SDXL 1.0 - SUPER FAST Render Times + Google Colab Guide

the fastest way to use sdxl1 only 9.3 seconds hello my friends how are you doing yesterday I showed you in this video how to use it in automatic 11 11 including a sick trick of only using the refiner and still get really cool results but today we're gonna check out the fastest method this is using com for UI but it will make it super easy for you now for everybody who can't afford that sweet sweet Nvidia GPU gold I will also show you how to use comfy UI inside of Google call up let's get started so the first thing you need of course is gonna be comfy UI I will link this page below and when you scroll down a little bit it says you're immediately installing conf UI click on that and you will be super surprised by what you see here you have Windows and it says it is a portable Standalone built for Windows so you only need to download this 7-Zip file and then run it and then you have comfy UI it could not be easier then below that you will also see there is an AMD GPU version that needs Linux sadly but you can still run it so you can have conf UI and use sdxl today if you already have come for your eye on your system I would highly suggest to you that you go into the update folder there are two different bets in here the update configure iPad and the update conf UI and python dependence is bad so you can run either of them to update your com view I version after this has run through of course we're gonna go back to the main folder and in here you have run Nvidia GPU bad so you want to double click on that and open that up now when we go back to the confui GitHub page for a second you can see we have here link that is set conf UI examples you want to click on this and this will give you some powerful my mind-blowing sweet sweet examples of different note builds and one of them is the sdxl build when you click on that you're going to be greeted by two beautiful images now where is the build what are we gonna do with this well here's the thing if you create images with conf UI it not only includes all of your parameters it includes the complete build with all of the notes everything you build up this is genius my friend so download this image and save it on your drive and then when you start conf UI you only have to drag that image into the canvas like I will show you right now you're gonna be greeted with this this is called the canvas where you build all of your notes and on the right side here you have controls where you can save this in a Json file if you want to have it extra as a build there is a q prompt button up here that is for generating your AI images and so on but the most most important part here is you have your folder with the image that you've downloaded from the page take that image and drop it into the canvas and boom there we have it you have all of the nodes in here this is just amazing now as long as you are with your mouse in an empty part of the canvas you can use your mouse wheel to zoom in and out and this build is really useful also for beginners because you can see here that a lot of these parts have descriptions of what is actually going on but also they are here in placeholder boxes that have a title over them so you're on top of that seeing what is going on now here it will already say that you have a checkpoint loaded in here but this might only be a placeholder text so you still have to click in here and load it from your models folder either inside of conf UI or inside of automatic 1111 and at this point you will realize something that is absolutely magical you you can use several models at the same time in the same window by the way one side note here if you want to load your models from another folder like from automatic 1111 you have here description on the guitar patreon how to do that it says see the config file to set the search paths for models in the Standalone Windows build you can find this file in the conf UI directory rename is this file to extra model Pathways yaml and edit it with your favorite text editor because you can download it from this link here so we can click on that you see on the right side the download button so you can click on that and then Simply Save that remove the example at the end so this is ending at Dot yaml and save it inside of your confui folder then when you right click on this file you can see you can edit this I'm using notepad plus plus for that this is a free editor and it's really useful when you open it up it looks like this so you can set you the location for your check Point files for the vae files for the loras and so on and then Simply Save it and restart conf UI so now that everything is done like this and we have this node built here for sdxl and you have actually selected from the list here on the left side your base model and here on the right side your refiner model you can basically start rendering down here you have the text prompts positive prompt negative prompt everything is described for you it is really easy you don't need to rewire anything or change anything just put in the prompts here you want to use and test out sdxl the rendering is going to be really really fast and here you have to render time while I'm not recording it is only 9.3 seconds with a 4080 Nvidia GPU but of course on my channel as you know you get a quick rundown of what is actually going on so you have a better understanding on how this is actually working out so you have here the positive prompt and the negative prompt this is used both for the base model and the refiner model but because the base model and the refiner model are treating text differently we have to encode them with the specific model so as you can see here for example for the positive prompt this is going here into a clip text encoder now this clip text encoder is also connected up here to our base model and of course this has to happen for the positive prompt this is here in green and then the negative prompt which is here in red now then we have a render section that is basically what you have as the settings inside of automatic 1111 so you can see here it's called a k sampler here we have the information of the model the positive and the negative prompt you have the latent image which is basically the resolution of what you want to render you will have the amount of steps as an external input and then also the end step the reason for the end step here is because the base model is only rendered an amount of the total amount of steps and then the rest of these steps is done with the refiner model this is very different from what you are doing with automatic 11 11 because in automatic 11 11 the process is actually not the correct process because it is creating a pixel image and then rendering this pixel image again in image to image while in comfy UI it is staying as a latent image a latent image doesn't have image information in it it is these kind of latent data points that the AI is creating and only at the last step it is converted into a pixel image now you have seen here that we are loading latent image steps and end at step now where are these cables going you can see here we have