Stable Diffusion-Master AI Art: Installation, Prompts, txt2img-img2img, out/inpaint &Resize Tutorial

foreign [Music] here welcome back to an all new tutorial this time featuring The Amazing stable diffusion which is both free to download and open source before we start I'd like to just give a quick tutorial overview for those of you new to the channel I'm a senior 3D artist working in the video game industry since 15 years and in my tutorials I teach people how to create 3D art for video games so naturally I'm also looking at stable diffusion through a video game art lens if this is the first time you have heard about it I'd argue it's one of the most revolutionary developments that we've seen in recent years and we are now entering a new era of AI art in which we have only seen the tip of the iceberg so far for the future think AI videos music 3D models and animations for now stable diffusion enables us to input any tax prompts that we can think of and create an infinite amount of pictures references or even textures for our video games in photorealistic or stylized ways based on what we want it to be stable diffusion is a literal Game Changer especially for smaller Indie Studios that don't have the resources to employ dozens of artists with that said and also addressing a bit of ethical controversy the way I look at it it's not meant to ever be a replacement for artists but it is a mighty tool that elevates artists to even higher levels while also enabling people that have no art background to open this beautiful world up to them we will go over the installation which takes no more than five minutes and then we jump straight into the action and learn about the tools different model files how to get great results both based on text and image input as well as where to get the latest news and updates we also take a quick look at potential use cases for video games such as creating procedural materials as well as updating stable diffusion to the latest version okay enough talking let's get started happy prompting and I hope you enjoyed the tutorial cheers Tim hey everybody Welcome to the tutorial make sure to expand the YouTube description for all the relevant links we're going to be making use of here and the first one would be GitHub so we're gonna download the latest version of git and let's just download everything before we install it let's go to the next Tab and let's download the latest version of python and then we can go straight to the stable diffusion related things gfpgn is a face restoration model we're gonna make use of let's download the 1.4 model here so next thing let's go over here to the compass page on hacking face this requires you to register an account on hugging face to actually download and on here for now we're gonna be downloading the 1.4 model for stable diffusion maybe there's gonna be the 1.5 by the time you see that but for now let's stick to the 1.4 and the same thing applies here to the waifu model which we're going to be downloading next the 1.4 version here in that case is the one that is not available yet that's why the one we want is the 1.3 so I'm gonna be cheating a little bit here throughout the tutorial and I'm gonna fast forward the downloading times and that means that whenever that is done on your end you can then follow here by going to our download folder and let's just also start with GitHub then so let's open that up and then we're just gonna click next on pretty much everything that follows here there's a whole bunch of things that are already recommended here and if you just go on next that should make sure that everything works as intended so after that let's install python and here when the installer asks us we want to make sure to check add python to path I was reading that is the recommended setting here for stable diffusion so let's just make sure to have it so next we're going to be opening the URL here for the automatic stable diffusion web UI this is a fork that has all the features that we could possibly wish for and that makes it the perfect web UI for us let's go over here to see and let's create a folder called stable diffusion let's go in it and now let's right click git bash here and we're gonna type in gits space clone and now we're gonna be pasting that same URL that we have in the browser automatic one one one slash stably Fusion web UI and press enter and when this is done let's then open up that new folder that is now in here the diffusion web UI and over here let's just expand it and then right click anywhere in the free space and let's type in git pool so since we already downloaded it just now it says already up to date but otherwise this is how you can check if there are any new updates good pool will do that for us so now let's just wait for everything here to be finished there it is now we have all the relevant files that we need for the installation let's open up here our download folder and on the left side I'm gonna have the stable diffusion web UI folder that we just created so let's take these checkpoint files here I'm going to control X and then open up the models folder here so then let's paste those in there that means that now we have our first two model files there are a whole bunch that you can try out later but these are the ones I'm using at the moment let's then rename the 1.4 model here and let's rename it into model checkpoint that makes sure that when we load stable diffusion the first time it will load everything properly so then let's control X on our gfpgn file that we downloaded and let's paste it into our main web UI folder so just to make it clear this folder here is the one where the gfp Gen file goes and then if you want you can simply delete those two XC files that we downloaded for GitHub and python let's expand the main web UI and before we launch the webuser.bat let's edit it and in here we can edit this line for a little bit faster loading time with not much notable difference in quality so I did some tests and I think this is a good setting to put and then we save that and let's launch the web UI user.bat so that will take quite a bit don't be surprised if you think that it's stuck and you have to restart just wait I was of course cheating here once again because I'm not gonna make