Stable Cascade in ComfyUI Made Simple

in today's video I'm going to show you how to use the new stable Cascade model in comfy UI show you where to get the models where to install the models inside comfy UI and some tips and suggestions for experimentation so let's get [Music] going okay so the first thing that we need to do is we need to get the models and save them in the appropriate spot so you're going to head over to stability AI hugging faed repo for stable Cascade I'll include the link in the description below and we're going to need to download a couple different models and there's also some different options depending on your graphics card and what it can handle so uh we're going to be focusing mainly on Mid to upper level graphics cards for this workflow but I'll also walk through what you can do for lower memory graphics cards so you're going to need to get uh stage a and this is going to sort of function like a vae in this workflow you're also going to need to get a Stage B uh if you have 12 gabt or more on your video card I recommend uh just getting either Stage B or stage B16 I haven't really noticed any perceptible differences between Generations done with uh Stage B the 16 bit float model here versus the just stage so I recommend to save space and generation time uh to get these Flo 16 models so you're going to need that and you're also going to need the stage C if you've got a lower memory graphics card you're going to want to get the light versions of these models rather than the default versions of these models so you need to get those downloaded and then you also need to head into the text encoder folder and get the model. saafe tensors right here so you're going to download those and then we're going to head into our comfyi folder and make sure that we get them in the right spot you may also be using an alternative spot to store your models depending on how you have that set up you'll want to use the related directories that you have set up for yourself so we're going to head into the comfy UI Windows portable folder we're going to head into the models folder and we're going to move these in the appropriate spot so I actually do have my model stored in a separate folder um but I'll just walk through these with you and tell you what you're going to need to put yours so the first thing that you're going to want to do is put your stage a model into the VA folder so stage a in the VA folder then you're going to want to put your Stage B and Stage C models in in the unet folder and finally you're going to want to take that model that we downloaded here from the text encoder directory and place that in your clip folder and once that's all done you're going to want to head into comfy y if you have comfy y manager you're going to want to go there you're going to want to update comy UI and once that is done restart comy UI and then once you have all that done you should be ready to go and so I'll go ahead and include this workflow for people to use we'll just go ahead and walk through this real quick and then do a generation so you can see how it looks so first thing that we're doing is we're loading our Stage B and C models as well as the text encoder if you have multiple models in your clip folder and you already have a model named model. safe tensors you want to rename the model used for stable Cascade to anything else could be Cascade underscore model uh really anything that seems appropriate and then you'll load that in and then in type you'll go ahead and make sure you have stable Cascade selected uh right here this is this is something you can play around with and see how it affects your Generations two and three tend to be good values to experiment with see what kind of different results you get then of course we have our positive and negative prompt and this is a new node that was added which is why we needed to update comfy UI which allows for the creation of latence using the stable Cascade method which uh starts with a very very compressed generation and then basically takes that and decompresses it in a way that allows for Less memory usage and faster Generations while still getting good sharp quality in the final generation and all of these values are are taken from the stable Cascade repo just to match what they've done there but you can still experiment with some of these values and see how that affects your generation and then basically the uh the stage a is used as the vi in this workflow uh so let's go ahead and run it and see what kind of generations we get so we're just we got a happy panda holding a sign that says hello there we go so you can see that this model or this method rather does text pretty well you only have one that doesn't look so great uh and pretty good image quality definitely some flaws but the prompts I use were very simple no negative prompt no real specifications so there you go that's stable Cascade in comfy UI uh this is definitely still a work in progress so the models and Method will likely improve and the implementation in come for UI will likely change and improve over time as well so yeah have fun with it mess around there's definitely some interesting things about it some things that it does better than sdxl some things that it does not do quite as well but certainly the future of this method is promising and it'll be really interesting to see the fine tunes that come out over the next couple months and what people do with this so thanks for listening catch you [Music] soon

Channel: How Do?

Views: 10,078

Keywords:

Id: yAZgeWGEHHo

Length: 6min 56sec (416 seconds)

Published: Mon Feb 19 2024