Intro to AI Art — Lesson 2: Creating AI Art
Speaker 2 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 peak, inclusive, inspired, in the know, in this together.
Speaker 5 Hi, GM. Welcome, everyone. Happy to have you here again. Benjamin, welcome. How are you doing today?
Speaker 5 I am good. Thanks for asking. I see a lot of people now coming in. We're welcoming you towards our second session of of, I believe, a five part story.
Ben Five, six. Fine. About eight hour. There we go.
Speaker 5 I think it's five. Please head on over to the poll section. That way, we get to know you better, and we see if you actually paid attention last time you were here. We have some trick questions in there for you. But for those of you who do not know Benjamin, Ben Meadows, he has over a decade of experience working with blockchain technology. 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, he has over two decades of experience with music and photography, mostly as a hobby. He has also been working with AIR tools for about nine months and has sold multiple NFTs. We saw some of that collection last time. And in his web two career, he has over two decades of experience in sales and management.
He has successfully started companies and held executive and managerial positions at multiple companies. And he's currently working in sales and consulting and lives in Phoenix, Arizona. He also sits on the board of directors of a search fund and a charity. Overall, he's a complete package, and he's here today to talk to you about basic tools for creating AI art. So, Ben, the floor is yours.
Ben Thank you. I figured you were gonna do an intro again. So but I still had a placeholder again for us just in case I needed to talk. Here's a piece by Claire Silver. She just announced that she's launching a collection with the Louvre. Like we talked about last time, one of her pieces is in LACMA. So just thought it was a cool piece to throw up while I kinda got organized and figured out where we're going. So let's dive right into it. I was saying before this that this session will probably be a little less tech heavy and a little more show and tell. So, hopefully, for the those of you that are visual like I am, you know, there'll be lots of pictures, and we'll be looking at a little more practical side of things. So let's dive right into it. Why am I here?
Right? What are we doing this session that was different from last session? So last session was kind of a review of AIR in general, the history of it as a concept, culturally where it's at, culturally where there's controversy. This section's a little bit more down and dirty. What are the actual tools that are currently being used to create images? So we'll look at DALL E two, mid journey, and stable diffusion. We'll talk about ways in which they're similar, ways in which they differ. We'll talk about how do you actually use them. Like, how do you actually access DALL E two or mid journey or stable diffusion? And then we'll look at, you know, where are they going as much as we know. Right? What's some of the new news and future developments? And then we may briefly take a quick look at some of the alternatives that are out there both now and on the horizon.
So that's our goals for today. Let's see what
Speaker 5 ahead and continue, I'm gonna go through the polls so you get to know who's sitting in the audience. So did you attend our first session? 50% said yes. 33% said no. And wait. Someone changed? Okay. So no. No. 60%. 30. Yes. And 10% watch on demand. So I'm gonna share that link with you guys in the chat. That way, you can get to watch it on your free time. How long has AR been around? 30% said three thousand BC, twenty seven percent 1973, twenty seven percent 2014, and '18 percent 1953. How knowledgeable are you on AI art? 70% a little, 20% a lot, 10% not at all. Have you ever created a piece of AI art? 50% yes, 50% no. Have you ever used the following to create an AI output? 40%, none of the above. Sorry, my dog. Toby.
Ben I'll I'll read it. I never used the following to create an AI output, but none of the above, 31%. Midjourney, 23%. Dolly, 15%. Two of the above, 15%. All 15%. And stable diffusion, 0%. So let me run through and quickly give my 2¢ on the polls. So did you attend the first session? Don't worry if you didn't. This isn't dependent upon it. Like I said, it's a great background. It'll sort of enrich the context around this, but you're not gonna miss out on anything today. How long has AI Art been around? I actually had this discussion. We talked about last time that you can look at it a whole bunch of different ways, and I went to school originally a long time ago for history. So there's a lot of ways to look at it, but I had a relatively well known watercolor artist.
Me and him were sitting, chatting last week, and he said, okay. But which one do you think? And I said, alright. If you put a gun to my head, it's 1973. Harold Cohen had his Aaron machine that created artwork. And I said that's probably the most what we're thinking about today if you kinda traced it back to its origins. '53, you're talking about more computer style artwork, generative plotter artwork. 3,000 BC is really interesting from a trivia night standpoint, and 2014 and 2021 are certainly major milestones and iterations. But I would say 73 would be my answer if somebody really forced me for where, you know, what we're talking about, the type of AIR. How knowledgeable are you on AIR? I would put myself somewhere between a little and a lot. It changes so fast that, I mean, even just for this session, had to go back and review some things, look up some new things.
I haven't even had time to try every new development that we're gonna talk about today. Have you ever created a piece of AI or, you know, where I stand on that one? I've I've got one up for auction right now on foundation. There's an active active auction going. And have you ever used the following? So I have used all three and then some others, which we will talk about. Now somebody's come in on stable diffusion. We got 6%. So and we'll talk about that in just a minute. So alright. Let's dive into it. Let's see. One quick note for the purposes of these lessons, we're gonna focus on two d nonmoving images. Right? There's a lot going on with three d modeling with AI, sculpture, video, music, and who knows? Maybe session seven or eight or 10 or something will deal with some of that.
But for today, we're talking about nonmoving images created
Ben AI. Alright. So what's next? We did intro to AIR. We're gonna do basic tools, give you a quick kinda what to look forward to. The next two sessions are gonna be diving into the tools and workflows. So today, if we're talking about what they are, how do you use them, we're actually gonna use them in the next two sessions. I'm still trying to figure out the exact way to do that efficiently. Maybe we'll do it like a cooking show. You know, you see the turkey go in, and I tell you it's gonna be thirty minutes, and then magically, we cut straight to the turkey being cooked and ready. But we'll actually dive in and use them and talk about, you know, prompt engineering and negative prompts and textual inversion and and really dive into, you know, making a few pieces.
And then we'll we'll see past that in terms of GAN and gen art and some of these other things where we go. So let's dive right into it. So the three major tools we're looking at today all use a diffusion model essentially to create images. So what's a diffusion model? Really simple, and we'll we'll stick to really surface level because this can get really crazy, but we'll stick to, you know, a diffusion model. Well, let me go back. Text to image models are based on deep neural networks, which are essentially complex mathematical functions that learn from data. Right? Neural networks process different types of data such as text, images, audio, video. They can perform different tasks. Right? You've read about neural networks that can do classification or detection or translation or generation. So to generate images from a text description, text to image models need two different types of neural networks, an encoder and a decoder.
