SDXL 1.0 IS HERE! Were to get it, how to use it, and what to expect.

so this is the one dot o release it's been baking in the oven for about five months or so and it's actually a combination of a couple different models we've been working on with internally uh so once again this is Scott Detweiler the QA guy over at stability and we just got done with our launch over on Discord of the wandetto announcement so I wanted to cover some things with you that kind of went over there as well as where to get this model and how to use it so first of all this model has been made to be very easy to train so it may not be out of the gate like wow it's just not a amazing or ideal in every way no but it's much easier to train now training is going to take a little bit longer than it did with previous models but it should also be a lot more flexible so she'll be able to get out of it more than you think you can also I would actually try things first before you go and train them I think you'd be surprised at what comes out of this model without required training so in the past you may have required a lawyer to do certain things and now it just comes Natalie to the model so make sure you get things a try before you go ahead and train them I've trained things on as few as five images and they work worked out okay internally we've tried eight and it's been pretty good so don't think you have to walk in there with a gigantic amount of training data to get something simple out of the model again it's a very flexible so let's start with first of all we always grab it over at hugging face and make sure that you grab the safe tensors version of this don't use checkpoints checkpoints can execute Python scripts on your machine and lead to horrific situations so always use the safe tensors version of any model you download this includes the stuff that you might find over at civet from this point forward now if you would like to try the model for free without fine tuning you can go ahead and try our Bots over on Discord those are free to use and if you already have a dream Studio subscription dream studio is another place we use internally to play with different things and different ideas feel free to try that that's dreamstudio.ai if you've seen my recent videos you obviously know that comfy UI is my favorite tool here now it's much I say much more difficult but it's a little bit more difficult of a learning curve than say automatic 11 11 where you simply type in some stuff and you get an answer but what's great about this tool is your ability to actually break down the flow of how the model the noise and other settings are applied and in which order so you can do some pretty exciting things in here this is for those of you who really want to geek out again you can just place that model in the vae which you can also download flying face and there's also a Laura that was created internally by Katie So Cal Yuga so you can download that as well to try all these things out and Link them in and play with them as much as you'd like now automatic 1111 also supports 1.0 however I wasn't able to get it to function perfectly it did have some issues so I'd say give it another 24 hours or so before I would go ahead and Venture into that so this is a great opportunity for you to go ahead and try out comfy you're going to notice with this model that you have the base model itself and you have another pass called a refiner you do not need to use the refiner the system will work fine without it but it does add another level of detail that you may find interesting now some of us may just feel fine that we're good with all the refiner and we don't need that additional pass that's really up to you and again the amount of strength you'd want to put into it if you're using something like comfy you have the ability to kind of use it a little bit instead of using it at full strength the one recommendation I would make is that you pass a somewhat noisy image to the refiner don't try and finish it all up and then go through a font refining and then go through refining pass is you're going to find that it adds like a stippling to it or at least that's what we noticed so I would go into it with a little bit more noisy image to get the most out of the refiner the other interesting thing about this model is it's been trained on a lot of different sizes so you don't feel the need to start small and work your way up or use something like a high-res fix because you can be able to do a 10 24 by 1024 native image with a lot of different aspect ratios so this won't trip up as easily as previous models did in trying to give you duplicate heads and things like that speaking of which negative prompts are really kind of a thing in the past for the most part you can still use them from time to time to push something out of an image that you didn't want to see but don't go gangbusters and thinking that you need to paste an entire Priestess of things you don't want to see because a lot of stuff is junk anyway like what are extra fingers it doesn't know how many fingers are in a hand to start with so why would extra fingers matter so a lot of this stuff has evolved quite a bit since that first started the other thing you're going to realize is that the weights that you were using before if you had something say a weight of two in your prompt you can find that just way more than you need now we actually find it easier to lower the weights of some things than increase the weights of things that are more important it may seem counterintuitive but for some reason it works better than the other way around and even it's something that's a 1.5 or greater as far as a weight goes is way too much so just try smaller increments this means that older prompts are going to have problems using the same weight methods as before anyway this model should now be available up on hugging face for you to go ahead and grab and make sure again that you download the safe tensors version again it and the refiner and then the vae that Cal yugamate are all available for download free to give a try so let me know what you think in the comments below obviously we're going to continue to move forward with scxl on this channel and maybe a Sprinkle in some other models here and there I'm excited to see what you guys do with this thing where it's going to go we also have other features that are coming out very soon as well as our own ability to train them up on our server so if you're someone who would like to fine-tune a model and you just haven't been able to do it because your computer is a potato we're going to be offering fine tuning Services very soon on the dream studio interface as well as our platform interfaces so hopefully that's sooner than later and we'll see again it's always a moving Target with things we'll see how things turn out everybody take care stay safe and catch you all next time

Channel: Scott Detweiler

Views: 16,833

Keywords: midjourney alternative, sdxl, stable difussion, ai art, comfyUI, automatic1111 alternative, stable diffusion, stable diffusion tutorial, comfyui tutorial

Id: t5yAHGmqeqc

Length: 5min 39sec (339 seconds)

Published: Wed Jul 26 2023