Stability AI's Stable Cascade How Does It run On My Lowly 8GB 3060Ti?

well hello good people this is stable Cascade stability ai's latest model which is based off different architecture we'll get into that in a second here actually put in a prompt of an astronaut on an alien planet and we running this on a hugging face page so I'm not sure about the traffic but currently it's working pretty well so I'll leave all the links in the description below where you can check it out and uh yeah let's take a look at uh the results here here we go follows the prompt pretty well let's open this up in the prompt I have him levitating off the ground so aesthetically it looks pretty good I haven't compared it to sdxl but just you know based on what I know I wouldn't say it's better at this point so far from what I've seen but this model is supposed to be more efficient where you can run it on less steps that type of thing a few more examples here that I ran just a few minutes ago and so far it's pretty good I wouldn't say it's overly impressive but this is an early release mainly for research purposes and for non-commercial use now if we head over to stability ai's website you're going to see some information based on this model and it's actually based off a new architecture I'm not even going to attempt to pronounce this but you can see it here if you click on the link it brings you to the paper which you can check out yourself and again this is only for non-commercial use at this point although there was a Twitter post where emad responded that there will be a commercial version coming soon this part is super interesting where basically it says it's easy to train and fine-tune on consumer Hardware thanks to its three-stage approach here we see some example images and they look great I mean I couldn't tell you off the bat if it's better than sdxl but basically down here they talk about these various stages a lot of this stuff is over my head so for those of you that understand it you can go ahead and read it here and then there are some evaluations compared on prompt alignment and aesthetic quality and I was actually happy to see that they measured it against playground V2 which uh update will be coming soon for those of you that know I work for playground there's also comparisons from sdxl Turbo standard sdxl whatever this word is it's German and on Google translate it trans Ates to hot dog I don't know how accurate that is uh if you're German please let me know I'm not sure the reasoning behind the naming but maybe one of you can uh enlighten me and on this slide they do talk about the inference steps here where with sdxl playground V2 the inference steps that would take 50 steps Cascade can do it uh at 10 steps but I don't want to bore you with all those technical details I do want to see if I could run it locally on my 8 GB V Ram card I'm highly doubtful and I actually found out about this we're going to go into Pinocchio and if you don't know about Pinocchio I've got a video you can check out but basically you could install anything stable to fusion with this installer we're going to go to discover and you should see it on the very first page here stable Cascade we're going to click on that click download and the reason why I like Pinocchio let me just save this as stable Cascade I'm going to click download uh the reason why I like Pinocchio is because if you're not familiar with manual installation it takes care of all the installation for you git python any of those things I know can get pretty confusing so now I've been using Pinocchio to manage all my local platforms automatic 1111 comfy UI Focus whatever the case may be but it's not limited to stable diffusion there's a lot of other things that I'll show you in other videos so once you click on the stable Cascade icon you're going to get this option in Pinocchio to install Let's uh cross our fingers and see uh if it's actually going to run on my system as I said I have a 8 gbyte vram 360 TI currently I can't afford anything better at the moment running a ryzen 5800 x 32 GB of RAM I'm very skeptical that I could run it and if I can't I probably won't include this in the video now while that installs we go back to the hugging face page if you click on the Advan options you have very similar controls you've got your negative prompts prompts up here you can input those there let's zoom in a little bit more here you have an area for your seed your width your height number of images basically the same stuff but the differences you're going to see here is you have your prior inference steps decoder guidance scale here and decoder inference steps not that I'm famili with this technology but it kind of sounds like this was set at 20 and this was set at 10 by default I figure it's kind of like having a refiner stage where you have two diffusion models and the last step's the vae vae is usually the last step that converts the actual noise to pixels right so my gut feeling is that it runs you know for 30 steps the last 10 steps it does this decoder inference steps Mak sense to me again I could be totally wrong okay well it looks like the installation is done it says done here and we'll go ahead and click Start and fingers crossed now at this point it's going to download some dependencies to actually run the application so let's see how this works no errors so far so good now we have an open web UI option we're going to click on that and uh yeah we get the same interface but let's see if I can get it to work uh I'm I'm going to actually pop this out of Pinocchio let's pick the same prompt and uh same settings negative prompt I don't tend to use too many these days we'll leave it at 1024 by 1024 you can adjust it oh it only goes up to 1536 that's fine let's leave the defaults for now guidance scale at four prior inance steps 20 decoder inference steps at 10 let's run one image for now just to see how long it takes hey we're getting a preview it's generating that's a good thing so 5 minutes for an image with my GPU if you're in my position you might as well run it on the hugging face page to try it out I don't think it's worth waiting 5 minutes at this point but at least it runs on my system and I expect the open-source commercial version uh will be a lot more optimized and faster so I ran it another time just to double check the times and uh yeah so 5 minutes the first time the second time 4 minutes and 38 seconds anyway folks if you happen to try it out let me know in the comments below what your experiences are if you have a better GPU you'll probably have a better experience than me but again you can run this on hugging face too um yeah until next video I'll see you when I see you

Channel: Monzon Media

Views: 12,252

Keywords: stable cascade, stablity ai, ai image generation, text to image, ai image generator, how to run stable cascade pc, how to run stable cascade locall, pinokio, stable cascade install, how to install stable cascade

Id: FbJ6w4xaeBo

Length: 7min 20sec (440 seconds)

Published: Wed Feb 14 2024