The encoder converts the text into a numerical representation called an embedding, and the decoder converts the embedding into an image. And there's different ways to design and train text to image models, but like I said, one of them is called diffusion, and that's sort of what you're using in a simplified way when you're using Dolly or mid journey or stable diffusion. So what is diffusion? Diffusion is a technique that gradually transforms noise into an image based on a text prompt. It works by reversing a process called denoising, which we're probably all familiar with. Denoising is a technique that removes noise from an image while preserving its content. For example, you might use denoising to improve the quality of an old photograph or a blurry scan. A lot of us have used these tools. You know, you're scanning an old picture in or maybe you're upscaling something.
It's it's, you know, taking the noise and guessing at what the image should look like. Well, diffusion does the opposite of this. It adds noise to an image while preserving its content, and then it goes backwards and it undoes that. And it goes back and forth a bunch of times, these diffusion models do. But you also have used diffusion models, just the diffusion side. Right? Maybe you create artistic effects, blurred out an image. That's kind of the fusion side. The fusion we're talking about in a really, really simple term are
Speaker 3 going back and forth. Noising and the that give us outputs that we see with these models.
Speaker 5 Benjamin. Are you there? I think we might have lost him. You can see me and hear me though. Right? Can I get a thumbs up? Okay. Yes. Okay. Perfect. Okay. Let me check on the back end. The diffusion model got him. He got diffused into the background. Give me a few
Speaker 4 seconds so we can troubleshoot over here.
Speaker 4 Well, for those of you who are just joining,
Speaker 5 please head on over to a poll sections. I'm I'm hoping that we will get an output pretty soon with his face. It might be easier for him to do audio only correct. The thing is I do not have direct contact with him right now, but I'm trying to get ahold of him. Wout, would you like to come up on stage and cover this while we're we're here? He has lost connection, so it might be something on on his end. Ned, you can try to generate him. You want me to bring
Speaker 4 you up on stage to generate him back?
Speaker 4 So just please give us a few minutes while we find him.
Speaker 4 So I'm trying to get a
Speaker 5 Whoever gets it first. We can do a raffle. Well, what do you think? Landline. I wish I wish
Speaker 4 I only had his phone number.
Speaker 4 Invited you up on stage while while Ben is back. He will be back.
Speaker 5 Hi. I'm sweating here because we're live and we don't have a speaker, but it's good in you.
Speaker 0 I have I have to improvise now with my AI skills.
Speaker 5 Well, you have 96 more slides to improvise then. Good luck with that.
Speaker 0 You know how to operate them?
Speaker 5 Yes. I can move them back and forward.
Speaker 0 Okay. Actually, I I I can tell this. I'll take this opportunity once more to share with anybody who is just coming in and is new over here. We've been running a challenge called the GoalGetters in the Discord for six weeks with amazing results where we basically leverage the power of our community to to grow for personal growth. And you can check-in the Discord if that's something for you because next Monday, we're gonna kick it off with a two week challenge to hold each other accountable towards these goals. And, actually, Ben I is back, but Ben created an artwork for the goal getters, a special artwork for everyone who completed the six week challenge. And we're air dropping those this week to everyone who finished it. So that's been amazing. And, yeah, thank you, Ben. Glad you're back. Try to fill up the time a little bit. I
Ben am glad to be back. I'm sorry about that. The Internet just dropped and dropped on me. Like, hard just cut out. I had to go run and reset the modem. Of course, doesn't
Speaker 5 robbed by an AI. That's what happened.
Ben Well, that's, you know, the the reality of living with technology. Right? Sometimes it doesn't work. So alright. So let's get back to where we were. So we were talking about
Speaker 5 the you, Woug. Thank you for sharing that information, and congrats to everyone who made it on the GoalGETTERS. Very proud of you all.
Speaker 0 How do I jump back off?
Ben I I Oh, alright. So let's jump back into it. So we're talking about diffusion. Hopefully, y'all caught enough of what I was saying before I cut out. So let's go back to so now we're gonna talk about, we're gonna review, from last time. Just quickly talk about when how long have these tools been around for, right, that we're about to talk about? All of them were last year, essentially. Midjourney, went in private beta in March of last year. Dolly came out in April. In July, OpenAI gave commercial rights to Dolly. Also in July, stable diffusion came out in a private beta, and then in August, they went open source with their model. So that's the quick version. This is all super new. I mean, we're talking, what, almost a year maybe on the mid journey private beta. So very, very new. So let's dive right into DALL E first.
What is it? So DALL E two, we talked last time a little bit about DALL one, what it was. Dolly two is made by OpenAI. You access it through a website. We do understand, some of how the algorithm works. I'm not gonna dive into that today. If you go to the references at the end, there's some more info about that if you really wanna get deep into it. It's a little bit of a black box, but as we'll see in a minute, it's more open how it works than mid journey is. So if you go on to OpenAI's website, which is I can barely see because it's so small, but I think it's labs.openai.com. You sign up for an account, and you hop on there, you're presented with the screen right here, which is pretty much asking you for a text prompt or you could upload an image.
So in this case, I think it gave us this one, and I just used it as kind of our standard prompt we're gonna use for the rest of the session. A pencil and watercolor drawing of a bright city in the future with flying cars. So that's our prompt, and we're gonna tell it generate. There's not a ton of options, right off the bat with Dolly. And so here's what it generated. It gives us four different variations on that prompt, and we can see they look like a pencil and watercolor drawing, not particularly realistic, but there's flying cars. There's futuristic cities. So then what can we do with them from there? Right? What can we do with these variations of generations that we've got? So we can dive into one. We can blow it up a little bit.
And then from there, we can hit the edit button, and that'll give us the ability to generate more variations, download this, gives us the prompt that we used, but not a ton of options, not a ton of end up. Here, we generated four variations from this original one. So there we have on the left our original image. And then to the right, there's four variations on that one. Now if we go back and look at our original four images, right, you can see that the one on the right, A car rocket ship, cars driving on the highway. The variations all resemble that one a lot more closely than the original set of four images that it gave us. And can kinda keep going down that rabbit hole for a long time and keep picking variations and generating more variations until we're happy.
But we can also do other things with it. So we can go to edit, and we can out paint. So what is out painting? It's taking that original image, and then we're going to add we're going to paint outside of that image. Now, you know, we can talk more about the details of that. There's some strategies for how much do you wanna paint outside of of it. It. You You wanna wanna give it some context. You have to leave the box generally. Some inside, you're bound by resolution and sizes, so you can't go create something five times the size of this at once. You you you can only do so much at one time. So we out painted, and it filled in the photo to the outside. So if you look, that's what we had originally, and you go to the left, it essentially guessed at what should be in that painting.
