DreamShaperXL Turbo - The Fastest Way to Excellence in Stable Diffusion

hey guys this is Kevin for pixel.com and in this video we're going to be taking a look at a new version of dream shaper so dream shaper Excel turbo just launched and I want to take a look at the capabilities of this now the old dream shaper for stable diffusion 1.5 that one is such a cool one that has become sort of my go-to recommendation for people starting out with stable diffusion and I want to take a look at the Excel version now if you want to follow along you will need to download the file just come over here to the uh civit AI page you can click here to download you can come to this section here to download as well now because it is a turbo version it uses the adversarial diffusion distillation method and that's add and I like saying that cuz it makes me sound like I'm super educated so what we need to do is to follow the instructions for using this one it's a little bit different from using uh a standard sdxl model now we covered this before when looking at the first sdxl turbo similar inst well it it's similar in that it uses a very small number of sampling steps and we really have to follow the instructions here what I'm going to try to do is to actually use the workflow inside of comfor UI now some of you may use automatic 1111 comfor UI is very easy to set up and to use and we should be able to get it working just by using an example workflow which is what I'm going to try to do for this video we'll see how it goes now I'm just going to set up my my comfy UI so that it's running in the background I've got a video which covers the basics it's a beginner's guide and I'll link to that below that will cover comfy UI and I have another one for automatic 1111 but we'll try to cover both techniques in this video please hold for a very important message for comfy UI you should be okay to just come down here and then to copy the uh actual file here which is the workflow so let's just click on that and see what we get so what I would be doing here is coming and just downloading this and this will give us a Json file and we should be able to use that Json file inside of comy UI so let me just grab the Json file and see what it looks like see if we can actually use it now this is comy UI open and what we have here is a workflow from Sego this is a new technique which I'm looking at and this new technique is fantastic it takes us in the opposite direction to the turbo versions of sdxl it takes us into the uh section where we require more powerful Hardware in order to get things done it produces uh better results however with sdx XL turbo we should be able to get very good results even with limited Hardware so let me go ahead and grab the J Json file and if you're new to com UI all you have to do is to drag the Json file into the workflow or into the work area and it just brings up the the new workflow so this is the workflow that we've been given um we've got a possible load Laura which is currently disabled we have got the clip encoders the text encoders and then we have the case sample let me just make that a different color uh me make that blue and uh we've got a second case sampler oh my God so we should be able to make this work just by clicking on the Q prompt let's change the color here to Blue as well that's just my preference and let's prompt this okay we haven't got the save tensor now I've downloaded it and put it in the right place but I've got a specific place where I put my Turbo versions and I'm just going to have to go and find it there so it's this boy here and what you'll find if you drag this workflow in you should have the correct settings for using this obviously we've got some room for maneuver in terms of changing the settings let's see what this produces wow that's pretty amazing so we've got the results here we've got two images 2K Samplers and I think one of these is much larger than the other one so let's go ahead and just open this up and expand this whoo that is a large image and you can see this oh my God that is a huge image so you can see it is creating this in five steps in the second Cas sampler and uh let's go and take this save image here so we've saved two images we can compare them and this one is going to be a bit smaller but still I imagine fairly large as well W those are two two pretty large images that's our first image and that's our second image and the second image I think does look a little bit better in terms of quality um you've got some a few issues happening there which have been fixed in the second image so it's really really cool and it's super fast uh given that we're creating an image which is going to be larger than this latent the first latent is 768 by uh 1024 but then we have an upscale Happening Here uh and this is the first time I'm seeing this so we've got an upscale method using by cubic scaling it up to 1.5 and it's going to be significantly larger for this second case sampler so the second case sampler takes quite a bit of time uh D noise of 0.5 means that we're getting about 50% of the original uh maintained in the output and the rest of it is kind of like an enhancement of what's there already um