How to create AI art with Disco Diffusion

1. Course Intro: Hi guys, welcome to this course, creating AIR with discrete diffusion. So what is this weird fusion sounds a bit funky. Well, it's one of many programs out there that you can use to create AIR. How do you do that? Well, this course we'll be going into it and a lot more detail. But essentially you can take a line of text or input a line of text to create some unimaginable crazy artwork. This course is going to, first of all, take a look at some of the work has already been made with disk diffusion. Then we're going to actually get into the nitty-gritty and look at how we use the program. All you need to use this fusion is a Google account. So as long as you've got that you are ready to go. We'll finish off the course by looking at a few tips and tricks and also a project assignment at the end where we will all create different art work and share it. So I hope you'll join me on this adventure and I'm looking forward to seeing you guys in the course. 2. Lesson One – Existing work made with Disco Diffusion: Hi folks. I thought in the first lesson it would be great to start off by looking at some work which has been created with disco diffusion just so we can kind of give up and get ready and see the potential in this software. So I'm gonna spend two minutes going through stuff that other people have created and then I'll reveal at the end what I've created with it so far. So as a few articles here, and what I'll do is I'll link these in a Word document or Notepad document, just so you can have a look at these links as well. But if we scroll through here, it has been used descriptive fusion. Now you may notice it says disk diffusion 4.1 or 4.2. Basically there's many iterations as time. So that says 5.2. So that's quite recent one actually. And it just gets updated all the time. 5.4, I think, is the one that we're using at the moment. So anyway, this is a bit of a run-through on it. Here's some images PER here, so that's a bit of a mad one. And don't think there's too many in here actually, but let's have a look. This was a good link. So this artist, Theodore McKenzie, created a range of different Roman villas. So we gave it some text prompts. And I mean, that's a bit of a weird one. You got some stuff float in there, but still pretty cool. A lot of Fargo, I reckon that you would have used fog is one of the keywords, but a range of lovely, lovely images created with that. Then we've got over here on Reddit 17 images generated with disk diffusion. They've used enough another AI pieces, piece of another way I program called swim IR to upscale the images as well. So you can see obviously a different style that this artist, that this person has gone for. And this will be due to some of the texts prompts that the guy gave. So I think there's some more over here. Yes. So you've got some more sort of environment stuff over here again created with Dischord of fusion. There we go. Um, so yeah, I have a scroll through. I'll send you the links for all of these. Now let's have a look at what I've created. So I'm going to start off by talking through some of these, this one really, really didn't work. I said something along the lines of my text prompt for this was something along the lines of Yellow Submarine, psychedelic, Savile, civilised and really trying to get something like The Beatles album cover, but really didn't work that one. Then I'm gone for ice. I think I've put a colony on Mars, Red Planet, futuristic city, those kind of text prompts. And this is the kind of stuff we've got really, really lovely. I think I put concept art and digital art as well. So this is really, really nice. I love some of this stuff, particularly this one. Then for this one, I put in and we look at prompts later. But I've put in something like slimy insect monster and Cyberpunk, I think. So you can see attempts on some of the sliminess here, some wings, they're from insect and so cyber punk, it's all in there. So that kinda got me some quite interesting stuff. Yeah, slimy insect, cyber punk, monster kind of stuff. Yeah. I try putting into words like Pandora and Avatar environment. So I want you to get an avatar kind of thing. Some of us a little bit messy, but some of it is quite nice. One that I did find that really worked was the name Simon style and HOG that worked in the text prompt. And it got me some stunning stuff. So just to have a look through some of these, I think I'll put like futuristic cyber punk city in the style of Simon style, style and hard, which is one of my favorite artists. And just look at just the detail. Look at the detail in that it's just stunning. And gave me some really nice images. I'm just loving this. I mean, the reflections here, reflections aren't actually totally accurate. But still it doesn't matter. Just look at this composition of that. So these are some of the results. I mean, that one, again, just lovely. As things get further away into the distance, you can see the haze, atmospheric haze coming in. So quite realistic as well. These are just some of the examples which you can go ahead and create. I've got some other stuff in here as well, which I kind of went through. It doesn't really like faces, but I did put this in the style of an artist who does do sort of weird obscure faces. And I put the word psychedelic in there as well. But yeah, I mean, I've got spaceship, spaceship crash landed in a desert with some of these ones But here and kind of mixed reviews. But particularly the Simon, Simon style and hogs stuff here was just honestly unreal. So what I'm gonna do is I'm going to finish this little less than that. But there, again, this just looking at this on screen is mad to think that this is created by AI. But I add the only thing I gave it was a sentence of texts and just look at how stunning assist. Yeah, cool. I'm going to leave it there. And what we'll do in the next lesson is we will start actually taken a look at how the Colab notebook disk diffusion works. What we need to do, and I'll get you started. So I'll see you in the next session. 