, essentially, I wanna use mid journey and stable diffusion, which are the two that I'm most comfortable with, to create one or more works in front of y'all. We're just gonna go through the creative process and just kinda talk our way through creating a couple works and using those tools live.
The next lesson four would be tools and workflows part two. There, I wanna look at other ways to use AI and art, and we'll talk about that in a little bit. But there's a lot of collaborative work that can be done other than just tell me to create this image. There's a lot more that can go into it, so we'll talk about that. That could turn into two lessons, and we could talk about some other interesting things. Lesson five is getting familiar with GAN, which we'll talk about GAN more in a little bit. But there are current uses for GAN. There are historical uses for GAN, and so we'll look at some of the current uses and how it's being used. And then the last lesson or the last lesson for now would be how to create basic generative art. This could end up being split into a couple lessons.
My plan today, you know, two, three months in advance is we'll look at p5.js and and collaboratively create a very simple piece of generative art. But we may break that up into two and kinda start with what is generative art, how does coded art work, and then create one. So that's kind of the big picture where we're heading. Let's see. Next slide. So what is AIR? Right?
Speaker 2 Before getting into it, but I think we should go over the polls.
Ben Oh, great. Where are the polls? There they are. Alright. So I I put up a few poll questions. So if everybody could, please answer them. The first question is how long has AIR been around? And we're about to talk about that, but I'm I'm kinda interested to see what everybody's thoughts are before we dive into it for the next ten, twenty minutes.
Speaker 2 Well, 44% of the audience said 1973, 33% said 2014, and then 11% said 300 BC, and another 11% 1953. So no one thinks it's 2021.
Ben And I'm I'm getting disagree with that. We'll talk about it in a minute.
Speaker 2 We'll we'll see. We'll see. I'm not gonna say how I voted, but how knowledgeable are you on AI art? 56% a little, 44% not at all, and no one is a lot. So we're all newbies here today. Excited. Looking forward to learn more about this. Have you ever created a piece of AI art? 67% said yes. 33% said no. Does TikTok filter count? That's my question. And then have you ever bought a piece of AI art? 56% said no. 44% said yes.
Ben So I'm gonna run through them and give my opinion real quick. I would say there's not a a, you know, perfectly right or wrong answer. So how long has AIR been around? I think any of those are valid answers, and we'll go through it in a minute and see why. How knowledgeable are you on AIR? I would put myself in the a little category even though I probably know more than a lot. There is a lot to know. I mean, it just it's expanding so rapidly. It's almost impossible to stay on top of all the new developments. So it's wide, and it's a deep field right now. You can also go really deep with how these models work and and some of the controversies. Have you ever created a piece of AI, Art? I would say it may be closer to a 100%, and I'll explain why in a little bit.
But, you know, if you look at AI as meaning artificial intelligence, not human intelligence, I would argue probably everybody in this class has used AI to create something, whether it's a TikTok filter, a Snapchat filter, or modifying a photograph with an algorithmic, you know, modification. And have you ever bought a piece of AI art? I would also argue, and once again, are opinionated polls, right, that you may have bought a piece of AIR and not even realize it. So let's run through what is AIR. And I started with some examples from me, not out of vanity, more just because it was easy to kinda show some some breadth and know how I made them. So these first two pieces or pieces I made, These are sort of what you think of probably when you think about AIR right now today. The left is an output from mid journey.
The right, is actually a combination of stable diffusion and mid journey, kinda going back and forth, taking images and doing image to image prompts. So these are both AI art. There's been some more work done to them after the they're not raw AI outputs, and we can talk about that more later. But these are both AI art, I think, most people would agree. Next. Couple more. These involve characters. Both are out of mid journey. Both are pieces I did. Also, probably kinda what you think about when you think about AI art right now when somebody says that. Next. So now this is a little more not probably what you think of. The the one on the left is a piece of coded generative AI art. I used p five dot j s to create this. There was a seed image. There's an algorithm on top of that.
The right is actually it was once upon a time a photograph, and then I actually ran it through, I like to say, some special algorithmic sauce, a couple different algorithms to come up with this. But it started as a photograph, but I would argue it's no longer a photograph. It's more influenced by the algorithms that touched it than the photograph that I originally took. Now these look like photographs. Right? Is this AIR? I would argue that it is, and here's why. The picture on the left was taken back in I think it was 2009 or 2010. I took several images, put them together, then took several exposures to get a greater dynamic range than a iPhone four at the time could take. Then I've run it through some algorithmic interpolative upscaling software to increase the resolution, essentially use AI to guess at pixels, which isn't that different than some of the latent diffusion models we'll talk about in a little bit.
So I would argue that the photo on the left is a photograph, but it's also, you know, an AI collaborative photograph. There's been a lot of AI that's gone into that photograph. And the photograph on the right is an image out of a out of my iPhone 14 Pro Max, but then I've also done some interpretative upscaling algorithms and some other AI, functionality has been done to that photograph. So while these aren't, you know, what you would think of maybe when you think of AIR, I would argue that these are, influenced by artificial intelligence. So, that's my quick what does AI art encompass. Now let's dive into the history of AI art for a minute. So I would argue that if you really want to go back as far as you can, Art using artificial intelligence, nonhuman intelligence, goes back to 3,000 BCE.
The ancient Inca used a system called, and I'm gonna butcher the pronunciation, capu, talking knots, to collect data and keep records. And it's really interesting when you look at these. It's essentially a very aesthetically intricate system of knots, but it had internal logic that was robust enough that it could be seen as a precursor to computer programming languages languages today. Today. So I would argue, you know, if you're at trivia night and you just want, you know, to know when the absolute earliest AI artist, you could probably trace it back to 3,000 BC. In '9 in 1763, Thomas Bayes came up with you've probably heard the term Bayesian inference. So artificial intelligence requires the ability to learn and to make decisions, right, often based on incomplete information. Just talked about upscaling algorithms.