Now I'm gonna show you a a far more advanced version of that. This is what you could accomplish without painting. It would take time, and it would take trial and error, but really exciting, really cool technology out panning. So let's see. What's our next thing that it can do? Ah, how much does it cost? So DALL E two, I will admit, is the one that I'm the least familiar with. I used it early on, but I didn't like the lack of ability to fine tune. It creates, some really great outputs. If we were going to compare these three tools, kinda like operating systems, I would say DALI is the Microsoft. Right? Midjourneys may be the Mac, and stable diffusion is more of the Linux. So DALL is definitely sort of the more corporate, more realistic, more, less options. So how much does it cost?
So they charge based on credits, and $15 buys you a 115 credits. One credit lets you submit one text prompt to the AI, which is gonna give you back the four images that we saw. So in other words, $15 gets you 460 images, short version. On top of this, users get 50 free credits in their first month, and then you get 15 credits a month after that. So you get 60 images. But as we'll talk about next time or later, 60 images doesn't get you that far. That sounds like a lot. But as we saw, if we go back to those four images, what if I don't like any of those? You know? What if that's not what I had in my mind for a pencil and watercolor for futuristic city? I can easily burn through 60 images just trying to get one that's somewhat close to what I want.
So so paper play, as you go, you have to pay. There's no subscription model there. So how does it work? This goes back to what we were talking about a little bit earlier. The DALL E two has a little bit of a different model than stable diffusion does. It's a little bit older technology. Essentially, it's using a prior and a decoder, and then there is a separate network neural network, which is what's shown on the right, which is called a CLIP, a contrastive language image pretraining. And the CLIP's a neural network that's job is to give it a caption, to give an image a caption. And the two work hand in hand, one generating images, one sort of captioning images. That's the short version in the references. You can dive way deeper into that, but it's significantly different than the way stable diffusion works, which we'll talk about a little bit.
So what's next for DALL E two? There's a new model that's an alpha testing literally as of, like, five days ago. Somebody on Twitter who was it? Emma Katnip was kind enough to share this with us. And you can see on the left is this new alpha model that almost nobody has seen, and on the right is the older DALL E two. Couple of the interesting things, visually, there's a whole lot to unpack there, but it gives you more cohesive images. In some sense, it gives you more artistic images and less realistic images, which is something that DALL E was known for was more of the realistic look and less of the artistic look. So it'll be interesting to see how this develops. Like I said, it's an early alpha testing. There's just not a lot known about it at the moment.
They're pretty sort of sealed box of information OpenAI is. When they wanna tell us, we'll find out, and we'll learn about this new model. It'll be interesting to see also, will this old model live alongside the new model? So mid journey and stable diffusion have always allowed you to go back and use older models. It'll be interesting because of how sort of walled garden Dolly is in terms of their options and what they'll let you do with it if they will, you know the new model will supplant the old model or if they will still allow us access to the old model. So we'll find that out. So now I
Speaker 5 see know. I mean, I I assume that was a possibility, but I didn't think they would close the doors like that.
Ben I I don't think anyone knows
Speaker 5 right now. Migrate though into other?
Ben I mean yeah. I don't know. It'll be very interesting to see how it works. Whether they force you to new use the new model, whether the new model incorporates the old model in some way. Like I said, DALL just has less app options shown to the user, exposed to the user. So it'll be interesting to see how they handle that. Midjourney, on the other hand, we'll talk about it in a minute, has had, depending on how you count, four or five different models, and now they're getting ready to come out with a fifth or sixth one depending on how you count it. So what is Midjourney? So Midjourney is an independent research lab that produces an artificial intelligence program under the same name, Midjourney, that creates images from textual descriptions, same as DALL E two. Now this is what's interesting. Nobody really knows what the underlying technology is.
Most guesses think it's probably a fork of or based off of stable diffusion, And there's some really good guesses that point to that direction, but nobody really knows for sure. So very interesting. How do you use it? This is also something that sets Midjourney apart from OpenAI or stable diffusion. Use it through a Discord bot, which is kinda crazy. If you go to midjourney.com and you click on at the bottom, there's some options there, join the beta or sign in. Join the beta takes you to a Discord invite. And if I remember right, it's the first Discord to have over 10,000,000 members. So kind of interesting the way this works. So you go into Discord, and there's two ways to access this. You can access it through mid journey's Discord server, and you kinda can generate in public, or you can interact directly with the bot in your own message, like a private message, which is generally how I prefer to work with it.
It still will output the outputs in public so people could find them or search for them, but I like to do it just because there's less noise. Right? I've got this one single chat history essentially with the mid journey bot that I can look at. So you call up the bot to generate an image. In this case, slash imagine, and then it asks for a prompt, we and use the exact same prompt in this case for consistency. A pencil and watercolor drawing of a bright city in the future with flying cars, and it produces these four images for us. Now you can also see as we're looking at these images that it it tells us what options we use. So upscale beta, quality of two. It's essentially, doubling the quality because that's just the default I have it set on, and then it's using version four of the mid journey model to create these four images.
So now that we've created these four images, what can we do with them? Well, we can upscale one. So in this case, I upscale this from I think it was 500 to 2,000 pixels. Now if I wanna download it, and it says on the side here, open on website for full quality. I click this web button down there. It would take me to a website. Albert, I see your comment. Yes. I will show you how to use private in a little bit. So we've got this one image. We can download it, but we've also got other options. We can do an upscaling redo. We can make variations. We can remaster it. I'm gonna hit on remaster for a second. We won't dive into all this today because we're doing an overview.
We'll we'll dive a little deeper next week into what all these options mean, but we'll remaster it, and we get two images that are very close to the original image. They're they're definitely based off of that, but they don't go too far afield. So that's the remaster option. And then there are a lot of options in the journey. So you can call up, first of all, you know, with the slash button just like anything in Discord, blending settings. So if we call up settings, here's some of the settings we've got. We can go back all the way to mid journey version one. We can look at mid journey version two, three, four. There's this Nidhi mode that's essentially a custom trained model based primarily on anime artwork, but it got some really good results and came out well enough that they made it its own mode.
We'll talk about it in a little bit. They're getting ready to come out with version five. There's also a couple of test versions, test and test photo. And then we'll talk about cost in a minute, but you can you can change the cost model by going half quality, base quality, high quality. You can also change how much freedom the AI has stylistically. So I've left my style at medium. You can go high, very high, low, and it will kinda give the AI freedom to go further afield from your original prompt. And then there's some other options with upscaling and remixing and all sorts of things that we can talk about a little bit later. So here, I've given it a I've set the style at very high. So if you look and you see the dash dash s seven fifty, that's sort of the manual way to set the style.