you sit through the 15 minutes here So eventually it will look like this here and that means it's finished so I'm just gonna close the command line for a moment right click the web UI user back and let's make a shortcut on our desktop in order to launch it from there let's double click it and run again and you will see how this time it goes really fast we're gonna copy that local URL and paste it into our browser of choice you will see that it says loading here if nothing displays here let's open that menu and select it manually but I think that's only the first time that it did that for me usually it should look like this so let's test if everything works let's generate something and then let's check the command line you will see that here it shows us the progress so the command line always needs to be opened in the background or else stable diffusion won't work so hopefully everything works fine on your end as well the installation is now finished and it is time to take a look at the text to image prompt so before we start with the prompts it's also important to understand how these prompts are actually working and what they are based on so in short these model files that we downloaded they were trained by millions of images and they are all text attributes this website here gives a small insight into that so for me it was of course interesting to see if any of my things pop up here and as you can see I have this helicopter in there so my helicopter is actually a part of this scrape and then let's go over here to my art station website and let's take a look at this helicopter so as we know these scrapes are based on alternative text and in my case here that helicopter that I made for a Splinter Cell Blacklist I named it accordingly here Scout helicopter and that's why it shows up here so the better job you did giving your pictures a name the more likely it is you're part of that big scrape that is now stable diffusion let's go on lexicon.art which is a great way to get inspired such as for example I like this one here and that's why I go on explore this style which will then reload lexicon and give us pictures in the same kind of a color range and in the same theme and that way we can then also see what prompts were used to generate those images so this is a great way to get started and to also see which prompts give good results for example you see the Cinematic Masterpiece term appears a lot and I like that one so I'm gonna copy the prompt and then simply paste it here into our own prompt field and we can then adjust it further so I want to generate a secret cat Civilization now a futuristic landscape of a secret cat civilization I want to crank up the width a bit 768 pixels and let's put the batch count to 10 so that we generate 10 pictures here at the same time I'm fast forwarding once again through the generating part of the images or else this would bloat the video up way too much so here we have definitely a recognizable cat planet quite hilarious results and nice colors and as you can see this is how we got some nice results really fast simply by adjusting the prompts to our liking so if you wonder where these images are saved you can find them here in the web UI folder under outputs and then categorized whether we use text to image or image to image or also extra settings for now let's stick to text to image and let's adjust that here a little bit further this time I want to have a bit more of a darker style that's why I add HR Giger to it and before we generate that let's take a look here at this image which nicely shows these different generators and also what the sampling steps are doing so in some cases like the K Euler there is not really big of a difference whether we have 16 or 64 steps so we can definitely try out different sampling method but in general the Euler a the default one is the one that I'm also using most of the time so I think that the bombastic color prompt that I used for this doesn't really go good together here with that HR gigger Style so I'm gonna put the sampling method also back to Euler a and I tell it in the negative prompt field not to use any blue tones so let's see what we get now and by taking the blue tones out we get very interesting result here mostly in the Reds now and that also fits the HR Giger themed style way better HR gigger is usually using the same kind of color palette and that also shows here very well by using this as a prompt if you ever want to go back to a particular image that you generated you can simply go here to history if it looks like that we can renew the page and it will then load so that way you can always find older images if you like one particular style you can go back to it see the prompts and reuse it so now let's pretend we're working on a video game we want to have a particular style of a futuristic city in a mountain but we're kind of lacking the idea how that could look like so we can go back to Lexica or we can also use this page here called Korea but I prefer Lexica personally I'm gonna search for City and mountain and see if we already get some nice interesting results so this one here for example looks quite nice you can see that usually these pictures have something like painting by and then there's a name of an artist same as we just did with HR Giger here is a website that shows a lot of artists and pretty much what you can do is browse down on it and if you like a particular style it even has some video game companies in there if you like a particular style then you can say paint it by and give it the name of the artist and it will then try to replicate that style so in that case here I'm gonna copy this prompt out and what I like to do is have a read me always in the background where I copy and paste prompts that I find on lexicon and then I can mix it all together here quickly in that readme where I just have a bit more space than in the prompt field you see I already have some prompts there so now I'm going to piece that together here for example the high towers and the Cinematic afternoon light I want that to be a part of this prompt and we have a whole bunch of terms here like glorious deviantART 4K UHD image Unreal Engine and these ones I also want to add to it and I like to usually put those