So in 1763, Thomas Bayes developed a framework for reasoning about the probability of events using math to update the probability of a hypothesis as more information becomes available. Thanks to his work, like I just mentioned, Bayesian inference would become an important approach in machine learning, and it marks one of the earliest milestones in our timeline of artificial intelligence. The next significant development in artificial intelligence was in 1842. English mathematician Ada Lovelace was helping Charles Babbage publish the first algorithm to be carried out by his analytical engine, which was we would consider the first general purpose mechanical computer. However, and this is where it ties into AI, Lovelace saw opportunities beyond the math. She envisioned a computer that could crunch not just numbers, but solve problems of any complexity. At the time, it was revolutionary that machines would have applications beyond pure mathematical calculation, and she called this idea poetical science.
Now this is okay. So here's your next trivia question. That's a picture of a Jaccard Loon. And this is this is really interesting because I would argue that you could view this as the first digitized image. The functionality of the analytical engine was inspired by this loom, which revolutionized the textile industry around 1800, so some forty years before that, by taking in punch card instructions on whether to make a stitch or not. So essentially a binary system. A portrait of the loom's inventor was woven into a tapestry on the loom in 1836 using 24,000 punch cards. And so in a sense, you could view that as the first digitized or artificial digital image. Let's see. In 1929, OCR was first developed, optical character recognition, with a device called a reading machine.
And I would argue that, you know, as advances in artificial intelligence and computers go, this got us talking about what does it mean to look through machine eyes? What does a computer see? The computer is recognizing characters. In 1950, I think a lot of us are familiar with Alan Turing and the imitation game. He developed the Turing test, a benchmark test for machines' ability to exhibit intelligent behavior indistinguishable from a human. And then in 1953, cybernetician Gordon Pasks developed his MusiColor machine. It was a reactive machine that responded to sound input from a human performer to drive an array of lights. At the same time, others were developing autonomous robots that responded to their environments and put out what we would see as very primitive art. So I would say this is very, very early history.
None of us would really look at this and think of AI or art nowadays, but it's good to have the early context for where we came from and where we're going. So now we're gonna get more into sort of the modern era and what we would think of when we think about art and specifically AI art. Let's see. I do see we have a question, and I will answer that one in a little bit. So history of AI, nineteen fifties. So let's see. So this is a picture called Untitled from 1967. It hangs in the Tate Modern. And I would say so this came out of Max Bensie's lab. Artists artists have been using computers to create work since at least the late nineteen fifties. There was a group of engineers in Max Bensie's lab at the University of Stuttgart. They started experimenting with computer graphics.
There's a large school of artists that came out of
Ben And, essentially, it was mainframe computers, plotters, and algorithmic art that mostly created visually interesting artifacts like the one you see here. In 1959, the term machine learning was coined. Arthur Samuel coined the term reporting on programming a computer, and this is interesting as much for AIR as it is for artificial general intelligence. So he reported on programming computer, quote, so that it will learn to play a better game of checkers than can be played by the person who wrote the program. I would say this marks a historic point in our AI or timeline with the coining of a phrase that will come to embody an entire field within AI. 1968, there was an exhibition known as cybernetic serendipity. Artists in the sixties essentially made artificial life artworks that behaved according to biological analogies, and they began to look at these systems themselves as artworks.
So there was a light sensitive owl or painting machines, which were essentially kinetic sculptures where visitors would get to choose the color and position of a pen and the length of time the robotic machine operated, and it would then create a fresh abstract artwork. So this really is starting more to look like the AIR we're familiar with. And then I would say 1973, when we look at Harold Cohen's Aaron machine, is really getting much, much closer to what we would consider AI art. So the work of artist Harold Cohen, he developed a piece of software called Aaron, and it's probably one of the most famous examples in the twentieth century of AI early AI generated art. Cohen trained this Aaron computer program to create art, and the results were often eye catching and expressive. The interesting thing is there's this question. Right?
The pictures that this program made all look like Cohen's earlier works. He was a in the early sixties, he was a successful exponent of color field abstraction. And if you see the photo on the screen, that's one of the pieces of work that this Aaron program produced. They never ventured too far from his own style and his own works. So let's see. So now we we're gonna jump ahead from 1973 all the way to 2014. It's not that there were Those things happening during the time, but there wasn't anything substantially different that looks like what we think of today when we're talking AIR. So we're gonna jump ahead, to 2014, and we're gonna start talking about GANs, generative adversarial networks. So this particular photo that we're looking at here comes from 2014. Again, generated these faces.
The right most column shows real photos used to train the system, and then all the ones on the left were created by the network. So let's talk a little bit about GANs and what they are. I like to call the GAN movement, which started around 2014, first wave part. And I'm I'm cribbing from clown vamp on Twitter. He came up with this. He said, everything before GAN was kinda wave zero. It was it was early. It was experimental, and it was niche. 2014 till about 2021, he considers first wave because most of the art created was created using GANs. After that, we're now in second wave, which is create created mostly using diffusion models, and we'll talk about that. But let's start with what is a GAN. So researcher Ian Goodfellow coined this term in a 2014 essay.
