ComfyUI img2img workflow using SDXL Turbo | AI art generator

hi there and welcome back so let's keep exploring comy UI so in the previous videos we show how to install confi the confi manager and also sdx sdxl TBO which is mindblowing so in this video we're going to work again with sdxl turbo but now I'm going to show you the image to image workflow pre previous we work in the text to image but before doing so let's recap a little bit so remember that you can install confy is very straightforward the interface is a little bit uh intimidating but the installation is is extremely easy we it's just download and that so and then configure something then you have on UI manager that strongly recommend recommend to install it and actually to use XD XL turbo un need to install this one because you need additional notes uh then also here remember that in here you have a link uh with some workflows Er specific workflow so you have here some examples that are really good that this is your starting point and from here you can complicate s uh how you can download this so here you can download the image or sometimes a Json file and just drag and drop here so this is our default workflow and from here we have we are expanding everything so starting from this one now let's say that I want to do the Excel TBO I already have my templates and in the video description you have the link for for instance the one that we did previously was this one text to image so it's very very similar to the previous one let me load the theault so look at that in default load your checkpoint your model this model needs your prompt okay positive and negative then you have in this case that we have text uh this we create an entty lat space but if you have an image you have a a a latent space where you read your image but the rest is pretty much the same and you can add prompts here the sampler in in the sampler you get you connect to the model with the prom and then you do all the magic here no to get your image then connect to the uh out encoder the decoder part of the of the encoder here the out encoder and convert latent space to uh pixel space face and preview your image easy peasy okay uh for those interested in the the just go know and is you're using stable diffusion like using here also you have the link and some explanations and some papers and so on but I don't want to go into details what I'm going into that now because I want to use it for something very specific but the normal user that wants to generate image of get more control un need to go into too many details so this is your standard one then uh the extended one for xdl T we look at that it's pretty much similar I changed now the position of the boxes a little bit but you need to rig your checkpoint remember to rig the checkpoint for sdl H sdxl TBO the the one was sp16 extension and you have your promt you connect it to the sampler but now the sampler is a custom sampler why is C custom sampler because you need to get access to these two options this is aular specific for SD turbo and this is why you need to have the latest version of confy UI but also the confy UI manager so if you go here you double click here and you search turbo and if you don't see this means that it's not updated your installation so you can go to the manager and update confy UI if after updating I'm not going to clip because I have it all up today if after updating you don't get it uh just download the latest version of of comfyi but that is very important to use SD turbo unique that so this is the only the only difference and the rest is pretty much the the same so something important also using uh TBO is that don't save the image is you save the image because everything that you're going to do is you have this it's going to be safe and it can take a lot of a space so just put a preview okay so it's when you get image preview and that so and to remind you to show you what we're going to do so now we have this and it's interesting in the website you have a few examples let me use what they use here and read what they have here now because it's very important they mention here for instance text to image all this models you're using the turbo model is train in this Dimension they claim that you can use larger no 10 24 but honestly my personal experience and don't get good results so use this and then you can use scaler which with be the next video also keep your CFG to one okay here the mention zero but probably this is in the python stuff but keep it to one and then if you are using image to image also you have this prod here the number of steps times the strange which is the the noise try to keep it one or more okay to get good results and as I mention you have a a prom here so let's use this prom here for the text so as you put it here Q then usually when you read the first time the library it will take some time and then everything goes and there you go we have the results and super fast as you can see so this is sdxl turbo in action so remember text to image you need to use this new interface then the CFG keep it in one and the number of steps honestly no more than five in my personal experience idea lot of benchmarking two is more than enough but this product this one times this one should be more than one to get the best results and remember always your Laten space 512 to get the best results so now let me go to this one so we're going to move to the next workflow image to image and I'm going to use this image okay already generated so this is the next example that you have here in their paper so I just copy and paste this this description this prom and I'm going to