How to Install ComfyUI - 2023 - SDXL Special!

hello and welcome to More nerdy rodent geekery sdxl 1.0 has been with us for a short while now and while it sort of works in the automatic 1111 web interface it is far from a perfect experience for example swapping models leaks memory which you need to do to use the refiner and that can lead to low RAM Systems slowing down or just plane crashing plus there isn't really a good way to use that refiner yet so if you're looking for something which works a whole lot better then you may wish to take a look at comfy UI one thing which could put you off comfy UI is indeed the UI it being anything but comfy you've got spaghetti nodes everywhere and lots of really tiny text unless you're zoomed right in you also have to have some idea of what it is you're actually doing should you wish to change anything on the bright side though we do get examples from people people who do know what they're doing and so we can use and follow those even better still is the resource usage for a 1024 by 1024 image I recorded just under 16 gig of RAM usage and around 12 gig of vram usage when I was using the high vram option compare that to automatic and I was often using over 32 gig of RAM and sometimes quite a lot higher with the low vram option in comfy UI I saw 24 gig of RAM usage with just six gig of vram on my GPU being used so even low end cards should be all right here however even better still was that the comfy UI was using the refiner there too rather than it being a separate process later in image to image all in all if you're looking for something which doesn't eat all your system resources and can use the refiner in a much better way than automatic 11 11 then comfy UI is an excellent choice install income Bui is incredibly easy and there are a couple of installation options for novice computer users starting out on Microsoft Windows they provide a portable Standalone build to use this ideally you'll need to be familiar with downloading and unzipping files on your computer as they've used the awesome 7-Zip software to reduce the file size you will also need to have that installed on your computer too in order to unzip the file that will give you the comfy UI Windows portable directory on there somewhere on your computer wherever you've saved that and then you just need to run Nvidia GPU but assuming you've got a Nvidia GPU and you will start comfy UI if you want to run it on Google collab instead no worries there they've got a link to open with Google collab I'm going to be running all this locally and of course we have the normal install here for Windows or Linux if you scroll down a little bit you'll also see there you've got things for AMD gpus on Linux only there's installation instructions there for Apple Mac silicon if you're one of the very rare people that has that and also direct ml for AMD cards on windows so let's just pop back up to this manual install and we'll run through this basically this is just a typical install the same you would do with any other Python program for me I'm going to be using Anaconda so download and install that if you want or use minicondervn for any other tools you fancy with my anaconda prompt open I'm going to create a new environment of course I've already created mine so I'm going to say n there but you would say yes and that will create your environment all you need to do then is activate it and there you go you've got a nice empty environment ready to install comfy UI into just like they say there get cloned this repo you can now start on that so I'll do that git clone I of course have already downloaded it you will download it if you haven't already and of course then all you need to do is change directory into your comfy UI meaning you can then copy and paste whichever instructions are applicable to you for example there if you're an AMD card user Linux only then you can copy and paste that one or if you've got a different card I don't know what all the AMD cards are but there so obviously I'm using Nvidia so let's copy and paste the Nvidia one in there pip install of course once again I've got everything installed but that will install all the bits and pieces for you once you've done that just one more command to copy and paste so we'll copy and paste that pip install paste that in there and then comfy UI is installed whichever option you chose for the install you're first going to need those sdxl model files as well Links of course are down in the description but there it is hugging face over to files and versions and you've got the sdxl base and also the offset Laura as well if you want that plus the same thing for the refiner over to files and versions and you can download the sdxl refiner those should be downloaded to your comfy UI models checkpoints directory as you can see there I've got the base and the refiner and it says put checkpoints here if you already have a load of other models then there's an extra model path yaml file example which you could also edit if you fancy there it is you can just rename that and put different paths in there if you've got a load of models already okay so with the software installed and the models in place we are ready to run like it says running there python main dot py okay we can do that but what I'm actually going to do is main.py and also add the minus minus help option because that gives us a whole load of options that we can have a look at and some of these might be useful for you for example you could set different ports in there and extra model path configs and auto launch and force fp32 most of these are going to be debugging things for the most part it does just work with the defaults perfectly well I however don't like clicking open link in browser because that's too much effort so I'm going to use the auto launch option there okay now got comfy UI running you've got your nodes you've got your spaghetti what more do you want you just click queue prompt and and then away it goes you can see there it's loading the checkpoint it's doing these things it's done a case sampler get a little bar going up there which is nice it goes through the decoder and then it saves your image I've got a thing in a bottle let's just zoom in a little bit with the old scroll wheel there you go see image and that is saved in the output directory so we'll just go and have a quick look over there over there and date modify just so we can see the very latest one there is the bottle that's actually 1024 by 768 what what's that oh oh you want a slightly more chunky sdxl1 workflow like I showed at the beginning with those examples okay no problem let's go and have a look at those examples so here it is this is the link that you can also see on the comfy UI GitHub page there it is comfy UI examples and that takes us over here and then when we go to sdxl then that