the empty latent image this is basically a noise set up and here we have a resolution of 1024 by 1024 batch size one we are going to render one image here is an extra information of the sizes that are recommended to be used with sdxl 1.0 now here on the right side we have an extra box for step control you can see here we have 25 steps these are the total steps and then down below we have end step 20 which means that the base model is rendering for 20 steps and then the refiner is rendering the last five steps of that process so when we go on in that process and again you have here a big text box that is going to explain to you a little bit more about the case sampler read that if you're interested in that you can see that we have here an output as a latent and we're going to go into the second sampler because here everything is happening in one go because this is based on notes and you can build crazy AI machines with this so this is going to be the process for our refiner again we have a model input from the refiner we have a positive and negative prompt this is now rendered with the clip model of the refiner so when we look over here you can see that we have the input here from text from the positive prompt in green the negative prompt in red and then the clip model coming out here of the refiner model so that this is encoding the text with the refiner model now this is going into the case sampler and then of course as the last step what do we need we need to convert it from a latent into an actual pixel image this is done with the vae decoder so now just by looking at these steps you understand a lot more about what is actually going on with AI how AI is actually working so you can see here from the samples this is the 8 and it's coming in here and the vae we are using is actually also the refiner model and this only at the last step is turning it into this beautiful pixel image to start the rendering over on the right side you click here on the Q prompt and wait until it is finished now the amazing thing is if you don't have a strong GPU if you don't have a good computer you can use Google call app to use that you have a certain amount of GPU time that you can use for free but this is going to be a slow GPU so it's not really going to be much of a help but you can also go with the Google call up Pro Plan now here they have a subscription I would only suggest that to you if you really use it a lot on the other hand you can buy computer units here I was told that about 13 computer units is one hour of computing time for the GPU with a really good GPU like the a100 and these units will stay active for three months so you have ample time to use them up now the way to use this is also very very simple when you go back to the GitHub page you see here the call up notebook here is a link to open with Google call up simply click on that of course for this you need a Google account and on the right side here you can connect to a GPU so you want to click on this to start the connection you can see on the right side that right now I'm connected to a V100 GPU and this means I'm paying approximately 5.36 units per hour so I'm getting even more time out of that of course the V100 is going to be a bit slower than the a100 now on the left side you have your lot of text but you don't have to pay much attention to that you only have to click here on this play button and wait until there is a green hook next to it this is going to give you warning but click here on run anyway there's no problem because this is the official link from the conf UI guys so this is the first step here you can see it took me 25 seconds now below that you see a long list of different kind of models and things you can use here and this is selected as the 1.5 model so you don't want to have that but you can see that all of the lines with a hashtag in front of it are green which means they are not used so in this case we're gonna put here a hashtag in front of the 1.5 line instead we're gonna download the sdxl base and refine our model so we're going to delete the hashtags here so that the lines turn white also you want to scroll down here a little bit we don't need a vae because this is included in the sdxl model but we maybe want to use down here the Laura for sdxl offset which is this line so we want to delete that now again click on the play button and and you can see that this is starting the download now because this is a virtual server from Google the download is happening incredibly fast next we're gonna need to install the local tunnel don't worry this doesn't install anything on your drive it's still running only in the browser so again you go here you click on that play button and wait for it to finish and down here this is how we get our interface for the UI there's a link provided here for this page and then also you have here this kind of IP number so you need to copy that and then click on the link this will bring you to this link here and then you put in here your IP this is generated every time you're using this and we are in the same canvas we have seen before and this already is loading some stuff here but we are not going to use that because I'm going to provide a file for you linked below this video so you want to go over here on the right side to this part and here is a load button click on that and then you want to load the sdxl1 workflow Chase and that I'm providing you this is the same as from the GitHub page so load that and now we have exactly the same setup that I just told you about with all of the settings in here now here's one thing you need to do for this to work you need to go in here and even so this already says sdxl base model you need to click on that list and select it again and then over here do the same thing with the refiner model so that they are actually loaded then you go down here write your prompt and negative prompt and you already are done and you can start rendering so over to the right you can actually zoom in here to see the image a little bit nicer and simply click here on cue the prompts you can see that the image has rendered here and it is really beautiful very detailed look at that we have captured a landscape in a bottle here is the beautiful image I've rendered on my computer even while I'm recording this video that's it for today subscribe to the channel if you want to see more like that thanks for watching and see you soon bye oh you're still here so uh This is the End screen there's other stuff you can watch like this or that's really cool and yeah I hope I see you soon uh leave a like if you haven't yet and well um yeah

Channel: Olivio Sarikas

Views: 50,883

Keywords: oliviosarikas, olivio sarikas, olivio tutorials, stable diffusion, sdxl 1.0 automatic1111, sdxl 1.0 comfyui, sdxl 1.0 install, sdxl 1.0 release, sdxl 1.0 huggingface, sdxl 1.0 model, stable diffusion sdxl 1.0, sdxl comfyui, comfyui sdxl 1.0, comfyui google colab, google colab stable diffusion, google colab tutorial, google colab pro, google colab, comfyui sdxl, sdxl fast, sdxl best, sdxl free, sdxl no gpu

Id: TYk6taV-gVc

Length: 15min 27sec (927 seconds)

Published: Fri Jul 28 2023