ComfyUI : Hiding words in your images + fixing faces. QR Code ControlNet Pareidolia Hybrid Image

hey everybody Scott detwar here today we're going to play with the QR code monster but we're not going to make QR codes with it no we're not we're going to do some subliminal advertising uh you can you can use it to drive symbols or logos and there's that spiral that everybody sees or Reddit and so on but we're gonna actually drive the word shop into a shopping scene using the outfits and the environment which is pretty neat and uh what we're going to do is we're gonna do one additional Step at the end we're gonna fix all the faces automatically because unfortunately at this time this control net is a 2.0 and obviously we would like to do the faces with sdxl so we're going to do is we're going to load SD Excel later and then have it do all the detailing on any faces that appear in the scene automatically so this is a really great little workflow to kind of combine multiple checkpoints to use their strengths so let's get going so the first thing I'm going to do is obviously load a checkpoint here so we're going to load our 2.0 checkpoint in this case I'm going to use this one now you can use whatever 2.0 checkpoint you would like it's totally up to you and then obviously we need to do our standard thing so we're just going to convert this to a text here I'm not going to make this super fancy because we've done this a thousand times so our positive are negative and our vae is there now what we want to do is we want to introduce the control net so the control net comes after the conditioning here and we can double click and type in control net and we see we have a control net loader right here and we're going to load in we're going to swap this out for the QR code monster which I have right here and if we take that we drag that out we can see control net apply and control net apply Advanced we're going to stick with Advanced today because some other things we're going to want to do with this and then we'll just hook up our conditioning here so our positive and our negative that's uh almost the majority of what we need to do now this control net here is it's expecting this this control here obviously is expecting a QR code but we're not going to hit an eqr code we're going to hand it an image so we're going to load an image in so just go ahead and pop in an image that you want to use for this and I've created a bunch of these they're just 512 by 512 white background in Photoshop is pngs with logos and text on what have you and I created quite a few of these just playing around I did one with cats which was actually pretty fun because the word catch is very subtle and that's the part of this that I'm trying to do is the subtlety piece not the obvious swirls or other things like that but hey you do you but I think this is a lot of fun and this is the image we're going to use here to drive everything so we'll go ahead and hook that into our control net and it's going to use this as the base for everything there and let's go ahead and get this going into a case sampler just like we would normally so nothing special here just hook her all up like you would use an empty latent and however you could use image to image here if you wanted to bring another image in and convert it so again super flexible whatever we want to do here we're just going to throw an empty Layton on here and I'm gonna I'm gonna actually bump this up to 768 here we go I bumped it up to 768 it's just what I prefer in this case and we're almost ready to go here so we need to take this latent and convert it back into an image so we'll decode it that we're going to need this bae you obviously you've done this A Thousand Times by now but it's always good to see a rehash of it right and there we go it's actually a fairly straightforward there's not a lot to it need to load the correct the correct control net and then you're going to apply that control net to the conditioning in here it goes in the middle before we get to the sampler so keep that in mind some of these things aren't exactly common sensical and it's really interesting to see where the control net sits like a lot of people think that it's probably something it does with pixels but this is way before we even get to the point where we're going to go mathematically go get those pixels I think this is pretty cool so let's go ahead and give it a prompt here and I'm just going to use women shopping in downtown New York and for a negative now this is uh 2.0 so we do need a negative in here and this is the negative that was handed to me um I really don't care about these things by the way I think it's really interesting that the artist name signature and Watermark are oftentimes used here and the reason it does this is if you think about every piece of art you've ever seen like a museum and so on is almost always signed so it's kind of learned to sign its art and it doesn't use the signature of an artist it uses part of the prompt usually which is really interesting low res that doesn't make any sense to me poor details okay um I don't know why it's more important than anything else and worst quality do we really think we have models trained on worst quality and poor details uh I don't know but we'll leave those in there just for Giggles let's go ahead and uh before we before we commit let's make sure we're good good to go here we have 20 steps which is fine um all these things are fine let's just see what happens we can kind of see it bringing in the people and obviously because they uh face is a very small part of the total pixel resolution of the image uh when they're smaller in the frame they tend to get a little wonky as we can see here but if we zoom out we can see the word shop kind of in there it's not great uh but we're getting close and if you watch it when it starts to appear you can kind of see what it's doing so let's let's bump this up to like 1.4 be aware that when you use the strength here this can start to destroy the image so we got to be a little careful with this but this is really the only way to do it and you see it starts with the word shop and then it starts to kind of build around it which I think is really interesting there we go so if we zoom out we now see the word shop is a lot more obvious in there so yes it's uh something from the movie They Live right away where