ComfyUI - Let's play around with ComfyUI to generate some AI images

good greetings Rusty dog here welcome to my live stream where we're going to do a little bit more AI stuff and uh we're gonna use this new interface well new newer new-ish new to me uh which is called comfy UI um and I'm going to try and explain what we're looking at on the screen here it's all node based so we're gonna check out some uh images and what it can do and stuff like that it won't be too long of a stream and I'll maybe be able to do a gaming one a bit later today depending how I feel um I thought um hi Andy welcome I think it's just me and you at the moment is it I'm not quite sure yeah well guess what I just thought not on AI in reality I just saw a snail eating cat food yeah out in my garden I put down cat food for the couple of cats who don't seem to have any owners and there was a snail inside the food uh oh guys I've lost I didn't bring my cup hold on I think it's here bear with me all right sorry need I need my cup of tea need my cup of tea Okay so before now on the AI streams I've been using an interface called automatic 1111 which has been running all the stable diffusion based AI models to generate the images but there is a rival and it's this comfy UI this is completely different rather than having a ready-made interface for you with this one well you kind of start from uh a fresh interface like a blank page and you basically go along adding nodes as you want in order to get AI to create your image so following a tutorial on the net I have created along with the tutorial this um all these cables let's call them cables or wires there is an actual way to kind of manage these cables there is a cable management on this in the form of this thing oops in the form of one of these cable routers anyway um yeah so I basically built this up following the the tutorial and I was learning a few things along the way now I don't know if I could remember from scratch how to set this up you know the the flow I know the first thing is to load the checkpoint okay which is the um the model the actual AI model in this case as you can see I have uh well you might be able to see it on my app to resize this but I have epic realism loaded up and then from there I have I don't know what that blue line is I just this blue box I've only just seen that I have no idea what that is anyway um yeah I love that you can do this the mouse wheel and you can so if you can't see the text I can just scroll it right in and boom so from here we have three outputs and the first one is model and what's cool about this is if you don't know what to send it to it can give you some Clues hello game damage how are you sir now in this case how you doing man might be good uh in this case the model is plugging straight into the model in the sampler here but I also have another one k sampler so there's actually two note two output nodes coming out of this or output outputs one's purple one's white and they're both going into the model of the K sampler the reason we've got two is because we've got two output images I'll explain that in a while and then we have clip now this is uh for this text encoding now we've got two of those one going up to here uh you can color these boxes how you want by the way you can just go to colors and then just choose maybe purple you can color them as you want but it tends to be for positive and negative prompts tend to go with green but positive and maybe red for negative right so we've got our positive and negative prompt so we we've got the the AI model that goes into that gets plugged into a positive a negative prompt I haven't got any negative here at all the clip is going to the to both of those and then the model is going to the sampler and the vae goes to the vae node back here which then links into the image I'm not quite sure what the vae is I think it's probably somewhat to do with tone image toning I don't know um so yeah so there's those and then you have the prompts themselves so if we look at the prompts you'll see that we have one ignore this connector that's just me messing about don't actually need it uh that's just cable management and I didn't mean to do that so that's cable management basically so I can make the cables look better so if this was say getting in the if this cable was getting in the way of something or I couldn't see what it was going I can actually just unplug it and stick it into one of these sockets and then just route it somewhere else just to keep it out the way make it look tidier anyway so the positive prompt is going to the positives of both the Samplers the negative is going to the prompts of both the Samplers the negative right and then we have the latent image which is down here which is depicting the width and height of the image and how many you want in a batch so if I tell it to create an image of what I have here which says fairy tale landscape if I set that to four it's going to create four images of fairy tale Landscapes and then I can just pick the ones that I want uh Hey Von doodle networks yeah these are nodes yeah I'm used to these from LightWave but not yeah it's very similar to LightWave actually very very similar and so we've got this um thing now in the Samplers they both have you can't see them here because I changed them they both have these are widgets okay now they both have a widget called the seed the random seed now you can see that they dot they're not here and that's because I right clicked and did convert to input so I converted these seeds to inputs so now they've become inputs seed noise seeds and the reason for that is I didn't want to have a different randomized number here than here right because I want to produce two images that look like one came from the other and if I have two completely random seeds they're going to be two completely random different images and I don't want that so what I've done is we've created an integer here which is set to randomize so it will randomize this value every single time you click go but that same number will be sent to both Samplers so they'll both be rendering off the same randomized seed number are you following me and this latent upscale all that is as you can see uh it is it's normally you the the latent would come out of K sampler and go into here straight across to latent image but we've got this right in between right in the middle and as you can see it's upscaling by two so what it's doing is it's producing this image and then it's producing this one this is the one that's done with the case sampler advanced and this is the one that's done with the K sampler so what this one's doing is it's producing an image twice the size of this one so if this one is 9 10 by 512 then this one will be 1820 by 10 24 yes got my maths right so that's what this is doing so why do images well for this one it's kind of experimental really uh you can see on this one that this sampler only does 12 steps so the more steps the more detail perhaps let's say it's using the scheduler normal I'll get I'm not quite sure what that is but it's kind of like um ways of generating the image I guess now Charis I found out is the name of a scientist guy who actually works at Nvidia right now and Euler or e Oiler I think his name is pronounced Euler is a German scientist or physicist or something mathematician perhaps from the 15th century something like that anyway so it's using 12 steps to generate this fairy tale landscape and putting the image here but then this one is using the prompts and the seed and the image size all plugged into here to generate 30 steps but it's starting at step 12. so it's taking over from where this one left off and generating 18 more steps but with this time we're using the Kara scheduler and that's producing this image as you can see they are wildly different and so this is how this is looking so the I guess the reason that we I've got it set up like this is this will produce what you could call a more simplistic image based on the prompt and then this one's going to go into more detail so if we look at this and we go to open image and I bring this one up bear in mind this is a smaller size you'll see that the fairy tale landscape has a castle and a tree and a bridge and what looks like a a girl there and it's pretty nice on its own and if you upscale that as well to add some detail uh because upscaling does keep the detail very nicely as well it's ai ai it doesn't just do what Photoshop does which would be resize now this is resize plus AI so it's adding detail and keeping it nice however if we go to this version well this is the one with the 30 samples and this is from a two word prompt fairy tale landscape that is all I said and I got this now the girls disappeared but it's been replaced by little people down here I'm hoping you can you guys can see that so we've got little people down here we've got a much much more detailed castle look at the difference here I'll bring this one up the bridge has got more complex and the wow I'm hearing some loud noises outside I don't know what's going on anyway yeah so the tree is looking okay there but on here it's much more detailed and these are just with very simple prompts now here's the thing that I like about this I have managed using a symbolic link on my folders so I don't have to copy another 12.6 gigs across to have all of my automatic 1111 Strokes stable diffusion AI models in here as well I just created a symbolic link right so I can have them all here and if I change to a different one it's going to take longer to encode the image because it has to load the model get all the details from it it does something called setting up memory efficient cross attention I have no idea what that is but I guess it's just making things memory efficient and that will take around a minute before it produces the image but then after that once it's got that model loaded in and you want to run you want to run another image it's going to take at the moment with these prompts and with this setup around six to seven seconds that is way faster than the automatic 11 11 interface way faster I think although I'm only producing one image but we'll see now um Mr loot is with us hello Mr loot uh let's have a look I hope they make an open source Network UI can be used for all sorts of things well this is all open source all this is all this is free comfy UI automatically M11 you don't pay for anything you don't pay for any of this stuff now the cool thing with this is uh you can load and save different workflows I've called this one Rusty one um well hang on I'll change the model in a sec the model here will change um instantly but it's in the command prompt let me see if I can pop that on the screen for you guys too I'd like to pop that up if I can if I run to the live stream and go to a window on top and choose the command prompt that is a little bit too big that's full screen yeah we don't want we don't want it let's bring it down see if I can pop it in a space something like that and just take it away from the edge so in fact let me just move this I'm just going to pop out of Studio mode so I can move it live right I'll pop it down there so this is how you see it you're nerdly excited don't worry we'll do all the model changing I just want to do something a little bit different for now now I don't see the prompt uh the command prompt like you do anymore so I'm going to change the prompt here from fairy tale landscape to future I'm just going to do a very simple futuristic City I have no idea what it's going to do the futuristic cities it tends to do neon lights everywhere so using epic realism I'm not going to use any negative prompt just futuristic silly so what it's going to do is it's going to generate it's going to change that random number the moment I press Q prompt that random number is going to change and it's going to plug that number into both of the samplers this sample is going to run 12 steps and produce this image this one's going to run up to 30 steps and produce this image at twice the size so this will come up quicker and then we have to wait a bit longer for this one go you can see it running through here the green and now this one's running through the boxes light up and you get the progress there's the first image okay so this is the 12-step image plenty of Lights there kind of cool yeah so the refined image what's that going to look like well it's nearly there uh I think did that one says it take two seconds yeah this took two seconds and this took 27. it's just doing the vaed code and then we'll it'll pop it in there boom Okay so let's pop into this one so we've gone from here with the 12 steps up to yeah with the 30 steps and that is at 1 820 by 1024. now I can go bigger I can go full 2K on this and I don't have any memory issues at all let me just get rid of the thing you see it's blocking the way now give me a second okay so there's our less sampled image and there's a bit more it's the same prompt everything is the same the only thing is there's more AI steps so you might think if I change it to 50 it's going to get even more detail and then but it doesn't really work out that way because if you take that number up too much it's going to start distorting things and uh it's not gonna look too cool so let's get rid of these let's try something like oh I know what we can try I'm trying to zoom in here we go all right so I'm going to do a couple of prompts here so macro I want to macro photograph of Colony amounts that's it macro I haven't suggested whether it's an illustration a cartoon or whatever I think it's going to do a photo I hope so but if we don't want a cartoon we can put that here if we don't want load quality we can put that there too and we don't want a 3D render type of thing we don't want an illustration medium quality don't even want that you don't but I yeah sometimes you don't really need too many of these things I've seen some prompts where they're just ridiculous like it's just there's some keywords are repeated over and over and I feel like if the AI could interpret that in English it would be saying yeah Masterpiece realistic high quality yep High detail photo of a woman in library Masterpiece