He theorized that GANs could be the next step in the evolution of neural networks because rather than working on preexisting images like Google's deep dream, which we'll talk about in a moment, they could be used to produce completely new images. So without getting too technical, let's talk about what a GAN is real quickly. So first, the generative part. The programmer trains the algorithm on a specific dataset. In this case, let's say a picture of flowers until it has a large enough sample to reliably recognize flower, the idea of a flower, what a flower looks like. Then based on what it has learned about flowers, they instruct it to generate a completely new image of a flower. The second part so that's the generative and the and the GAN. The second part is the adversarial part.
So these new images that are created are presented to another algorithm that's been trained to distinguish between images produced by humans and those produced by machines. So, essentially, it's a Turing test for artworks, this goes back and forth until the discriminator is fooled. So that's what a GAN is at a very basic level. So let's go from 2014 to 2015. So in 2015, Google's deep dream was born. This is a photograph. There's actually two photos. One in the back is a picture generated by deep dream from an art form called inceptionism, which we'll talk about in just a second, and one more to the front and to the left is from Kyle McDonald, and he calls it Deep Dream FBO Glitch. So in June 2015 I'm gonna butcher his name, but Alex and Google's Brain AI research team published some fascinating results.
After some training and identifying objects from visual cue clues and being fed a bunch of photographs of skies and random shaped things, the program began generating digital images suggesting the combined imaginations of, you know, Walt Disney and Peter Brugel the elder, including a hybrid pig snail or a camel bird or a dogfish, and you may have seen some of these in the news at the time. This birthed a new form of art called inceptionism, named after the inception algorithm in which a neural network progressively zooms in on an image to try and see it within the framework of what it already knew. So one of the first artistic applications, this deep dream from 2015, is actually not distinctly a GAN nor is it a later diffusion model. It's its own system of visual transformation, technically based on feature visualization.
Now DeepDream, and and this is relevant for where we're about to go, has mostly been used in conjunction with something called ImageNet. ImageNet could probably regard be regarded as the most important image dataset in machine learning today, originally created in 2009. So it predates GANs and DeepDream. It consists of over 14,000,000 images of everyday objects, people, animals scraped from the Internet by Amazon's Mechanical Turk laborers. Most commonly used, however, is a subset of this dataset created for the two thousand and twelve ImageNet large scale visual recognition challenge that consists of a thousand classes with 1,500 images each. And were it not for this large dataset, this labeled dataset, we wouldn't have had any of these algorithms that kinda came after it and used it for training. So we looked at 2014. Again, it's just able to make blurry faces. We see 2015.
Google's deep dream is starting to create things that look almost like art. 2016, GANs improve. So by 2016, this particular picture is a GAN that digested 3,000,000 photos of real bedrooms, and then it generated these rooms on its own. So that's what you're looking at is a bunch of computer generated rooms output by a GAN from 2016. Now these first GAN generated images had issues with them. If you really look, there's some coherency issues. In 2016, researchers from Facebook and a startup called Indico made an improved version of GANs able to create more realistic, although still imperfect images, such as interior scenes and realistic faces. That same year, there was a team at the University of Michigan on the Planck Institute in Germany that demonstrated how GAMs could generate relevant images in response to a specific text prompt.
And I would say, you know, that's kind of where we're going and what we're thinking of when we think of AI today. So the next slide and one moment. I need a drink. In 2017, GANs learned how to remix. And so before, they were generating images based on other images. Now well, let me just dive into it. In 2017, there was a project called CycleGAN, and it showed that algorithms could remix visual components from different images. So researchers at the University of California at Berkeley showed that GANs could also be used to modify images. For instance, you could add zebra stripes to horses, or you could convert a photograph into a painting style of Monet. The research demonstrates that algorithms can remix different elements or styles encountered in the training data, which is what we're used to today when we talk about AIR. Let's see.
Also, the University of California at Berkeley said that this particular project also showed the scientists that more data and more computing power could significantly improve the output of an image generator. And this wasn't something that anybody really realized before, but essentially throw more GPUs at it and it made better images, which kind of led the way for where we're heading with OpenAI and mid journey. Essentially, deep pocketed tech companies could exploit this and quickly scale this technology up. Now 2017 to 2018, let's talk about the art side a little bit. So you get a movement that I would call GANISM. So what we're looking at right here and I would say this is more you know, before the art that was created by these GANs was more experimental and fringe, now it starts to become more accepted as art.
Multiple artists are working with it, and there's actually movements using this technology. So the photo you're looking at right now is from an artist, Helena Serin, known as Pretty and GAN, this particular piece. So Serin was a traditional artist who's been using GAN to transform and enhance her own pencil on paper sketches. She's been building commercial software for a long time, but recently, she combined coding and art in this way. She almost exclusively uses or used to in 2017, 2018, this particular piece of software, CycleGAN. It's a GAN variant at the time that did image to image translation. And, essentially, she fed this GAN her own work to change her work. So she trained the network to transform images with the form of one dataset to have textures of another dataset. So, for example, she would translate her photos of food and drink into the style of her still lifes and sketches of flowers.
So this is an example of early GAN artwork. So 2017, 2018, let's look at the image on the left. This is Anna Riddler's tulips from her mosaic virus series in 2018. So for the project Mosaic Virus, she used something called spectral normalization, which was a technique that helps the algorithm generate better quality images. So, essentially, she created her own training set by taking 10,000 photos of tulips over the course of a tulip season and categorizing them by hand and then feeding them into again to create new flowers. So that's what you're looking at on the left from 2018. Photo number two. And this one's got some controversy around it, which we'll talk about in a moment. But, essentially, in 2018, October 2018, Obvious, which is a Paris based trio, fed an algorithm with 15,000 images of portraits from different time periods.