use this one to do image image to image so basically image to image we use this image and then we can connect it to a prom but also to the sampler and you can modify okay so you can get the any style or modify morph your image so it's kind of very cool to do that okay so this is it now let's move to the image to image uh image to image here and this is it already load the image so look at that the workflow is pretty much the same uh load your your safe tensor your library so remember to load sdxl turbo F fp16 and a reminder that to download that one just here in the website just look for that one then files and versions and here you don't Lo so as you saw these are big files then load this one and put it uh I put everything in automatic 1111 but it's up to you so you have your library so this is exactly the same as the default now look at here that we are not connecting anymore to the custom sampler we can use the normal sampler with no problem okay so here we're not taking uh an N Laten space our Laten space we have an image okay so remember you load the image and just to show you everything Laten image here let me go here and put there you need to create a a VI okay so this is to convert the lat the latent image the pixels into latent space and you put here and then you have low image and you can load your image okay so easy peasy this just knowing I think that's why it's very important to know how things are working and the steps and so on so here I'm loading the image so just click there and you can look for your image so usually you put it here in the input folder here you convert you go to the sampler and the sampler here works exactly the same way here you have again latent space to pixel space and off you go so let me click q and see that reading the image we have a prom this prom is going to give some weight to the image you use the same rules in the sample and then you go get this so basically what I'm doing is that taking this original image remember that we generate this image text to image SD T but you can use anything but I want to do since fast and now what I want to do is kind of keep the same St but morph the image to something else so see here now I'm morph it to a girl with glasses long hair and in my style so here you have my prompt okay so as easy as this so something that I have to mention that this is a very a big modification sometimes you can have problems in your IM and let me put it here so look at that they are quite different so this is much when I morphing from this to to this for sure you're going to have problems in the anatomy and namely the hands so look at this hand here it's a little bit weirdy hand and that can be a little bit tricky to control and some people tends to do to use the negative prone to control hands and so on so remember sdxl turbo doesn't read the negative prom but if you're using negative you can use negative Proms and so on so later we're going to talk about that but honestly I think that doesn't help much or as I already mentioned that was more something the initial uh models today models and they're going to get even better that stuff is not going to be a problem this negative is used for some Nails okay so we have this and to show you a little bit how to control what is happening uh did the noise level so this is very important and now we're the nois in the image so if you keep this Den noise in z .5 and let me redo this Prett much you are not going to change your starting image or you change it but not much so see that you have the glasses and so on so if your goal is to keep the same image but at some kind of of nois and lat in space kit this one 0.5 and then you can increase steps so as you go here in steps you will see the difference and and sounds okay but see that you increase to too many steps now you do the transition to this so you're giving a lot of weight to your prone if you keep it to one or two I like to use it too I want to remind you now again that I'm using xdl turbo if you use the standard one you can do the same but it will be a little bit more uh time consuming so this is a perfect case and invite you to play around with all these options and you will test the influence but in generally speaking this is the idea 0.5 you keep the same starting 0.6 you kind of give more noise more weight to your latent space here in the sampler and with the model and then you morve to this okay so now let's talk about this positive en color so actually it's not it's not compulsory to to have know this PR you want you can have an empty PR if I go and I generate my image nothing here and basically you are just mod in here sampling from your model using your sampler now to gather the information and you get this okay and this is then again you can play with these options and you will see that you will get different results but now you are conserving a lot of the style of your starting image but I want to show you the importance of the pos positive from here okay and that the negative is not don't use the negative not to erase like hands and stuff but Dr own faces and things like that so just to show you what happens so I go I want a gear and look at how how heavy the weight of this positive prom so whatever you put here immediately will have an influence okay gear animate style would have an influenc in your output image and you start to put it so see that these two keywords there enough to change everything then try to keep no do prom similar to your image or what you want so try to cap the characteristic that you have there so put keywords that can cure that so for instance let me add that I want uh glasses okay you put