I detail and he's going yeah you said that before you know and I'm just thinking of the AI just complaining like why'd you have to keep saying Masterpiece realistic high quality I get it I get it and then in the negative clip it says low quality medium quality but do you need to tell it don't do low quality when you've already told it up here to do high quality anyway that's the that seems to be the trend so let's run that and we're going to first see how quick it can do this first image boom just done uh that didn't take long so wow that's with 12 steps that's a macro it's not really a colony however brutalist future city yeah is that the prompt we should do well futuristic city will be my first prompt is it possible to do 4K images yes but here's the thing though you can do it at 2K and then upscale by two to get it to 4K and even with uh even with low memory like four gigs of vram you get your gigabytes of RAM don't count here okay I mean they they probably will you've got CPU mode in Nvidia GPU mode here but what it's going to be using is your is your vram your graphics card memory but even on four or eight um gigs if you use the tanning method which I don't know if you can do it on this but if you use a tanning method where it's upscaling your images by just breaking it up into square tiles and doing one bit tile at a time that way you can do it where you might have to wait a while if you want an 8K image of this right if you want an 8K image you can just I can tile it I can you know tell it to do it that way but increase it with tiles increase the size by tiles and that way I don't worry about memory but I might have to wait a while but yeah and that's in automatic 11. how this one scales I don't know but I've so far I've told it to do native 2560 by 1440 images uh and I have not had any memory problems of course it just takes longer to render the image especially with uh you know 30 steps but still now this is 12 steps this is 30. I actually prefer in this case 12 steps because this coloration here on the back of the ant looks as I've I've taken macro photographs of ants I've no idea what the hell that's supposed to be and this is the color that they are this one doesn't have well it does have a little bit but look at the detail in the soil as well now we can change where they are here well it's not that it's the AI that doesn't know what an ant is it might depend on the AI model that I'm using if I switch from epic realism to something else it might do a better job plus you know you can put in a negative prompt you don't want to have too many legs too many arms too many whatever so the prompt can also change and also if I render another one with a different seed it might produce a better looking ant so here I can put High detail I think I don't need to put eye detail because as far as I'm concerned I got high detail already so if I if I choose a batch size of two it's going to produce two images two lower ones and two more sampled ones and these damages aren't being saved by the way but if you want to save let's say I want to add a save image instead of me going right there's a couple of ways I can do this right I can go right click add node image and then save image and there it is and then I just plug this into there and then I can right click here go title change my prefix name oops no that's the title of that's the title of the Box sorry it's that one I can change the prefix name here to whatever I want um and now it's going to save it so what will happen here is when this is rendered because I'm that means I'm not saving this one because I haven't put a safe prompt on there but I'm saving this one and then it'll come up with a Windows prompt saying where do you want to save the image okay so that was a way to add it you know doing it like that just by clicking right clicking and but there's another one if I just drag out a cable and leave go now it's giving me all the options that I can plug into for that cable I'm going to call it a cable okay and save image is one of them bang and it connects it how freaking cool is that it's the same as if like you you weren't sure um let me get rid of this if you weren't sure where to put latent for example you could drag out a cable and lead go and you can say right which one do you want to put it on latent upscale latent upscale buy which you've already got one of those that's what I need and you just do that one and it'll plug it in also if you have prompts like these that you don't need see there's nothing on here you need to change so you can right click and do collapse and it will just make it a little bit smaller make and make it a little bit tidier if you want to do it that way um I've seen people collapse this one as well but sometimes I like to change this you can uncollapse it obviously and this one do you always want it to be twice the size what if I said four times the size or three times let's say three times like if I chose my my native resolution which is 2560 by 1440. but who were to half that this would be 720 on this side and I believe 1 280 would be yeah so that'll be half the size so the this top image was going to be 1280 by 720 and this bottom image will be twice the size of my native resolution so we'll try that see how long it takes to do the image and we'll change what the image is I could say Colony events on a leaf or let's stick with nature I like nature so how about um what would you call all like a load of ants uh you know when they walk in a line I'm just going to put lots of ants climbing up a tree in the jungle that's it um yeah and I like that this thing stays wherever you can do whatever you want with this and this stays where it is but we'll cue that up and see how long it takes to do a 720 image and then a 1440 image good afternoon angry citizen oh you're welcome one doodle is that a column it would be going upwards wouldn't it uh okay that looks kind of cool those ants are huge let's just pretend that's a twig and not the actual bark of the tree those ants are massive although they do look kind of cool but this is the native size so that's the 720m jeez let's just pretend that's just a little bit of the tree and not the actual or it's a very thin tree a very small tree because those ants are huge maybe there are certain species of ant that are just like you know Mega ants so you can see we are now this green box is lit up so this is where it's at and this is the progress that is taking so it's not only it's not taking this long because this is how long it takes to double size the image new new no it is double sizing the image right and there's various ways of doing this as well I don't know what by slurp is I guess that's when two people are eating soup maybe I don't know so you got the upscale two and then it's also increasing the amount of steps to 30. from 12. oh so much what's that heart thing all right sorry I was looking what the heart was in the chat there I've seen that before what happens if you press it oh they're like Auto emojis okay that's a new thing yeah but I I like this image look at the detail on the tree and everything wow and this is just with 12 steps normally I think I would like in stay in automatic 11 11 I just set the steps to around 20 25 maybe even 28 up to 30. and then just say produce four Images based on that well this one we've got it doing two different stages which you can't do that in automatic 11 11 at the same time I don't believe and oh and also this isn't the be an all and end all guys because if you remember down here we selected a batch size of two so up here it's actually produced two images and you can see one of two and there's the other one again I didn't realize it it just yeah so what it's doing here is 30 steps twice the size plus it's doing two images this is why we're having to wait a little while so now we got this bugger on the girl so it it's actually rendered that one and that one I love the depth of field how it's gone blurry in the background that's really cool and that's what it that you would expect that on every macro shot anyway macro shots will blur the background there's no way out of it uh although I have produced macro photos where everything is in Focus but you need software for that and multiple images and a still subject and software controlling your camera but you can do it I've done it and it looks freaking impressive at the end anyway yeah so these ants I don't know it's got quite a long I don't know whether they fused or if they're just confused okay so it should be jumping over to V8 Ed code very soon does that light up while it's come yeah collapsed it does so it's doing the decode and then it'll pop the final image there yeah I've seen I've Elon Musk seems to be a Target loot um where I've seen videos of him saying things and you look at it and you believe it there was a video of him going around I'm assuming it's AI where he's saying if they if act if Call of Duty or Activision don't sort out oh I don't know he was whinging about something if they don't sort such and such out I'm just gonna buy the company I'm gonna buy Activision and fire everybody and yeah fire everybody that and get rid of get rid of um multiplayer matchmaking or something like that the one that the one that matches you with people based on your own skill that stuff and I commented on it saying how cool it is to have so much money and power that you can just if you don't like something in a game you can literally just buy the whole freaking developer and suck people but I think it was an AI video of him saying it and you can't tell I mean well you can if you if you look close Maybe all right so this is that's one image and then we've got the other one right so these are my full screen we had to wait a little while and yeah this sounds okay but some of them don't look quite like ants so it's not perfect there but hell oh and it also tells me on the command prompt that it took to generate this one was 300 and let's call it 340 seconds so 5 minutes 40 seconds to generate this at Nate at 2K yeah and you can see that the upscaling hasn't taken away any detail not even on the ants bodies because you know the AI sticks all that in for you it doesn't yeah it doesn't do anything like that you don't lose any quality but if if you were to take this image and upscale it in Photoshop you'd see a lot of quality the only time you could do it outside of here is a software like topaz gigapixel AI which would quite easily scale this up to twice the size and not lose quality but yeah it's nice to have it just done Insider here so I would say that we should reduce this down just a tad um for quickness probably go 512 here and then the 69 variate variant of this would be 910 oops how has it got 912 there is that what I typed in nope I typed in 910. okay it doesn't want to do that I think if you divide 512 by 9 and multiply it by 16 to get the 69 I think it's going to be 9 10. yeah I'm sticking with that anyway so you want me to change the thing we're going to change okay let's change this a little bit around I'd like to have a little Switcheroo thing here but we we don't have that where you switch it and it swaps the width and height round we don't have it so let's go 512 here and we'll just do line 10 here actually Let's Do It Let's do let's play with some standard values 768 because the reason why it's good to stick to these as well is because these models are trained at these resolutions so if you go a little bit larger let's say you want to render a person and you the height if I put 1024 here so it's a really long portrait you might get people with two heads or two bodies and stuff like that and it kind of kind of have some deformity sometimes so if you stick to these resolutions that it's trained at you don't get that so much if we switch to I like this one called humans because it does people but it doesn't normal people it doesn't like every time you generate a picture of a woman or a guy she doesn't have to be extremely beautiful or the guy doesn't have to be some rugged handsome dude they're just normal people from everyday life which I kind of like that I like that that we have that but there's one which I can't seem to get to called photon so there's Photon so now if I generate an image by the way these little circles here that's just me that's just my Firefox it has a an add-on for grammar and spell checking and stuff like that and this is just correcting me on my things so anyway yeah um using Photon we're going to do a portrait of somebody I can't guarantee that it will be safe for work um actually I'd probably try actually if I try and put this in I can so nude they could not save for work there we go we'll add that into the negative and we'll just do make it simple keep the prompt simple some of these um some of these AI models work better with some more simple prompts you know so sometimes you can actually make it do images of celebrities other times you have to just use certain special trained lawyers I don't know how to add a Laura into this yet I'm assuming it can be done though um what would it be yeah Loda Laura but that goes from model to model so does it come out of the checkpoint into the Laura out of the Laura and across if I pull that across there and leave go checkpoint loader sample simple checkpoint load is simple yeah if I add that that's one of these so it can plug into this but all my loras are in the automatic 11 11 folder I I didn't set up a symbolic link on my network drive like I did with the other one so I would have to copy those overall create symbolic links so I don't have any loras as such but yeah we can add those so this is good for adding detail to renders and making it do specific tasks within that render so photo of actually we want portrait don't we portrait this is photons pretty good for this but portrait of come on guys help me out here what do we want male female age background I mean what's in the background well portrait of let's just do 35 year old woman and leave it there nothing else and I'm gonna add bokeh because that's the background I want um okay now just because I put nude naked not safer work it doesn't mean it won't do lingerie and stuff like that we can handle it right