I've set the style very high, and I've told it if we go back to our original image that we generated, this one right here, I said make a variation, but do it with a lot of ability and a lot of freedom to change the style, and this is what it came up with. So you can see there's still a similarity. It's still a variation on our original image, but it is significantly different in in terms of artistic freedom from the original image. We could also take that image, and we could put it through the Nidji model, and it gives a very different flavor, different color palette, different drawing style. So you can tell it's a very different model. Let's see. Here is we're gonna do so now we're gonna look at downloading the image. Right? So we've got this image. We wanna download it in high quality.
How do you do it? You actually go back to the website. So the website that we couldn't use earlier to generate images is what we're gonna use to retrieve images, look at our history, because going through, one long Discord thread, doesn't work real well. So you click on the website, you go to the website, and it gives you the option to view this image. You can download the image in high quality. You can download the image with the prompt. You can go back to your archive and look at previously generated images. I can go all the way back the first images I I created. I can look at what model I used to make them, what prompt I used. I could make variations based on image I made eight months ago using a new model, which is pretty wild. Let's see.
You can also download a bunch of you know, bulk download images using this option. So maybe I've been working with the bot for thirty minutes, and I've made 50 different images, and I don't remember which are the four that I really want or I wanna play around with a dozen of them, I can go bulk download a bunch of images and then start processing them, whether that's upscaling or Photoshop or, we can talk about some of that workflow a little bit later. So what else can you do with mid journey? You can do this with Dolly. You can do this with stable diffusion, but you can use an image to make another image. So in this case, because we're in Discord, I have to get the image somewhere where it's publicly accessible on the Internet to give the bot access to it.
So I can either upload that image into the channel in Discord, or I can, you know, upload it somewhere where it's publicly accessible on a website. So in this case, I took an image that I already had made on the left. It's three mountain climbers looking in a mountain, essentially. And I said slash imagine prompt image, which I had uploaded into the Discord, And then I said, do it in the style of John Constable, famous English landscape painter. And it came up with these four images on the right, which we'll zoom in a little bit just so you can see it. Now sometimes the AI has its own ideas and it does weird things. Right? It thought John Constable wants a house in the photo. There's no house in the original photo. There's now a house. Certainly, in the style of John Constable, sometimes there's now two people, sometimes four.
So it certainly has its own ideas about the image, but it does guide it. There is a general thematic sameness to the image that comes out of that. And I'm gonna stop and look at questions real quick before we keep going.
Speaker 5 Yes. We have two questions on the questions tab. They're not from Sony. They're from members that put it on, you know, the regular chat.
Ben Alright. So can we animate images in mid journey? The answer is no. However and we'll probably talk about next time. You could take a bunch of images that were output in Midjourney and then run them through some newer models, generally stable diffusion based models or a couple others. So, no, not directly in Midjourney. And let's see. Can you please cover how we can use Midjourney for creating POAPs? Okay. So to use Midjourney to create a POAP, you're creating a square graphic. You're asking for what you want. Let's say that image on the left, we wanna make a POAP out of that. Or on the right, let's say we want one of these giant constable ones. Once once you're done downloading it, you could now use this as an image to make a POAP out of.
Now you would have to format it in such a way that POAP's happy with that they have their own guidelines in terms of size and resolution and roundness. And so you may have to do some editing after that. But, essentially, you could use mid journey to create an image. Poepathon actually does it every week on their weekly call now. They are generating based on the topics in the call. They take the topics in the call. They input them as a prompt into mid journey. It outputs, you know, a cat eating pizza or something, whatever somebody talked about that week, and then they make a POAP off of that. So it won't do it soup to nuts, but it'll certainly help you make the image for the POAP. And then let's see. So Albert asked, writing is not possible on mid journey. Do you have some workarounds?
So that is true. Both DALL mid journey stable diffusion struggle very heavily with text. You can tell them, make me a stop sign that says stop, and it will look like some sort of alien language. So text is definitely a huge issue right now. There's not really a great answer when I need text on an AI generated image, and we'll we'll talk about this sort of in the next couple lessons where we really get our hands dirty with this. There's still a lot of manual work you have to do right now. If you want text in an image, you'll end up deleting whatever gobbledygook text it created, and then you'll probably end up having to write your own text to get it onto the image. Let's see. Wootways asked, can you ask mid journey to create the images in specific sizes? Yes and no.
So you can ask it to give it to you, and this this gets into the weeds a little bit. So it depends on what model you're using, what version you're using, what you're using for upscaling. So Midjourney will allow you to do certain things. It might let you do 16 by nine or 16 by 10 or three by two, like a portrait. So it doesn't constrain you only to square. It will let you do, you know, some different sizes. However, there are constraints. You know? Like I said, it might say you can do three two portrait, but it won't let you do, you know, sixteen nine portrait, only landscape. Or so it certainly has and then whatever size it outputs it in, you're kinda stuck with. You can upscale, but it has limits. If you want an eight k resolution landscape portrait, you're gonna have to go somewhere else from an upscaling standpoint.
And there's certainly AI that does that. We'll deal with that in the future. But right now, it's very disjointed. You end up having to use different tools to accomplish what you want. AI is not a one stop shop in that sense. And let's see. Hire or hayer? Let me answer this, and then we will keep moving on. Are the AI platforms you mentioned distinct from one another each with their unique strengths? Would you be able to recommend one platform over the others, or are their performances similar across the board? I will get back to that one. So hold on. I actually have a slide on that, and we will discuss that. Alright. Let's hop back into it. Alright. So you can blend images. Right? We just saw that you can use an image to create another image. You can also blend images.
So to tie back into what WootWays was talking about earlier, I created an image for the goal getters, and it took three to 500 generations of images to get what we wanted for this one single image. And some of these were sort of rejects. Right? Not because they're bad art. They just didn't, you know, show what we wanted to show for NPeak. So the one on the left has multiple people. The one on the right only has a single person. I really like it artistically, but it didn't speak to the goal getters and the group process the way we wanted. So I ended up in sort of the reject pile. Right? It's done. It's a raw output that I've then done more work on. So I said, what happens if we take these two images and we tell journey to blend them? And there's more options.
