Mastering Stable Diffusion SDXL in Automatic 1111 v1.6.0 | Tips, Tricks & Workflow

[Music] hey everyone welcome back to alemy with zero phase I'm Eric and in this video I am going to go through my workflow just for some basic stuff um talk a little bit more about stable diffusion sdxl and what you need to watch for and um really how to get some of the best results out of it so as you can see I have my stable diffusion interface up and right now I'm selected on a model called sahastra I don't know if that's how it's pronounced but it's a fabulous model there are a ton of models out there uh for sdxl now um I mean really it's just a matter of pick your favorite what does it do you know what do you want to do go to civit Ai and you know look and see what models do what and they have a ton of them now uh it's great to see the involvement people have and and what they're uh building for this it's absolutely phenomenal especially the ones that are building the refiners into the models I kind of still use the refiner once in a while or a lot of my stuff I don't think the one I currently use like this the sahastra sahastra kai I don't think it has a refiner built in so I do use a refiner on that one um so it really depends on what you want to do uh this particular model is uh what I call a generalist model it's very well trained on a lot of different things whether it's digital illustration abstract realism um photography that kind of thing it does a really good job okay uh with um with most of the newer sdxl models that come out you don't need a heavy negative prompt um most of my negative prompts are pretty heavy and sometimes I will use them but they are not entirely necessary another thing to note about the sdxl models is they are very good at interpreting uh your prompt um they will give a pretty good level of detail even with simple prompts you could say photo of a um goat on a mountain and I'm not going to use a negative promp I just want to see what it does uh so we're just going to hit generate on that okay and you get a photo of a go on the mountain the uh sdxl models are are much better now at uh legs uh you're not getting you know five legs or six legs on animals that shouldn't have five or six legs um I do have my settings uh set to by default now go to 1024x 1024 the sdxl models that's what they're trained on they are trained on different aspect ratios uh you can look that up um I have some of the bu in ones uh as part of my aspect ratio helper here 1:1 3:2 4: 3 54 and so on okay uh it's good to have that uh as it really helps yeah that's a pretty good goat not bad did a lot of good detail on the ground and everything lighting it I didn't specify anything about the lighting I didn't specify anything about the environment I just said on a mountain and so the AI interpreted that pretty well okay um you can choose pick and choose your your samplers they have some new Samplers out with the new version of uh automatic 1111 uh the more more notably is the DPM plus plus 3M series uh the DPM adaptive now I haven't played a lot with these I still uh find myself using the uh uh the DPM Plus+ 2m or what is it uh Plus+ SD cross I'll use UL a a lot still but I am starting to move into the DPM Plus+ 3M SD cross okay all right so um you don't need a high sampling Steps either as you notice I did uh 25 with this um I think anything above 30 to 40 I found I did some testing a little while ago in one of my videos uh really doesn't net you much in the way of better results um if you really want the details you know leave it down at a lower One enable the refiner uh just so a lot of you know um they did change something on this so a lot of us and it took me a while to figure this out uh when you go to one of these extensions usually they have like an enable detailer or enable whatever and you have to click the enable to enable it okay when it comes to the highres fix and refiner all you have to do is open it okay open it select your refiner model uh I have one in particular I use called the refiner VA see it's sdxl v1l refiner VA fix okay so there's there's uh I don't know if this is still an ongoing issue but uh the original refiner they came out with had some issues with memory management and so and they had to have a vae for it and so um this particular one uh implements those fixes okay with a refiner um you have a switch at slider bar and what this means is that at what point in the rendering does it switch over to using the refiner and adding more detail and I've heard people give different opinions that uh for every step here um sampling step for every excuse me for every five sampling steps you would want or excuse me for every step you'd want five steps here or something like that I don't know exactly what the ratio is but um I like to stay between 7 and8 um now I have had some very very interesting results and some very interesting images when you slide it all the way down you tend to get because the refiner is trained on a lower resolution um you'll get a lot of duplication of things like if we were to switch this down to say 02 let's go ahead and just generate that um you'll kind of see what I mean here we'll see what it see what it does here so we hit 20% Which is the 02 and then it switches over to the refiner and starts doing the refining as you can see it uh does some very interesting things now when you're working in abstract art sometimes that's really cool you know you get uh I I did some really bizarre HR Giger uh Landscapes very