Stable Diffusion SDXL 1.0 Released! | Automatic1111 WebUI

hello and welcome to More Than dirty rodent geekery yes the model we've all been waiting for is finally here stable diffusion XL 1.0 it's all of a few hours old so this is hot off the press and works beautifully in the very latest automatic 1111 stable diffusion web user interface which is currently at version 1.5.0 what do we get with this model well as the name suggests this is now extra large 1024 by 1024 in fact so twice the resolution of that stable diffusion 1.5 release if we have a look at this model card here we can see it also comes with a refiner and allora model as well sdxl consists of a mixture of experts pipeline for latent diffusion in the first step so there we've got this base model and we've also got an optional refiner as well that we can use in image to image and then they've got a little chart that shows the user preference for sdxl with and without the refiner over stxl 0.9 so there we've got 1.0 base 0.9 look look it's better it's better and with the refiner yes it's better so there you go sdxl it's better than all the other ones yes the chart says so so I believe it okay so this is brilliant I want it how do I get this working in my automatic 1111 web user interface well it's dead easy go over to the files and versions here of course links are down in the description and there you've got the base safe tenses and the offset Laura so you can download that just click on download file and you want to put that into your model's stable diffusion directory there it's got the text foot stable diffusion checkpoints here and that's where you want to put it so there you can see I've got both the refiner and the bass model as well for the loras we just go up one and there in the Laura's directory I also have the sdxl offset Laura okay so now you've got those downloaded you can hit the little refresh button and in the stable diffusion checkpoint selector you should now be able to select stable diffusion XL base 1.0 now you may find that some of your previously installed extensions do cause an issue with sdxl and in that case I would suggest disabling them until you find the ones that cause problems for you so I'm running here without any extensions just to make sure I don't get any issues and also the fastest possible rendering for reference I launched my web interface with just the xformers option and there's all the options here on the automatic 1111 Wiki command line arguments and settings if we go down here it's got a load of them and one of the ones you may want with a very low end graphics card is med vram along with no half vae obviously you'll be used to using those already if you do have an entry level card alright so we've got the model now and we are ready to rock and roll so just select your sampler of choice I'm using oil array there now you'll have to bump the resolution up a little bit there so I've got a width and height of both 10 24 by 1024 if you leave that at 512 by 5 12 then do expect to get a rather messy output with a batch size of one and a batch count of one that used about eight gig of vram for me obviously I've got a whole lot of recording equipment going on here as well also using vram I noticed with the med vram and no half vae options it was around six gig but well your mileage may vary right so with the requirements covered it's time to take a look at prompting now the first thing I noticed was prompting seemed to follow what I asked for a little bit better don't know could be just my imagination but let's check out some of these examples and see if you agree with me now styles were the first thing that I tested so here I'm going to generate a pixel art style rodent detective there he is let's have a click on that make it a bit bigger I think that turned out quite well that's not bad for pixel art style all right let's have another style let's go for a 3D render of a rodent detective instead and once again that has come out quite well that's a fairly nice 3D render of a rodent Detective okay how about if I go for a rodent sculpture made of glass can it do that well we'll find out in just a second and yes it can there is a very fine glass sculpture of a rodent I quite like the reflections through there that's quite nice playing again with materials let's try exactly the same thing but this time we want our sculpture to be made of chrome and there we go look at that that is a rather fantastic Chrome rodent can it handle watercolors now watercolors in the previous stable diffusion models I don't think looked very good but how does sdxl come out with my watercolor rodent astronaut that's not too bad you've got the definite watercolor art style there okay I like it I like it how about sketches let's have a look at a sketch of a cyberpunk rodent wearing a cool jacket all right yeah that's that's not a bad sketch he's got a cool jacket he's a bit cyberpunk you can tell by the glasses and just one more for this little bit here a realistic photo of a humanoid rodent Druid can it make it realistic well let's have a look there you go he seems to be holding some sort of nut and yeah that's not too bad as I guess it's fairly realistic as far as rodent Druids go then true nerd style of course I made some earlier so let's have a look at these Styles they are of course listed up there at the top so there we've got a traditional wood block style post modern art style as just a painting inspired by Joseph Turner it's not too bad there's a that's meant to be impressionist art style cyborg rodent but yeah not I don't think that's really impressionism there's a graphic novel art style uh futurism it's not too bad it looks a bit more like Cubo futurism but never mind there's a folk art rodent for