gives us some examples and oh what's going on it's just some pictures I want some workflows well don't worry because the workflows are in the pictures exactly like it says there you can download this image and load it or drag it onto comfy UI to get it let's just do that we go over here go on load I've got the sdxl simple example PNG there I select it and then it loads our spaghetti for us look at that now just hit Q prompt again there see what comes out from this and once again we get our image of a thing in a bottle which is what we're asking for over on the text prompts there now that's quite spread out so what I'm just going to do is I'm going to load a slightly different version of that there which is exactly the same thing I've just squashed it all down into one little area makes it a little bit easier to read and understand what is going on here okay so basically everything works from left to right we can zoom in and see things a little bit more clearly remember I mentioned lots of tiny text unless you zoom in well you're gonna have to get used to that when you're using this interface but don't worry because most of the time you don't have to zoom in quite to this degree and you can just sit there and play with your prompts all right anyway so here we've got the base model it's loading the sdxl base and it's loading the refiner then we've got our text prompts here so we've got a positive and negative prompts all of these have little notes as well which you can read at your leisure then we've got the base prompt there in the refiner prompt so that's going through doing some conditioning going into the Samplers and we've also got the empty latent where we specify the width and the height as well so let's scroll out a little bit again so we've got the models we've got the prompts they go through those conditioning and the latent into your sampler and then you've also got step control down here so we've got two different step controls here 25 and 32 and those refer to these Samplers here so we've got one for the base sampler and again you've got all the standard options if you've used stable diffusion before that you will be familiar with so you've got noise seed and the guidance scale and which sampler you're using and a scheduler of course you can just click on these to get a little pull down list or you can click on the little arrows to scroll through the options what's going on here well we're using that base sampler first we're using the base model then we're using the refiner and the steps here are essentially we're doing 25 for the base and then we're finishing up on 32 for the refiner so just a few steps on the refiner at the end personally I recommend doing no more than 10 refiner steps but that is entirely up to you and what you want your output to look like all right so let's scroll out again whoops and we'll pop over here into that positive prompt and we'll put our own prompt in there and scroll into so we can see what I've done there so that's a professional photo of a mystical rodent Druid with an HDR cinematic fantasy film style all right let's hit that Q prompt and that will go through all these things and eventually generate our image now you'll see there's an extra options there as well so that's handy if you want to do loads you can click extra options there set your batch counter whatever click Cube prompt and then that will generate 46 images all right let's just turn that down there we go we've got our rodent that's quite nice isn't it I think he has come out very well now I suggest playing with this basic workflow for quite a while and also remember you've got the second image there where you can give a bass and refine a prompt so download that second image as well you get the two different workflows there's some sample notes there so that gives you all the things that you should play with and it's basically these up here so the guidance scale the sampler name all those things play around you know get get used to it and then you can move on to the next stage which is this one here the saitan sdxl comfy UI configuration file so we just need this one file here that Json file click on that and then go over here and click on the download raw and save that wherever you like personally I'm saving things in comfy UI workflows I made a directory called that and that's where I'm saving them so save it wherever you like and then back into your comfort UI and we'll load this one up and this one is the saitan sdxl workflow select that opens up a little bit higher than normal but don't worry about it okay so here we've got much the same thing as you can see it looks really complicated with all these spaghetti things so let's just go through the extra options you get with this slightly more advanced workflow now the first thing is over here with the prompting as you can see we've got two positive prompts we've got a linguistic positive and also supporting terms there's a prompting guide there so that's the first Advanced option in here you've got two positive prompts the sdxl refiner also has some little tricks up its sleeve and that's in this gen settings over here so you've got a negative aesthetic score and a positive aesthetic score as well so these are things you can change those values on six and two and this workflow also has some options for upscaling and contrast fixing as well let's just zoom in there so we've got some upscaling models in there so you can choose an upscaler you've got a contrast fix and an upscale mixed diff option there as well so that will take your original image and then upscale it let's just run that prompt and see what comes out okay there we go let's just zoom in on those so we can see a little bit better and there we've got our two rodents so the original version and then the upscaled version as well it seems to have a little bit more detail it's it's quite nice but yeah that is a much more advanced workflow to get used to of course you will need an upscaling model as well much like with the automatic 11 11 you've got models and then upscale models and there I've put some real ESR Gan models if you don't have any real ESR Gan models of your very own then there are a whole bunch that you can download and use and for even more nerdy Road and geekery do check out this next video

Channel: Nerdy Rodent

Views: 35,386

Keywords: nerdy rodent, ai, ai art, comfyui, sdxl comfyui, stable diffusion sdxl, stabilityai sdxl comfyui, sdxl 1.0, python ai apps, generative ai, ai generated images, sdxl free, Generative ai, Generative ai tutorial, How to make generative ai, comfyui install, how to install comfyui, stable diffusion comfyui, ComfyUI standard install, ComfyUI Linux install, installing comfyui, comfyui for beginners, comfyui installation guide

Id: 2r3uM_b3zA8

Length: 14min 9sec (849 seconds)

Published: Thu Aug 03 2023