3. Lesson Two – How to use Disco Diffusion: Okay, so here we are. I mean, disco diffusion 5.4. Now I'm going to provide a link to this however, or you can just copy this, but I will provide a link as well. You will want to take note, I'm recording this on the 23rd of June 2022. There's new versions being given all the time, so you'll want to check that. Now this looks pretty confusing, but I really don't want you to kinda get overwhelmed by it. Okay? First thing I'm gonna do is just this little drop-down over here. I'm going to click this. And then I'm going to click the one on number one and number two. And I'm going to click that one and number three, number four, and number five, right, so we've got a little bit more basic setup. We're gonna go through each of them one by one. In the setup. I don't need you to do anything with that. You can leave that what will happen is we're going to run this. We're going to go up to runtime and we're going to click something. And it's going to run through each of these little 1.11.21.3. So the notebook is going to install an important dependencies and set up runtime devices. Then it's going to define the Midas functions. I don't even know what that is. The beauty of this. I'm not a programmer, but you can use this Google Colab to do this for you. So we're going to come into number two. We're going to start to come down and I'm going to go through a few important things. You don't need to do anything in this one. And in fact, I think let's just come into custom model settings. Note number two, you can actually leave. So yeah, that's all good. Number three is the one we want to come into. Number three. Okay. First of all, batch name. I've always left this to time to disco and arguably I should probably change it. But if I go to my Google Drive, you'll notice that when it connects, it'll do the AI folder, disk diffusion and images out, and time to disco. It's your project name. Okay. I'm a little bit messy and I've left all of these Kelvin there, but probably shouldn't do that. So you can give that a folder name. Let's just do what should I call it? Skill Share? Skillshare lesson. Okay. I'm going to put an underscore instead of space because I don't like spaces. Steps. This is how many iterations that goes through. And I'll explain that. Well, if I go into my time to just go and go into partials, what you'll notice is, let's scroll down iterations as double-click this. First of all, let's start off with a bunch of noise on screen and then step-by-step. And I asked you to do it in step to 25. It gradually gets rid of that noise and starts to reveal and build detail into an image. Now the reason I save out all these partially done one zone may think, actually, I like this. I just like it quite blurry and Bland is. So you've got, you've got every 25 iterations I get it to save. Now, what we're going to specify over here is how many steps we wanted to do. 250 is a probably a definitely a good start. It depends how detailed you want it. Now, let's come to the final one. Let's say, for example, I'm getting to here and I'm like, Whoa, as too detailed. Well, I'd always argue, give more steps than you need because you can always go back and take one of those partials out. However, the more steps you do, the longer it takes. So as a rule of thumb, I normally go for 450. Okay. It takes a little bit longer, but it gives you more detail on it. It goes through these, you can see, starts noisy and gets better and better and better. That's just how it works. Okay? Width, height. What you need to know about this is this needs to be a multiple of 64. So if I get my calculator out, just to illustrate that, let's bring this onto the screen. 80 divided by 64, some like 20. Yeah, there we go. Quick maths. And then 768 divided by 64 is 12th. Okay? So I, this ratio, 1287768 is widescreen. So if I just type in 1280 by 768 bits, you should hopefully get an image of what that looks like. It's saying that that's a wallpaper of that resolution. Let's just type in resolution. And you're going to see that Twelfth Night or less put ratio. So I'll show you the aspect ratio then. That's gonna be a ratio of 15 by nine. What I do basically is I up this, I put this generally to 1344, so I add an extra, I add an extra 64. Now, you don't have to, but if you do adjust these, the final image, width and height, they've gotta be multiples of 64. You could do 640 by 640 if you want to, to, to get a proper square. The rest of the stuff and you don't need to worry about. So it's just these three. So let's just scroll down this bit down here. You can give it an initial image, but it doesn't work nearly as, anywhere near as good. So we're not going to look at that. Don't worry about