Exploring ComfyUI Latent Noise Control with Unsampler!

changing president is extra easy when you just make some noise yes hello and welcome to More nerdy rodent geekery where today I'm exploring the power of this awesome node called un sampler which is from The Comfy UI noise custom node pack if you tried the old imageo image alternative test in automatic 1111 it allows for pretty similar changes while I didn't have much luck using that old imageo image test myself ampler certainly seems to do the job in comfy UI just as with the examples they provide for automatic 1111 there you can make changes such as hair color or expression and even more things too let's start off right at the very beginning with a workflow similar to the example one from the nodes web page the test here is to ensure that we get the same image out as we put in because then it means we're making the right sort of noise all I've done here to start with is tidy up that example workflow a little bit well there are some extra nodes underneath which I've got bypassed at the moment but I'll explain why those are needed soon here we've got the flower girl going in and that's going through the ampler along with a little bit of a prompt and we can see there we've got the noise that noise is then going into your normal kampler as the latent image and we get a pretty good representation of the original image out in some ways it's a little bit like when you use the vae for image to image but as its Noise We Now get some control as we can use the prompts as well another pretty cool thing is the output resolution if I open this image up in a new tab we can see there it's 1136 by 1440 Which is far above what stable diffusion 1.5 would typically handle without a hyra fix as you can see the output is almost identical we've got some fairly decent looking hands there the flowers look nice although there is a little little bit of writing on the handle there that hasn't come out properly if I zoom in here on the original you might be able to see it there CC Ingam whereas on this one it's just a little bit more blurry let's try and zoom in yeah you can't really read that can you so there are some finer details which are lost but otherwise it's a pretty similar image I've got a new image here which I'm scaling down for Speed and like the examples from the imageo image alternative test let's try making some changes all right we'll zoom out there so we can see first of all we've got the woman she's turned into noise and we've got the exact same image out as far as it can be but what happens if I change this checkpoint up here from the standard pruned EMA only stable diffusion 1.5 to something like oh I don't know let's pick epic realism oh what's happened there all I did was change the model and now the image doesn't look like the input at all well no worries because that's why the control Nets are there got those selected so let's un bypass them and run it through again this time you can see the output is definitely much more like the original although it is now quite bright on the plus side of course it isn't quite as bad as it was before okay so now for those output control prompts what is it that we want to change about that image I don't know how about let's give her um blonde hair and a red hat and we'll try something in the negative prompt as well to see if we can do something about that brightness if we run it through with that new prompt then hopefully we should get yes there we go she's got a red hat and blonde hair that's absolutely excellent I think it sort of feels a little bit like instruct picks to pick there in a way now another thing you can do is increase that guidance scale if we pop that up to something more like two then it should follow those prompts even more there it is very red hat very blond hair and I think that's a pretty decent output now as you may be aware the higher you raise that guidance scale the more like your prompt it will be and so the more this image will change let's give a little example here let's put that all the way up to six and then we can see how the image changes then okay quite a lot we've got a very red hat and lots and lots of red there very blonde hair so the more you increase that guidance scale the more like your prompt it will be in the example from the beginning I've got the output control prompt there as Morgan Freeman so I'm turning that old picture into a picture of Morgan Freeman got a quite a low weight on the depth and it's a fairly High one on the line art control net there and as you can see it's got the face there and the stars but even though that control net is quite High thanks to the high guidance scale there of 2.5 it's still come out as a very decent picture of Morgan Freeman Morgan Freeman 2024 okay so initializing the latent noise in this way looks to be fun but what else can you do with it well one thing I've been having fun with is turning paintings into photos and people into rodents all sorts of things like that for example we've got a classic artwork here I've got a very simple prompt which actually matches the original blonde hair and blue eyes but because I'm using that epic realism model when the image comes out we get this which I think is quite nice so there we've got a more photorealistic version of that painting for another classic artwork I've used The Prompt photo realism here with a guidance scale of 4.5 which is pretty high for this sort of thing and I think it's turned out quite nicely so she's got like the original hair but let's turn that into a photo realistic person nice and weird the sort of thing that I love and how about for this beautiful lady here with such a stunning smile well in the output control there I've got a rodent and there she is she has become rodent tized much more normal don't you think this node looks to be both powerful and fun as it gives you a little bit more control over the noise that you're putting into your Cas sampler so then you can be slightly creative with your prompts and yet still have something quite similar to the input image maybe this could somehow be used with animations or masks so many more things that I need to try do let me know down in the comments if you discover even more exciting ways to use this node as I'm sure it can be used in some really creative ways plus if you're looking for even more comfy UI inspiration then why not check out this next video

Channel: Nerdy Rodent

Views: 8,777

Keywords: unsampler, bnk unsampler, comfyui, comfui, comfyui unsampler, stable diffusion noise control, generative ai, ai image editing, Stable diffusion, Ai art, ComfyUI, ComfyUI tutorial

Id: qW1I7in1WL0

Length: 8min 12sec (492 seconds)

Published: Sun Dec 17 2023