at the end I'm not 100 sure whether it makes a big difference if something's written at the very beginning or at the end but I think so let's paste this prompt that we have here I'm gonna put the sampling rate up to 30. I notice there's not really much of a difference in loading time for me I'm gonna put the width to 768 so that it's not a square image that we're generating and let's see our city in the mountains and that just shows how powerful these prompts are a tower built into a mountain on the side of a cliff and it just shows same as the high towers that we also have in there some of these buildings are quite High and once again it just shows how extremely powerful these prompts are two little words can make all the difference here in our composition and that's also what makes it so fun to experiment around so let's assume we have one picture that we really like and now we want to upsize it because right now if we look at it here in our outputs just to make it clear we have a 5 12 height and a 768 width and if we zoom in on that you can just tell it's not very high risk it actually has surprisingly many details for such a small resolution but that is of course not the final resolution that we would want to have here in case we want to further work on that image so for us to upsize this image we're gonna have to select it and send it to extras here under our text to image input we see our current resolution and here under resize if we put that to resize by four it gets multiplied by four so the original resolution plays an important role for our final resolution and I'm gonna try here with the upscaler put to ldsr and the second upscaler we can kind of blend it with a different upscaler method but I'm gonna put it to none so I'm gonna hit that generate button and you can see here in the background it takes a very long time which once again I'm fast forwarding through so just wait a few minutes till it's finished and then it will also show up here so let's go over here and compare that you can see here the resolution is now four times the one that we had before and if we go here to full screen and now do a zoom in on it we have all these nice details and it's not just a simple rescale it's an actual upscale which means that details were added while they are also being preserved on the original image later we're also gonna have a look at making different variations let's say this one here we want to have different but similar Styles we're gonna look at that later for now let's go back here to our prompts and I want to show you this really cool prompt Builder tool which is a website and basically it lets us generate prompts without having to use Lexica or this other website let's start by going over here and let's search for a city this one here is looking nice I like it with the river and the high Towers and essentially we're gonna just reuse the prompt that we had before and let's see what this makes out of it so the first one here is the one that we already had and also here on the second one let's add these high Towers back to it but now it becomes interesting so this one here is simply the style in which case we want to have a landscape and here under select more details we can really go crazy so we open up these fields and simply click on the ones that we want and it will then end up on our prompt that we have there at the top we can also tell it by entering a weight how strong this particular style should be applied and because this is too time consuming I'm just gonna fast forward through it just select any style that you like and click on it and keep scrolling down until you're pretty much at the end of it so one thing I noticed is that the picture link that we put in the first field somehow disappeared I think when I put the weighting and pressed enter maybe that's a buck so I'm gonna add that back in here and let's just copy the whole prompt now into our readme so let's just copy the part here without the image URL for now and then let's see what that generates for us after that we're gonna paste it with the URL it's actually quite interesting and without that website I wouldn't have known that you can also paste image URLs into the text to image field so this one here is without and I'm quite liking it and I think it shows that this prompt Mania tool can be very useful in addition to using Lexica but here we can actually be more precise and really generate something based on what we want it to be now let's try the same prompt but let's add that image URL at the beginning and this is actually very interesting to see if you remember that was a picture of Chicago with that River and this gets applied now here in our text to image generation before I thought that we have to purely rely on image to image but we can also add the image URL here to The Prompt field in the text to image and you can definitely see that there's a little bit of this Chicago in there so the goal is to find that one image that we really like and that can be achieved by experimenting around here and then maybe who knows we see this image here and it looks so cool that we could imagine basing our whole video game on it or on that particular style so now that we looked into some landscape shots let's get over here to the character creation I'm gonna make use here of a Thai actress and I want her to be a queen on a tarot card so this is a prompt that I found on Lexica and I quite like it and maybe it's a little bit Overkill but I'm gonna paste a whole bunch of terms here into the negative prompt Fields so hopefully we don't have bad hands or any text or missing fingers but I noticed it's not really always working but you know it's worth a try so putting this here to restore faces and the high-risk fix should already try to make the best out of it here when it comes to characters especially for the faces but we can always try to fix it further with the gfp Gen modifier that we downloaded so here we have the tarot cards which came out really nice I don't see any problems here with the faces either and even the hands look good but sometimes it's worth counting very often the hands are having like six fingers instead of five or like here it looks kind of weird so apparently the hands is something that is not