we are influencing your behavior subliminally so let's take and let's fix the faces now to fix the faces we're going to need to load in a completely different model and I want to do that with sdxl so we're going to go up here to the top and just load it in like we would normally in fact you could actually do this uh with a shortcut but I think just to kind of Keep It Sane let's just do it the old-fashioned way so we're just going to load in the sdxl checkpoint here I'm going to load in the refiner as well because the refiner uh actually is much better at faces than the base here we go and we will need prompts for all these things but again we're going to uh we're going to do this the cheap way in fact let's just do EXL we'll use the we'll use the refiner clip and we'll use the other one I'm gonna need two of these for each so we'll just do this real quick again you take your time and make your template once uh then you're good to go and you don't have to do this every time but for for today we're just gonna do it this way just kind of get this done I'm not going to worry again so much about prompts because we're just going to prompt for faces this time right so in here we're just gonna put beautiful and we don't need a negative prompt at all I'll put the same thing into refiner so that's good enough for me to make this a little easier we're going to use part of the impact pack which is the pipes uh so if we go in here we type in two detail we have a 2d Taylor pipe sdxl we're going to use that because it have all these things and we're just going to hook them up real quick okay now the detailer pipe also needs a couple more things down here it needs a bee box detector it's the the bounding box and we're going to use ultralytics for that so we're going to type in ultralytics again you should have had this installed when you installed the impact pack so it should be included and it's set to face there's a bunch that come with it face hand in person we're going to stick with face here I'm going to use the beatbox detector here put this over a little bit and then we're going to use we need a a Sam model as well if you drag that out there's load Sam model or Sam loader and the one that it comes with again is fine so all these things come with the impact pack there's really no reason to to worry about going and finding them somewhere they should come with it and I'm just moving this so the lines don't cross these nodes because if it does this I might get confused is what I'm looking at so I'm always kind of nudge it out of the way a little bit maybe I again we can sit and play with it all day long and let our OCD win but we're not going to do that Army no we're not okay now from here we can do is pull this detailer pipe out which contains all of that other information and we'll just do face detailer pipe here so this will go ahead and find each face and you can look at each face if you'd like to so it'll show you the cropped refined as it's working or the mask or all that kind of stuff I don't really care about these because it will show a display down here at the bottom if under your manager here you have your preview method set to this one it will show a preview under the Samplers here so we're going to go ahead and do that so we don't have to use any of these so I'm going to put like maybe let's do like 35 steps into the faces uh 50 denoise is fine and the rest of the defaults are pretty good to go we do want to force an impaint though in case it's too small of a face it would normally skip it so we want to go ahead and make sure that it gets caught here at least that's my understanding of how this works okay now we need to feed this an image now the great thing about this is once we have pixels and we're not dealing with conditioning anymore or latent that's come from another model once we get to pixels we can do anything we want to do so let's go ahead and drag this image up here and we'll just effectively start over right from this point here I'm not going to save this one anymore we're going to use this one as um our final so we'll put this here and then for this one I do want to see it though just for Giggles so we'll do the preview image and we'll put them kind of over here together something like this so that should do it let's let her rip here and see what happens kind of see the word shop here but um faces are going to be a little wonky you can see it going through and correcting each face will expand it s face detailer is very nice again it comes with the impact pack and if we don't like what we get and we think we want to do better you can always run this into another face detailer there's no reason you can't just chain these together and do whatever you want to do again that's the great thing about comfy is you can really kind of play with the whole workflow until you get something you really like and there we go so we have good looking faces ish and for the most part again it's gonna be random every time at least they're not floating or another problem again it's the 2.0 model I'm really hoping we have an sdxl version of this one pretty soon so there you go you have a workflow that takes whatever mask Type image you want to and applies it in a subtle way and you can do image to image if you just want to replace this empty latent with another image and then encode it and pass it in so you can do whatever you'd like to do in here but then we added the ability to fix the faces using sdxl and the refiner and this one node here from the impact pack to kind of make all that happen so again you can stack these if you want this one to kind of rough it in and then do another pass to kind of make sure you're getting all the details that's cool but you can do whatever you want to do just save this off and you're good to go so thanks everybody for watching I really appreciate everybody especially those people who are helping to support the channel once again couldn't do this without you guys and I really appreciate it everybody take care stay safe and I'll catch y'all next time

Channel: Scott Detweiler

Views: 5,359

Keywords: stable diffusion, stable diffusion xl, ComfyUI, stable diffusion tutorial, AI art, SDXL, comfy tutorial, sdxl tutorial, ai art tutorial, qr code, QR code monster, qr code comfy, qr code controlnet

Id: 6vc_a4aS19A

Length: 10min 59sec (659 seconds)

Published: Tue Sep 19 2023