animation settings. Let's have a look at extra settings. Now, intermediate saves. So above, I told you we're gonna do 450 steps, 450 iterations. Now, if I take this to 0, it's going to save out these steps which I've gotten here, these partials. It's going to save none of them. But I may say, or what I normally do is say, save an image every 25 iterations. Just going to do 450 iterations, but it's going to save an image every 25 and intermediate in sub folder. It's already ticked. Perfect. Advanced settings. We don't need any of this. Don't need any of that. You can, if you want to. You can turn on symmetry. If you turn on vertical symmetry, it will make your, it will make your final image have a certain level of vertical symmetry or horizontal symmetry. I don't tend to do it. I'm not going to bother button lock. This is all about playing around and having a go and seeing what works and what doesn't. Then we need to come down to prompts. Now, I'm going to put in something like so let's delete or less. You want to keep the colon's each side. You can go ahead and just use this if you want. But I'm going to go a war torn city in the future. I'm going to put by Simon style and hog, again, one of my favorite artists. So Simon style and hug. If you have a look at his stuff, it's just stunning. I love love his stuff. Okay, So let's come back. A war-torn city in the future by style and solid style and hog dusk. Let's just put dusk by here. And I'm also going to put trending on Park Station. I'm going to go to rendered in Unreal Engine. Now, there's a particular reason why I'm going into these specific prompts. Now that will come in the next lesson where I give you some tips and hints. And I'm also going to link you to some YouTube videos which can go through it in more detail because the more information you have the better. So I'll leave it at that. I'm just going to do that. So a war-torn city in the future. Let's put in the distant future at dusk by Simon, start by Simon Stone hug or in the distant future with nebulae in the sky by Simon style hog training on art station that may throw it off. I don't know what this is all about. Folks. This is a 100% experimentation. So what we wanna do is prompt to Dan and all that. So once you've done that prompts your little intermediate saved and your stuff up here. To be honest, you can minimize settings. And the next thing we wanna do is diffuse. Now, just have a quick look in here. Display rate. Remember I said you've got 450 iterations and we're going to save out an image every 25. You can also show how often it's going to display it. So I'm gonna put a display as, as it's doing it every 25 as well. I normally keep those numbers, the same number of batches. Click at 50. Now, because I'm on a free version, it'll never normally see the 15th through. It will normally timeout by the time it does that. So I'll always keep it to 50 and I'll get as many as I can if we just have a quick look. Normally I tend to get around. So 12345678910123456712345612312345123456. So January around six seems to be the average. Again, it depends on time of day and how busy it is. You can get more by updating to I'm going to cancel that by updating to Google Colab Pro. I haven't done that because I don't need that. If you do update it, you just go faster and you get unlimited resources and stuff like that. But I'm not in a position where I don't need it. And so what I'm gonna do now, that's all set to 2550. And don't need the video of stuff for this. So what I'm gonna do is scroll back up to the top. I'm going to click one time. And then I'm just going to click Run all. Okay? And what it's gonna do is come up with a few things. First of all, is can say this was not authorized by Google. It's been loaded from GitHub. Do trust it and all that stuff. And I'm just like, yeah, run it. Then it's going to ask you the notebooks, awesome. Hi Rambo of largest click. Okay. And then it's going to wait probably about 1020 seconds. And it's going to ask you to connect your Google Drive. I think it'll do it with mine. It may have or may already be linked up, but I'm pretty sure that it will ask me there we go, connect to Google Drive and I'll just click, Yeah, cool. And allow. And then what's going to happen now is you're going to leave it. You're going to see each one of these. I'm not going to leave it recording. It'll take a while. You'll see each one of these will kind of go through where it's got the dotted line circling it, it's waiting. Now it's executing all these commands. Just tick that one is now preparing folders. So it's saving folders, making folders in your Google Drive. And I wonder now if I go back to My Drive, ai disk diffusion images out, hasn't created those ones yet. Asked by. Now, doing installing an important stuff. And basically it's going to work through all of them. Now what you'll wanna do is when it gets, it'll go through 123 in about four or five minutes. And then what you'll wanna do is come into diffuse. And as it starts executing this, your image will pop up just by here as it starts going through. So what I'll do is I'll come back to the recording in a minute once mind starts going through all the iterations. Okay, So if we have a look at mine, number one's done too is done. Custom models done, three is done for its active. Okay, so if I scroll down, I have to wait. It was active and I had to wait about a minute or