SDXL - BEST Build + Upscaler + Steps Guide

how to get amazing results with sdxl hello my friends how are you doing we had an amazing live stream yesterday you guys are legends we talked about the best settings how to upscale the images quality reviewing images online so today I will compress all of that information into a tutorials so shout out to all of you but also special shout out to Winston wolf he's part of my Discord Community he helped me yesterday in the live streams he helped me today so let's get started here first we are going again to look at comp view I because this is an amazing lab to experiment with the different methods and see how everything works but also set up automatic testing of different methods so here is what is new let's scroll in here on the side and you can see that we have multiple instances of prompts so here we have positive R and negative R and when you look at the wording you can see that this only defines the setup about the quality about the style of the image so here we have detailed photo wide angle shot realistic 8K UHD high quality and then in the negative prompt it's also 3D render anime smooth plastic blurry grainy low resolution deep fried over saturated not sure what the deep fried is there but we all love the chicken wings so that's okay with me and then we have here the model Lotus for the refined model for the base model we will go where this is linking in a second but below that you can see we have here another prompt now here you can see it says in this case beautiful young French man with a beard and curly hair standing in a forest and then in the negative prompt we have 3D render anime then below that this is the next important thing here we have the total steps and then we have the steps on base model because there's a difference between them so for the base model in this case we are using 20 steps and then we link that to start from the 20th step upwards to the refiner which is using 25 steps so there seems to be a golden ratio of 80 20 so that you have a 20 percent of refiner steps and then 80 percent of the base model steps below that we have just the image size so we have seen that before 1024 by 1024 batch size one down here we have the CFG scale and we have the seed in here which in this case is fixed but of course you can also switch that to for example randomized over here fixed is good for testing of course the same results with the same seed with small iterations in your settings and then below here we have another positive prompt for L and the negative prompt for L now this is the same prompt as above but it can also be a different prompt than above and this again focuses on the technical parts of the image so is it a photo wide angle realistic 8K and so on and so on so now we come to the next step and this is the base rendering where it's not refined yet and don't get confused by all of these different cables here between the nodes it might seem intimidating but it's really not and also I have this as a download for you so you don't even need to set up anything of that but of course you can experiment with that you can expand on that build anything you want from that so here we have the two clip text encoders specifically for the sdxl model and this upper one is for the positive prompt and this here is for the negative prompt now in this case you can see here we are coming in with the clip this is from the base as the XL model then we have here two text inputs one is text G the other one is text L now when you follow these lines you will see that the text Chi is coming here from our input prompt that defines what is in the scene the content of the scene and then the L input comes from down here where we have positive L negative L that defines the technicality of the style and so on about the image so here we have the input for both of the positive prompts and here of course we have the input for both of the negative prawns both of them are set up with a width of 4096 by 4096 but don't worry the image is still going to be 1024 when it comes out next we have here the K sampler Advanced here we have the input for the model this is of course the base model for sdxl the positive prompt for here and the negative prompt from here then we have the latent image in this case is an empty latent image that is coming from what I showed you before 1024 by 1024 batch one below the that we have the input for the steps the end at step the noise seat which is basically just a seat and the CFG scale from the notes that I've shown you before and as we have seen this is going from zero steps to 20 steps as the base stage The Next Step here with the vaed code is not necessary for this step this is just so you have a preview of what the base render looks like in this case this is our result of the base render which is not refined yet that is why there is still so much noise in there because this is not rendered to the finish so now at the next step we are going to look at the build up here for the refiner stage now again we have here a clip text encoder for the sdxl refiner in this case you can see it is a lot simpler because now you have here a clip input and a text input and that's it so the clip input in this case of course comes from our refiner sdxl model and the text up here comes from up here for the positive R prompt for refine negative R prompt for refine so those go in here but not the content description of the scene just this kind of technicalities here so those go into here get mixed with the clip model and then get output here