The album generated its own the algorithm generated its own portraits attempting to create original works that could pass as man made. This particular image, which is called Portrait of Edmund Bellamy, sold for $432,500, so over $400,000 at Christie's in October 2018. It was in the news, very significant event. Also in 2018, what we're looking at is AI generated nude portrait number seven by Robbie Barra. So Robbie Barra was the first minted artist on SuperRare. Back in 2018, he was encouraged by Art Nome, who's well known in Web three and in the art community, and they decided to give away 300 pieces of AI art at a conference in 2018. Only a handful were claimed, leading to the moniker, the lost Robbies. One of these pieces has sold for over $1,000,000. Now here's the controversy that I was talking about.
So Robbie Barrett created the algorithm that obvious used to auction off the painting, and there's a whole bunch of controversy later when we get near the end of this. My references include a couple interesting articles about it. But here, you can see the tweet from Ravi where he talks a little bit more about this. Let's see. So the lines between sort of wave one and wave two AIR tend to get a little bit fuzzy. And, obviously, you know, this this is all happening quickly, and this is recent and modern. So I would argue that there are also artists that kinda straddle the line. So one of the distinctions between sort of, for me, first wave AIR and later wave two AIR is that the first wave of AIR is tended to focus on creating coherent images.
While they were worried about aesthetics, the coherency of the image was probably the most difficult problem that they worked with at the time. And so this is this is a picture by Rivers Have Wings, the lead developer at Stability AI, Stable Diffusion, the the company behind Stable Diffusion, talented artist, and this is their CTS test number one, which is their first minted NFT, but it actually combines GAM and CLIP, which we'll talk about CLIP next time. But CLIP is used in some of the diffusion models and and wouldn't be considered a wave one technology. But this this particular piece is created sort of with some wave one and wave two technologies. Let's see. So let me see what's going on in the chat and see if there's any questions to answer before we keep going. So one second. Alright. So let me answer the question before we go too long.
So we have the question. No image a modern smartphone takes does not have a huge amount of post processing. Does it help the space to define AI art so broadly? So to answer your question in short now, and I'll get more to it later, I would say there's there's two answers to that question. On one hand, no. If you make something too broad, it loses its meaning. Right? If we just call everything that's been made in the last hundred years AIR, that doesn't really help us define what AIR is at the and so I do believe there is a there is there is sort of a broad meaning of AIR, and then there's a more specific meaning. And so when it comes to the specific meaning of AIR, I do think the broad meaning informs us.
As we get into controversies a little bit later, I think it informs us in the sense that this isn't something new, and this isn't something controversial. It's iterative, and it's something that we've all been dealing with. And so if you'll allow me to, I'll come back to the question a little bit later with that in mind, that there's both a broad definition that helps us to understand that this is technology we've been dealing with and using for fifty plus years. And then there is a a more specific narrow definition, which is really what we're looking at right now. When we think about AI art, it's the narrow definition, right, that we're that we're talking about, things that are created with algorithms and and diffusion models. And however, when people get up in arms about the narrow definition, I think the broad definition can inform and help us.
So, let me know in chat or questions if that didn't touch on what you were asking, but we'll touch on it again a little bit later. Alright. So now we're kinda jumping from I would still include this in sort of wave 1.5. We're gonna talk about the original Dolly, not Dolly two that we're using now, but the original Dolly. So here's a prompt. Well, let me let me start again with this phrase. So far, so weird. Right? Most of what we looked at was very strange looking. Like, it wasn't art like we would think of art a human makes. Maybe some sort of abstract impressionist art, but it wasn't you know if you think of an avocado or you think of a hedgehog, those are very abstract pieces. So in January 2021, the company OpenAI announced Dolly, a system capable of generating impressive images from a text prompt.
The name is a combination of the name Salvador Dali and the Disney character, Wallet. So Dali and this is why I say it's 1.5. It's kind of in between. Dali actually made use of a transformer, which is it it didn't use a GAN, and it didn't use the diffusion model. It's essentially, bear with me for a minute, a deep learning architecture that surfaced in 2017 and has been more of the de facto choice for text encoding and processing of sequential input and the variational autoencoder. Now we'll talk next week when we talk about the diffusion models about how this transformer, which was a which was a language model, essentially, got us to where we're at today. There there's a huge shift in going from GANs to transformer models, which were essentially translation tools and language to making this leap to text to image.
But we can get pretty deep in that, and so I I'm gonna save that for next time. We're really diving into diffusion models so that we don't get too far down the rabbit trail. Essentially, though, a model is trying to encode an image into a low dimensional probability distribution and then decode off of it. And that can then be used for generating new images by sampling from the intermediate distribution and passing it through a decoder. So it's very different than a GAN, but it's also not what you're thinking of with the current DALL E two or mid journey or stable diffusion. Also, a key ingredient of Dolly's impressive performance says one of the researchers from Berkeley that we're talking about in the previous slide was simply the huge amount of data training data that OpenAI fed into it.
And this goes back to, you know, now you go from scientists at Berkeley to large tech companies that can dump a ton of data and use a lot of processing power. And he said Efros from Berkeley said that they're using reasonably simple algorithms that have been done before more or less, but they really scale them up in a way that, you know, magic starts to happen. So the real magic with with DALL E was essentially this transformer, this this translational text model that they use, and then just the sheer amount of processing power and data they fed into it. And you can see here the difference from DALL E one to DALL two, which is what we're about to talk about. So DALL E two. Now we're getting into latent diffusion models and kind of what you think of when you think about AIR today, probably.