we've got a fixed uh control after generate and just manually change the seed here and here as well and it's using the same seed which sometimes produces problems in in terms of how uh the things render but let's see what this one produces and you can see how quick the first one is and I've got all kinds of software running in the background my screen recording software there's a lot going on this one is going to take a bit bit more time let me pause that and here we are that's the first image that we got that's the 76 8 5 10 24 and then the second image the slightly larger one you can see once again we're getting really nice looking image with much better quality uh and still also try at a pocket there which wasn't quite successful but we've got better looking hair on the arms we've got a slight uh better looking face as well and I think overall this kind of sample and then resample method seems to produce really good results and we're only taking five steps so even if you haven't got a really powerful powerful graphics card you can still get things done pretty quickly now the other thing to take into account is that we're using the same text for the second uh for for the first and the second case sampler and we are using settings which are recommended we can go down to about four in terms of steps for this particular method CFG stays at two and the sampler has to be this guy here the schedule has has to be this one here as well uh normally I think it's a good idea to use different seeds uh on different renders should we try that and see what happens so we just change the seed on this guy here and see what happens so this is what happens I Chang the seed to just one ahead of the previous seed and we can see the results that we got here uh it doesn't damage the image changing the seed actually produces sometimes slightly better images so that's the only thing that I would change with the with this workflow and I would just change that seed just change the second seed to a different number you can of course change this to increment so that you get different uh different result each time you hit the Q button maybe we can go ahead and try a different prompt and what we can do is actually come here and uh maybe go ahead and uh choose the prompt for this image here what kind of results are we going to get now if you're using uh if you're using automatic 111 you can actually copy the generation data and uh there's a way of pasting it inside of automatic 1111 I can't actually show you right now because comy UI is open but uh what we'll do is just copy The Prompt and go into config you and go ahead and copy this guy here uh inside of U conf inside of automatic 1111 you can just copy manually the data and make sure that your setup has the same data as we've got here uh let's go ahead and paste okay and let's see what that does now hopefully this gives us a very similar looking image uh we we've got the woman showing up here but she's going to look slightly different because we're using a different seed and oh my God that is awesome that is awesome and then we'll get the second image my PC is running a bit slow because uh the graphics memory is just totally used up at the moment well that is pretty amazing that's a really nice image so what you can do here is if you want you might be able to increase the steps maybe 2 6 7 8 uh if you want slightly better quality but even with a small number of steps this uh particular this particular one produces really good results I'm very very happy with this now if we wanted to get exactly the same results as the previous one we would have to change copy the seed as well and change the clip skip uh let's go ahead and copy this bad boy and uh paste it oh my God and we're just going to paste it in the first one and see what happens it's slightly different and the reason why it's coming out slightly different is because there's no clip skip so there's a technique for creating a clip skip oh my God uh there's a technique for creating a clip skip which is uh this thing here this needs to be set to two uh and it's slightly different at the moment inside of uh comfy UI there's a technique for doing that but that will be another video or you can just sign up for the course and that will explain all the clip skipping and all that kind of stuff uh but for the time being that is how you use dream shaper inside of comy UI hopefully you'll give it a try and find it just as amazing as I did guys that is going to be it I think this uh turbo technique is one which uh I think is really going to produce a lot of very good easy to use very efficient um models and yeah guys I'll see you in the next [Music] [Applause] [Music] video

Channel: Pixovert

Views: 970

Keywords: DreamShaperXL Turbo, Stable Diffusion, AI Images, Comfy UI, AI Technology, Image Generation, Adversarial Diffusion, AI Art, Image Upscaling, Bicubic Scaling, AI Tools, Pixovert, AI Efficiency, AI Hardware, AI Software, AI Image Quality, Generative AI, AI Tutorials, AI Demonstrations, Graphic Design, AI Learning, Image Processing, Digital Art, Creative Technology, AI Innovations.

Id: vUQdD9kS7Uc

Length: 11min 19sec (679 seconds)

Published: Thu Feb 08 2024