can't we guys we can so what it's doing now if I get the window is it's loading it loaded the checkpoint Photon and now it's doing all of this memory efficient cross attention so I guess it's just analyzing I don't know I'm just making making it it's to do with resolutions here query Dimension is 1280 context Dimension is 768 which is the one I've specified so I don't know it's kind of just doing I don't know what he's doing oh and there's the image coming the first one looks a little bit like Minnie Driver and it's squeezed both onto the one prompt that's kind of cool I hope they are treated as separate images though yeah they are she looks a bit Mini driverish except she hasn't got freckles or this what color is a mini driver's eyes I think they're brown anyway and now it's generating the other image you can see here and it's doing two images yep and that's it I think it's done them right yeah it has Kaboom boom done took a minute and a half to do these two and these two not five minutes but I will get the batch size down to one but yeah so we have her what did we say 35 year old look at the I mean I just unexpected by anything else I just said I didn't specify hair eye color whether she was going to wear lipstick what kind of necklace she was going to have on if any jewelry at all what she's going to be wearing I just said bouquet which is what you see in the background these lights uh and now we've got the other I don't know which is here so let's put them in order so we've got image 48 image 49. right kind of cool it invents the clothes it does all of that loot yeah it depends on your resolution and settings and all that stuff I found this comfy AI to be very bleeding quick though so if we look now at the refined images if we go so this is the equivalent of 40 8 which is image 50. you can I mean the difference is absolutely noticeable open that oh two there so we have the 12-step version which looks okay in and of itself it looks fine nothing wrong with that the hair looks fairly detailed obviously it's it's there's a very shallow depth of field here to the point where even had even her neck is out of focus I can change that with the prompt I can tell it to use a different camera aperture so it's more so she's more in Focus this must be like a 1.4 or a 1.8 aperture to get her shoulder out of focus I mean her shoulder isn't far away from her face and it's already defocusing maybe that's to do with the bouquet I don't know now the these 30 sampled image is this one he's looking a bit older there I think but certainly more facial detail the hair detail is even more I think the fact that it's jumping up and down isn't helping there we go does she look older than 35 though one quarter What's one quarter does she look older than 35 the necklace has changed if you look here feeling a bit dizzy here right so the necklace has changed there just looking at the clothes as well yeah look even this just by sampling a bit further from 12 up to 30. we're getting different clothing even more detailed I would say and then we've got this lady that was at some up to step 12. and now on to steps 30. we don't have to have it at 30 we can have it at 20. so we're only adding another eight more steps on top it doesn't have to go all the way to 30. but look at her eyes here they look a little bit well they're large and she looks like it's all pupils the eyes here are much much more detailed man alive yeah look look here at the here we've got a looks like this got a bit missing here can we go in a bit more there we go I don't know if I've got them at the right sizes yeah I think we're just not in the right place it's hard to line up between tabs yeah and look at the hair just in this area here let me get rid of the uh the window on top here so if you look at the hair here just in this batch this area here and then you see here look at this the difference is just crazy so you might think well why bother with the image 12 or the step 12 image then you know why bother with that and my answer would be I haven't got the faintest idea oh yeah I guess here what you can do as well is you can change uh different Samplers different schedulers and see how it looks here like I could put these both down to 12. and then just use different samplings right I could use let's say that one and then this one I could start this one at Step Zero and step it up to 12 so it'll be the same as this but all or make them both 20. and then just change the sampler name so we'll get two of the same images only with different samplers so if we set this to I'm going to go to 25 and then I'm going to set this one to 25 but it has to start at zero I would say and we've got sampler deep DPM and schedule a normal this one is using let's put the schedule on this one to Caris as well to match the other one now these two are now matching it hasn't got this function but these two are matching except the sampler name is now different so now we can test two images one with the Euler or Euler and one with DPM p f p M thing yeah so so we're getting the first one first with dpmpp and now we're getting the second one which I think should be the same image but with a different sampler and this way you can test different samplers if you really wanted to you could set up more than one case sampler here you could do eight of them two or six of them to do you know six different samples but the cable management would be nuts I mean look at this whoa okay you know what we should do first is set them both to the same settings and see if we're getting the same images because that's wildly different obviously this one is getting twice the size I can't believe her neck is out of focus already I think we're going to fix that in the prompt I want to get rid of the bouquet and we're gonna get camera lens of f eight yeah f8 should be enough do I have to specify aperture f8 I don't know Maybe so let's set these two the same and see if we get the same result both 25 steps I'm just wondering whether the advanced one is producing a different image the only thing that's happening on the advanced one that's different that I can see is that it's being upscaled we can set this to one so it's effectively not doing anything the decodes here are both the same so both Euler Charis the end step and start step I don't know why that's these are different prompts but the denoise is set to one there or a hundred percent I guess you would say yeah these are two different okay let's just cue that and see what happens if we don't get the same image then it's not it's not really a scientific experiment oh it's gone black and white on us this time it's not really an experiment to change sampling there we go are they the same well they're close but they're not no they're not uh so this is the 12th sample well I wonder why it went black and white on us they're not quite the same okay they're damn close I mean the hair is a bit different she ain't 35 guys oh at least 45 I wonder why it went black and white I want to do another one I want to see if it's going to do black and white again I'm just wondering whether it's so much to do with with the prompt no it isn't okay okay these are radical differences yeah so this is the advanced one that's why I've got it changed I've changed it to final preview here and this one just says preview image wait that's another thing you can do as well and double click on this you want to add something and you you know you want to find the name like save image you can double click here and just type save and there it is that's just double clicking on the canvas here all of them work 20 years in direct sunlight she hasn't even got a tan though and it's given her like this red purpley coat on both but wildly different though right let's get rid of Mrs black and white there I'll open up this one it looks like vaguely she looks vaguely like somebody I know so even just using 20 steps I'm trying to think which one I prefer hmm interesting I actually thought these would be if I change this from sample Advanced to sample case sampler like this one is I I would certainly expect exactly the same image definitely oh God you want to upscale by 1.03 we can upscale by 1.5 or by or to three and two is good we could even plug this into how do we do this then oh I I could do another one of these upscale Laden I could do another one and plug it into here so they're both twice the size yeah we could do that I'm trying to figure out how to add something hello zonta uh it's not dangerous game I'm trying to figure out how to add a node of something that I could use mask load images and mask like the face-changing thing that I've got for my automatic 1111 oh the other thing I want to check out it might take a few minutes to test though guys I hope you okay with that is if I tell it to generate an image here with a specific model like Photon which is what we're using take these prompts run it through the K sampler let's forget the advanced one for now run it through the case sampler with these settings right or a a particular set of settings to get to this final image and then go back into automatic 11 11. same prompt same everything and see if it produces the same image using will capture the seed number as well to make sure it does exactly the same and also how long it would take to render that image how much if it's quicker slower but we can only do it from this one I think because some of these settings I don't have an automatic 11 so we wouldn't be able to really capitalize on them so I think I've not tried any of these other ones cune is in automatic 1111 some of these DPM ones are but they've got slightly different names I think I think they're the same ones yellow and yellow ancestral are both in in uh automatically level to put this one back down to 12. and also the CFG as well which is I'm going to tell you what CFG is because I keep forgetting um second classifier free guidance classifier free guidance basically the more higher the number the more it's going to stick rigidity to your prompt the Lesser number the more art artistic license you're giving the um the AI so let's get a let's get a guy let's do a guy uh uh same same age perhaps 35 year old man portrait let's make him handsome I don't want some ugly old goon uh and no offense to ugly Goons there's a place in the world where optic Dunes definitely right uh we kept these the same this wasn't the case before though was it we had normal on here before okay but in a minute I'm gonna set it to Karis because I can set that in automatic 11 11 as well so we're gonna run this let it let it all happen and then I'm going to capture the uh prompt which is Portrait of a 35 year old handsome man aperture too I'm gonna have to put those into a text file actually because I need to know the okay do a rich text file then your little tool oh don't do a rich text file if I wanted to reach a text file do a do a poor text file text document there we go I actually prefer the other one first one uh right so let's so we have gonna take these I think I can yep we can take that that's a positive then we've got we will come back to this as well got that that can be the negative and then we've got oh I only use 12 steps I probably should have gone higher on that one oh and this is yeah this is wrong okay this is wrong this should have been starting at 12. so go again well this image doesn't matter in this case I'd like to have used higher steps here but it's okay so steps 12 CFG is eight sampler is Euler or Euler because he's German it would be pronounced Euler I would think and Karis and the denoise is set to on I wouldn't do that in automatic 11 11 but okay we'll try it I don't normally set denoise up to one in 11 11 that's just way too much yeah this one's not working out too well add noise turn with lower leftover opening here what's happening with these dude here I mean he looks alright but why is he pointing at his neck and how come this image is looking better even there's a white line down here I've never seen that before this little guy this is like if they rebooted Happy Days he'd play the funds okay so we need to save out this this dude so let's I'm gonna just save the image I'm gonna call it that comfy ai60 there we go so we can leave the interface here but we're going to shut down the engine I think I've got all the info I need 512 by 768 yeah we can always come oh no no no no no no no no no no we need this yes this is very important the seed otherwise we're going to get a completely different person is automatic 11 11 going to produce this well we're going to shut down comfy UI completely now good night I'm just going to shrink this down a little and we're going to start up now the good thing about what's cool is with comfy AI is comfy UI when you launch the batch file to fire it all up once it's got to the stage where it's ready for you to use the browser it knows what your default browser is fires it up and takes you straight to that page this one one two seven zero zero zero one eight one eight eight takes you straight there so we're just firing up automatic 11 11. and we're going to get into that interface it's on a different port number so we can run it here um completely separate but it's not going to be it's not going to be ready yet it's just booting up in fact in fact in fact can I can I shall I yeah it's booting up there guys obviously if you have a better card than me uh better graphics card with more RAM you know with automatic 11 you can you can jump straight into those bigger um images natively instead of resizing okay so there it is running on local URL um so let's Twist Again just like we did last summer we'll put that down there get shot of it bye bingo okay this is more familiar isn't it so we're using a photon right and this is what we're using if we go into comfy AI using Photon V1 save tensors and we're using the same one up sure now the vae I'm going to set it to none I'm guessing that's what's going to do things I've added a couple of things to my in devices so actually what I can do is take this and just copy paste it in bingo and in the negative prompt we put in what we don't want to see now the sampling method we just we didn't use the ancestral one we just used