So here, I would say, is where we really get to the second wave. So if you look at these two images, the first image is Dolly one. The second image is Dolly two. And you can see that there's a a huge change in accuracy complexity there. So before, we were talking about January 2021, DALL one. Now we're talking in 2022, OpenAI announced a follow-up, DALL E two, that was improved thanks to more data and more computing power. However, it was also a new and more powerful type of generative algorithm known as a diffusion model, which was inspired by math used to model phenomenon in physics. This works by challenging an algorithm to learn how to remove noise that hasn't been added to an image, and we'll talk more about that in just a moment.
Later in the year, in June 2022, an independent project inspired by OpenAI's work known as Crayon now, kind of became an online sensation for a little bit. You probably saw a lot of these works. Very surreal and comical images were made by this free, now it's called Crayon. I forgot what it was originally called. It was a play on Dolly. And several companies made AI image generators similar in power, which is what we're about to talk about. Let's see. Next slide. Alright. So so the bottom image here is actually made by stable diffusion. And I would argue that you can see another jump in complexity from Dolly two here at the bottom to stable diffusion. Now we're gonna talk a little bit about diffusion models for a minute. Why did we move from them, and why are they, you know, better?
So why move from a GAN to a diffusion model? Why not just keep using GANs or make more advanced GANs? Or so here's why. GANs are famous for being really hard to train, where either they they've run into one or two problems. The generator either just straight up doesn't learn or it falls into what they call a mode collapse. AI AI I e, they learn to generate the same image every time, and I would say those are kind of the two biggest issues with GANs. Diffusion is an entirely different process in the sense that it's essentially a denoising process. So models are trained on how to denoise an image till it gets so good that with some guidance, it can denoise random noise into an image.
To get a little bit more technical for a moment, diffusion models consist of generating a chain of increasingly increasingly noisy images by gradually adding Gaussian noise to an image and then training a model to predict the noise that was added to the image from one step to the following one. If the steps are small enough, one can ensure that the image obtained at the end of the sequence can be approximated by the same Gaussian, the noise noise that the image is being sampled with. So it essentially allows us to generate a completely new image by sampling from the same distribution and passing it multiple times through a trained model. Now this is where back to the question from a little bit ago, there is a difference between the diffusion model and a picture that's been upscaled.
However, in a different sense, if you take out the, you know, sort of transformer translation aspect of it, the text to image side of things, using an AI interpolative upscaling algorithm is not so different. It's guessing at data that's not there, and it's guessing on noise in an image in a similar way to how a diffusion model is. And so on one hand, they're very different. On the other hand, you know, they're they're not totally different in type. So now let's talk a little bit about second wave AI and how it really how it happened. And most of us are probably experientially, we remember some of this. So 2022, last year. So in March, mid journey came out in private beta. People started to hear about it. People started talking about it. It was very kinda hush-hush in March. In April, DALL E two beta came out.
But OpenAI was very specific in not giving clear commercial rights for their outputs, while mid journey in March already gave explicitly gave their users clear commercial rights. So July, DALL E two includes commercial rights for their users probably as a response realistically to mid journey, and that was on July 20. Then in July, stable diffusion enters private beta, and people start hearing about it very quietly and using it. In August, only a month later, the stable diffusion 1.4 model is open source, and this has all sorts of implications that we'll talk about probably next week and in the weeks to come. And that, essentially, I can run that model on my own computer in a way that I can't with Dolly or Midjourney and and create my own AI art without being dependent on a Discord bot or OpenAI's website being up, which was sort of a game changer, but we'll get into that more in the future.
Alright. So now let's look at some pieces created very early on in the second wave. So we're talking sort of summer of last year. So right here, you're looking at a couple of early mid journey works. And one of the giveaways for early mid journey works was that they often have this sort of melted wax aesthetic. So on the left, you're looking at a piece called the everlasting impostor syndrome of the AI artist by Von Doyle. And on the right, you're looking at grafted mind by Phosphor. So these were early mid journey works shortly after it came out, so maybe March, April. The next slide is one of Claire Silver's works, and this one's interesting because I talked about on 07/20/2022, OpenAI announced that outputs from DALL E would now come with commercial rights. Literally the same day, like, she was just waiting.
Claire Silver minted this piece on SuperRare. She used Dolly two to make it, minted this piece. I think it originally sold for 35 east eth, and then it's gone on to resell, I think, as high as $50.55 ETH recently. Let's see. So here, we are looking at so this is also from July. This is with within the first two weeks of DALL E two. So after Claire Silver, but within the first two weeks, these are a couple of pieces that were meant to using DALL E two. These are weird Nikita's steel and skin on the left and special one series on the right. And the really interesting thing here is you can look at the texture and the realness of the textures in these particular pieces. So next here, we're gonna look at a couple of the earliest stable diffusion pieces.
On on the left, a piece by nocturnal lady two, and on the right, a piece by third channel. Also, really interesting thing here is not necessarily the realness of the textures like Dolly, but the interestingness of the patterns and the textures that came out of stable diffusion. So if we look at these, you can really see a difference between early mid journey pieces, you know, Claire with Dolly, more early Dolly pieces, and then Sable to Fusion. They each have their own flavor or look to them, and that gets changed and modeled as we go. But especially early on, they definitely had their own intricacies and and flavors. So as we talk about sort of who is an iconic artist, I'm gonna defer to people that know better than I do, essentially.
And so I would say TenderArt has a really great page on, you know, iconic AI artists where they go through and and walk through some of the best pieces by some of the best people. Interestingly enough, they also divide in decoded and AI works, and we can talk about that down the road a little bit. Let's see. So current events. And let me see if there's I don't think we have any questions. Let me just look at chat, see if there's anything interesting going on. So current events. Let's talk about what's going on today. And this will be kinda interesting because this will be outdated in six months. Right? We're talking about where we're at today, and this probably while it will be historically relevant, will not be current events relevant six months from now.