biomechanical with bodies and stuff like that and this doing this technique actually netted some really cool results that I loved so you have to kind of just experiment with it and see how it works for you um let's see there's the refiner um slide that back up to8 for right now as we go through this I'm going to leave it open um something that most sdxl models really excel at now um taking in a lot of the details now if you add a lot of detail to your prompt um if you use the break command it will interpret those really well so but what I'm going to do we're going to set up a scene I want to do I want to do another nature shot we'll just say a really good nature shot of a goat but uh we're going to we're going to really implement this go through my workflow so you can see this so um let's get rid of this here and we're going to leave all the other settings we're going to and switch this over to ur because it does tend to render faster um I think with a refiner it does tend to negate some of the softness that UL has um and then we're going to go up we're run we're going to generate some prompts here and I want to see uh professional um nature photography of a goat we'll say mestic goat on a rocky Pinnacle at Sunset okay uh let's make sure we select our model okay so just so you know uh some of you may have seen my previous video this is the an extension I built around my prompt generator um you can get it for free off GitHub uh it doesn't come with the prompt generator but you can add your own prompt generators your own promp pre- promps or seed prompts to it uh so you can start utilizing this inside of automatic 1111 okay so we're going to just do four prompts and uh I want to get some style in there okay okay we're going to submit that takes it just a few seconds to generate those prompts what this will give me is a lot more detail uh I've given the uh gpai a lot more creative freedom uh to be able to um add elements emphasize the proper uh and uh features of each prompt so we're going to grab those head over to here now to utilize multiple prompts you want to go down to script and select prompts from file or text box okay and it adds this field down here you can paste those in there and as long as there is a a paragraph break you don't necessarily have to have a line in between them I do I like being able to see each prompt you just they just need to be there needs to be a I guess a carriage return to indicate the separation of the prompts but I like to see the space okay so we have four prompts in there um I want to see some deal I'm going to bring this actually down to 7 uh we're going to leave everything else the same here uh we are going to change this to uh I want to see what it would look like a really tall aspect ratio so we're going to flip that so it's going to be a 9 by 16 vertical essentially uh for this I am going to add my Prime negative prompt and then we're going to just reduce the negative prompt effect down a little bit to3 you're still going to get a lot of effect from the negative prompt but it won't be as strong so it gives the AI a little more uh flexibility and what's in the scene hit render on that you'll start to see the images come across I'm only doing one image per prompt so we should see it one image at a time okay there's that so what I've done with my prompt generators it does add in camera information it'll add information about light and coloring um it'll add details to the scene so you get something that's a lot more natural a little more cinematic almost or just uh details to the background and something that looks like it could have been a professional shot okay I'm going to let it go through this and we'll be back in just a second all right there we go so as you can see we've got some nice vertical images I love these images because they tend to give like these really cool longdistance Vistas and uh it did a really good job uh maintaining the prompt I wanted the goat out on a Rocky outcropping and with mountains in the background had a lot of good detail uh added detail more about the lighting too um go through these got a lot of detail around the rocks and what's going on there this one even did it as a black and white I'll bet you this one included the uh name anel Adams in it let's see oh France lanting okay and then that one's a little interesting cuz the goat's got its head turned all the way back around but I love the atmospherics you got going on here the storm sty Sky uh the light from behind it maintains I love how the AI just knows like if the AI if the light's coming from one direction to maintain that consistency throughout it's just fascinating okay so again play around experiment with different models um one thing you may want to do a lot of you are still using 1.5 models and sometimes this menu gets a little messy so what I've done you can go into the models directory uh under stable diffusion and just create a directory like for me I created a directory called sdxl and just that's where I put all of my sdxl models so that when I come in here um and I want to find a particular model let's say we want to find um anything sdxl so we can actually type in sdxl there's all your sdxl models right there whether they have sdxl in the name or not but because I put that as the folder name and automatic 1111 automatically kind of adds that to the beginning of the name uh I think that's a really cool feature for organizational purposes you know something you may want uh something that is nice about that is