1000% FASTER Stable Diffusion in ONE STEP!

I'm going to show you how to speed up your stable Fusion up to 10 times and all you need to do is just download one file and put it on your computer oh and just between you and me I have a fear of speed bumps but I'm slowly getting over them AI so these are live renders these are not sped up at all and this is 1024x 1024 SD XL images we're going to be downloading something called an LCM lore all the l are in the description below so you're going to go to this page you're going to download depending on what you want you have sdxl SD 1.5 and SSD 1B we're going to download first the sdxl and the 1.5 so go here's the sdxl one press files and verions you have the save tensors here just download that you're going to go into your stable Fusion folder models and then Laur here you're going to save them you're going to go back into that folder find the file that you saved we're going to rename that so you know which one it is so this is the LCM sdxl we're going to do the same for the 1.5 and the SSD 1B if you are using that so we downloaded the SD 1.5 we're renaming that sd15 the SSD 1B Laura is only going to be if you are using that specific model so if you're not sure you can skip that if you want to if you're using comy UI they're just going to be in the same models lauras hey guys remember to like And subscribe down below I do the research so you don't have to so I've loaded a 1.5 model up here I loaded my Laura LCM sd1 here if you don't have this you can go into settings here go down to user interface and scroll down and here quick settings list if you change this it's just going to see SD model checkpoint here you can go in here and add SD Laura and then you're going to have it same as me you're going to need to apply the settings and probably reload the UI then you can load the luras from up here now I have medieval Viking man War you're going to set this to eight steps so eight steps is all you're going to need for this I've loaded some of my Styles I have digital art fantasy art and I've also made a XYZ plot here so you're going to see the difference between the Samplers a lot of Samplers here then I have CFG values between 1 1 and a half and two I would say the max you're going to use here is two so use between 1 and two if you're using comy you can use the LCM sampler as of recording this video is not available in automatic 1111 but probably will be very soon so I'm generating now this in real time so here's a lot of images and we can see here if you can see the speed we're getting about 20 sometimes 27 10 iterations per second here so these are a lot of images uh I would say 1 2 3 4 5 6 7 7 11 30ish time 3 so these are about 100 images just popping out here now live in front of your eyes now bar in mind I have a very fast GPU I have the 4090 but even on a lowend GPU it's going to be much much faster and even on a Mac this is going to help tremendously I think Mac has some of the best gains uh so the grid is complete now I'm going to get it up here for you I'm just going to zoom in and a lot of these images are going to be messy and you don't want that so my favorite sample like the 2m Caris for example isn't working great with the LCM luras as you can see both from CFG SC 1 1 and A2 and two is not fantastic however some of the Samplers like Oiler a here for example are all looking beautiful in in all the cfgs let me show you some some sampl that do look good because most of these here you can see will not work if you're in comy the LCM sampler is going to look amazing but like I said not available in automatic 111 yet so just keep moving here a lot of these are terrible but we have here DD no not great uni PC here at CFG to2 is looking okayish but it's not fantastic and you can see like DPM fast here and some of the 3M On's SD are really terrible but we have the dpm2 here that's pretty good and the oiler a that's pretty good we're going to do a second one just with a photo realistic prompt here I'm going to take another model epic realism instead of the dream shaper one we're still using stab Fusion 1.5 and um you're probably going to see that for this model the results are going to vary a little bit we are again generating this in real time so about 100 images popping out here at lacing fast speeds total progress time here was about a minute for around 100 images and let's check here again for this model we have two M caras not fantastic Oiler a here looks pretty good Oiler pretty good DPM 2A this one here see if g 2 turned out to be um pretty good DPM plus plus SD okayish at a higher CFG most of these are not great similar as previously however more are working now with this uh shorter cinematic prompt as you can see some of these are actually looking pretty good DD IM around 1.5 to CFG is okay uni PC starts looking showing some promise but isn't really getting it so what's my recommendation then well if you are in automatic 111 try oil for now but make some test of your own and make sure you test with um the preferred model that you like to use so let's hop over to confin this is where it's really going to shine first of all we're going to go into the manager make sure that you update your comy that it's it's the latest version because we're going to need the the K sampler with LCM sampler here so we're going to change this to LCM we're going to add a Laura Laura loader we're going to get the one we want which was the LCM SD 1.5 in this case I'm just going to take this load the Lowa here in between and to drop the model here go and keep going into the K sampler I'm going to change the steps to eight and we're going to CH change the CFG to let's do 1.5 for now and we are generating so this is 1.5 and we did a quick generation here that looks much much better than the results that we got almost all the Samplers in automatic 1111 so whatever you generate now we're going to get some pretty good looking images as you can see the LCM sampler here is doing a much greater job so if you want these insane speeds without the need to do messing about with settings and other Samplers in automatic 1111 you're going to need to use confy UI for now we can just keep queuing these up you can see the speeds that we're getting here live I'm also going to show you this with an XL model here I'm going to change the Laura to the SD XL1 I'm going to change the size 1024 4 1024 here and I'm just going to queue this up so it takes a second to load the new model and now we should see live renders coming in here so these are live renders these are not sped up at all and this is 1024 by 1024 sdxl images pretty good if you ask me if you want to learn more about this check out the blog post I'm going to put that in the link description below so this is basically explaining how all this work works it's magic for most of us but for a lot of the AI researchers it's just what they do every day it says here here's an example the speed difference we're talking about generating a single 1024x 1024 image in on on an M1 Mac with sdxl base takes about a minute using the LCM Laura we get great results in just 6 seconds and they used four steps using a 4090 we get almost instant response less than 1 second this unlocks use of sdxl applications where real time events are a required ments so that's kind of cool and I know people in my Discord have been playing with this uh I know Kiku play with it to get um he used his face in a webcam and got real time renders from the webcam so that was kind of cool to be honest you can see comparison here from the number of of steps so between 1 1 2 3 4 5 6 7 eight steps here and you can see at start starting at step four five here the image really starts taking shape so you can play with even lower values than eight you can play with like four or five especially if you have a lowend GPU or an older computer try four steps five steps and you can also use this with animations so if you want to get some speeds into those anima animations make sure to try out the LCM lorus and talking about guidance scale here like we talked about the CFG so between one and two but they say if you have a CFG of one it effectively disables negative prompts if you have a guidance scale between one and two you can use the negative prompts and like they said we found the largest values don't work so remember that set your CFG to one or two but if you have one negative prompts are going to be out of the question they have some speed comparisons here sdx lur LCM four steps on the left here we have sdxl standard 25 steps on the right like we said on a Mac it's going to be hugely difference it's 6 seconds 6 and 1 half seconds versus a minute 280 TI 4 4.7 seconds versus 10 seconds 390 you start getting to to some real speeds you can get almost almost to a second here 7 Seconds that you had before and if you are not 49 like me you're getting subc speeds which I mean it was fast before but this is blazingly fast it's going to help with especially with my animate diff renders and you can even use this on a CPU now this Intel here is a is a quite beefy I9 it says they're using one out of 36 cores and it can still get to 29 seconds compared to the 219 seconds used previously this is good even for you potato pieces owners out there hey thanks for watching I hope you learned something today check out this video here and as always have a good one

Channel: Sebastian Kamph

Views: 52,354

Keywords:

Id: kVrd_VDAtUM

Length: 10min 10sec (610 seconds)

Published: Mon Nov 13 2023