New Model! SDXL Turbo - 1 Step Real Time Stable Diffusion in ComfyUI

hey everybody Scott dwe here so stability has just released our sdxl turbo model so this is a distilled version of the sdxl full model and it's going to have some restrictions but it can do images in nearly Real Time with only one step now it's going to be limited of course that's really great goal and again this is a research release so these are things that we're developing as we're going and this is a chance to kind of play with that and kind of see what we can what we can do with it so it does some restrictions uh it is only a 512 x 512 Square model so realize that um it's not going to do text very well it's probably not going to be very photo realistic and it's probably not going to do faces extremely well but it is again a research release or a stepping off point to show how we can speed up this process in the future so kind of keep an open mind to that now I'm going to show you in comfy how to use this it's very simple but we're going to take the case sampler and we're going to rip it apart into some of the subsections or other things that drive the case sampler so this has been available and comfy for a while now but I guess a lot of people really haven't gotten into it very much but this is an opportunity where we really have to rip into it now a couple things to be aware of your CFG is going to have to be a one and you're going to need one or maybe two steps at the most anything else and your images are going to look pretty trash so if you're wondering why that'd be the reason also make sure your comfy is up to date because this just came out yesterday uh so if you're wondering why you don't see it that'd be the reason always make sure you're comfy as well as all the custom nodes are up to date that will eliminate 90% of problems everyone seems to have when they put in the comments after these videos and saying it doesn't appear in my image that'd be reason why so anyway let's take a look at it okay now to get started with this we're going to go ahead and load in the sdxl turbo model just like you would into a regular checkpoint I'm going to do this very quickly we're going to create two text encoders now remember that the sdxl turbo model does not use the negative prompt so we're still need it uh but we're not going to be using it for anything else now here's where it gets really interesting normally we' load a case sampler in here but you can also load in a custom scheduler Uh custom sampler I should say so that we can kind of break it down into the individual ual components that make up the case sampler and this is kind of a opening the door to other goofing around as well instead of just you know using this model but using this method with other models as well so if we pull this out uh we can see we have our typical case sample here and I may have other options that you do not have but don't worry about that but if we type in if we double click and type in the word custom we can see sampler custom comes up right and this is pretty obvious from at least a couple of these so we have our positive or negative and so on again we won't be using the negative so we'll just kind of collapse that now the other ones here that are interesting are sampler and sigmas so if we pull sampler out you can see we can load in case sampler select this allows us to pick which sampler are we going to be using I'm going to go ahead and pick the DPM Plus+ 2m this is the one that I'm preferring right now so for the latent image we're going to be using a 512x 512 empty latent this is ideal for this model so go and stick with this for now and then for the sigmas this is where it's really interesting there are a lot of different sigmas that you can use now now if you pull this out and look at it you're going to see there's a basic scheduler a Caris scheduler an exponential scheduler a polyexponential scheduler and so on but you'll see this new one is the SD turbo scheduler if you do not see this then you need to stop and update your comfy because it is brand new and if you're using one that is even a day old uh you're now behind the times and this is the number of steps is going to allow now in this case two is probably the maximum one is ideal uh two might be allowed uh but uh any more than that you're going to see the images is going to break down into something silly so this is basically it so from here we can use a veed code just like we would normally and borrow our vae and then I'm going to highly suggest to use a preview image here and not a save image because you're going to see it gets insane pretty quickly so if we go ahead and put in a prompt of anything okay now if we hit Q prompt on this you're going to see it's very burned out and that's because the CFG is set to eight you're going to want to set this down to one try it again and there you go and you see how fast that is this is insane there's actually something pretty funny you can do here so this is this is a little uh little fun uh but this is really where I think this is going to shine I'm just going to copy and duplicate this whole thing by holding down my shift key and I'm going to take and just let these U be whatever they're going to be as far as the seed goes what I'm going to do is I'm want four different ones and you're going to see here what I'm going to do is a pretty wacky so I'm going to take these four here and I'm going to create kind of a a gooey for myself a graphical user interface so here's my prompt right so what I'm going to do is I'm going to take my prompt and I'm going to drag this over here I know this it's going to be messy U just don't worry about it right now this is going to be pretty fun and this is where I think this model is really really interesting again this is a research model uh so it's more for a goofing around and not so much for serious work uh but again this is pretty fun all right so once you get these all in place and you little OCD there once we have this kind of arranged what I want to do now is go up and set these all to fixed we're going to make sure they're different seeds but we want them all to be fixed and the reason is if we don't do this it's going to be spraying images at us too quickly so we get something like this one other thing you want to do is go up and hit Auto Q here and then click Q prompt and it's going to go ahead and start the process and as we type it's going to update all four of these in real time so as you work your prompt you can kind of see how it's affecting things again the seeds we're only looking at four seeds here uh so if you want more obviously you can go ahead and expand this but this gives you a little bit of an idea of how the engine is interpreting your prompts which I think is the thing to really walk away with here is learning how the engine is looking at your prompts because as you change the prompt you'll see it reflected immediately so it's a much more interactive process so hope you found that helpful again this is a research release so let me know in the comments what you think of this and how you're going to find it useful again there's lots of great things we can do as points of departure for this model and I'm really excited to try it out everybody take care stay safe and I'll catch youall next time

Channel: Scott Detweiler

Views: 27,716

Keywords: stable diffusion, stable diffusion xl, ComfyUI, stable diffusion tutorial, AI art, SDXL, comfy tutorial, sdxl tutorial, sdxl template, ai art tutorial, comfyui tutorial, stable diffusion beginner tutorial, LCM, latent consistency, lcm sampler, comfyui lora, SDXL turbo, custom sampler, stable diffusion turbo, comfy turbo

Id: DZ2dfq8ljrc

Length: 6min 27sec (387 seconds)

Published: Wed Nov 29 2023