two. And if I scroll down, I can now see, boom. So this is the first, it's sort of 2% of the way through and it's saying it's going to take about 38 minutes, 42 minutes to do one. Okay? So you'll see batches 0 per cent, 0 to 50. So it's gonna be doing 50 images and it won't get there. But it's saying that it's prep to do 50 images, which is what we set earlier. And it's saying that it's done iteration 15 out of four hundred and forty four fifty, I think so. I don't know why he's kinda gone with that google Colab notebook. It can be buggy. But what you'll notice is when we get to 25, I've asked her to save out an image every 25 iterations and also update every 25 images. Remember we did that earlier. We'll likely see this when it gets to 25 out of 440, will likely see this update. So let's just have a look and see if it does that. Hopefully should update. And if it doesn't, well, it's not the end of the world. I don't know why. Again, it can be a bit buggy and maybe it's just gonna do it in a second now, or maybe they are. There we go. Perfect to it. Waited a couple of iterations, but it has updated, so that's good. I'm really glad that worked. You've got to do now is you've gotta leave it. Now. Time is gonna take, will go up, it will go down. It depends on how many people are using the GPUs and the servers. Now you can go and update a goat Colab Pro or Colab ProPlus. For me, it's not worth it. I don't mind waiting. I can leave this one as I do other things. Okay. So what I'll do is I'm gonna leave this go. It's ten to ten. And now I'll probably recall this tomorrow morning because normally I get about four hours worth of GPU time. So what I'll do is I'll revisit all these images tomorrow morning with you. Then I'll also record a lesson on prompts. And then I'll set you a final task, which I'm really excited about. So cool. Hi guys. So I'll just show you the results now from disk diffusion. So you'll see I've got my AI folder. If I double-click and go to disk diffusion images out, you'll notice I've now got that additional folder Skillshare lesson. So if I double-click in there, you'll see that I only let it do fall are snakes ago to about 12 o'clock and I want you to go to bed and didn't really have a computer on. So first of all, we had this run. So let's have a look at this. So this is quite a nice one. I like that as a little bit of weirdness or over there. But for the most part that's really, really nice and a lot of detail in there, I love that. Then we've got this one. Again. It's a bit weird that you've got that on the floor and then this up there, but you can see what it was trying to do and still looks, regardless of it being quite surreal, it's still looked very nice. You've got this one, which is I think one of my favorites. Again, lot of that plumped really coming into work there. And then we've got this one as well, which is similar to the first one. I may have actually had a fifth. If I go into partials, I may have not let it complete. So we come all the way down. You can see we've got all these partials and this was the last one. I really I didn't let it go all the way, but I really like that one. That one's probably actually my favorite. So you can, if you were to know Simon style and the parts work, you can really see some of his, some of his kind of influence coming across into this. So yeah, those are my results. So like I said in the next lesson now we'll go on to look at some prompts and some tips on that. Yeah, I look forward to seeing you in the next lesson. Cheers. 4. Lesson Three – Prompt Tips: Okay, So I'm going to go through a little bit of a lesson on prompts with you right now. Okay, so first of all, what I'm going for something like sci-fi and fantasy and concept art or anything really which is kind of coming up in here along the top, are generally like to put in a trending on art station because art station is the home of concept art and just look at this stuff. So if these are the things that are currently trending, then it's going to be looking at this and trying to replicate it in some of these styles. So I really like this. Another one which I used to use is 500 px. Now, that's generally a bit of a photography one these days. But what you could do is if you knew a website. So Deviant Art could be another one. I don't use Deviant Art anymore. But if you, if you use or you could look at topics in here to help generate ideas. You could put in anime and manga, you could put in and throw or artisan craft or comics, or Fan Art or fantasy, fanfiction fractal. You could use this as a prompt generator to be honest. Pixel art, I never thought of trying that. Street photography, science fiction, traditional art. That actually, I think I'm just stumbled across quite a good one. Quite coincidentally, putting in 3D is quite nice. So using trend in on art station or trending on Deviant Art, or finding a style in here which you like. Really, really good idea. Also, I've done another course on something called Night Cafe. Okay, so go and check that one out, because this is quite a nice one. But also if you just go ahead and come to the website, go to Google type of the income year. You can obviously go through and click where I've just clicked and typing prompt. But you've got all of these, all of these styles. So synth wave, cyber punk, heavenly, surreal horror. So you can use them as keywords. Let's go and type into Google just