SUPIR: Best Stable Diffusion Super Resolution Upscaler + full workflow.

just on the back of my tutorial teaching you how to create a magnific slre scaler windin developer Ki has released a super wrapper what is super it is a super resolution stable diffusion based upscaler that's going to take an sdxl model and spit back out and upscale image all in one node let's go have a look at it so to get started you're going to navigate to your config UI Windows portable install or your install you're going to get to custom node you're going to hit CMD when you're in that directory that's going to pop out a terminal you're going to go and copy the link from KJ repo then you're going to go back to your terminal window with the text copied you type git clone you paste it press enter and it's done so now you need to go and start your instance of config UI but when you do so you need to go in the log and check for those two lines go check if it says no module X formers it stands for Transformer it would not be uncommon if that was the case for you and if you're missing that well you need to install them so we reopen file explorer we're still in config UI we're going to get back to the root folder and we're going to find the subfolder called python embedded if you're using the portable install you're going to click this type CMD and it's going to open a terminal window within that folder just like last time go check out the description of this video I've put the text in there you copy paste that into the terminal and you press enter it's going to take its sweet time to install that's installing the Transformers okay so while I have you just keep in mind that doing so is not necessarily going to break your Cony but if it does you can always uninstall it it's not that hard and in general it's good to have them theoretically they give you a Boos in performance whether or not that's true really depends on your pie torch it's a little bit complicated well now at least you have them go in the console logs or while it boots and check that it does say using X forers cross attention and while I still have you it's also a good time to remind you that D on the bleeding edge tend to bleed a lot so unfortunately we're not quite yet finished we need the actual models for super it's super v0q they're available on a Google Drive which is linked from the Ki page currently I understand that they're a little bit under load because all the cool nerds are trying to download them if you can't get them go and find them on the hugging face link that I've included in the description of this video as usual and one last thing this thing is non-commercial guys so please be very careful and mindful of this and finally we just need to place them into the right folder so we can place them into the models folder uh we're going to drop that under say checkpoints there you go I'll just put them in the root folder but you can put them anywhere you want in a sub folder there all right so this is going to be very straightforward we're going to delete the default workflow we're going to need a load image and then we'll go and pick an image I'm going to select the same image from the previous tutorial so you have a reference point to compare it to what's best a more complex workflow or a single node well I suppose we'll see it depends next we're going to need an interrogator of some sort we're going to use moondream as usual and I'll ask it to tell me everything it knows about the image because my life depends on it you know the usual story go check out the other tutorial if you haven't checked because you can 100% blackmail these things or even bribe them tell it you're going to give it 200 bucks anyway oh yeah you absolutely need to choose CPU here you'll see why in a second and make sure you agree to remote code execution again make sure that you do so knowing what you're doing next uh we're going to add a text box we use the one for python gos but you can use any text box you like and after that we need the actual super upscaler and that's obviously a single node which is really nice we're going to convert the aom to an input that's obviously the positive prompt we're going to connect it and now we can resize it a little bit let me go down there because I can't read my model names so the first thing you need to select is a super model you have two to choose from one is called that's the one we just downloaded by the way yeah obviously that's super v0f and v0 Q what's the difference you ask well the v0 Q is the default training settings that associated with the white paper if you want to go check that out it's on the superior GitHub it's General it's pretty much the one you start with it's the best quality for most cases the second is what they call train with light degradation settings in other words it might work better in some use case and not so well in others personally I found that it was pretty good with humans actually and also with some type of scenery like Vehicles we're going to look at one in a second once we've chosen that next we need an sdxl model I said sdxl I didn't say 1.5 I didn't say lightning I didn't say turbo I said sdxl so here you can choose whatever you want uh but I would prefer personally epic reality because I had good results but I'm going to give a chance to Proteus because I heard it was pretty good so let's try that next we need to pick a seed guess which number I'm going to pick yeah the lucky seed and of course we want it fixed because we don't want we don't want to rerun this every Quee trust me uh then after that let's go and change the scale four is great it can go up to six eight I mean go crazy as much as vrm as you have steps need to be in line with what you would usually put in the case sampler so it needs to match your model parameters and same thing for the CFG scale when it comes to the text for the negative promp we're going to leave it to default there's an option for turn and noise turn is related to sigmas and it's also influencing noise noise is noise you can play with that if you want to insert details personally I played with it and I found that it just made the picture look worse for some reason so it's very likely that what you're going to see is that it's been optimized for certain models this is indeed the bleeding edge my friend for the color fix you can choose wavelet