you forvism Cubist R style Road a little bit better cartoon I do quite like the cartoon art style Art Nouveau which used to be quite popular in stable diffusion 1.5 and Anime rodent he's got a rather trendy hat on do quite like that academic art style that's meant to be which is I I guess it's close and they're abstract expressionism okay so it can definitely do a load of styles some better than others now of course obviously the one thing we all need to know thanks to these various technological advances can we now generate perfect human hands and the answer is yes of course we can here a realistic photo of a human hand and what do we get we get an absolutely perfect hand I did a load of tests with the different Samplers and guidance scales as well there we've got the guidance scale five going up to 12.5 as usual the higher your guidance scale goes the more baked that's the word I like to use to describe the look of it it gets so generally speaking sort of 7.5 to 10 seems to be okay before it starts getting baked or overcooked another thing I like to test these things on is does it have problems understanding positioning and all that color bleed you know remember when you could ask for a blue bear and a red forest and it wouldn't quite get it well let's see how well sdxl does so here I'm asking for one of my favorite tests of fish riding a bicycle and I think that's done really rather well all right let's let's put that on random and we'll we'll crunch up that batch size just so we can get a few more examples all right yes those are actually all all very very good those are there's a fish that's there's a fish riding a bicycle brilliant it's got it right I'm okay with that I'm pleased with that all right let's let's take it up a notch shall we let's make it really really difficult so now we've got a red box on top of a blue box how will it cope and the answer is not very well at all we've got lots of red boxes and blue boxes and not really any boxes on top of each other there we've got a little tiny box on yeah no I don't think it understands uh you know Red Box on top of a blue box so let's make it even more difficult a red box on top of a blue bench and we want it photo style at midnight so we've given it you know two chances I'm saying box is on the bench really I want that box on the bench I want it to be a photo and midnight as well because obviously previous models have had a little bit of an issue with Darkness sometimes and there it's it's not too bad we do actually have one image that's okay that's a red box on top of a blue bench it's done it but all of the others are just mixes of red and blue benches and things but midnight it's done very well as well so that's come out quite nicely okay let's give it one more complex test so here I want a huge green man standing next to a tiny blue alien and I'd like it auto realistic and once again it's it's interesting so we want huge Green Man tiny blue alien and we've got well that's that's interesting that's interesting and uh yeah I mean it's close it's close but no no banana there alrighty so let's move on to the next thing which is that Laura offset so here I've got exactly the same prompt there we've got the huge green man standing next to the tiny blue alien but this time I've put the Laura offset in it as well you can click that little icon there to show and hide your lauras and then click on Laura as you can see there I haven't got a preview on it because well I'm not actually sure entirely what this Laura does if we take a look at the images there it's it's pretty hard to tell how it's changed I mean some of them seem a little bit darker they're definitely different but as to as to what I should be doing this Laura not entirely sure so if anyone has any ideas do let me know down in those comments okay so the next thing is that refiner so this at the moment you might think perhaps you could use it in high-res fix but no there doesn't seem to be a way to do that as yet maybe something will turn up later in automatic eleven eleven but for now what I do is just send that over to image to image and then in image to image I select the sdxl refiner model in there pick the denoising strength you want here I'm going to go down to all around 0.44 that will do and now when I click generate it's going to give me that image but with a whole load of extra detail there it is I click on that we'll zoom out there we go that's that's quite nice it got a little alien there got the man and lots of extra detail on that large head as for that denoising strength you certainly don't want to go above 0.7 let's just have a look here so that's denoising 0.2 these are all examples so you can take that as being very close to the original image 0.4 there we've got a little bit more detail as we go up to 0.6 the sort of shape of the face is changing 0.7 the the very definite maximum I think there 0.8 the face is completely weird 0.9 it's stretching and by denoising one we've got multiple noses so yes not a very high denoising strength on that refiner would be my suggestion so there you go nice and easy to install and use and you get loads of really interesting outputs I know I'm going to carry on prompting to see what I can create do let me know if you've discovered any cool prompts down in those video comments of course just before you watch this next nerdy rodent video

Channel: Nerdy Rodent

Views: 56,629

Keywords: nerdy rodent, ai, ai art, best ai tutorials, sdxl, sdxl 1.0, sdxl stable diffusion, automatic1111, automatic1111 webui, stable diffusion sdxl webui, stable diffusion tutorial, sdxl howto, automatic1111 sdxl, stable diffusion sdxl 1.0, stability.ai sdxl 1.0

Id: KKJlkHz8kI8

Length: 13min 8sec (788 seconds)

Published: Thu Jul 27 2023