the normal Euler one sampling steps well it was set really low wasn't it we set it to 12. and the CFG scale was eight now the denoise you don't have it here we don't have it here as a as a function if we wanted that we could maybe use Well Control net would introduce that but to do other things with that you can't just add denoise here but there is a noise C Delta which is not the same thing and a detailer this is my new way of fixing faces now I don't use restore faces when they're mangled I don't bother with that anymore um I think this a detailer is better it can do faces and hands and people and I don't know what media pipe is but media pipe face I don't know but yeah it's very good it replaces monkey faces with normal faces it looks to me like it just gets them from stock images I don't know but it works right uh the other thing was which is really important here really really important seed so let's just quickly grab that again and pop it into here cool so we need that size and we need a 768 CFG scale has jumped okay so let me just you were here that we've got everything right so cfga sampler is Euler then we have it says here carrots as well but then Charis is down here I can't have both of them so it looks like there is we are going to get differences here because we've got a denoise on comfy UI which we don't have on we've got this denoise factor which we don't have and I don't know what the default is on I don't know whether you know it's already set to one in in automatically level and we've also got the sampler name and the scheduler we don't have a scheduler in here all right everything else is the same now the question is can we get this guy or will it go anywhere close and how long will it take it's doing a preload thing I guess to get the accurate time you'd want to render it twice because it may be for the first time it may be having to create some it's going to have to do some scripting here let's just take a look at it I have learned that when it looks like it's not doing anything it is so leave it I it's gonna do I don't normally you can generate some decent images with these amount of steps but I don't normally go this low when I'm generating stuff but I've got so much more here as well all of this cool stuff and all of these add-ons down here like uh see I've got this new control net thing here which is pretty cool you can actually just get a canvas here scribble a drawing and it will do it and it'll render a real one or a cartoon one we've got a aspect ratio calculator which is powerful as f group um which is good for face swapping segmentation which is great for masking out and replacing parts of images and some of these scripts which was for doing upscaling and this XYZ plot is just awesome if you want to do comparisons over different settings XYZ prompt is just Supreme for that what the hell is taking so long here I know I can see it's just sat there doing whatever it's doing and the interface still says interrupt and Skip so it's definitely working I'm confident it'll kick in I'll know when it does it because my graphics card all the fans on my graphics card will start to spin I wonder if you I think you can have both comfy and automatic 11 running at the same time you can have them all loaded in at the same time I guess if you've got the ram but I wouldn't attempt to click on render on both at the same time no not unless you want a melted graphics card I wouldn't I wouldn't do that no not be bad what's happening here then a detailer isn't enabled uh there's no scripts running is it just taking that long to process the photon model Bae this time clip skip is set when you're generating one yep no issues as far as I can see let me guess come for UI didn't unload the memory it when it quit and now there's none left check up here everything went okay right yep no errors this one that we get couldn't find Bae that's normal I don't know why it does I think that's a bug uh it doesn't load the vae which is this thing up here so I you know even if this says that it is there up here if you want it to work I think you've got to set it to none and then go back in and set it to Bae and then it'll work I'm not going to give this too much longer guys because this normally doesn't take this long I don't know that's interrupted if we can not if I quit out of here yeah something's wrong here okay I'm gonna fire it up again this is the you get when you go live all right let's fire it all up all good I'm gonna have a new tab I haven't done an update for a few days on this troll net and group all going in Photon load it up proton here load it up fine E1 inference running an EPS protection well this is all normal it's all good textual inverse numbering oh okay I know what they are could have set up yeah done okay cool it's not gonna let's see if I can bring the prompts back okay cool it's got all the settings in place yes yes on the seed yet this is what's good you can quit this and then you go back in you just click this little tick and it'll put all your last settings back in I am going to deliberately switch out to a different model which because it's on my network drive will take around 30 seconds I'm going to switch back into photon just to make sure you know it's loaded I've never seen it just hang before it was probably when I when I abandoned it it was probably like 0.3 seconds from kicking off that would be uh so typical used to take five seconds to load these but because I'm now grabbing them from my network folder instead my network drive I it's taking like you know this amount of time this is a little bit too long it's normally 30 seconds I wonder why taking so much longer 30 to 40 seconds I would expect it to take okay 60. 58.1 seconds well now we've got to go back to photon it won't be quicker because it had it there before this will take just as long unless the orange chill mix model was just bigger could be that let's see how long this one takes so this is the downside to having your models on a network drive but the upside there you go 29.5 seconds but the upside is that I get to save 112 gig on my normal drive my local Drive plus if I wanted to use the same models again in comfy UI I'll have to copy them to that folder as well so that's 224 gig where this way oh here it goes it's just going boom all right let's get rid of the window now guys it's is this this all right he's tanned here is this the same dude in different clothes he's a scent well is it the same guy it's not the same render is it even though the values of the seed and the cfgs are all the same this is not the same and that's not a bad image for considering only used 12 steps that's pretty freaking good actually depth of field is just so good though the old one looks seven years younger this guy and the skin tones he's more tan there isn't he yeah Let's uh how long did that take by the way does it say oh shut up Alexa oh it was around the same six six seconds I think it's claiming this aperture f8 is that what it thinks this is at I don't think so f8 would have put his coat in focus and definitely would have put his ears in focus so I'm gonna run this again see how long it takes oh it is pretty quick isn't it that was four seconds yeah that's pretty quick he's starting to look a bit Ross from friends now and everything's in focus so I guess it just didn't like aperture f8 he was happy with me just saying f8 what's with the white bar at the top yeah four seconds so it is quick it is really quick it can produce Ross from friends in four seconds oops if I just shut it down um hold on I didn't mean to do that not that it matters we'll jump back into the comfy in a little sec I'm trying to think if there was something what is it on here that I had new oh yeah so I don't know how we can do this to demonstrate it but if I do an image let me do two images because I'm going to need a couple of tries on this one up the sampling to about 25. happy with that Euler a is fast it's quicker than any of the DPMS so I want uh photo of a soldier that's it full body oops full View in other words don't give me a head-only shot okay the reason is because when if you have a head only you don't need to use restore faces it's normally quality enough but if you have the whole body and the face is kind of small in the image sometimes it gets a bit distorted and you need to use restore faces but restore faces isn't that great so let's just render one and what we're looking for here is a bad face okay it's it's picked the era you know I've been doing uh soldiers like in dressed in stealth Gear with uh would you call it rifles and things like that so forget this image all right it looks like action man doesn't look realistic at all that can be changed with prompts of course by telling it I watch it realistic so if we come down to batch one and recycle the seed and generate again we should get this exact same guy cool awesome now let's I don't know if it's going to be better or worse it really can be better but concentrate here on the face and what we want to do now is do the exact same image again but this time we're going to use a detailer and we're going to use I'll keep it on this one face yellow of 8N and I'm going to copy positive prompt down to there and the negative which is no crop down to there and now enable and what a detail is going to do is it's going to detail the face apparently so to speak same image but with the additional detailer now it's concentrating on the face see oh has it improved it I should have kept the other one shouldn't I I don't know does it look better he's in a weird stance though but if we change oh I've just done photo so I've just done photo ah so let's do photo and now let's add some stuff like High detail let's add all the crap that people like to put in Masterpiece DSLR and then I'm going to pick a Nikon d850 that seems to be the popular one or the Fuji film but I can't remember the brand of that one uh the the model number so Nikon d850 and we want a photo realistic here we want uh no crop we want low quality medium quality which I think some of these a lot of these are redundant uh uh medium quality what's the other one oh uh worst quality they're kind of you know if it's AI it should know what you thought what you want from local you shouldn't need a low quality and worst quality really because if it's told not to do low quality it sure is not going to say well I'll do worst instead then when you've asked for high detail and a photo anyway so worst quality and then we want um I haven't specified any hands or anything like that you know fixing but it seems to have done a decent job there the worst quality cartoon uh another one there's another one that's not coming to mind right now um pretty low quality car too normally you would put like you know not too many fingers not too few fingers but these this render this Soldier seems to be okay so we'll just add those and see if it makes something a bit more realistic and just to go with this we'll go to the Laura's go to add detail okay cool actually not cool because it put it in the clicked it six million okay so Laura add details so it's going to just add overall detail and we'll now copy so down here but even though I'm copying these down to a detailer for the first render I'm gonna turn it off so it's not none of this is going to be used go I've kept the seeds so it should be producing hopefully a better version of this image and it isn't same seed totally different image and yeah Ma you do have a huge weapon don't you alright let's go again I can actually change his pose as well cool can I do it here though huh oh no like he's got this action action man kind of pose but I can't remember is if I can do it straight from this I think so if I go into open pose what would be a pose for a soldier hands on hips I guess so remember this he's got meat here as well so Hands-On hips would produce a slightly dropped shoulder right I mean I'm only guessing unreal six right so that I think is going to be so send to text to image that's not what I want to do though is it Julie I just want to save the Json file wait a minute if you send it to text image no it has to be done with before we do that let's just run the same image again but with with the a detailer this time on the face it is concentrating on the base there's another one you can set it up you can do it for hands there but I found that if you do that it compromises the first one some for some reason I don't know why see if you can see the face change in real he's looking a bit younger there and he's looking a bit less he looked a bit more Asian last time now not so much okay uh right let's begin a detail for the moment now I think control net we have to use open pose but we don't need a pre-processor because we're already going to have the pose so if we go open pose here and I think do I use control now is more important or can or balanced interesting also I don't think open pose looks like it can send to you can't send it to control net so we're going to save it out as a PNG and now I'm going to go into here I'm going to load it up pretty nice graphics that's not me but now I'm Gonna Leave It balanced for the moment and see what it does uh there's no D noise is there right so control type so we've got control net enabled and it's using the open pose model using that drawing with the weights we want a hundred percent here so it should now re-render this guy with that pose right right guys I don't know I think so is he still a soldier though this is a totally different Soldier oh okay so he put him in that position with his hand on hips and now he's just got him holding the gun in that pose okay that works for me this let's just say it's his jacket I don't know what that is in the background how about we do something a bit wacky here guys just for fun before I go back to comfy there's other pose um apps which I did show in a previous uh stream where you can have these in 3D and rotate it around and stuff this is kind of flat but if I save that out as a PNG and overwrite the other one come back in I don't know I've just got his leg in there it's probably gonna rather than just have him stand in there with his leg in the air like terminate the two it's