or Ada in because my image is not really warranting any color modification I'm going to leave it to none you definitely definitely want to keep the model loaded trust me and you definitely want to use til V Because unless you have 100 gig of RAM if you don't run TI it's going to it's going to crap out even if you're using a 512 x 512 image personally I had good results up to 1024 after that I ran out of vram and have 24 gig right so now we need to clean this up a bit this got stuck whatever it's a common bug with confy and we're going to put a water mark because I like to remember what I've done with my image my 100s is a image on my drive 3 months from now let's call it superar fun fact super in French means sigh aren't you glad you're watching this channel you're learning so much stuff it's amazing you know what you should give it a like anyways let's go and save the image um it's always good to save your images you can always refer to them later and an image comparer from RG3 we drag this over there we rename the slot to guess what superar and we rename the first slot click the button if you're not familiar with this and we call it reference so now we drag the image to the upscaler and then we drag it again to the comparer all the way there and we cue The Prompt and we can look at the results now obviously that was fast forwarded guys it can take a while uh but yeah on first impression without tweaking it without adding noise without messing with the settings I'm really happy with the outcome because that was just a node last time it took us an entire workflow to do this it's a one note solution is it perfect no I mean it's not redoing the faces you're going to need a face detailer obviously uh you also see that we have some very strange artifacts there I probably don't want to know what that is and we have some funny labels from the shopping card stuck in in the belt I I have no idea what it's doing anyways I think it's a good first try we definitely need to try some more model and different images all right so I think a little thing we could do here to improve the quality of things and to also help those at home that don't have a 49t is to resize the image before we pass it to the upscaler as then we size it down so there's a million ways to do this in the tutorial files that you can download from the link below you'll see I've done it a different way by counting megapixels here I give it an arbitrary size of 512 x 512 you can use whatever you feel is right another thing that we can do is add those groups here I'll put downscale and query and I'll create a second group which I'm going to call Super yep there you go and another thing we can do is we can add a group bypasser or group mutter in this case it's probably best to control whether or not we want to run the Superior upscaler or if we only want to run maybe the down scale first to see what it looks like uh make sure there's no mistakes and maybe add some stuff in this group as well again this is something that's in the full tutorial that's in the download file okay so that's what it looks like when it's all done and what I've done is just downscaled it to fit it in vram played with the settings a little bit uh played with the control Nets and so on but I wasn't happy with the results you see this uh compression artifacts I mean I know it's really hard to clean up but it's transforming into snow I was hoping it would just remove them entirely but it didn't so I started to play with it a [Music] lot [Music] okay so after many many many attempts what I discovered was yes there is a control net in there but if you set it at anything but one you really lose consistency on the image especially with things like vehicle and in addition it just wasn't possible to get rid of those mountains they just would appear every single time so what I ended up doing is I used a DRT Factor if that's the word which if you watch my other tutorials you know how to use it it's simply an upscale model that doesn't upscale anything it removes the artifacts it's really useful is it cheating perhaps does it work yes and that's what matters I'm very happy with the outcome so then I started to think what if we took an image that's not essentially a trap right not the worst image on Earth but instead if we use the proper image of the SC for example now again I picked this cat because I used it in the previous tutorial so you can refer to the previous video for comparing results spoiler alert I think that this version is slightly better and definitely because it's a single node there is no doubt that you get a major win in terms of time spent building the workflow I didn't have to use IP adapter the control Nets built in and it didn't mess up the eyes which I really like look at this output I think it's excellent single node it's probably the best of scaler when it comes to these pictures also use this one which is actually the magnific test image that I downloaded straight from their website and look at what it did on the skin it re establish all the pores Etc that being said just like the last tutorial it did not add details but I mean come on this is a 6X zoom and I think we can't complain I think this is a brilliant tool kiji once again delivered something beautiful it's constantly updated so be very mindful of that if you want to use it for example I had to reinstall it 3 hours ago because it was updated again but that's part and parcel of using brand new stuff the final image that I used as a trial was this parrot because I know how difficult it was to work with this parrot last time especially with things like unclip the branch would disappear Etc and the results were fantastic so I hope you enjoy this video if you want to see more there should be more on the screen right now see what I did there pretty clever isn't it and there's another video on the right which I'm sure you're going to enjoy because YouTube set so I'll see you next time guys take care

Channel: Stephan Tual

Views: 14,294

Keywords: AI, comfyUI, svd, sdxl, sd15, ipadapter, controlnet, animatediff, loras, models, checkpoints, tutorial, stable diffusion, sora, open ai, ArtificialIntelligence, MachineLearning, #technology, SUPIR, kijai, Stability.AI, super resolution, moondream, llava, vram, xFormers

Id: Q9y-7Nwj2ic

Length: 12min 3sec (723 seconds)

Published: Sat Mar 02 2024