You can you can tell Midjourney to blend three or four different, pictures. You could tell it to blend two. You can tell it to blend two and then give it a text prompt. But in this case, I just took these two images and said blend them together, and it came up with the four options on the right. So If you look, you can see, it took some of the, stylistic inspiration from both, photos. Right? On the left, we've got our three people. We've got three people on all of these. You've got more color than the photo on the left, more like the photo on the right. So it will blend images. You never quite know what you're gonna get. You may have to go through several iterations to get what you want, but you can use it to blend images, which is pretty cool.
So what does mid journey cost? Right? That's the next question. We went over what OpenAI Dolly two cost. Mid journey is more of a subscription model. Starts out at $10 a month. You can go all the way up to $60 a month. There's different options. They generate images based on sort of how fast you're using the processing power. So whether you want the image fast, slow, they've got some different options. I usually bounce back and forth between the 10 or the $30 depending on how much work I'm doing that month. I know, Sami, a while back, was looking at the pro plan because it includes some privacy options, where people can't see what you are generating. Right? And there's different views and different reasons around this. Right? As an artist, I don't know that I care that much. You know? How is somebody gonna front run my art?
I guess they could. But as a business, that's a real concern. If you're branding or using this to create, you know, some sort of IP, you know, you want that privacy and that and, essentially, that costs more with Midjourney. So what's next? This just came out yesterday afternoon, evening. Midjourney is getting ready to come out with version five. We talked about they've got version one, two, three, four, and the Nidji model. They're getting ready to come out with version five. But first, they will have the community help them train this model, and then they'll release it. I don't know how long it'll take. Nobody does, but that'll be the next thing that we'll see with mid journey is their version five. One of the interesting things with mid journey, because they've had so many versions out over the last year, we can see how these versions have improved.
And so version five, I have no doubt, will be really interesting how far it's come just like version four. So if you look, the first image is version one. The cat's face is deformed. It doesn't really look like an airplane. In the second image, it it has that sort of waxy appearance. It's starting to look like a cat in an airplane. In the third image, it's it's finally starting to look like what we asked for, an aviator cat that pilots the Red Baron. And in the fourth image, it now looks like an aviator cap that pilots the Red Baron. Here's also on the left is version one asking for a stainless steel sink in a kitchen. On the right is version four. There's all sorts of strange visual artifacts and deformities in the outputs on the left. The output on the right looks almost like a photo.
And then here is DJ Khaled as a menu, version one to version four. So you can see in six months, the huge leaps this, mid journey has taken, just in in six months of iterating. So let's see. Stable diffusion. What is it? So stable diffusion is a company that essentially ran, the world's fifth largest supercomputer, and trained it on a bunch of datasets with a latent diffusion model, which is a different type of model than DALL E two is using, a little bit more advanced. And in August, they launched stable diffusion. And the interesting thing about this is they open sourced it. Now you can't run DALL E two on your own computer. It's it just takes too much processing power. In fact, most estimates say it took I forget. I was reading this the other day.
It was 300,000 GPU hours to train Dolly two and hundreds of thousands of dollars just to train the basic dataset. Whereas stability and stable diffusion, you can run it on your GPU, and it's and a lot of that is attributable to the difference between the latent diffusion model and DALL E two's diffusion and clip. And so, anyway, they open sourced this in August and allowed people to start building with it. And this is interesting. At one point, I think it's in here. Yeah. Four of the top 10 applications on Apple's App Store were powered by stable diffusion. So as I said earlier, if we were sorta comparing these, for sure, you could argue about whether Dolly two or Midjourney is the Mac or the, you know, Windows. Stable diffusion is definitely the Linux of AI right now, which is good and bad. It's
Ben It is definitely more accessible in the sense that you don't have to spend as much money, but it's not as accessible in the sense that you have to nerd out a little bit more to use it. And so we'll we'll dive into that. So how do you use it? And the answer is there's a lot of different ways. So with the other two, we're talking about how do you use it. You log into this website. You log into Discord. With stable diffusion, we're not gonna be able to cover all the ways you can use it because you can run it on a Mac, on a Windows, on a Linux machine. There's a bunch of websites you can access it through. The official way you can access it from Stability AI, is through this website, Dream Studio, and you can tell right off the bat when you look at this that there's more options exposed to you than there are with Midjourney or Dolly.
So it's asking you width and height, how many steps you want, how many images you would like. It even asks you which model you would like to use, whether you would like clip guidance or not. So in this case, we use the exact same prompt, and it gave us these four outputs. And then once again, there's a million options, and trails you can go down with this. But the next question would be, how much does it cost? And in their case, they go off a credit model, same as DALL E two, where credits create images and you pay based on credits. They also do give you some amount of free images per month, a very, very low amount. So then let's move on. I don't know too many people that actually use this to create. You know? It's more of a test thing.
Try out a few things. I'm sure there are people that use it. There's just cheaper and more powerful ways to use stable diffusion than using DreamStudio, but it's kind of the official way. If you just wanna hop down and tinker with it, that's a good way to see what it can do. And all
Speaker 5 of these options are fiat? There's no crypto options for any of these?
Ben I don't think there are. I think they are all fiat right now. So you can actually hop in the Hugging Face, which and you can actually use a demo version of stable diffusion for free right now. The link for this is also in the references that we'll share in a minute. But you could ask it to create an image with a different model. This is the 2.1 model. There's also sort of a crowdsourced version called the stable hoard, where it's sort of a quid pro quo form of creating images. If you give them processing power, they give you processing power. It's a giant cluster, and they have several different options for creating images. Through this, you go to stablehoard.net. This is the Artbot, which is one of the versions, chock full of options, different models, and you can create images on the website, and the speed will be based on essentially your score.
How much have you given back to the network? You can pay for it. They offer it for free. Let's see. Another option. This is pirate diffusion. It's actually a telegram bot where you can make images. So in this case, their their web access is scum.co, and I logged in and asked it the same prompt, and it gave us some different outputs along with all the options that used at the bottom down there. There's also scribble diffusion, which uses a version of control map. We'll talk about that in a minute. Very new, but, essentially, you can hand draw something, tell it what you drew, and it will produce an output. Now this cat wasn't that impressive, but it was pretty quick, and it was a pretty bad scribble. So pretty cool, though. It's free. You can log in and and try it out and see what happens.
Let's see. You can also run stable diffusion on your own computer. We talked about that. This is probably the number one way to access that right now, this automatic repository. The install instructions are not super easy. You gotta dive into the terminal. You have to do some Python things. Sometimes upgrades break things. So I wouldn't say it's user friendly by any means. You also need a powerful enough GPU for this to actually be beneficial. I tried it on an older computer a few months ago, back around Thanksgiving, I think it was, and it took, like, an hour to make an image. So it's not terribly practical at that rate. It also has a GUI access through a local host website on your computer. Pretty cool. Not as polished, obviously, as Dolly or Midjourney.