only for artists but also for AIS apparently a bit of an issue or maybe that's exactly why it is an issue for the AI because it learns from actual humans who knows so now let's try here with uh yossi from blackpink in case you don't know her she's a famous K-pop singer and let's put her in here and see what that gives us I want to put the height up a bit further I think to roll cards are actually having a bit more height here than what we had before so let's take a look at this and a complete different look here probably because she has pink hair in some of her performances I could imagine that also reflects here then on outer row cards very nice result also great looking faces but let's just assume that there is some issue with the face let's take a look at how we could further fix that with the gfp Gen we sent then that particular image here to the extras and over here once it's showing we can scroll down and we then enable the visibility of gfpgn I hope I pronounced that correctly we Crank that all the way up and here is an image where I already saved it out it's a different one but you can definitely tell here the difference it makes it actually makes it a bit more detailed and high-risk and it eliminates out these weird eyes that we had before so this is working very well and now it's time to take a look at a different model file that we downloaded earlier that was the waifu model so here on the upper left we can then simply switch it over to that model file and then once it's done loading we can go ahead and use it take a look here at this website which is absolutely great for keeping up to date here with the latest things that develop so this one I would recommend bookmarking and checking if there are any new developments and those can happen really fast here at the moment I'm going to put the sampling steps to 30 as you can see this time I wanted to create a female winter goddess I already put the terms in there and now we have these to roll cards based on this anime picture trained model file that we downloaded so it might be interesting to create characters with this pack but it's actually also working quite nicely for Landscapes which we're gonna do in just a moment let's just generate something similar here but instead of ice let's try with fire and let's see what that gives us I quite like to add the term fire sometimes to the prompts it can really make some impressive pictures simply by adding that and I think those came out really nice as well here I don't see any issues here with the faces nice Style and here we even have two fire goddesses very nice so now it's time to look at the image to text tool we're gonna remain in the waifu model file for now and we want to try generating some images here based on a kid's drawing so let's load something in here and as you can see I already have the prompts up here I pretty much describe everything that I see in that picture and I add some extra terms to it as we did before I'm gonna adjust the width and the height so that the canvas is fully captured here by it and now when you compare it to the other model pack that we had before you can definitely tell that there's a bit more of this comical style which is very nice that's why it's also Worth to experiment and switch around between different model packs so here for the kids drawing I think this is perfect with the waifu model file and now let's once again take one picture here that we like in particular and what if we want to expand it to either left or right or really any side top or bottom in that case we send it to impaint and now let's say we want to add more to the left side here we can then scroll down here to where it says script and in here we have two different out painting plugins the poor mans and the out painting MK2 so let's choose the other one here the MK2 and if you scroll down you will see then that we can expand and simply choose into which direction I'm gonna just add something here to the left side we have a few more settings here such as the color variation I'm going to crank that up I want that something interesting is there on the left side and then let's expand it to 256 pixels here to the left side and also I'm gonna increase the blur otherwise it looks to cut off if we have only a few pixels of blur size you can always hover with the mouse over these things like the denoising strength for example that I just put to 0.8 and it will give you a quick explanation of what it does but very often it comes down to just you know adjusting the sliders a bit sometimes even extreme into One Direction just to see what happens and that way you get a good feeling here for it so here we now see that on the left side we have these 256 pixels added I'm not really liking it here with those pink flowers or Violet one thing to keep in mind when we do this in painting is that the terms that we have above that we already used for the main image they will be cramped into the new space that we add so actually it might be a good idea to get rid of some of these terms here or some of these prompts but you in that case I'm actually quite liking this the only issue is that we have some text now here on the upper left side so what we're going to do is send this picture that we just created to impaint once again and now we can paint over it here so before we were doing the out paint by expanding pixels to it and now we have more control here with our brushes well actually it's just one brush and we can change the size but we can then tell it here what we want that masked content to be like and also we want to have a bit of a mass blur once again so that we don't have too much of a cut off border let's remember to put the script To None because before we still had that script active so in order not to expand and just do the in pain we have to tell it no script so now we generated nine different versions which actually is a bit Overkill just for this small area but then as you can see no more text and now we can pretty much choose the version that we like the most here where it was then painting over the Mast area so this one here let's send it to impaint again because let's take another look here at those tools for example if we're not happy here with these details here these flowers