there you have the glasses and hopefully we'll add glasses there and there you go so the other part of the pr let me see it was G long black long hair so let me put it here black long hair cupron and then there you go so it start to see that you is it's morph but it's a little bit strange you have those extra hands and artifacts okay let me put here EX stand then okay so now it's sitting kind of in the tables I put it a standing you have it there much better and see that now all those extra hands disappear here and still you have one here but my point here is that it's better to fix those strange filters in the positive prone rather than putting negative prone because this negative PR they don't have a a heavy influence it's quite tricky to to solve those issues using negative PRS and sometime the negative PRS also they do change the style later we're going I'm going to show you not working with the turbo model because here is disabled that you will see that if you start to put those crazy proms on let me go here that I have one and I find it super funny those those those proms like yeah like stuff like mutated hands mutation mutilated uh extra lean fingers a little bit no funny the keywords and honestly they didn't have such a heavy influence so my ADV advice try to get it right in your positive prom so let me go here and what I want to show you in this image so I have here now this is by the way this is a Sher case as you see I work it a lot to understand every since how it's working so here uh this is the one that I using so let me go here and put it here look at that I'm putting just the keyword so you can put a complete sentence so here like a color for 2D blender anime portray whatever or you can put keyword by keyword I prefer to put keyword by keyword because then you can erase one and and keep pledging so let me go like this so look at that I'm going to put put anime portrait so anime is a very specific uh style so usually it will add a character so you have it there I mention about all these Styles remember that you can download librar is here and I recommend C AI is the place to go then you have hugging face also but here you have all those Styles and just to mention that there are many anime Styles and well the people that train those models kind of their fixed with this female characters it C my attention know that all female characters all around okay that is the people training that so any case so see that we have anim portray portray and then let me add here then I want a girl with black long hair and glasses CU there and this is this is what I usually do okay I like to work with the tool with super fast I can get my prom right here and then when you have a good prom you can go to the full model and use the prom that you already prepare here or you can use the image and retrain so look at that here now we have this and you have your extra hand here so that's the thing I say that you can go the full model and start those weird keywords in the negative but honestly it's quite difficult to to erase this one so it's better to get it in the right no in the positive PR so here basically what we can do for instance I can say St and maybe and let me put the auto que here so Auto Q this is also you get the auto que only when you install the the manager so see that now it's standing problema solve you have that extra hand then here probably you're going to have some extra fingers but it's still you can work it out with the positive spr but let's say that I want it there in the table okay so okay I didn't need it to to do that so so how can I fix that but probably I can say Crossing arms it didn't have much an influence but I say writing okay there you have closing hands that's well it was delay okay but see the cross crossing hands it have an influence now writing and there you go much much better so now I can argue that this hands is not probably physically possible probably you try to scratch your elbow your your shoulder you can do it but it's not very nice but it's not a problem you can get an idea that you can fix that with your prone it's extremely the weight of this prone is really really really large so use your positive prone to fix your image instead of focusing into the negative and let me show you an example probably have here of that negative prone let me see if I have of those two images uh I think now later when we work with the O scaling I will show you that so that is to to that negative PR you can use it to to fix array something added by the model when composing the image so here we have the image and this is our composition so the main composition is this character so putting a negative prone is not going to anything from your main character your main composition but maybe when creating this one uh reading here maybe we add let's say a clock here so you can say erase that clock that you added that so it's what is being added in the composition rather than what you have from from that composition or from what you have from this keyword so you put here a gear and then you say gear here yeah you cannot remove that because it's part of your composition okay so we have this is quite nice we have our image we work it out and then I invite you also to play with this so again my B remember this is uh the turbo the turbo you don't need to do many steps but here you can see the influence of this so as you keep increasing okay more steps the turbo doesn't work very well you keep adding more steps so in my personal experience five up to five you can get good good result but see that now that is probably yeah probably can be okay for you know I know but personally