probably gonna put his leg on a box or something like that so if we Crush that out and just do drop image you're looking for it I don't know what the hell it's gonna do now well it's going to put his foot on something or make him you're just gonna leave it dangling in the air really yeah it is uh awesome oh but what if I say foot on box oops box foot on box just add that to the prompt go surely it it knows to put the Box come on AI be smart I think it's doing it we'll see isn't AI great a cardboard box but look at that it's got It's take just taking this pose that I we just did very simply and bang I pose the guy what a poser that's a weird weapon but it works if you want to specify the type of soldier I guess Russian American Korean whatever you want to specify those you can I'm sure but it actually knew to put that foot on the box AI is quite clever I do it's clever but so far I don't know if you can like do this kind of stuff in company AI because this is some powerful stuff and you can add more control net units to do different things um it's so bleed and powerful and if you had an image here that you wanted to use control net to to manipulate your resultant image and they were a different like let's say the image here was a different resolution to this one maybe one's portrait one's landscape or one is four three one sixty nine blah blah blah then you can just click on pixel perfect and control net will fix all that for you now the other good thing with control net if we clear everything one last thing now because I can't draw for turkey and because the pen nib is so big um I'm gonna try on my own and then I think somewhere I still have it if I don't have it I'll download it I'll download one show you can't see it there I may have actually deleted it just looking through some images guys yeah I think I'd yeah I deleted it but let's give it a go uh I'm not guaranteeing this is gonna work get rid of that seed I'm gonna make no guarantees here so we're gonna change the model to scribble I think scribble the hell uh do I have to do it here with a preprocessor this isn't right then nope this isn't right because I need a canvas so it must be image to image that does it or is it in paint we've got sketch but sketch isn't what I'm after really no it's no no no no no no we do not want to drop an image okay the Bell has rung the clouds have cleared right so what we do is we click on this and create a new canvas and we want 512 by 512. let's go 768 by 768 make it Square on this one I would love to have this bigger let's see if we can shrink this down to the smallest point that is way too small I can't draw guys and especially in a canvas this size what the fleek there is an app or an extension you can you can add on to this which will which you can use to extract this to full screen and blah blah blah this is going to be really bad but here we go this is when you need a freaking what you call those things with a pen pad those pad things the drawing pad thing it's supposed to be a fish give me a break this is ridiculous I can't control the mouse so all I've got to do here is just say fish and I'm expecting it to draw one that looks like this that would be bad let's match the canvas size come on that's so bad I'm gonna have to download some some art of my own uh or off Google anyway not my own art is no good but if I if I download some art see not even close not even close so let's set it to control net is more important yeah very nice thanks generate again he's a fat it's a dune Navigator oh the guy yeah right this is a bit more close to what mine was I forgot to put the fin on the top as well this will do line drawings or scribbles that are because you've got sketch as well which is think a little bit similar um it did a lot better job than what mine was and we nearly found Nemo there so if we go into goggles here if I go to GI and I just go to simple sketch and we look for something that is a sketch of something but like this is a slice of pizza UFO that's actually that would actually be a kind of cool one but oh okay this one right so it's a horse its back legs are right out there and he's looking at us okay uh uh same image and we'll just call it how to draw a horse that's what it's called let's see if we can load that up here in control net let's go to desktop because that's what I saved how to draw a horse and I can't put fish now of course that's it that's all I'm gonna say but it should draw it should render the horse and it should kind of look like that sketch yeah man see the see the back legs out and the tail down oh yes come on baby now what I could have specified was that I wanted a photo of a horse not cut some kind of artistic illustration there you go so you can have like your kids just sketch something fairly decent pop it into control net and it will actually do a real more real version so if I said photo but not only that professional photo now I don't know any particular animal photographers I can't specify artist names here but let's regenerate that again but the thing here is as well if these are other these are features these are things I can do here in automatic 1111 that I cannot do in company UI yet I know comfy UI has got a lot of power do stuff especially with the nodes you've got so much control when you're using nodes but yeah so it's kind of put the tail over there we'd need to do a few images I could have specified a background like in a horse in a field but you can clearly see it's based on that sketch no two ways about it come on I'm going to try this one and then we'll switch back over to save link oh no don't want to save a link do we it was a chain maybe but save like so that is let's just call it uh ping I think that went onto my desktop what if you what if you uh upload a picture of a penguin on the sketch and then in the prompt you put a horse how is that gonna work and how is it going to do a professional photo from this well I guess the only way to find out is to ask it to do it no what is this going to be like from Pixar or something yeah it you can't do a photo of a penguin like this because Penguins don't actually really look like this it doesn't have the other leg or foot here on this side well this is it to put there I don't think the AIS realized that so it's it's gotten away with it by burying it in the snow and it's got a weird flipper thing coming out here but look at that face we could generate obviously four images of this until it produces one that doesn't look a bit weird isn't this cool though I I could easily just up here just specify a cartoon or like a in the style of Pixar or something like that it's not going to realize that this um but this down here is it's on the foot I don't think it's ever going to realize that penguin walking in the snow now let's just say penguin walking in the snow right and then just put in the style of Pixar right so there's the let's get oh I forgot to invert oops my bad yes the image has to be black on white so I should have set it to invert this is what the image should be if if you've got a a white background with black pen you have to use the invert preprocessor anyway let's do now that we've got that set let's generate the penguin walking in the snow in the style of Pixar I was hoping that using walking in the snow would make it realize that this was its left foot but I don't know is this Pixar guys spaceship Interiors uh they'd have to hire me first you know if the AI knew what Elite dangerous was it could probably let's have a play with that yeah I don't think it knows what it was to be honest right so this is what the AI should be working off of has he got a little furry hat on it definitely pixarish isn't it and all that from a drawing that or a sketch that somebody made but yeah you do need to invert it has to be white on black yeah white as if it's like a Blackboard like you've chalked it on a Blackboard that's what it has to be but the horse still came out okay even though I didn't invert now I don't know whether I don't know if it knows let's have a look asp no way I mean just no way ASP Explorer and then I'm just going to put Elite dangerous oh no I'm actually gonna tell it from the game Elite dangerous Okay so we need to obviously not make the ASP Explorer if it's even gonna do one it can't look like a penguin so that has to go so we've disabled that there's nothing else running here seed is on minus one put the batch I don't think it's gonna do it ASP Explorer spaceship from the game Elite dangerous I'm going to just say no chance no way it does this it can do stuff like from movies but a specific Niche game like Elite dangerous I don't think so no no no not gonna happen buddy not even close oh I can do a ship interior or on the bridge of a ship but it won't necessarily match what the Anaconda would be or any such thing like that it clearly hasn't got a clue what an aspect Explorer spaceship is I mean that's more creative anything but no no no no no no no not not at all so let me take out from the game and I'm just gonna put Elite dangerous there's no way it can know this it will know things I would think it would know stuff like the Starship Enterprise and stuff like that that's that's a massive franchise and it would know that it looks better than the actual asp well there's a lot of ships in Elite dangerous that look better than the asp and we could also add the uh the ad detailer on on this as well yeah this is not good however if I was to suggest TIE fighter Star Wars maybe right I mean is there a bigger franchise than Star Wars Star Trek Star Wars they're huge it's not going to show me two two ties in a tailor shop having a fight early well something's something's Brewing I don't know if it's a TIE fighter but something's Brewing no it's got the idea but no definitely not it's certainly A variation on a theme no question about that but no so it's kind of like well it's almost as if it has a very vague memory I guess we should maybe you could maybe use different um different models here like perhaps cyber realistic would do a better job um quickly dry it I guess oh Babylon five yeah I don't know I haven't watched that let's try let's try with this and see what happens TIE fighter should be it is it's an abbreviation isn't it so it should be should be capitalized although I don't think it makes any difference t-i-e the Imperial Eagle okay so cyber realistic same thing same prompt Google and I might want to add this back in as well at some nothing I change now is affecting this this is already on the road doing its stuff oh okay well you can see how it's closer but it's completely different I mean it's there thereabouts but need some work I tell you what I do have though I have Laura's which those those little trained prompts oras which can do people in all the different Star Trek uniforms you know so you could do a guy in a Star Trek uniform that looks kind of like you but doesn't have your face but just got the similar hair and then you can replace your face with that and then boom that you got a picture of yourself in a Star Trek uniform I'm gonna render one one more so yeah I got all that got that recently I'm not sure if I've actually oh yeah I have I have passed them over to here oh my God what is this made of like big thick pieces of wood I it's creative it's like it's an alternative Universe Star Wars oh that's with the vae added and the add detail I think it's just gone a little bit too far now do we have this as a Laura I don't know oh TNG Star Trek TNT right so Star Trek strange new worlds so TNG the Original Series so we got TNG there we go so photo of a man on the bridge of the Starship Enterprise and I'm gonna put TNG there just to see if we can get NCC 1701d and then we've got the Laura for TNG here with TNG 2 which is to hopefully give him a Next Generation suit negative prompts for people oh we can go with that one I'm expecting oh bum hole spell Starship wrong it shouldn't really matter there we go and I spelled photo wrong as well but thankfully the AI is also pretty clever at spotting typos and saying yeah I know what he wants so if Sulu and Riker had a baby right now you see how the face is so if we'd go back into Ed a detailer now and enable and copy the prompt I don't think we need the full negative prompt but I will and the reason we've got the dark lighting as well is because I don't know could be this vae here maybe so now we're going to just fix the face basically so if the face looks a bit naff we're going to fix it let's see what it does to it to the fias oh he's aged a bit now here we go sorting it is it making it better better better you look at the face there I'm expecting better results on that here well he's got the uniform shiny skin though is that just because of the lighting on the bridge that it wants to do it that way and it does look like the 11701 DS desk area there it looks like it um just a good lighting trying to think of this Cuts I don't know I don't know why the lighting is fairly dark I know it's out in space and the bridge isn't exactly the most it's fairly well lit though to be honest and we don't have what I would call photorealism here either I would probably use a different model for that wow these are very very identical that that uh I'm not into the lighting at all yeah so oh and I think you can specify the colors as well if I put yellow here I think it changes to the different yeah I think I think it's a trigger word for the uniform color so we've got like a Miles O'Brien type of dude right yeah there you go so you can actually use a color as a trigger word for the color of the suit you want and then obviously there's blue as well so a blue red and yellow right and then you've got the other TNG uniform which like the captain has with the ribbed top part oh what's it doing with the faces man I think we just need to use a different uh yeah this is shite so epic realism would be fine and realistic vision four would be fine but let's go back to Photon for now I want to see what kind of image it produces I'm gonna keep everything as it is I don't know what the difference is here between face eight n and 8 f don't know maybe the S is for sport sweat like hell on the ships yeah seems they are let's