So now there are some more polished things that are coming out, and I'll talk a little bit about the Mac side of things just because that's where I've found the best success with stable diffusion because I've got a new Mac with an m one in it, and it can produce images pretty quickly. It can put out four to eight images in thirty seconds or a minute. So the first option is diffusion b. It's a GUI. You download it off of a you go to I think it's diffusionb.com. It's also referenced in the references. But you go there, you download a Mac app, and this thing installs on your computer, and it's got a nice sort of Mac ish GUI, and you tell it what you want. Now there's some intricacies to that, and we'll talk about that in a minute.
When you first install it and tell it what you want, you're not gonna get the same results as you would get out of a mid journey or a Dolly two, and we'll talk about why that is in a minute. But it does immediately expose more options to you. So it allows you to do image to image. Right? Like we're talking about, here's an image. Make me something out of that image. So in this case, the fellow uploads a picture of himself, and he says, I want me drawn as if it was a painting by Van Gogh, and it outputs. But you can also see it asks a whole lot of, questions about how to how to make it, input strength, number of images, steps, batch size, guidance scale, seed. You could also input a cartoon drawing, photorealistic render of people on grass with sun and clouds, unreal image, and it'll produce that off of that image.
So it has an image to image system. Now this is cool. So this is inpainting. And to I know Midjourney doesn't offer it. I don't know if Dolly does, but stable diffusion allows you to inpaint. So you can take a part of an image and tell the AI you only wanna change that part. So in this case, this fellow wanted to change his pink hair. He tells it to do that, and it does that. You can access it through this image to image, or you can access it through a separate inpainting section in this app, which is where I say things are a little less polished with stable diffusion. You can do the same thing in two different parts of the app. It's open source, and people are maintaining this for free. So in this case, they want a dog sitting on this bench.
So they mask off the portion in the middle of the bench where the dog needs to be, and the AI generates the dog sitting on the bench with changing the rest of the photograph.
Speaker 5 So do you think, like, this will make people kind of not wanna learn Photoshop and just come straight here?
Ben So in my experience, at least thus far, it's not good enough. Like, you tell it to inpaint something it's really good at deleting things. It's not so great at adding things. It takes a lot of tries. It it creates some weird malformed looking things. This is a great example, but that is not how it usually works. So, yeah, not yet. I don't think we're there yet. I do think we're gonna get there, and we'll talk about ControlNet in a minute. It's getting scarily close to that. So here's out painting. We looked at Dolly's example of out painting. Now we can see, stable diffusion. It's gonna show you as it works, and it may take a few minutes. And then it now very different, right, than the output Dolly gave us, but it still does outpainting. Let's see. So, here's the fusion b on my computer.
It first came out, September 9. So shortly after, it was open source, and we're gonna give it the exact same, prompt as we've been giving for all the others. So a pencil and watercolor drawing of a bright city in the future with flying cars, and it gives us four options. Now this is using the default model. So this is stable diffusion. I forget 1.4, 1.5, a little bit older model, came out last year, and it looks okay. I would say less impressive than the default results we got out of mid journey or dolly, but decent. Now here's where stable diffusion gets really interesting. You can apply custom models. And so in this case, this is a site called StableRes. I picked this one because the models are generally safe for work. There's other websites that have way more models available like civet.ai, but you have to be careful because some of the models are not safe for work.
So in this case, though, we can see there's some different models here. There's a double exposure model up in the top right. There's one that specifically works on making characters look like Vulcans. There's one that makes art look like paper cuts. There's one that does complex line art well that sort of specializes on that, and that's because these models have been trained on a specific subset of images. We'll look at that in just a moment. So you can install these models into your local install of stable diffusion. So in this case, diffusion b, you go download these models from a site like stable res or hugging face or and I install them into the program, and they will produce different results with the same prompt. So in this case, we've got our same prompt in our original set of images.
Here's that same set with a different model that's, attempting to copy mid journey version four. That's sort of what this model's been trained on is a bunch of mid journey version four images, and it produces these outputs, which you can see are different than the base model. Here's another model. Let's see. That does complex line art specifically well. And once again, you can tell that even with the same prompt, the pencil and watercolor drawing, it definitely adds its own style to the four generations that came out, the variations. Here's another one. Synthwave, it's focused on sort of cyberpunk art, and you can tell these have that brighter, more neon cyberpunk look to them. Now this is interesting because you can this one is a vector art model, but we told it to do pencil and watercolor drawing.
So in in a sense, we've kinda we're trying to tell the model to do something it's not meant to do, and you can tell it's struggling with that. So you do kinda have to keep in mind what what the model is trained on. And if you get outside of that, you're probably gonna get some weird results. Here's one that's supposed to do paper cuts. It's trying to do, like, a advanced paper cut look, and you can tell on the far right picture, it's trying to do exactly that. It's taking the photo and putting it inside of as an paper cut model. Now you can also go back and look at your history, and you can see what prompts you used, what settings you used, even what seed you used, what sort of random seed you used when you were generating this work. So that's diffusion b.
It's been around for a while for Mac. Decently fast, but very slow development wise lately. The maintainers are moving a little slow. Kind of the new kid on the block for Macs is Mochi diffusion. Just came out in the last month or two. It can't do as much as diffusion be yet, but they're moving very fast, and it actually runs faster because it's been optimized for the new Apple silicon and the CoreML architecture. So this actually generates images quicker, but it's just not quite as polished yet as diffusion b. So here, once again, we're gonna use the exact same prompt, but we're gonna use a different model down here. It's a version of DreamDiffusion made specifically for the Apple Silicon, and it produces some pretty high quality images, probably two or three times faster than diffusion b, and this is mochi diffusion.
All these are in the references as well. But this gets a little technical for the models. It requires a very specific type of model that's been produced for Apple silicon to work correctly, and it's very easy to get it to not output a picture at all if you don't set up everything correctly. Alright. So what's next for stable diffusion? So the next thing kind of the next so this maybe answers the question about can this supplant Photoshop? And I would say ControlNet would be the thing that that kinda starts getting us there. So control net's a neural network that controls diffusion models by adding extra conditions. And what do I mean by that? It modifies the text prompt embedding with another embedding derived from an additional input. So the input can be anything that provides more info or guidance for the image generation process.