speaking the IDE and know of the turbo also is just few steps generate good images I would say something between two and four okay then the cfd also you can see the the influence so also look at that you cannot add a to a cfd too high because then you are going to oversaturate but you can fix this probably increasing sampling uh steps okay so it is a combination so here is a very nice tool that you can play with these combinations and you see the influence it is the same effect using the standard model so here super F super fast standard model is slower also the standard models the the the values are higher but the IDE is the same these combinations so generally speaking I would say keep your CFG one 1.1 0.9 this example shry Peck I'm putting 1.1 but I will say one is okay and this one two we have the cage and then we go here and you can play with the noise so look at that if I put c0.5 you keep your original image the main features and then start to add what you put here in in your PR so see that basically the the most important feature that is putting of this PR is the glasses okay you have it there did morph to the gear and as you get increasing you will see that 0.7 0.7 adds more noise and it gets farther and farther away from this style because basically you want to capture that style to morph to something like this and you can argue also that okay you can increase the steps and then you get closer but they did I know I go back to this concept they is to keep the steps low to get very fast outcomes 0.6 and this is what we have and there you go we have our image okay simply we start from text to image and let me reproduce and remember that this quite easy here so we go text to image I generated this prom here cat wizer using sdxl turbo then here is the preview as I say I try to to bypass always the safe image onless you want to save so remember right click and you have the option there okay then if you have the pr views you are happy you can save the image and from here I move to the uh image to image okay we have this one so image to image it's pretty much very easy you go here and read your image conver lat in space then the sampler which will mix the image with your model and your prompt and get your outcome and there you go uh let me disable the outo Q be careful with the AO queue always do not leave it on because otherwise it will be always keep running and it can give you problems it will save if you saving images it will save a lot of images but then it can take know a big chump of your video card memory and there you go we have the image so at this point remember all these images are 512 so very important this XD XL turbo was train 52 so keep that same resolution as you put larger you might get in some cases good results but personal experience uh is not the case like for me probably 70 or 80% of the time I don't get good results on cases I get good results so for instance I have this one let's do something funny because also we have uh this promp okay uh of the baby raccoon that is in the website know that you get from here they test this PR so we can do something instead of using this gar here let me put that PR so basically what I will do now more the cat to a baby raccoon in this specific style and let's see what happens and there you go so this is very good work okay I'm quite happy so see that very easy very fastly table and then when you move now to the full mobile Prem much you're going to say get this out will be more time consuming of course if I put 0.5 you will try to keep this sty let's see what happens okay it's a smoother now there you you're morphing in a very smooth way so let me go 0.4 let me put the out to there generate 0.1 so lowest it keeps very close to that then you go 0.2 okay 0.2 have that 0.3 and now you see that you start to morph and probably you can get all these images and then you can do the animations you want to see the morphing and whatever same effect will be for with your full model and 0.7 it is very heavy now this is very heavy so as you recall this is this will be the output from the text to image so avoid very high for me 0.6 remember that then you can play with CFG and so on so that that will have some influence and increase in here but yeah generally speaking Yeah that's what is happening so let me keep it 0.6 and there you go so finally to here let me put it again you have it there but I want to show you okay this is enabled see that is you leave it enabled every time it's going to start to do everything so I'm happy with this you can save this image and then you can do some up scaling of this image that that will be our next video so I just want to show you what will happen there because now that you have everything remember low resolution you have you can go high resolution and I have this workflow here and look at I'm reading exactly the same image so you can go always load you load the image so here you have image image loader look for the image and basically where just load the image then connect the image to and O scaler so here I'm scaling the image to uh in pixel space I'm not going to Laten space so I'm doubling the image here so I'm using this scaler if you don't have this scaler you will need to install that so in the manager you have here and you can look for O scale scale lers and you have okay and the wrong place okay it's model o scaler and you have many o scaler so you can install those there so I'm scaling two times so this is 512 now it will be s24 and using a full mod so this I don't need any model everything is in this scaler and then I'm proud you will see that this is a little bit different from previous