see if this one can produce a better face something a bit more realistic idea I know we're in space and shadows are always like hard that you don't get but we are inside as well it's not natural light who's this guy is he intent forward I'm sure we could get the correct like the the lighting on the bridge I'm sure we could get it correct let's take him outside of the Bridge photo of a man on photo of a man that's it forever man nope that's not gonna be it whatever man on planet surface that's it I want to see if it gets rid of that over bright look that too contrasty look that's not on a planet surface Doodles that's not on a planet surface wow that's kind of blown it out okay and we're gonna get rid of this I think it's got something to do with this TNG Laura thing then because I think if I remove the uniform we might get a proper image because Photon produces better than this much better so now they're out exploring beam me up Scotty wow contrast is like just crazy there's nothing else set here right that we've got set control net not doing anything surely nope nope all right let's remove the TNG and see if it just fixes everything now he's not even a trekky guy now he's just a dude on a planet oh God it couldn't be more different that's because it doesn't have to show the unit the uniform anymore Raymond rayton great yeah Earth is that close right it might not be Earth it doesn't have to be Earth this is like an elite dangerous shot isn't it you've got your Maverick suit on and you've got a planet you know close by coming up SRV in the background I've taken screenshots like that in Elite dangerous very much yeah actually I didn't take the TNG out of okay so meanwhile back on comfy it says reconnecting it's waiting to connect back into comfy so we can arrange that to happen I think foreign just fire up the engine so comfy blah where are you have I gone past yep there it is okay it's running like hell okay so if you want to run this you have to install python version you have to go to the website and install Python and go to an earlier version you need 3.10.6 can you install that and then you go and look for comfy UI I've got the portable version and then I think you just need to extract that into a folder and put one of your checkpoints your AI models which you can get from I'll I'll do a video I'll do a video on it on how to win how to put this in if you guys want one right so we are using Photon here and I don't actually need this I just put this in as a test because if I remove this it's going into the positive right there you go so rather than have this cable just dangling like that I just added a can you add one just like in here yeah so I think it was I don't know what it came under now ah here reroute you're supposed to reroute and then you can root it straight into to go but this way you can keep your cables a bit more tidier you know what I mean so it's now it's down there it's the same thing it's just that it's just tidy and you can do that with more if you want to if you have bit of OCD that you need to fulfill I'm okay on this as long as it doesn't get too messy I like it a little bit tidier so I like to try with just like very very simple prompts we go believe it won't give me 9 10 will it so let's just I want to see if it's going to put fish in there as well we won't need this prompt be not this bit I'll leave it in any way as long as it doesn't start putting clothes on the fish because Naked Fish are all right but mine fish oh so now we're gonna have to wait a minute for it to uh to load all this in you can see it all happening on the screen right here so this is what it has to do when it loads the AI model for the first time and there's the image on its way Jesus that's pretty damn good just for a 12-step image and now we can see here the sampler Advanced working away to get the next one damn see sometimes I like to just like this is realism right you could just say well he could have got that image off Google that's not how it works it doesn't do it that way it is actually generated um in real time but I like to do something like prompt it with something that it it couldn't possibly get from anywhere else you know um something maybe even you have to even if you have to go into the Realms of Bizarro world Starships I've got aircon it's broken on that one that's why they've got on my day going out so it's just doing the decode now that's lit up and then suddenly bang will get the final preview at twice the size now is it me or is this a bit more colorful see this see this yellowy bit of coral here that's kind of been got a bit more I don't know it's a bit more bleached and it doesn't look quite as colorful it's a bloody good image don't get me wrong but it looks a bit less saturated yeah yeah this one has a bit more color in it I wonder if we can tell it to not mess with that too much I wonder what it would be here like if we took if we change that to disable I don't know what the difference is and then just change the seed to fixed not going to change now is it if I hit Q prompt that number I'm hoping it's going to stay yes so it should generate the same again but this time with the add noise disabled okay forget that forget that it's a whole different image this is actually a better image as well I kind of like this one there's more variation and if you were to generate this at my native resolution then I'd have a nice wallpaper but I don't use them I like dark wallpapers it's I find uh light colored wallpapers to be too of putting they're distracting so they have to be dark if I'm gonna have a wallpaper it has to be really dark or something that is just not gonna stick in my peripheral vision all the time anyway let's see what this does is it going to desaturate this first image or not what what will the odd noise have done or not done and are there any nodes where we can correct holy crumbs a lonely way yeah okay it looks like it just didn't do anything didn't render any steps or anything so yeah we need that enabled holy oh what a difference that's not good it's like yeah here's a basic image I'm not doing anymore no that's not good all right so do we have any nodes that can mess with the image so the upscaling post-processing image sharpened image quantize image blur and image blend so is this for blending two images together maybe now if I I don't know where this goes if I pull out this image here ah vaed code so it goes before the decode and after the image we can add an image sharpener to both images but Sigma Alpha I I just want a clue so we can disconnect that pop this down here and we need to uncollapse that and then take the image input up output to this input and then this one into there see with the default settings it might just be these one these both set to one that might be too much it might be too little is it even needed do we really need an image sharpener it'd be interesting to see what it's going to do with it what else have we got mask advance and there's obviously uh well I say obviously it's not obvious but there are a lot of nodes that have been made by users third parties loads of them apparently and there's new ones coming out every single day I haven't looked at any tested any don't know where to find any but one day I'm sure I'll check them out I don't know what these are checkpoint save there's lots of stuff here yeah I think there's one where you add it it's called custom have a custom here and then you can just loads that you've downloaded I'll have to have a look so now this is going to double size the image and then sharpen I don't know might be too much it's going for the decode let's see how long it takes on here because this will be the next box to light up but for how long there we go oh that didn't take long at all all over Sharp okay yeah that's too much and with coral you notice that that's way too much holy crap so you would need to bring this down not 0.3 perhaps on these but as I don't know what the hell I'm doing with that we're just gonna ditch it I don't know that that would be something I'd need to play with and then decide when I would and wouldn't need to use that so okay what else do we have there's so many groups at a group don't need another sound plot I know Laura I couldn't do that right let me go to uh place if I know where it is uh and see if I can take some of the Laura's from or all of the laws actually because they're not too big I don't think are they oh okay eight gigabytes well let me just take a few so add detail uh blurry light is a good one sometimes I've got one for The Simpsons but it makes people look like the cartoon Simpsons characters that one that one would be cool that one would be very cool that one would be necessary that one that'll do we'll take those see if there's a folder side of comfy to put these so we've got comfy models Laura's there we go and paste done so now the question is are they already going to be here okay what if I add that particular node again let's add it here Laura and it goes oh yeah that's right yeah okay I got you so you can have the Laura the add add detail Laura on one image but not the other and again you can just use comparisons that way so we know that it's going from the model or to the model from the checkpoint load so from this so where is this plugin in this is plugging in over here before the sampler so we can plug it in right there so if I unplug that and did I unplug it oh shoot um did I just do something I shouldn't have done I've got two outputs from there to there no I don't think I did okay right so we can unplug that and plug it into there and then take that plug that into there and then I think if I press alt and slide this down we get another one and I want to unplug this cable plug it into there is that right it's it's before this yeah if I if I unplug this and it says case sampler yes cool so and now this goes to there but wait though that means we've got a load oh hold on a minute we've got two Aura loaders here right but I I want them both to have the same Laura name so we right click and convert Laura name to input and then we right click convert lower name to input and then we go right here oh double click here can we just do a text no how is that not gonna work is it I was gonna well I was gonna I wanted a Laura box where I could choose the name and output it to both of these at the same time so I don't have to have the same one on both okay so I don't have to you know select the same one on both it's not going to do that so convert lower name back to widget and I think I'm just going to have to refresh and if it doesn't load them yay cool all right but that means actually it means I can load two different loras though but Okay so we've got Forest background on that one and I want to choose the same on this one all right so it's a lore that's been trained to produce an image with a forest background so I don't need to specify it here so I can just say photo of a man standing outside that's it and then oh clip okay yep yep yep yep yep yep no problem I say that but I guess the clip has to come from here so we've got a third cable coming out and a fourth cable coming out this is some messy cabling and this has to Output two wait there's another one no that's exactly the same no point doing that where the hell does the output go so you can have it on reroute clip text encode oh of course you need text right because you need to tell it how much no you don't because that's done here okay I'm gonna leave it out see if it renders it okay so there's a clip input but there's no output yeah it works it's done the forest Forest background cool it did it it still works don't need the clip and now this is getting more complex I'm going to save this out as Rusty too Dot Json and the Argonauts I'm going to put that into my documents come for UI folder that I created save it in yeah it does work it's made it very contrasty though like he's pretty much in the shadows there I think these are the kind of loras like a forest background you can just leave it at one because if you put it at 0.5 what are you suggesting that you only half want a forest uh I just not really sure it's the first time I've added my own nodes and some things worked there you go so if I okay so we do have this is nine oh okay see why it was using a weird resolution there so photo of a man so we're gonna just add one word here which is going to be portrait that portrait should produce a close-up image like face face and shoulders maybe half the body something like that and no Forest definitely no Forest there a tree yeah that doesn't count was this just a fluke then oh boy where the hell would you plug this clip to clip is like this text and stuff right this is I mean no it hasn't worked how else would I test this blurry light Maybe very light we'll switch it to blurry light and render if it puts any blurry lights in the background of course he won't get the forest nope nope nope nope nope so it's got to be to do with this clip but if I output it there I can eat I can only send it to clip set last layer stop at oh okay that's okay that's not what I not what I want though because the output of that is also a clip wait a minute though I think I might have an idea is it coming out of here through the Laura and then this clip is returning back into the prompt and is it would would it only be the positive prompt and not the negative yes Laura's going positive prompts and you have to add it to that prompt so yeah it would go there there you go and this one I guess we can plug that into the positive prompt as well and now these cables are just getting crazy oh it's we can only allow we're only allowed one yeah actually they're both the same so it makes no damn difference which one we use or does it yes it I think it might because it might mean that one image will work and one won't let's run it see what happens oops I think I've run two yeah that's right so there's the blurry light in the background I've got a feeling when are we no way it's fine right because if this one obeys the Laura like say with the forest in the background this is just a refinement of that no on it wouldn't this just be a refinement of this one what if I put a forest on this one but not on the other right because the only one we've got a blurry light here that's going into the prompt so the forest background should be completely ignored I don't need I don't need this right actually I do I need it to go into no I I might be able to get away with just the one thinking about this let's get rid of this so this one is all plugged in but it's only going to this one which is producing this image it's not going to this one however if I can take another cable out of this one also to here without destroying the other cable bingo now this is plugged into both samplers and it's going into the prompt which is going also into the sampler Kaboom boom we're done so if I change it now to Forest background guys oh no the hell do I have to reboot the engine maybe that's I don't have this right I shouldn't have to specify Forest because the lower should do that for me do I then surely having this here though I wonder I see I'm of the of the mind that having this is a box means I don't have to specify it here in the prompt however that might not be the case I mean surely the name of the prompt he's being outputted the name of the Laura sorry is being outputted as a clip over to the prompt and it gets it gets added into the prompt in the background and then that goes to the positive there positive there and then gets decoded okay let's switch it back to blurry light go again if it puts blurry lights in the background of this next image it's definitely working it's just oh there's that dude again all right so the strength of the clip is going to be 0. 