even though they actually look quite nice we can then paint over it so let's say we want something new generated here we simply mask it out and then we tell it that the impainted mass region should be regenerated of course this is going to be based on the prompts that we have at the top so depending on what you want that to be for example if you wanted to be a hiking trail we should write something like simply hiking trail there at the top for now it will just regenerate based on the same prompts that we had for the main image but still you can see it worked exactly there where we masked it and that means that if you're not happy with one particular region of your image you can simply paint over it and regenerate so as the next thing once again let's assume we're working on a video game and maybe it's supposed to be something cyberpunk related so as you can see I keep bash something together here in Photoshop and it's not really a masterpiece but that's exactly the point if we have a vague idea of what our game is supposed to look like we can put it together like this here in Photoshop save it out and then we will let stable diffusion run its magic over it of course once again based on our prompts here so I'm gonna simply reload the interface here the web UI and also I want it to be the original model pack here not the waifu this time even though it's also worth trying both so I'm gonna load the concept into the image to image tab and then I already prepared some prompts here which I'm gonna paste now and one thing that I wanted to emphasize here is friendly colors and coziness I didn't want it to be this cold style that we very often see with cyberpunk games I wanted to look like something where you could actually imagine drinking a beer or just walking through it and also one of the prompts is bars and shops so I wanted to look like a place where people like together so let's see what that gives us and it's really nice I love it so you can see that it actually looks a little bit like anime and that is probably also because of the artist that I added there and that way we can in record time create a whole bunch of art for our video game some mood boards and just explore different styles and maybe even base our whole game on it if we like something in particular so if we like a bit more the stylized style then maybe this is the right way forward here for our game and if not we can simply change the prompts to something more realistic and gritty till we like it so I'm gonna send that one here to the extras once more and let's just do the upscale so I'm gonna put that all the way up here to four and then for the upscaler we can mix it the upscaler one or the two but for this one I'm just gonna use one here so gfp jam visibility in case there are any humans on the picture so once we have that we will find the generate button here at the bottom again let's generate that and that might actually take a while because I resized it here to times four which is the maximum based on our original resolution so if you have the original resolution higher then you will also get obviously an higher output image here if you resize it by times four in that case here we have now a resolution of I think 4K and this is really nice and high-risk and basically it's just a stunning piece of art generated in a few seconds or maybe minutes based on your computer and I just thought it would be fun to Overlay it here with the original one that I quickly put together so this is where I see a very very strong use case here for stable diffusion for smaller Indie Studios to make these mood boards and pretty much generate art to show something for the direction you want to go So speaking of video games let's take a look how we can create tileable textures and for that I was just installing the demo version here of substance 3D sampler and let me show you some materials that I already created here and then we're gonna create one ourselves to see the process real fast so here we have some maybe gold nugget material here we have some ornament some mystical Cube almost looks a bit like the Hellraiser Cube and in substance sampler we can then have full control over it back in stable diffusion you can here see some terms that I was experimenting with and that is also a good opportunity to show you how we can make more variations so I like that one here in particular I'm gonna send it over to image to image and I want to copy the seed that we were using so here in our settings let's select tiling most importantly and then we want to paste our seat if we leave it at -1 it will just generate pictures based on the prompts if we put the seed it will be more closely to the picture that we have on top the same goes kind of for the denoising strength if we bring that down it will be more close to the image we have on top so let's imagine you're working on a video game and you want to texture this Viking room of a king or something like that you could pretty much generate a whole set here of these Textures in that same style bring it over to substance sampler and then make PBR textures out of it and we're gonna take a quick look at how we do that before we go back to stable diffusion won't take long but the process is that we scroll down here in our output folder we find that texture that we want to make a PBR texture out of and then here where it says image to material you can see it also actually does something here with AI so like I said earlier I think we really are in a new era here of AI and this is just the beginning so now we pretty much have that texture already finished it generated a roughness map a normal map a height map and an Albedo map so in case you have some issues there with the tiling like we have it here if we zoom in we can use that generator called make it tile and that makes sure that all the borders are perfectly stitched together so if we bring the tiling down to one then here we see exactly the output that we got from stable diffusion and if we want we can of course play around a bit here with the settings how much height we want to have even though the height map we might not even need for our engine later on but substance sampler enables us to create PBR materials here based on only that one image that we