So it could be another image, but it could be a sketch, a post skeleton, a depth map, a scribble, and we'll look at a couple examples. So on the left, there's this input, just the edge drawing. Right? Next is the photo that we're modifying, and then we have the prompt, the automatic prompt that the AI would come up with. But then on the right is the user prompt. Right? So in this case, on the top one, it changed this man with a child to a woman with a child. The second one, it changed the man in a gray suit and tie to a white suit and tie. The bottom one, it changed this cat to a different type of cat in a different environment. So it's it's pretty impressive. It can be used to enhance diffusion models in various ways. Right? So you could refine an existing image.
You could transfer Styles between images while retaining sort of most of the original image. You could edit specific parts of an image at a very fine tuned level with this in a way that's far more accurate than in painting. You can control poses. We'll look at that in just a second, and you can add effects. So let's see what are next. So here's an example of the scribble use of ControlNet. So somebody scribbled a picture of a turtle, and you end up with some you know, a masterpiece of cartoon style turtle illustration on the right. The bottom with the balloon is really impressive. Right? Somebody draw drew a hot air balloon, and you end up with a magic hot air balloon over a lit magic city at night. Let's see. So this is also cool. This is using so boundary detection.
So ControlNet's HED model so, essentially, it took this photo. Let's let's take a look at the bird photo. Top left, there's a bird photo, and the model itself looked at the edges and then was able to create the images on the right after auto detecting the edges in the image. So pretty wild when it comes to slightly changing an image. This is what I was talking about about pose detection. So you could take a pose. In this case, this child's leaning back, arm crossed over his body, and then you could apply that pose to other generations while maintaining the exact same pose. So ControlNet's really wild. I haven't had a chance to play with it as much as I wanted to yet, but I kinda think it's it's a game changer, in a sense. Also, Laura, which there's actually two Laura's.
There's Laura, which is large scale open domain realistic artwork, which is itself a collection of diffusion models that are fine tuned using low rank adaptation. What does that mean? Short version, it's a technique that reduces the number of trainable parameters and speeds up the fine tuning process. So a LoRa model can be permanently fused to a diffusion model or dynamically loaded. It replicates art styles such as anime, cartoon, painting, sketching in a way that loading new models every time and training new models doesn't. So this is also gonna be something very significant we're gonna see. This is pretty darn new, but I think we'll start, seeing this and ControlNet together, creating some pretty wild outputs very soon. So back to one of the earlier questions, which one is the best? So I asked Bing AI which one is the best the other day because I figured what better way to learn about AI than Ask AI.
And it said, don't have a subjective opinion on which one is better as they all have their strengths and weaknesses. It depends on what kind of images you wanna generate and for what purpose. In this context, a human might say DALL CHU is better for realistic and detailed images, and I would agree. Midjourney is better for artistic and creative images, and stable diffusion is better because it's open source and because it's more accessible in the sense that you can install it on any computer and really fine tune it and tinker with it. Now I posted this on Twitter yesterday, and it made me laugh enough that I wanted to share it with y'all. I said, Bing, please show me some examples of representative images from DALL E two mid journey and stable diffusion for the same prompt, a cat wearing a hat, and this is what it gave me.
Obviously, we're not there yet. Bing AI is not gonna take over the world because this is what it gave me for an answer.
Speaker 5 I mean, it tried. It tried.
Ben They they call this, I think they're calling this a hallucination, you know, an AI hallucination that it just made up a BS answer and gave it to me. But I did find some representative images. It's nice to see them next to each other after going through all these different slides, and you're trying to remember, well, how different were they? Now this is also gonna change. Right? This is already out of date. This is what mid journey version two or three looked like. This is what stable diffusion one model looked like at one point. DALL E is about to have a new model come out, but this gives a very rough overview of the differences. Here's another one. I think the one at the bottom does a good job of kinda showing where things were two or three months ago.
Stable diffusion would tend to give you almost a photograph that wouldn't always make sense. DALL E would give you something that looked much more realistic, and mid journey would give you something that looked like an oil painting. Like I said, that's changing with new models and with advancements, and I don't think that's as true as it was in the past. Now I wanna talk for about five minutes about some alternatives and what's coming up, and then we'll wrap it up. So Google has a model called Imagen, and Google's model is extremely good at creating realistic photorealistic photos. And you can go to the website. It's also linked in the references, and it will show you some examples. But right now, it is a black box. You can't access it. You can't use it.
People that have said it's amazing and it's awesome, but nobody knows if and when Google's gonna open it up for people to use it or experiment with it. But it's out there, and I would say it's significantly better at creating realistic images than DALL E. One of the reasons I've heard they haven't let it out into the wild is they are concerned about sort of the deep fake concerns with this technology. It's so good. We won't dive too far into the details, but, essentially, it can fool most people most of the time into thinking it's a real photograph, including other AIs. So very interesting, something to watch out for, and know is is out there.
Speaker 5 I mean, at the end of the day, as we were speaking with other AI, you know, specialists, it's about changing the humans. The AI, the tools is not the problem. It's us behind the tools, how we create them and use them.
Ben Oh, if we were talking about video, there are some scary, scary deepfake video technologies out there. So Crayon is an alternative that actually came out in June 2022. It was an independent project that inspired by OpenAI. It was originally called Dali Mini. I think they got in some trouble because of the name. They rebranded it Crayon. It generally produces some weird outputs. These are the outputs it gave us. Very simple, very artistic, but it's free. Anybody can go to Crayon. I forget what the ending is. It's in the references also and use it, and play around with it. Another alternative and most of these alternatives coming up are going to be based on stable diffusion. Because it's open source, somebody's got a server farm. They've got a website, and you can access it and use stable diffusion without installing it. And oftentimes, they'll give you some amount of free credits.
So this one's called lexica.art. It's obviously a front end for stable diffusion. But back to Wuwei's question about dimensions, it gives you some fine tuning over the dimensions of the photo. In this case, it gave us four portrait photos. They look good. It's using a good model, so that's an option. Dreamlike is one of the earlier stable diffusion front ends. It uses a custom model called Dreamlike Diffusion. I actually use that on one of the images we looked at earlier. It's a model that you can download and, you know, install on your Windows or Mac build of stable diffusion, but they also have a website. You can hop in and use it. They give some free credits. You can buy credits. Also made some good looking generations here. Nightlife is probably one of the most interesting front ends out there right now. I'm sorry.