7 that's the strength of the of the the the the prompt the model strength I'm going to leave that as one because I don't know any other way and just cue it again it should still have blurry light but we should get a different image there we go oh the seed is fixed damn you not helping hello aquatic how you doing sir welcome to this live stream I'm just experimenting at the moment carrying out prompt experiments so can I now say that the forest will work come on that's a no what is going on with that Forest okay another way we can test this is this one photo of a Cube okay so I want to see what what the cube is uh and if you didn't want the Laura to work how would you get rid of it like there's no none here so I'm just wondering whether you'd use Ace a strength of 0.0 let's just run this Cube I have a particular Laura which should completely change how the cube looks okay that's the one it's decided to do today this has come for yeah it is and it is it is quite comfy I've just I created this from scratch today but I did it with the help of a tutorial and I'm not even using sdxl I'm just using my normal my normal um AI models okay so that's kind of cool very nice very cute very pretty but I want to change this now to Honey Tech so now it should make the cube look like honey and would you call those things in the hive like that the hexagon shape things yeah those anything at all nope nope so this this thing's not working does it does it light up at all if we go into VQ and we can cancel okay I want to see if the Laura lights up I doesn't right this is not lighting up at all why not Laura's get put in the positives right so if we know that the checkpoint comes before it we know the model must get plugged in and then this model would go to these two right a sampling and the clip goes into there gets added and then comes out to go into these as Sam I don't know it's not doing the honey Tech thing that's 100 sure Let's Go 1.2 if this isn't all orangey and honey looking it's not working weird I mean I've only guessed how to do this but I know that when I type in Alora it's always here up here but truly I don't have to actually type it in the prompt now because that's what this clip out is for to pop it in there right or or or no it can't be hold on let's double click here let's clip oh it could be any of these couldn't it so it could be text and then you would send that to what to the positive prompts of both is getting no that's for Excel we don't want that clip Vision loader dual clip loader the hell's that okay I'm lost I don't know how to add the Laura into this workflow and make it work yeah it's just node-based aquatic it's very similar to uh LightWave LightWave works very similar to the it's it's uh it looks exactly the same actually okay so I know it's not this no it's not that what's this glimpy thing where's that gone this clip different colors different color inputs and outputs it won't be that so the output from the clip can go to what's this conditioning oh this is even worse width height though where else can it go reroute but that's just a box you that that would apply to any output surely okay so it's not that it's it's not lower loaded because that's another one that's so it has to be this text encode but why would I type it in here to send it to both samplers no point when I can just all right let's unclip okay wait a minute no this is fine yeah okay let's unclip this cable which means that this has to go back up to there and we're just running it through the model come on oh it's Laura I'm trying to guess what the format is here so it's honeytech-20 or do I have to hold on can we do no I haven't I it might be a trigger word that I'm missing I think it's the trigger word let's take this okay shrinky shrinky take the clip down to the Laura and move the clip up back into here but we need the keyword trigger word which in this case is honey I think would work on its own but we'll do honey Tech Dash 20. let's see if that works some of these loras need trigger words talk to me yes that's what it was it needed the trigger word yes that's all it was the honey Tech trigger word for this Laura the honey Tech was the trigger word for this Laura yes baby that's that's what I'm talking about so if we go back to the forest background the trigger word for that will be I need to do it on another browser because otherwise I'm going to be going to a website which I have got the not safe for work images set off I'm not going to put that on yet so we're looking at uh Forest back ground no not back round background voice background okay oh no results found I'm not buying into that one son the forest results what's happening with the search on this website it's all a little bit different okay right we'll just take a guess then so instead of the honey Tech so it's Forest background underscore V1 normally if you you can type like with the honey Cube you can just use the word honey or or the actual name of the Laura so this one was what Forest background underscore V1 so we'll leave that as the keyword which matches the name oh buggeration which matches the name of the Laura for his background B1 underscore V1 and the clip strength it doesn't need to be 1.2 when one will suffice and photo of a man outside foreign Forest in the end yes sir so that Forest that we got the first time around must have been a complete freakish coincidence so let's get the double-sized more advanced version up and running there it is cool now the guy there I preferred this dude with a cap on spell this foreign let's put a fish on his head cocky clothes and a baseball cup go I've got a completely different image than before though yep if what I was able to get over to the preview image it already rendered the damn thing yeah so The Laurels are working now it's just a case of adding in the trigger words because we're not adding the the actual Laura um itself in the brackets like you would in automatic 11. so you need the trigger word some dollars don't have trigger words though so I don't know how we're going to get around that I think just have to specify the the Laura name in The Prompt anyway so there you go guy in the forest he's not really got khaki clothes on not really they're kind of just green but he does have the baseball cap and it is definitely a forest now here's a one adding this is gonna be a killer for cables adding two loras like what if I want the honey one as well as the forest background I'd have to add another one of these right but then I think I would just bring the outputs of this one over to this one and then wherever these outputs are going would be where this one's going to go right but then what happens then it would do the as long as I put both trigger words up into the positive prompt it would work however what if I then change my mind and decide I only want to use one on the next render it would be hard to just take this one out of the Loop and there's no non here there should be a none see and there just isn't or if we go all the way back is there isn't there's no known here the only way you could get around it is to change these numbers to zero well I've got this I've got this saved so this is going and clipping into both models right oh wait a minute yes because the models coming off this checkpoint okay so flip this put it into there and same with this one this one we can just wipe out take this clip drop it into there right so now I don't think we can have blurry lights on Forest background but I can certainly do honey Tech that would be weird but okay and then the clip come on now so the clip would then go through here right oh dear now I'm hoping it's going to use both and the model would be plugged into this sampler this one so honey tech-20 and a forest background so photo of a cube come on come on come on come on it probably worked with a guy as well photo of a cube outside so let's get rid of the baseball cup because that'll be weird and we want the other keyword now oh wait a minute to have the other Laura not working if we don't want it we just remove the trigger word from the prompt here so Forest background underscore we want and honey Tech Dash 20. correct right so if we don't want honey Tech to work we just don't use the prompt up here just don't use it and then this information should flow through here and out so are we going to get a honey honey Tech Cube in a forest oh uh allows you to do more I don't know with control net and all those other extensions in automatic 11 11. there is a lot you can do in automatic 11 11. what this allows you to do though is just have control with nodes and stuff and so in certain ways yes it does allow more things but there are things in automatic 1111 that this can't do yet I don't think I mean how do you how do you load in things like control net or root to to do poses and replace faces and sketches and scribbles how would you even do that on this I'm a bit surprised that the cube is floating but nevertheless it's a honey Cube and it has a forest background so they both did work and the way to just cancel one is to just not use the prompt if I took out the forest background prompt um here we just get the cube even though Laura is active the forest background is active I'm going to say that this image now will not have a forest background but it will still have a honey Cube there you go hardly a cube and then you just want to pipe it back in if you want both so now on any image I can have two loras running so what I'm going to do is right click on this one change it to load lower one and load lower two there we go okay is on top make sure they're A Cut Above the Rest oh boy my time is limited so if we remove these completely and we just say four to over Cube outside there should be no honey cubes and no Forest that's a coincidence it's not a forest was it what is it part of a house now tell you what if wait a minute go back this traces of Honey there isn't there oh right because it's affecting the model oh okay I understand so by removing it from The Prompt all we've done is effectively said strength clip zero but not model this we'd have to change to zero so that it can affect the model either now it should be completely lowerless there you go and therefore that means if I wanted to keep the prompt in I can but if I change the the strength clip to zero then it should it will read those words but it will say how much strength of these do you want applied and you say zero so it will completely ignore it and if we run it again it should be complete ignorance that's not a cube it's not honey based but there is a is that just oh it's got me wondering now whether that is just purely coincidence that it's gonna a forest in the background I think it is coincidence because that's not it's one of the Woods let's go again yeah I think outside is just generating so if I say outside City right let's cancel the one that's running oh gee so how there can't be any influence on these if everything's set to zero maybe it's still yeah it's obviously still reading the word forest in the prompt it's still doing something with that word Forest okay well it's all a little bit uh complex but if you want to run the two loras or one Laura you can plug one in switch these on by putting these to one and adding the trigger word into the prompt it's a little bit more work than automatic 1111 but it will get you the results you want and this way we've got a double Laura which is awesome and there's many more you can do I've got many more Laura's which I only I only passed a few over so this concept WB is for wide body cars you've got a large window there which we can try so if we say photo of a woman in a room how did I get that start again I nearly typed womb there womb that'd be a week right imagine that having the AI to a woman inside a womb oh it's already done the window that's cheating you weren't supposed to render a window that was what I was gonna do and tell you to do later do another one cheat no window this time oh man another one all right we're gonna have to remove black and white from the list of possibilities here so comma BW comma B and W and comma black round white that should be enough why are you insistent on the window now you weren't before but I'll do we'll hold that one so let's change the random seed to fixed and I want to choose a large window set this to one and set this to one nothing here and it's large window zero five that will be the keyword bingo yeah I think that's right so it should render the same image again hopefully with a large window Maybe the same image I wanted the same image viewed well I got my large window but that's not what I was