gave it as an input so here we have a whole bunch of settings where we can then further fine-tune the textures if we want to have maybe more roughness information or less we can change our AO strength and we can also change the set duration here of our Albedo and speaking of color information we can also completely change the U tones so all the modifiers that you otherwise are used to from substance painter or maybe even Photoshop we have a whole bunch of them in here too so let's say we want to have a complete different color here on it we can then make use here of the color variation modifier and let's say we want that to be a white or maybe a blue it's as simple as going into these sliders here and picking the desired color so that is pretty much it here for the substance sampler quite impressive in combination with stable diffusion and since this is all still so new I think there will be much more coming in the future to make it even more useful and speaking of useful tools I'm gonna load in a PNG here of that shirt and I want to have a stylized look of it so that already comes with some Alpha information which unfortunately at the moment is not supported by stable diffusion we don't have this transparency layer yet the models were trained on RGB only so nevertheless we are able to generate something like that here and let's say that maybe you want to have this stylized nice looking shirt here in your game or maybe for the inventory you can simply delete the white background out and that's it so another useful tool here maybe to use stable diffusion for and also something that I could imagine in the future we get some nice plugins for transparency and pngs and speaking of the future while I was recording this tutorial stable diffusion 1.5 model dropped I want to close the browser and update that web UI that we're using and as you can see 63 files changed since I've been recording that tutorial I haven't updated it since now so let's download stable diffusion 1.5 model pack and also the in painting 1.5 so giving this page here a bookmark the rentry.org is definitely a good idea here for the stable diffusion model category and whenever we have that we can then simply load it in as a new model file I actually read some controversy about this whether that actually is stable diffusion 1.5 or if that is not it but you know it doesn't hurt to try it and personally I found that this in painting one here the 1.5 gives some very stunning results and it also I think is now faster than before and speaking of that in painting model you have to agree to the terms before you download it and then we can simply save it into the same folder where we also downloaded the 1.5 model which I already did so now you just have to wait for that to finish and once this is done we can then just same as before do a Ctrl and x and then paste it back into the model folder that we have here in the web UI same place as before we're just gonna paste it in here so updating stable diffusion consists out of git pull and downloading new models whenever they become available and now what I want to do is actually close the browser so that we make sure stable diffusion is closed in the background and now let's see if everything works fine here so we're not getting any error message that's already a good sign let's launch the browser I'm gonna close my tabs here and here we have the 1.5 model we can now select it but for some reason that I can't explain I'm not seeing the in-paint 1.5 model and I think maybe it is because when I loaded up the browser maybe it showed a cached version of what was there before so I'm gonna close it and simply launch the browser again and see if now we have the in pane showing and here we have it so I'm really not sure why that was the case but in some cases maybe it's as simple as restarting as it is so often so now that we are fully up to date here with the latest version and the model files let's have some fun this year was actually one of the first things that I saw when I learned about stable diffusion and it absolutely amazed me in case you don't see what that is let's bring the opacity down and now it should be pretty obvious everybody's favorite hate symbol and we are gonna also have some fun with it now my idea was to generate a Pepe based on guacamole and I spent way too much time on that as you can see I tried with different ones and it's actually not as easy as I thought so it's really a matter of playing around here with a different CFG and the denoising strings which by the way also serves as a great lesson in how stable diffusion works so the goal is that we generate an image that people that know the meme can definitely associate with Pepe that has the cigar here for example but for other people it's just an image of a tasty bowl of guac but of course we have to find the right Middle Ground so that it's not too distant from the actual meme we want it to be visible but just the right amount of visible and for that we have to find the magical combination out of the CFG scale and the denoising strength not everything is going to be meme worthy but it's definitely a fun way to explore these settings that we have here in the search of this one picture that we like and I'm still not really happy with it because this here is too obvious it looks like oh look someone made a face we want to have this Randomness in there it's not supposed to look like someone was trying to build a snowman with guacamole but it is supposed to look like a random thing that fell into place and happens to look like that so the more we bring the denoising strength down the more it will look like the image that we have as an input so the takeaway here is that if you want to create your own memes something like this here it's a great way to get familiar with the CFG and the denoising strength sliders and see how they work in combination so here you see my work in progress for the cover image for the video I'm not decided on it yet but the interesting thing is that I took that chili and turned it into a beautiful woman simply by using this chili as a input image and some prompts here both in the negative field and in the main fields so once again I found these prompts on some other page and in case you're curious