So on February 13, we have here the announcement of the first donation of on chain art from a collector to a museum and the largest digital fine art collection to enter a museum. So 22 works from the Cosimo dei Medici collection have been added to the permanent collection of the Los Angeles County Museum of Art, LACMA. The interesting thing is if you look down let's see. Dimitri's piece ringers, I think, was made with an algorithm, but definitely the two on the bottom left by Claire Silver and by Pender Van Armen are both AI pieces that were included in this announcement and in this collection. So very interesting from that standpoint. I know the lube also received some works right around the same time, but this is very recent. This is February 13. So, what, ten days ago this happened. Yesterday, I came across this article.
Viral Instagram photographer has a confession. His photos are AI generated. The story is also available on the references section if you wanna read the whole thing. But, you know, if you wonder how good is AI, it's good enough that, you know, it fooled 26,000 followers, essentially. And he finally came clean and said, hey. By the way, these were made with AI. These aren't real photographs. Let's see. This, I think, was also just yesterday. The museum in The Netherlands where Vermeer pieces are exhibited just hung an AI piece in place of the Vermeer wallets loaned out right now. So also a very interesting article. They received almost 4,000 AI entries into this contest. Obviously, there's quite a bit controversy around this. There's some people that are pretty unhappy about this, but also a good read if you wanna read more about that.
So here's a piece that going back a little bit further, but it's gonna start walking us into our controversy section. So back in March 2022, the US Copyright Office rules that AI can't be copyrighted, and I'm gonna read you about a paragraph for this because it's really interested. So the US Copyright Office once again rejects a copyright request for an AI generated work of art, not the first time. A three person board reviewed a request from Steven Thaler to reconsider their 2019 ruling, which found that his AI created image, quote, lacks the human authorship necessary to support a copyright claim. Thaler first brought the image created by his creativity machine algorithm in 2018. He describes it as a simulated near death experience where an algorithm repurposes pictures to create images seen by a synthetic dying brain. He was seeking to register his computer generated work as a work for hire.
And so both in its 2019 decision and in its decision last year, the US copyright office found that, quote, the human authorship element was lacking and was wholly necessary to obtain a copyright. Current copyright law only provides protections to, quote, the fruits of intellectual labor that are, quote, founded in the creative powers of the human mind. So kind of an interesting copyright ruling that was very anti AI last year. We have a newer ruling from, I think it was also, yeah, from yesterday that is not as anti. It's not a pro ruling, but it's not as anti. So, essentially, there is a comic book that was made with Midjourney, and the US Copyright Office ruled that the entire work could be copyrighted, but the individual images could not be. So the creator will keep the copyright registration, but it'll be limited to the text and the whole work as a compilation only.
So in one sense, it's ProAI. In another sense, it is as minimally ProAI as the copyright office could rule. So let's talk about controversy a little bit, and let me stop for a moment to take a breath and see if anybody has any questions or wants to hop on stage to ask me anything before we move on. Alright. Let's see. Would this oh, let me press the right button. Alright. Would this mean AI generated 10 k collections can't be copyrighted? Even if the traits are made by artists and put together with a program, Does that mean all 10 k collections are CCO besides logos? Well, Jacob, I would say you beat me to it. We're gonna talk about that in a little while, but I'll jump ahead to it. So when we get into controversy, I have that as an open ended question.
And my question is, same as yours, what about PFPs? 10,000 piece collections. Traits are not assembled by humans, but an algorithm. So back to your question, and this has been asked and by some people. Can YUGA really have the copyright on a Bordet? And the question goes kinda back to what the US Copyright Office was saying. What about what is the artistic provenance of something that was you know, the artist made individual traits, but the artist never made any single Bordeaux or CryptoPunk or those were created by an algorithm randomly. Right? There wasn't even the curation aspect. You know, that's part of the sales pitch of something, you know, an algorithmic collection like that is that it's non curated. So I would say that's an open question. I would say we haven't had the, you know, either rulings or lawsuits to answer that, but I think it'll have to be answered.
And as we go through controversies, maybe we'll find our way to an answer. I don't think we will. But this goes back also to what about code generative art. Right? As we're talking about, you know, pieces by Tyler Hobbs or Zancan or these amazing generative coding artists that are selling pieces for very also equally large sums of money. A a human wrote the algorithm. But once again, by nature, the outputs aren't curated only in the sense that the algorithm was modified to produce more desired outputs, but then the actual outputs are are random, you know, chance. So I would also extend your question to what about code generative art? Can you own a copyright on that? I think well, I'll give away my my stance on that right now. I think it can, and I think we'll get there.
It makes common sense that somebody put all this time and effort into this creation. But I think, eventually, the courts will rule, yes. But I also think we've got a long way to get there. So, hopefully, that answers your question. Alright. Let's jump into controversy a little bit. So here are some recent headlines, which y'all have probably read while I was talking, but I will read them real quickly. So recent headlines from The New Yorker less than two weeks ago is AI art stealing from artists. Last year, from a blog, the intrinsic perspective, AI art isn't art. Also, late last year, in The Verge, the scary truth about AI copyright. And Ars Technica from September of last year have AI image generators assimilated your art. So, obviously, this is the media. So these are, you know, attention grabbing headlines. That's what they're there for.