that everything is compact so later I'm going to show you how to do this because now you can put in in in in order so look at that let me click here and see that convert to notes so basically what I did all those notes that probably you're already familiar you can convert everything so you select everything with control all the notes and you can right click group no and you give it a name so let me call it GT and there you go so this is how you can simplify your in interface so later I'm going to show you that then you have the group so everything is kind of know you post everything here is attached to this group and then you can this group you can just bypass the whole group or you can uh or you can turn it on so it's very helpful this stuff and also here you have this stuff the positive negative Proms and so that I separated okay so I will show you how to do that later so this is just to put since in order so likely sometimes when you download load and let me go back to civit AA uh sometimes when you download something from civit AA you're going to get so see that you have Works confy UI so you can look for that and you can get no works workflows here so when you L some workflows here you will see that are super complicated or you will have the impression that they're complicated but basically what you're doing what they're doing is that they're colling everything into a single no Tab and that's all okay so and let me run this case later we talk about in details what is happening here so let me go Q prom okay so load the image then this image has been O scale pixel space and now converted to latent space use a model and look at that I'm using this model here and it's going to do ano scaling of this image using this model you will expect as previously that now it's going to add some noise it's going to be slightly different because you are getting the model and is the concept is the same so see that 0.6 and the image is slightly different okay but keeping the same feature so let me go and soon in here so see that this is your image see that the resolution here you have some pixelation so when you do the O scaling pixel space look at that it now you have more sharpet here and to avoid those pixels it will add some information so this is the thing that sometimes at the beginning with the first scaler I didn't like because the information that it was adding it wasn't too good but now these scalers are getting really really good and to show you something that I have here an image and let me go here that this one here of scaler not talking about that and this is what is happening so basically you get this so if you want to increase uh then you have this pixelation but then the scaler will add some extra information here okay so this is this is this this is very very impressive very cool so this is what happening here so now I scale to two two times be careful also do not add very aggressive of scaler so here you have different scaler 4X ax and so on so if you put ax when you pass to latent space which is not compul okay you can stop here this is a big file it will use a lot of memory so be careful about that so my personal experience 2x is enough then you go here Laten space here also you can O scale that so will be the same concept you multiply by two it will start to get expensive so here I will put the same image s24 and then apply this model and this is a very good result very impressive and look at that the hand here we know friendly turbo that is a problem but look at that by just controlling my prom you can fix the issue of your hands and here now your prompts okay are positive know so just to show you these are positive and negative and they're enabl and now you can control this negative prone so look at that let me show you something I generate this image and this original image not this one if I do everything and I don't put a smile there this is what we have so I'm losing that biggest smile so if you want to keep that put there a smile and you're going to have then issues that you will identify here you can control and my advice try to fix any issue in the image in the positive Pro so it seems like extra fingers back Anatomy stuff like that person that is experience not going to work okay maybe you will be lucky enough that it's going to erase something but it's not going to to work and you will see when you start to do that it will change the style so look at that now my style is a little bit different so if I erase this cupron my style you will see the difference in in styles okay it didn't have that heavy influence this case but if you add many you will see and then you have embedding so this will be part of the next video what we're doing here so you have that template also so here I put some comments and talking about that how you if you want to get the original image get your your your denoise to 0.5 and so on so in the next video we play with that but there you go so you have the sdxl turbo which I recommend you just to to to adj to use that one to adjust your your prompt it is fantastic so that is how I proceed I do everything in Turbo adjust everything and then when you have your good prom or you have a good image move to the full model so that's all for this video thank you for your attention see you next time bye

Channel: Wolf Dynamics World

Views: 1,860

Keywords: AI, AI art, stable diffusion, ControlNet, generative art, ai art generator, run ai local, air art local, SDUI, install ai, instal stable diffusion, ComfyUI, ComfyUI manager, SDXL, SDXL turbo, img2img, txt2img, upscaling

Id: -t98Njqk9p0

Length: 34min 28sec (2068 seconds)

Published: Tue Dec 26 2023