expecting the whole point of keeping the seed is at least to try and you know keep the same image it's certain we certainly got the large window and a woman who seemingly is in need of a sandwich in fact that is a really if you then reduce the strength down to say 0.7 so we get do we get a smaller window let's reduce it to 0.6 on both I also take requests if you guys want to see this thing render something specific yeah it's kind of it's a it's a Pro of the dice isn't it really it seems to be uh pretty well set on having a woman with a back to us looking out of a window with a very short skirt on so I'm going to change this portrait now what you're gonna do ugh what kind of mutation is this does that work cancel let's move on to something else right so let's get rid of these uh loras by zeroing this out and up here photo horror blood Gore oops Gore vampire demon and Nightmare I'll leave it to you right oh did I specify that I wanted it I did say photo but not so much a vampire though I mean his fangs are not really this is probably not the best AI model to be doing this with there are other ones uh which we can switch to Let's cancel this one oh don't need you know this man so if we switch this over to I would say dream shaper Maybe oh there's actually one there called uh oh no Dimension twist I did get something that was I think it might have been a Laura to be honest which was for um demons especially to do demons I was trying out some zombies and things like that before so this didn't take a bit longer because it has to load in the new thing oh have you had guys had this there all that time Alex I didn't know the window was there and I don't think is anybody here that window wasn't meant to be there all this time so another ruined portion yeah this one definitely gave a better result same kind of idea though the goth guy with the long black hair ai's got the same idea even though the AI model is completely different it still gave us the same yeah it could be the same dude in a different is it trying to work off the Twilight movies I haven't seen them anyway perhaps I shouldn't have specified vampire and Demon I'm gonna go height would be I have to do 720 or not guess we can it's it's rendering fairly quickly so I don't mind upping the stakes a little said the butcher no then oh d-moon please render the moon in the background oh zombie in a swamp go can I get over there in time just about yeah and now because I've gone widescreen it's managed to put two of them in quite an eerie looking image and now this is going to be double the size which means it's going to be 2K oh maybe I went a bit too big there I do yeah like I said I do have a Laura or I've got a textual inversion I think it is for demons is there such a thing here okay image invert but mask no there's lots of different nodes you can add here but latent Road tight flip and crop oh control net dudes no way yeah I don't have all the control Nets in the folder either and what was the other one and where does this output to control net apply okay there's one so we need conditioning that's from the prompt that's an image input and that goes out from the prompt to wherever that to the into the sampler then into there so we can okay but if we have control net sketch then it would work the sketch from the image input as well and also the pose interesting oh man I've got to say we're gonna get con control it inside this that is gonna be eight there you go I don't know what I've seen this before where it puts these round things on the heads but I don't know what they're supposed to be I'll save that one out oh so a control net is actually on here that's great because that's going to give it a lot more power but look at this guys if I go and save this workspace out all right so I'm going to call it Rusty three to Laura I'll know what that means dot Json and I drop it out into my documents folder now if we were to get rid of all of this and replace it with a new one let's clear the workspace first and load so documents now this is a this is one that's been created by somebody else locating the folder now I'm not saying there's a lot of complications going on here that there kind of is you've got a positive prompt a negative prompt and then you have a positive left a positive right a negative left and a negative right what do those do you ask don't know there are additional prompts so you've got female cyborg Chrome ledi perfect face in the style of Charlie bow water I don't know who that is Alejandro bodicio I've never heard of him either intense and then you've got detailed photo so this is the type of shot you want detail photo wide angle shot realistic so this is the subject and this is the environment I guess and that gets plugged into left and right and then you've got there must be another left and right over here somewhere these two I guess but two checkpoints this is using the new stable diffusion XL base with the refiner so what's going on here then this preview image is it refined or without and then you so you've got a save image for normal and a save image that's been upscaled and this is being upscaled by there's the refiner upscale it's using it's using this I I haven't got this so good luck with that I don't have this file either so good luck with that one if I click on just go it's not going to do anything it's gonna say hey you haven't got these see what happens okay it just says in the thing it's going to be ignored output will be ignored but the preview image I don't know if it's the refined image or the base image oh boy the way that this is grouped it hides all the cables I don't know what the hell is even going where it's it's not using control net either so that's not in you do need two checkpoints for sdxo because it does have the both holy crap I've just realized it's it renders natively at four four thousand stop stop stop stop cancel please cancel but wait a minute you're rendering it for and then you're upscaling on top of that dude I don't have a 40 90. graphics card here this is off the wall and the other thing here is if these both are this is a bad way to do it as well from what I can see if if this is if both these encoders here are meant to be the same which if one is for the base and one is for the refiner they should be the same and then here as well why why would you have instead of one two three four five six seven eight inputs you only need two you need a width and a height what's going on here I'm trying to work this out you've got the image size here is 10 24 by 1024. so what's this I'm going to show you how to get this eight down to two but it's going to be messy see the width and the height are not always going to be the same so you need a box here where you can type in a width and another one where you can type in a height and you plug all of those into all of the widths I can do it on these two bottom boxes I don't know what this is doing though this is the base these are the refiner this is the upscale uh thing the upscale model so is it upscaling from this 1024 to this 4K or 4096 I should say is it up is it doing that because yeah it's going to upscale from there to there using this but I don't have these files so when it says ignored because I don't have the files does that mean it's not going to upscale it I can't cancel this yet because it's still working on the memory efficient processing so it won't cancel yet so we can set this up let's say I want to make you've got four prompts we get that down to two and you can do the same for this so right click now double click here and I think it is integer we need an integer so it's which one was it now nope now it's a node that has int on it I know it does oh I don't know it does but okay I'm gonna switch to my other Rusty three has it ah primitive isn't it is it integer control yeah yes because to do it random would be just that'd be weird okay just to have it doing every time you render that it would do a random a random upscale of a different size every time uh but no if you want to use this like if you have to keep putting multiple inputs into like if you want to switch this to 2048 and you've got to do it on all these boxes to pain the ass so let's double click here and just type in Prim and well I I'm thinking this is going to work and we'll title this one global height width dude and then oh let's just do it this way alt and drag and then we'll just type title this one at Global height and connect to widget input wait a minute this is different mode how is this different where's the thing where's the uh where's the two controls this doesn't make sense now for God's sake not making any sense on my thing we have primitive but it has these two primitive node why why does this have these two prompts and mine didn't I don't understand oops that was a bad key combo okay I've just done a copy paste right is it okay maybe you have to plug it in first I don't know so what we'll do on this one is right click and convert width to input and plug that into the width right click I should have suss that one out shouldn't I convert height to input and we're going to plug that into the height see it's fixed and it's 496. there you go now on this one convert width to input and convert height to input we should be able to take I've been a title this a Global web let's take that and see we can also plug it in up here there you go and this one yep not bothering about capital letters and this can go to here so now it's plugged in this height is plugged into here here and this width is plugged into here and here and now I want to go to change two even though I don't want that I'm just I'm going to knock it down to 10 24. Jesus 10 24. right that's just going to match the original image so now that is forcing the width and height at 10 24 for both of these now you can do the same on these two right click conver T width to input and convert to input Target width Target height this is too complex for me but you get the idea you just plug all of these integers into all of those things let me just undo this before it gets out of hand I'm pretty sure it's not going to produce a 496 image because I don't have the upscaler I'd like to see what the image is going to be though and this is a complex this is why it's best to create your own because at least once you've done your own at least you get to understand what's got what the hell's going on okay so we are using where's the checkpoint okay it's done right so it's it's rendering so it's about to render the base now with the base is ready now done it's just decoding oh so the base sandwich is going to go here and the other one is going to go okay this might look bad but I think it's okay now it's loading up the refiner that's either gonna go here yes the refined image is going to go here of this one at 10 24. and then the upscaled version which would have been to 496 is going to go here you've got the base the refined and the upskilled gotcha right that looks a bit manky but that's just the base image so I'm gonna try not to worry about that maybe altering these is an issue I don't know I do have a much simpler layout for stable diffusion XL but it does not make use of these positive bars and positive and negative positive and negative left and right it doesn't make use of those but combining all of these prompts into um one might work well maybe steps on the base model total steps so there's 20 and 30 I have to understand some of this I need to like pull these out and see where they're going see where they're plugged in just basically move all the boxes out the way just rip them apart at the moment it's still it's still loading the refiner once you see this green box jump you know it's done that and it's on to the next stage thing with um sdxl is it does take it needs two AI models to produce the image normally the other ones I've got it's kind of got the bass and refiner built into one this one is it's got two people if you like to do the two separate jobs as opposed to one person doing both hello anybody on the chat that might still be there probably isn't those who have fallen asleep and we're away I think it's certainly just got past the next stage on the command line but still on little stuck here my sound just went off because it hasn't been the sound played through the computer for a while come on dudes this is where it's at it's loaded all of the memory efficient stuff and then it jumped and did that paragraph underneath the vanilla Transformers sounds like a robot changing into an ice cream so once it starts rolling again there we go it's away but now it's here on the K sampler which box will it jump to next I'm gonna say these two at the same time well I don't know the output is latent so is it going to jump here hmm yeah it's got no output to go to here it's pink the refiner does take a little while I wonder what it's going to do with this image wow it has work to do let's remind ourselves of female cyborg Chrome LED I perfect face in the style of Charlie bowl water Alejandra bordecio two people I have no idea who that is and intense Alejandro Alejandro I don't even know yes we could Google them that's Charlie bow water art okay and Alejandro Alejandro badicio hold on how do people know these people these artists I don't know any of these anyway I'm just thick wow it got the job done I don't know where the Charlie bow water comes into it but I guess maybe on the face dude that's the base and that's the refiner that refiner is pretty fine wow okay uh has it finished yes because it can't upscale because I don't have these so it can't do the upscaled version what the hell's over here why is there two I don't know what's going on I don't know why we've got two of those but it's complex anyway guys there's nobody watching so it's time for me to shift off so I will catch you next time until then not now okay have a great day everyone take care I'll catch you next time [Music]

Channel: Rusty Dog

Views: 662

Keywords: #rusty dog, AI, A.I., Stable Diffusion, Controlnet, SD Upscaler, artificial intelligence, image generation, rusty dog, how to, comfyui, comfy ui, automatic1111, automatic 1111

Id: DpNjPnjOx50

Length: 221min 50sec (13310 seconds)

Published: Sun Jul 16 2023