you can do that in a lot of different menus um you can come up here and in the Styles start typing and it gives you anything that starts you know like I typed color here's all my different uh styles that have the word color in the name and you can do that in other areas too even on the checkpoint here um o open that back up and uh you know anything where you would have a drop down man you can actually type and just get to it a lot quicker it's really cool love it another thing that I want to take note of is I'm not using the highres fix as much as I did with the 1.5 models you can still use it what I've found is that it really extends the render time um and so I only use it if uh I have an image I absolutely want to do it on okay and another thing that so again I I I don't use this as much I still use it occasionally okay um because it does it is nice getting that higher res uh but a lot of times I'll still just move to uh image to image and use the ultimate SD up scaler so the other thing is a detailer so when detailing faces uh um this I do use this still like when I'm rendering people it is nice the the sdxl models are really good at doing faces now there are still issues dealing with eyes uh teeth things like that once in a while so I do use this it you know increases the render time just a little bit because it's taking time to reload the the model that it uses to do the detailing and then moving back to the other model but it is worth it uh to get those really good eyes and and facial features okay just a couple more things that I want to point out here before we wrap it up um one of the biggest changes that I love in the new automatic 1111 version 1 point where we at 1.6.0 it deals with the Styles menu um uh so originally you had to go into the CSP file to edit the entries in there you could add the entry there but uh you had to um it wasn't very easy it wasn't very convenient they've added this button over here looks like a little pencil uh it's there little edit Styles button when you click that brings up a new window where you have access to all of your Styles and so you can come in here and select the style including negative prompts um and edit them directly within this interface face here now uh you can refresh it here and it looks like they moved the button to push it to the uh um main UI and the the fields there into this window but um this has been incredibly helpful being able to come in and refine your prompts all right sorry about that I had to take my little one upstairs so um this is a great function to have a great uh feature that allows you to really customize it and one of the things that uh is great if you look let's see where does it show this I don't know where it showed this but okay so one of the features this has now is you can uh let's say you let me see if I have one in here probably my glass sphere one glass Globe okay yeah so this is absolutely amazing this has been kind of a game changer on on a couple different levels uh so you can have your style your prompt that you have programmed in here and what this does it goes in the Styles and then you put in this variable right here uh so curly bracket prompt curly bracket and what that does is uh it takes whatever you put in your positive prompt and and puts it it inserts into that prompt that style so I'll show you here so I'm going to pick my glass globe prompt um not going to worry about a negative prompt we're going to try it and see what it does so we're going to do uh uh cute tiny teddy bear and we're going to just leave everything else there we're going to change this over to a uh um 5x4 this time we're going to shut off the script here we don't want that script anymore leave everything else the same I think and generate so what it's going to do is it's going to take this cute teddy a tiny teddy bear variable here inserted into that prompt and start rendering using that instead I love that prompt it's such cute such cool prompt so very cool feature so start thinking about um the different prompts you have and the different parts of that prompt you could change you know you could insert a variable in and uh uh I mean you can it's just really cool being able to do that and just render out anything you want so you get kind of the same thing but you get whatever modification to the prompt that you want that one's a cool one I love the book and everything on the table that one turned out good give it a second to finish up I want to see that one full size oh yeah so I I'm sure most of you have noticed that the sdxl prompts are much better at doing Dragons for those who are interested in doing Dragons uh it does a phenomenal job yeah I that's cool got a little funkiness going down here on the lower jaw but the rest of it is pretty darn cool anyway I appreciate you taking the time to listen um trying to keep these videos a little shorter than normal uh but uh appreciate all of you and the feedback I get make sure you like And subscribe and we'll start putting out some more videos talk to you later

Channel: AIchemy with Xerophayze

Views: 3,587

Keywords: AI art, Stable Diffusion SDXL, Automatic 1111 v1.6.0, AI art tutorial, AI art tips, Creative AI, AI art workflow, AI art techniques, SDXL guide, AI art mastery, Xerophayze tutorials, AI art community, Artistic AI, Advanced AI art, Digital art, Creative process, AI art creation, Digital creativity, AI art insights, Artistic inspiration

Id: nqZkm216Glk

Length: 21min 36sec (1296 seconds)

Published: Wed Oct 04 2023