and here into our K sampler for the positive and negative prompt again we have an input for the model this of course is now the refiner model now we have a latent image that is of course the latent image that is coming out of the base render so when we are following the line you can see this is the output of the base render going up here as the latent image and then over here and into our input now here you can see we have here the total step count and then we also have start at step now this is the note I showed you before that defines the step for the base model so we can also use this as the input for this case sampler because we want to start at the 20th step and then from then on render the other five steps after that we have here also a seed input but this does not have a CFG scale input so here you have to set the CFG on your own in this case and also the other settings in here like the sampler the scheduler and so on now in this case we actually do need a vaed code and in this case when we follow the line of the vaed decoder it goes over here over there and then into our refiner model to render the output of our image as a pixel image so to understand this better a latent image is just made of the data points that the AI is using to define the image but it's not an pixel image yet and then what did vaed code is doing it is taking all of these data points and actually rendering it into pixels that you can look at on your screen so at this point in the process we do have a beautiful picture but it is 1024 by 1024 and when we scroll into it it is very nice but it's a little bit blurry and it could use more details so yesterday in my live stream you showed me a process that is using a double upscale now the first upscaler here is a specific skin detailer this is a 1x upscaler so it doesn't upscale at all it only refines the image after this is done the image is then going into a second upscaling down here now I have split this up into two different methods one method is using the 4X nmkd CX upscaler and the second method you find over here is using the 4X Ultra sharp upscaler now here at the output for both both of them you can see the results side by side and when we zoom into the image this is the upscale with the nmkd model it looks very nice but it has a little bit of a graininess in there although hairs and other details come out very nice and sharp and over here we have the upscale from the ultra sharp model this also looks pretty nice but it's a little bit softer and it sometimes has the downside of being blurry in separate parts of the image so for example here you can see that this part of the image is rather blurry while here we have rather sharp lines on the face although the skin looks very dry and a little bit wrinkly I will link you all of these upscaling models below my video so you can download them you can experiment with them and when you download them of course you want to put them into your comfy UI folder there into the models folder there into the upscale models folder and then you simply want to reload this m image you don't have to restart comfy UI just reload the browser image and then you can select it here from this list I will also link for you a model database for upscaling models that are trained by the community for all kinds of purposes so you can download them you can experiment with them this will often link to a mega drives link you will see here a view of a folder and there you have a pth file that is the upscaler file that you need to download So for that go to the right side for these three dots you click on that you have download select standard download for that it will take a little bit for preparation of that and then it will ask you to save it to any kind of folder where you want to save that now here's another advice from my community you want to go to the official stable diffusion Discord in there is a show and tell XL room for the XL beta so here you can find very very nice examples to look at also in here you have a Parthenon of the winners of their challenges so this not only shows you the best images from these challenges but also the prompts that they have been using so you can experiment with these prompts and they are also relating to the sdxl beta so there's a lot of inspiration and amazing results in these two Discord channels I will also link the official Discord for stable diffusion below this video so you can join it right away finally I'm also giving you my image for download and you only need to drag that image into the canvas of conf UI and this will then automatically load the complete build for you of course for everything to work you still need to download the sdxl models and the different app scales that I'm using in here all of that is linked below my video I hope all of that is helping you getting amazing results thanks for watching leave a like if you enjoyed this video 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 leave a like if you haven't yet and well um yeah

Channel: Olivio Sarikas

Views: 43,012

Keywords: oliviosarikas, olivio sarikas, olivio affinity, olivio tutorials, sdxl 0.9, sdxl stable diffusion, sdxl beta, sdxl text, sdxl model, sdxl vs midjourney, sdxl prompts, sdxl huggingface, SDXL best settings, best ai tools, new ai tools, ai news, SDXL news, SDXL 0.9, comfyui, stable diffusion, stable diffusion tutorial, stable diffusion xl, stability ai, ai art, stable diffusion ai, adobe alternative, stable diffusion gui

Id: GrgA8ge5hf0

Length: 12min 15sec (735 seconds)

Published: Mon Jul 10 2023