randomly, art styles. All of a sudden, I've gotten that most famous art movements and styles come and look through here. Art Nouveau, avant-garde, constructivism, conceptual art, catechism, data, digital arts, Futurism, neon arches have a look at that. I mean, yeah. You may want something like that. Pop out. You may want something that kinda looks like that post-impressionism. Use these. Again, if you don't know what Rococo is, which I don't even know if I've just said that right. But I may use this as just see what it kinda comes up with. Surrealism. I like, again, if you've interior severe autism, just go Surrealism. Artists. Salvador Dali, obviously it comes up and you could start, but obviously I put in for my one which is still going but here, look still going quite nice. I put in by Simon style and hard because he's an artist that I like. So you can go and find an artist that you like. What I found because I like the 3D things is if I put renders in Unreal Engine, if I can actually spell it. Click this. I know that this gives you a sort of photo-realistic or can give you a photo realistic style. I'll often chuck that in there with my problems. Now I did put it in here. So we'll see how photorealistic ends up becoming. But that's another thing that I like to add. And then there's two other things which I was going to share with you. As well as using sort of like websites, finding art styles, finding artists add in Unreal Engine or photo-realistic. Or if you wanted something that looks like a photo, type in something like shot on DSLR. Why not? The other thing is short prompts versus complicated, or short prompts versus non short prompts, I guess. Let's just make a new Word document for second. You could go as abstract as the wind by Van Gough. The wind that is very abstract. Or you could go a countryside field of sunflowers with clouds in the sky. Right? Oh, let's go more. A family in a countryside field of yellow. Because you can have different colored sunflowers. I think maybe not. Yellow sunflowers with dark no, with white fluffy clouds in the sky. And some green trees in the distance. Then in the style of golf. So you can see, I think it's always worth trying both. This one is weird, short prompts. You can get some really funky like, Oh my gosh, I wasn't ever expected in that kind of stuff. Sometimes they can just be like a little bit like what am I looking at, but you can really surprise you sometimes, sometimes being too descriptive and giving it too much, it may just get like really, really busy. So for example, I've done this before where it would be a countryside and it's a field. But the sunflowers Look all puffy like the clouds and stuff like that. So sometimes it can be a bit too much. So again, it's about trying both but try, try both different things. Another thing, one final tip that I do, like using is I'll often add a time of day. I know as a photographer that shooting at the golden hour. Let me just show you what a golden hours, I'm sure most of you know, but I'm shooting at the golden hour. It gives you some stunning. It's when a sudden setting or just rising where the sun is so lower, interact with a lot of the atmospherics and creates a really golden sort of vibe like that. A lot of Netflix stuff shot around about the golden hour because it just makes it look more cinematic. So I'll often put either golden hour, golden hour, or I'll put at dawn or dusk. Dusk in the one which has generated now. So yeah, that's that's that's what I'll leave you with. So let's bullet point that artists search for artists twitchy like styles, websites like art station, trending on Deviant Art, stuff like that. Unreal Engine. I like to use shots on DSLR if I wanted to look forward to a real or just add the word photo real. And yeah, short versus short versus long prompts. Try and see what you get. And so we're going to leave it there. That is my short little advice on prompts. And I'm uploading quite a few of these AI generated tutorials. Each will have a less non prompt and each will probably be slightly different and more updated because this is growing fast. So yeah, cool. I hope you found that useful and I'll see you in the final lesson next, where I will set you your little project, which again, I'm super excited to see. 5. Project Task (Assignment): Hi guys, welcome to the class assignment for this course. So all I want you to do for this class assignment is I want you to create two pieces of AIR funnily enough using disk diffusion, which we've just learned. Now. Two pieces though I want you to create. The first one, I want it to be an environment. Never can be any environment you want to go wild, be creative. The second piece, I want it to be either a character or object, completely up to you. But what I want you to do once you've created these pieces of AIR, I want it load them into the class project, but I want you to upload them with the prompt that you used back way. We can all see what prompts get different results. And it can become a little bit of a gallery, a little bit of a source of inspiration. So battle and ask him to pieces of artwork more if you like. But I want to see at least one environment and one character or object. So go while to show me what you're made of. And the only thing that's left to say is thank you very much for taking this course. I really hope you guys enjoyed. So I hope to see you again soon. Cheers.