Night Night Cafe. The the spell check kept wanting to change it. Night Cafe. It's in the references. This one's really interesting because they actually open up different models. So if we look at the next photo, you can actually use it to generate using stable diffusion or Dolly two or some clip and GAN models that we'll talk about in a few weeks. So really interesting, little bit expensive, little unintuitive, but NightCafes really kind of a cool alternative. There is also some stuff coming down the pipe. I had a question about this the other day. Somebody messaged me directly, and they said, hey. I'm hearing about this aluna.ai. I think it looks really cool. I like some of the generations I've seen, but if you go on their website, there's almost no info about their image generation process and what's coming up.
He said, can you make something like this with mid journey, the photo on the left? And I said, sure. Let me take the prompt, and let me just run it through a few different models of mid journey quickly and see what I can dump out. So on the right, I think that was the Nidji model produced these four images, And then I did a couple versions of v four and came up with these eight images. So Eluna's interesting, but nobody really knows what it is. You can certainly get comparable results at the moment from mid journey and stable diffusion, but it's out there. We don't know what it is, but very soon, they will launch something. And then this is really interesting. This answer is something I've heard Somi ask a couple times. She's talked about, could you use AI to generate a PFP collection?
And Automender, which is a tool that helps to launch and generate, you know, 3,000, 5,000, 10,000 piece PFP collections just came out with an AI generator. It's in beta. It came out, like, last week. So you go on to their website. You click AI generator. You enter in a description. So in this case, a head portrait of a robot dog, and it says, which style do you like better? And so in this case, let's say we pick the one on the left. And it says, okay. How many do you want to make? What traits are there going to be? So in this case, it's 3,500, and then you add the traits and quantities, which each trait should occur. And so you could say, you know, I want dogs with a rainbow or with a butterfly, and I want it to occur 5% of the time to and it will sit there and generate 3,500 different, PFP NFTs using AI based on what you gave it.
So I I think that's coming soon and back to kind of the the humans need to be better trained. We need to have a good eye for, you know, was there a real artist that spent a ton of time on this, or did somebody use this tool and crank out 10,000 PFPs, and now they want your $100 and why? So really interesting. Not it's in beta right now, but it's coming. That's all I've got. I included a couple of, AI pieces that I've done and haven't released, along with a quote. I always love to have a quote sitting on screen. So, what questions does anyone have?
Speaker 5 Let's see. I see a lot of people are so surprised of how fast everything is moving and all the things we can actually do. I'm I'm feeling it for the artist right now. I know it's not the same, but still, like, the average person that gets their hands on these tools and actually works with them and creates the inputs that they're looking for, the outputs, it's gonna be incredible. I see a question coming in from Youssef.
Ben I was reading Albert's comment, which I'll respond to really quickly. So, Albert, you're a 100 correct that there have been multiple AI generated Ment drops. I think the difference here is, you know, before you had to go to a little work to do it, now there's a tool to make it really easy. The tool will generate them for you, mint them as NFTs, handle the contract. I think that's the difference is it's it's gone to a turnkey solution. You know, there's even less technical knowledge that you need to have to do that. Alright. Let's see. Joined the session a few minutes ago. Don't know if you have answered the questions already. Is AI 100% efficient? Are there known risks or limitations? So I will say this. Obviously, this is about AIR. I'm enjoying AIR. I'm a little bit of a voracious reader when it comes to technology in general, have been for a couple decades.
So if you go to the references that I link, there is an entire section on the bottom of it entitled epilogue, the last invention that man need ever make that sort of gives the bad, you know, risky side to AI, and it walks through what are some of the best technologists and philosophers worried about in terms of artificial intelligence. You know, what what are the risks? Where are we at in that risk timeline? So I would say if you wanna read more about that, dive into that. I love reading about it. You know? I don't know how much we can do about it, so I'll just use it to create art at the moment. But but it is out there, and it does it is a concern long term.
Ben Okay. Let's see. I can't answer another question. That one's stuck on screen. There we go. Alright. I'm a photographer and want AI, to make my style of photography. I would say stable diffusion would probably give you the most control over that. I have seen some really good results with mid journey, but I would say if you have a very specific style in mind with stable diffusion, you could train your own model based on your style, and we can talk about that. I don't know if we'll hit it next time, maybe the time after that. We're gonna talk a little bit about custom models. So you could literally take a bunch of your photographs and train a custom model and then put that model in disable diffusion to make more of your style of photographs using that. So, yes, I would agree. Let's see.
In the customer perspective, in case I go to a freelance platform and I ask a freelancer to create an illustration, how can I ensure that the artist didn't create everything on AI? So there are a couple of tools. I didn't include them in the references yet because I just ran out of time. But there are a couple of tools that do a pretty good job of this. They can they can try and tell you if it's made with AI, and then there are even some that if you know it was made or suspect it was made with AI, they will try and show you the images used to create, you know, that the model was trained with, or we'll try and give you the prompt that was used to create it.
So I will make myself a note to add those to that references, list, and I will let you guys know in the Discord chat when I do that because there are a few tools for that. It's not perfect. Like I said, the new Google, Imagen, is able to fool some of that. But, yes, there there are some tools, for that. Alright. What do we know about the images that feed the datasets in the respect of their copyrights? Alright. So there are some articles about that from the first session. If you go back to lesson one, I would say there's a lot of different viewpoints on that, and there's even contradictory expert technical opinions on that right now. So there are people that say that the datasets that were used and, you know, involved a lot of copyrighted material, and that should make the datasets themselves invalid and unable to be used.
There is currently people working on open non copyrighted datasets. So I think that concern will disappear over time as new datasets come in. There are people that say that you can essentially get, an AI diffusion model to spit out almost a perfect copy of images that it was trained on, and then there's people that say, no. That's a 100% not possible technologically. At the moment, I'm on the fence. I'm not, you know, trying to go out there and and put out images on, you know, Getty Images or, you know, I'm more having fun as a hobbyist making making art. But, yes, I think it's a real concern, the datasets, and I think that'll have to get figured out. I think it will get figured out. But, yes, I do think there are some concerns around that. Although I don't think, you know, it's a 100% either way right yet.
Alright. I think that is it. So next time, we will meet in a couple weeks, and I wanna dive into actually trying to make a few pieces. Let's get our hands dirty. Let's try and make a few things. We'll probably use mid journey and stable diffusion. We may even take something that we export out of mid journey, dump it into stable diffusion, take it back, put it back in the mid journey, and, we'll have some hands on fun next time. So thank you, guys.
Speaker 5 Thank you, Ben. See you next time, guys. I shared the link with you all on the chat, so please go ahead and RSVP so you don't miss our next session. And this one will be available shortly on demand. So thank you, Ben, and everyone have a great rest of the day. See you.