SDXL 1.0 - Img2Img & Inpainting with SeargeSDXL

hello and welcome to More nerdy rodent geekery sdxl 1.0 is a great set of models for generating images and works brilliantly in comfy UI however creating all those workflows can be a bit overwhelming thankfully the community comes to the rescue and this new workflow combines many of your favorite things into a single compact interface text to image image to image and image in painting are all available via an operation mode selector plus there are also options for both high-res fix and upscaling the sdxl Laura is also supported along with a variety of upscaling models plus there are a variety of prompt inputs and also prompt styles to choose from of course you'll need comfy UI installed and running already so if you haven't done that as yet then do check out my previous video which covers the installation if you've done a standard install of comfy UI installation of the new nodes is a simple case of running that git clone command there and then just restarting comfy UI so let's do that there it is we'll go there so navigate to your comfy UI custom nodes directory okay we're going to come for UI and then we'll go into custom nodes and then we'll run that git clone command there it is as you can see I of course have already done that download so it says it already exists then if you want to update exactly like it says there you can just run its pull I have got some updates there we go so I am now running the very latest version if on the other hand you're a Windows beginner where copying and pasting git clone commands isn't your cup of tea then there is the alternative installation method here you will instead need to have an understanding of downloading and unzipping files as that's exactly what you'll need to do for this alternative method there is a link there provided to a civet AI page and also here in the releases you can see a link to the zip file you will need to download just as we did earlier with the git clone you will need to drop that extracted directory into your comfy UI custom nodes directory and restart comfy UI if you want to update this type of install then just repeat that entire process again so download the zip file unzip it and overwrite all the files in that existing directory like it says here you will also need a number of files if you haven't downloaded anything at all then you'll need to download all of those chances are of course you already have the base the refiner and the offset noise Laura when you're downloading those files do ensure you save them into the correct directory exactly like it says there for each of the files with the installation and downloads complete simply start your comfy UI and load the latest workflow here we are so I'm going to go to load going to go across to where I've got it installed which is there comfy UI in the custom nodes and then across to the workflows gonna load the latest one there which is version 3.4 so there it is if we scroll out a little bit here you'll see just exactly how big and complex this is there we go it's absolutely huge loads of things going on but all the spaghetti is down here it's all hidden away from you and all you really need to look at is this section user interface as you can see it's got two very large preview Windows there that you can look at if you zoom out a little bit but for the most part you just need to concentrate on this one little area here now first up you want to make sure your models are okay this bit here let's do a bit of a zoom in now one of the things you'll probably have to change is the vae model by default it's got a little hyphen in there but if we just do a click refresh we'll actually see it's called sdxl underscore vae so there's a tiny change on that from the file that you download by default next up I'd suggest looking at these miscellaneous parameters up the top here we've got a Laura strength and operation mode where we've got text to image image to image and impainting and also the prompt style there as you can see there are plenty of options to choose for that one as there are quite a lot of prompt Styles they do each have a description over on the GitHub page there so you can go through that and have a look but I would basically suggest having a play yourself to get the hang of what each of them do but here is a quick rundown so they're simple that's the first one there simple this is basically as it says the most simple one and this just uses the main prompt and the negative prompt if we scroll out a little bit here just so we can see these so there we've got the main prompt and the negative prompt so those are just the two used imprompt style simple next up on the list there we've got three prompts that's very much like simple but that also uses the secondary prompt as well and this is quite similar to other sdxl workflows if we go to subject Focus there the main and the secondary prompts are more important than the style prompt and then we've got another one there style Focus which is basically the opposite with the style prompts being more important next up we've got weighted weighted allows a balance of the main and the secondary prompts but also the style prompts and on top of that you can influence those with the style prompt power and negative style power up there as well the next option there is overlay with overlay the main and secondary prompts compete with the style prompt which is quite a weird one the next one you've got there is subject style this one uses the main and secondary positives with the style negatives so that's that's quite a strange one and then you've got the opposite of that again style subject so that uses the style positives and the main and secondary negatives further down the list you have a style only just like it says that uses this style and negative style prompts and then finally you've got weighted overlay and overlay weighted the opposite of that so weighted overlay is a mixture of both weighted and overlay where the main and secondary prompts are