But still pretty shocking sort of statements some of these are. So let's go into let's start with the argument against AI art. And let me also start by giving away my bias right at the beginning. So I believe that as time goes on, art created with AI is going to be viewed simply as another tool that can be used to create art, just like any other tool that we can use to create art. Whether that tool is a paintbrush or a digital paintbrush or a camera, I think it'll just be viewed as another tool. That being said, you know, that's my bias. That's not where we're at today. So let's talk about some of the arguments against AIR, and then we'll talk about some of the arguments for AIR. So AI is stealing artist's works and is infringing and derivative.
And all I can say on that is I have researched this till I'm blue in the face, and I don't know. You you read some articles and studies by very smart scientists and programmers, and they argue that, essentially, you know, these diffusion models and algorithms essentially memorize some of the input inputted art and can be told to essentially output things that they've memorized that are very, very similar to the trained dataset. Then I have read other very smart people and programmers and scientists that have said, no. That is not true. 100%, the AI is imagining a nonderivative wholly new work. There's also, work going on in terms of datasets and pruning datasets and curating datasets to not involve artists that don't want their work to be in there, to only use CCO work. So I really see this as being sort of an unanswered argument and question, but I think it'll get answered.
Right? Nobody actually believes it's good to steal from artists and steal from people's copyright. And so whether the answer is no, these algorithms aren't stealing the art, or the answer is yes, They are, and we need to train them on a new dataset, I think we'll get past that. This is a short term argument against AI art that will get resolved. Number two, the machine is doing all of the work. Right? One of the arguments you hear is it doesn't require any any skill. Skill. I can tell the machine to make a a Monet, and it'll make me an amazing piece of art. The other argument is that it's unpredictable. Essentially, a person can't direct it to actually put out what I want it to. And there there's some truth to that. Right? But that's like saying a camera's unpredictable.
I click the button, and who knows what's gonna come in there. So I think this is sort of a false argument. The machine is doing all the work, and we'll get into that. And then number three, people naturally prefer art made by a human. I've heard that one. Right? The the people just want art made by people and don't want art made by computers. Now that one I wanna address for a moment because this is sort of interesting. So I'd rather look at art made by a human as the argument. But then I would say, but you can't tell. And I and I would argue that, you know, you a lot of you probably read the article that came out late last year about the piece that won the blue ribbon at an art fair. It was an AI piece.
The artist didn't disclose that it was made with AI, and it won the award, and then people were angry about afterwards. Well, nobody knew it wasn't made by him. Right? They couldn't tell an AI made that. Now this is even more interesting. Back in 2017, there was a group of people that worked on a project called AICAN, a I c a m, and they showed these images at Art Basel. Images and works created by AI and human artists. And for each artwork, they asked the participants back in 2017 and 2018 at Art Basel whether they thought it was made by a machine or an artist. They found that people couldn't tell the difference. 75 of the time, people thought that the AI generated images were produced by a human artist. However, it wasn't simply that they had a tough time distinguishing between the two pieces.
They genuinely enjoyed the computer generated art using phrases such as it has visual structure, it's inspiring, it's communicative when describing this AI work. Also, we talked earlier about the AI artist Robbie Barrett, who was the first minted artist on SuperRare, And he talked a little bit about is AI art Made by a human. And Ravi was talking about this back in 2018 with GANs before the sort of diffusion models we have nowadays, and this was his argument. He said, is it really made by a computer? And here were his four arguments for why it's not. He said, a human chose the dataset. He said, a human designed the network. He said, a human trained the network. He said, a human curated the resulting output. So in his mind, at least with regards to GAN generated artwork, it was a 100% wholly human creation in that sense.
But these are the primary arguments that I've heard against AIR. Now I wanna talk a little bit about arguments for AIR. Obviously, like I said, that's my bias. I've made AIR. I've sold AIR, and I'm teaching a class about it. So number one, arguments for AIR. Claire Silver says taste is the new skill. It's not just about generation. It's about curation. And so this is all I would say about that. There's also sort of Claire's argument is essentially a distilled down version of a information theory argument that says a tool cannot create art. Only humans can create art. And so the the best way I've heard this put is human artists have nothing to worry about. Human drawers do. Right? I used to be over purchasing. And, you know, I would tell people I want purchasers, not buyers. A computer can buy things.
It can fill an order. A purchaser uses intention and discretion. Same idea here. Computers don't do what we wish them to do. They do what we tell them to do. Essentially, this argument relies on the idea that a computer is a genie, not a genius. And so I would say this is more of an argument about what is art. I would say the downside of this argument is if we ever get an artificial general intelligence, this argument kinda goes out the window. But as of today, AI algorithms for art are simply a tool. They can't create taste. Next argument, it's a lot harder than it looks.
Ben this is what I mean by that. Don't take that in the wrong way. This is what I mean. There's more steps to it than I tell the computer to create something, and it comes out with a Monet or a DaVinci. So first of all, I would say this. Claire Silver uses the term AI collaborative art. CloudFam uses a similar term. I've used the same term. And I would say this, it's not just about generation. So I would say it's about variations, restyling. If you only think of modern AI art as a text to image tool, it's clear to see that, you know, why people are worried that it's coming for someone's job. But that's not really how it works. So here's a great example. So this photograph is by a piece by the artist Clown Vamp who's listed in the Tender art directory as a significant AI artist.
So the bottom left photograph, the really small little one, is the raw output from mid journey. Pixel size of about a thousand pixels. The final piece is the larger piece that you see. So this is the polished piece, which is 7,000 by 10,000 pixels. And to get there, they had to out paint, which we'll talk about out painting next time. They had to upscale. They did post processing and Photoshop and other things. And if you look, you can see the differences. The picture is out painted. It's wider. It shows more of the scene. It's been upscaled. Colors are different. Hairline's different. Minor subtle changes, but a lot more work than, you know, produce this one image for me and it pops out. So I would say there's a lot more subtlety there to creating AI artwork than just I tell the computer what to make and it makes it.
