ComfyUI SDXL Basic Setup Part 1

okay this is a bare Burns basic tutorial for how to set up your workflow for sdxl um I'm going to start basic and it's going to get more complicated as I add stuff so step one you want to add um your model loader so in order to do that you can double click on the background and it'll open up this search thing and you can just type load and then find it that way and you grab your model which would be this now I'm going to set it up with just the one checkpoint but the standard way of running this is to use refiner as well but it's a little bit more complicated so that's going to stay down here for now I might do it later um then we want to sort out our clip text all right so you can use the standard clip texting code I'd suggest to not use that one I suggest using the um instead to use the SDS sdxl specific one which is under Advanced conditioning and then sdxl and code now you could use just a double click thing and go SD Excel so if you see me doing any of these things you can just do this I have a habit of going in and right click finding things so don't remind me if I'm not faffing around with that a bit too much so clip text encode sdxl it's specifically for the base safe tensors model and it has it's got a whole bunch of options with it um there's some trickiness to do with what all of this does and most of it doesn't really affect what you're trying to do if you're not doing anything particularly Advanced you can just leave this stuff as it is and it should work fine but I'll go into this in a bit more detail later but you can use it to make your images look a little bit better otherwise it's not much help all right so clip G clip l so it's weird this has two prompts and the reason it does is it technically uses two different methods to the pass your prompt so it goes through your prompt and it goes okay what words mean what and what things can I find in my model that I can put in to this image so clip G in theory as I understand it it's not really been well explained by anyone as far as I can tell is for your primary source so it's like your subject your um your location um the say if you want it to be a photo you put photo up front that kind of stuff so it's like the the stuff that definitely is going to be the focus it's what's important and so it has to be in G and the less complicated stuff the background the um the quality like say you've got 70s photo or an old photo or film grain or something that stuff goes in clip l it doesn't need to be in clipped yet it'll work but it doesn't need to be so if you right click on the box anywhere that's not in these actual boxes you'll get different options there you can change these to a different input so this is something you'll find used a lot so if you see me do it I'm not going to explain every time but basically this is a very common way of sending information to your nodes differently than you would otherwise do it now the reason we break this out as text inputs is it gives us more flexibility we we will need to use the text that goes in these in more than one place later on like it will go into the refiner workflow so instead of copying stuff between different boxes it's easier just to use a shared kind of connection so uh standing we're doing this is drag it out and then you can if you have something in here that says text box or something it'll work these are custom nodes for me um more than likely you'll have to go into utilities and do primitive if you don't have any custom nodes you'll need to do this and it will give you this thing which says it outputs a string so the string is it's basically a line of text and that's all it is so it's the line of words and text and symbols and the cool thing about strings is if you connect them all together they're still a string so if you have something else with a string and you have a way of merging them with a custom node you can stick them onto each other so you can basically use multiple of these to build one prompt if you do it we'll eventually be doing but for now I'm just going to do one text box and it's going to plug into both because this works and it's the easy mode way it's like uh you know your standard kind of 1.5 setup is to do this um cat on a roof basically all right conditioning out re-route so I like to create what's called a bus which is a it's like a a highway for your nodes right and it allows you to easily access certain kinds of repeatable data that you'll always need so it's a good idea to have one of these it's basically a line where you have a whole bunch of these laid out as a kind of group thing that's the slow way of doing this you there are other ways of making this really quick but uh anyway so my way of doing it is to actually have this out here and then when I need nodes somewhere in here click on the middle Dot and go add node utility reboot and it'll stick one in the middle so you don't have to make a new one and connect both ends of it you know what I mean so this is the clean and easy way of doing it and it'll give you the option to put some between them if you want so anyway this bus basically gives us the conditioning for positive now so we're going to do a negative one a negative one uses the same text encoder the same kind of primitive thing comes into it and we'll say text um rocks yeah so there's no rocks in this image okay um and we plug in the clip and because we like to be neat and tidy shift select right click colors red and we change the colors here as well to Green you can do this on reroute nodes I'll show you a little trick here as well if you see this little connector here and you right click on it click rename slot you can say pause div and it'll put text in there if you right click on it and go title like you would with another thing it's not going to put any words in there but if you do it on the these little connectors it'll work so you can do that I it you'll be doing it a lot if you do it with these kind with the stuff in the bus that you don't really need to it with these but um you might do it if you build a little like you know like a little block which does upscaling or something you might want to label the outputs you know what each thing is all right so this is a typical bus so it's got positive and negative it's got your V output it's got your model and that's everything that goes into a sampler so we're going to add a sample in the middle here I really wish come for you I would actually have a better like that if you clicked on the middle thing it would just let you do a reroute it would be so much quicker all right so part one is done that's your basic layout for creating the loader loading model building a prompt and sending it through to a key sampler so you can use a normal K sampler which is the easy mode way you just plug everything into it positive to there negative to there it will need a laden image of some kind so you could use a valiant code and then load an image and that'll give you the ability to load an image as your source in which case this needs to be a lot lower you need to have a lower denoise but um if you don't want to do that and you just want to use a standard way that we do these things you pull this out and pull out an empty latent image which will give us the option of setting this it's set to 512 by default but you don't want that because 512 will not work properly with DXL so we do it with 1024 there's a list of acceptable resolutions for this floating around on Reddit um talk about that later if I do a more advanced tutorial but it it's got more more of them optimized like resolutions for this it'll work better basically all right sir I'm just going to leave this as it is and pull out your standard thing all right now you can use save image or preview image I use preview images almost exclusively because basically um I don't like saving stuff on my hard drive unless I have to it'll fill up with junk if I do that the amount of stuff I generate so if you queue prompt it should work to load the model if you've got your server window available you can see it loading all this stuff there you go all right so now it's generating a kitty cat on a roof and done so as you can see the image is not great there's distortions it has a gray sky for no apparent reason um there's some Distortion around the eyes the fur looks kind of wonky legs weirdly positioned the roof is just all kind of broken and one of the main reasons for that is the prompt is very kind of very simplistic which is good but the more complicated your sentence for building images the more it's gonna have an idea what it wants to make and it's less likely to make weirdness but one of the other ways we can make this look better is by using the refiner which I'll show you later which you can set it up pretty much the same way you just use a different text clip in code and you plug it in so all right

Channel: Ferniclestix

Views: 5,201

Keywords: SDXL, Tutorial, stable diffusion

Id: vABjDE40h8c

Length: 13min 5sec (785 seconds)

Published: Sat Jul 29 2023