weighted and the negative and negative style prompts are overlaid and obviously overlay weighted is the opposite of that so loads and loads of things in there I suggest basically starting with simple and three prompts so simple we're just going to be using the green box there and the red box there or three simple prompts you'll be using that green one the sort of turquoisey one and also the red one as well yes that is an absolute load of things loads of styles but perhaps the main thing to understand there is that not all of those prompt boxes are used all the time it all depends on which prompt style you've got selected so if you're getting something a bit weird happening maybe you've got some prompts in there not quite coming out how you'd expect do check and make sure you've got the correct prompt style selected if you're looking to generate larger than normal images 1024 by 1024 is the default here you've got a couple of options you've got high-res fix parameters there so by default it is disabled you can just select enabled on that and that will give you a 1.5 Factor upscale with a denoise of 0.25 and 25 high-res fixed steps personally I think those are pretty decent defaults but obviously you can change anything you want to in there the other option for making larger images let's just disable that again for now is down here so here you've got off by default for performance reasons so this will use the upscalers so you just connect that one to there there you go and then you will get an upscale if you don't want the upscaling you can just click on that and delete and then those two are disconnected again the upscaling factor there is controlled in the advanced parameters by the upscale resolution factor which by default is two finally you've got the rest of the nodes which are pretty common things so you've got the the generation parameters there which you've got your image width and height steps guidance scale sampler name all that sort of stuff the advanced parameters you've got the base ratio so how much of the base model are you going to use before you use the refiner so there it's 80 base model and 20 refiner refiner strength and all sorts of other options you can change in there as well denoise perhaps being the most useful one an interesting one should I say as you'll need that when you're using image to image or in painting okay so with the interface Basics covered it's time to put it all together and generate some images for the first example here I'm going to use text to image I've got a simple prompt there and I'm going to set the Laura strength to 0.5 got no high-res fix for now and I have turned the upscaling off in simple mode just the main prompt there and the negative prompt there I used so let's fix this up I like to use English so we'll just repair both of those quickly and then the main prompt up there needs fixing two as obviously we don't want to generate the enemy there with that repaired I can now cue The Prompt and see what happens a voila one rat and a hat okay let's crank this up a notch Now by switching to image to image so they're operation mode image to image I'm also going to change the prompt style to three prompts as well just because I like bigger images let's also enable high-res fix now your Source image is over here on the left it's got that one by default you can click that little box to choose a file to upload but personally I just like dragging and dropping them in there we go there I've got my image now the other thing you'll need to do is change the image width and image height unless you want that cropped so let's just do that now with image to image the other thing to change is the denoise I'm going to put that up a little bit as well so that I have quite a significant change to my original image with the three prompt style you'll also want to use that secondary box as well so let's change these up a little bit and fix my prompts there we go so now I've got my main prompt my secondary prompt and an updated negative prompt let's cue that through and see what comes out there he is our little rat in a flat cap if we can open him up as well there we are there he is in his full HD Glory that is 1.5 times upscaled so as we can see there that's one seven two eight by one three four four pixels awesome so image to image works brilliantly there which means it's time for the next mode so there we go operation mode we're going to change that to in painting and now for in painting obviously you're going to need a mask as well which is where this second box comes in handy so we've dragged the same image in there now when you do a right click and open up in mask editor there you get a much bigger version of the thing to fix okay so let's do our little mask we'll paint around the eyes there and when you've done whatever masking you want to do just click save to node again you may want to change the denoise strength when you're doing in painting all right so I've also changed the prompt here as well because obviously I'm just doing that area so I want these as realistic human eyes because that would be perfectly normal okay let's cue that prompt and see what happens and there we go we've got a lovely fresh set of eyes if you want to see those as well in slightly larger preview versions there they are don't forget those are above there as well if you want if you if you fancy zooming out personally I like just dealing with that area there so there you go that's pretty cool isn't it three different modes lots of prompting options high res fix upscaling all in a single compact interface with hardly any spaghetti and if you thought that was fun then why not check out this next dirty rodent video

Channel: Nerdy Rodent

Views: 30,159

Keywords: nerdy rodent, ai, ai art, best ai tutorials, SeargeSDXL, SDXL 1.0, easy AI art tutorials, simple generative art guide, Stable Diffusion, ComfyUI, SDXL 1.0 ComfyUI help, SDXL guide, SDXL inpainting, SDXL img2img

Id: _Qi0Dgrz1TM

Length: 12min 48sec (768 seconds)

Published: Thu Aug 10 2023