Let's see. Next argument. AI is simply a tool. So this is one of my favorite arguments. Best example is a camera. This is a unedited photo I took with a camera years ago. So I don't know who said it first, but the problem isn't a new one. Years and years ago, when cameras first came out, artists got up in arms because the camera was coming for their jobs as portrait artists. And you know what? To some extent, they were right. Nobody hires a portrait artist nowadays. In 2023, if you have a kid's birthday party, you're not gonna hire a portrait artist to sit there and paint a picture of your kid's birthday party. You're gonna use the camera. So on one hand, as a society, we gain this ability to capture moments instantly. But at the same time, it also elevated painters.
Painters don't just sit there and and draw portraits nowadays. They can actually make make art. And then the camera itself spawned a new industry of art. Right? Well well, operating a camera, anybody can do. You can press a button. Operating one well is still an art in its own right, and so photography is considered art and pieces sell. Back to clear silver, taste is the new skill. This has already happened before. We're not talking about something new. I would also say this, and this is, for me, one of the most compelling arguments for why AI is simply a tool. I've got a good friend that's a traditional watercolor artist, and a physical artist could make the same argument that digital art is an art as a digital artist makes the AI art is an art. Right? What do you mean?
You're not you're not even using, you know you're using Photoshop. You're not even using a a paintbrush in your hand. Right? And this is a great little graphic that kinda pokes fun at it. But I would even argue you can go further than that if you really want to. Right? Because if you took my friend who's a modern painter and you ask Michelangelo if he was a real artist, Michelangelo would say, you you bought your paintbrush at a store, you bought your canvas at a store, and you didn't even mix your own pigments. You're buying store bought colors. You're not even making your own colors. You know, you're not a real artist. So I think it's sort of a slippery slope, and it's sort of a false argument. You know?
Everybody always thinks the newest tool doesn't count as well as the oldest tool, and yet none of us are suggesting that the only real art is art where an artist mixes their own pigments and paints on an animal skin or something. And then the last argument is it's a social good, and this is what I would say about that. AI lowers the barrier to entry, and this is what I mean. This is a piece I created. I could imagine it. I could you know, I've touched this up quite a bit. It's not finished. But if I had to sit down with a pen, I couldn't do this. I had it in my mind. I know what I want, but I don't have the hours or the time or maybe even the the skill in that particular thing to sit and draw this out.
But with AI, I can. And not only can I do it, I can do it a lot more cheaply? You know? $10 a month, a computer, and I can output this. It still takes hours of work and a lot of tries and a lot of time and honing a skill set, but it lowers the barrier to entry and starts democratizing art. And it really does make it about style and taste, intention, and creativity, and not as much about a very specific skill set. And I would say that's true of some of the greatest artists in history. You know, you do have to be good at a skill set, but the person that's best at a specific skill set isn't always the most famous or the best artist in that sense. So at this point, I will stop, and I will ask, do we have any more questions?
I see we have one. Oh, quick aside. What you're looking at right here on the left is Edward Hopper's famous painting. On the right is an AI imagining the other side looking from the other window. Saw it on Twitter the other day. Thought it was really cool and would be good to leave up while we answer some questions. So let's see what we've got. Alright. By the way, Seth says, any thoughts on mid journey? Where is it heading? Is it a cash grab as it's on a paid basis or a legit AI art tool with future? So I will say two things. So I do have a couple thoughts. On one hand, I don't think it's a cash grab. It's one of if not, there are outputs I can get out of mid journey that I can't get out of stable diffusion and vice versa.
So I would not say it's a cash grab. It certainly has its own intricacies and its own style, but I do think it's a legit a r AIR tool. Now whether it has a future or not, I mean, you saw all the things that have happened in the last nine years versus the previous three thousand, five thousand years. I have a hard time saying where it's gonna go in the future. There are some people that argue that that essentially well, we'll get into the technology behind it next in two weeks. But I do think it has its place, and it is its own legitimate tool. Where it goes from there and where the future of AR goes, I don't know.
Speaker 2 And I also think as the course goes and we evolve, there's gonna be new things happening and things changing that will add up, you know, spy to the space. There probably will be new controversies, new art that have been generated, new artists, new tools. So I think that's also what's fascinating about Web three. It's happening at such a high speed.
Ben Yeah. I agree. And that was one of the hard things putting this together. Right? Was I mean, just in the current events, there were things that were happening yesterday. I was, like, adding last night and this morning, and that's gonna keep happening. And in, you know, two years when when we update this course or somebody gives a new course, you know, we'll be talking about DALL two mid journey and stable diffusion in the past tense, and we'll be talking about the new tools. So I would say as you look at the slide on the screen right now, in a couple weeks, we're gonna be talking about basic tools for creating art. We're we are gonna talk a little bit about DALL E, mid journey, stable diffusion, how are they different, how are they similar, how do you access them, how do you use them.
And if you're interested in some of the references that I used, if you click on the link, themeadows.xyz/np, there's both a copy of this presentation and a copy of all those references there in that GitHub repo. So that's it for the moment.
Speaker 2 Thank you so much, Benjamin, for being here with us. It has been a great pleasure, and we're looking forward to having you next time on Inpeak. Everyone, have a great night, a good evening, a good rest of your day, and see you tomorrow at our networking session. Goodbye. Thank you for being here.