why there are these brackets I actually have to read up on that a little bit more whether this makes a big difference but I think it's the same as we did before with the weights on that website that we use the prompt Mania if we put these brackets it simply means the more brackets we have the more strong the input is supposed to be so some of those came out quite nice others have some obvious issues here with the face and also some other parts but in general I think we could try to take one that we like and make some variations out of it so this one here I'm gonna send it over to image to image now we have it in here and also I want to copy that seed so that way we make sure it's as close to the original image as possible so let's see what that generates for us so this is just an important thing to repeat also here to know that we can always fall back to making more variations if we like one particular image so in between making the variations and the in paint and the out paint tool we have full control over what we want our image to look like those came out quite nice actually also the face looks good and now let's turn her into a cyborg girl with cute cybernetic ears in the middle of a forest to be exact so I'm gonna copy that whole line here including the red armor suit and override it with the dress so now she's gonna wear armor and I also want to add some extra information that I didn't have before such as trending on Art station which is a very popular prompt I see it all the time on lexicon and that makes sense if you're familiar with art station and the trending sections that's where the best results are always at and that's probably also a part of this scrape here so let's generate her as a cyborg and I mean you just gotta laugh how changing a few prompts already results into a complete different look here from realistic to science fiction same as before some of the images have some issues others are really nice so this sums up most of the things that for now I could think of that I wanted to show you and that I hope you also found interesting but I saved the best for the end what we're gonna do is we take a picture of ourselves or of our cat or our dog pretty much anything or maybe you want to impress your girlfriend or your boyfriend you gonna take a picture you load it into stable diffusion do a little bit of masking work and we get an absolute Masterpiece created around it so I'm just gonna show you what I already created here and then we're gonna repeat it from scratch I'm just gonna go over some of these settings I put the mask blur up quite a bit because once again if we paint a mask we have to consider that there is supposed to be some blur around it otherwise it would just be too sharp and here for a cat I want it to be quite blurry so that these hairs are nicely fading into each other from the AI generated content so I'm gonna open up a new instance of stable diffusion and just do it from scratch we load stable diffusion make sure to load the new in paint model that we downloaded earlier we go to image to image in paint and then bring in any input picture here whatever it is you want to portrait pictures are definitely working nice and this is also reflecting here on my prompts I'm telling it I want a portrait of a cat King so what I want to do here is not mask too much basically I want it to be the recognizable Parts here of the cat the eyes and the nose and since we have the mass blur we make sure that it's fading a bit I also want to take some of this information here on the left side because it has these nice highlights and then if we put the blur up it's gonna take more from the surrounding regions anyway I'm gonna put in paint at Full Resolution up and also this slider here I'm gonna crank it all the way up sampling steps let's put it to 40. let's put the batch count up and then let's generate and see what we get here so again I'm fast forwarding through this so now it becomes clear that I forgot to invert the mask but it's still interesting to see that it also did a good job here on the Mast region here of the ice and the nose but since we don't want to change the main look of muffin I'm gonna invert the mask and generate again so let's see what we get this time and here we have it a glorious cat King ruling over his peasants so these came out quite nice I'm also certainly liking the light that we have here that is also why I wanted to take the left side of the face into the mask because it already had some nice lights there and that also shows then here on these generated pictures let's simply try the same prompts here with this cat from my friends same as before I don't want to mask too much but the recognizable Parts here so let's see with this for a cat King I think it might actually be a princess but let's see and I like it a fierce look here it's missing a little bit the king part so maybe I should have added a crown here to the prompts but other than that this concludes the tutorial for this time I hope you enjoyed it thanks for watching and also if you want you can join us in the Champa Zone Discord server in here we are all about learning 3D art together we have professionals from the industry same as absolute beginners and anyone in between everybody's welcome and I literally just created a stable diffusion Channel everything around the subject of AI art is still relatively new especially stable diffusion so hopefully we can all learn together here by sharing the combined knowledge check out my other tutorials especially if you have an interest for video games these tutorials will teach you how to become a professional 3D artist so if you want to know how to create a revolver like this you can check out my latest revolver tutorial that will go over it from scratch make sure to check out the free tutorials that I have on YouTube you're on the channel right now most likely don't forget to subscribe and like I hope you enjoyed it and I'll be seeing you soon cheers Tim

Channel: ChamferZone

Views: 82,660

Keywords: Download, installation, stable diffusion, AI Art, prompts, text to image, txt2img, image to image, img2img, upsize, resize, github, git, inpaint, outpaint, gfpgan

Id: 1_dViOW3Vy4

Length: 58min 46sec (3526 seconds)

Published: Fri Oct 28 2022