ComfyUI Basics - Live Stream - Join me & Have Fun

[Applause] [Music] [Music] ch [Music] [Applause] [Music] [Music] is [Music] [Music] [Music] is [Music] [Music] is [Music] [Applause] okay hello my friends how are you doing I guess today we have a lot less people watch watching uh because it's about com your ey W um yeah so I just want to show to people how you can use it that it's not that hard you can do a lot more things it's more experimental these kind of things right so let's see where the numbers get I don't know I don't know let's see it's just right now the numbers are pretty low but let's see um hi deppt how you doing how Louis hi banish hi Olga hi roll how you doing hi niiki how you doing thanks for joining uh control all delete thanks for joining um Andre did I write it conf oh okay that's okay hi Andress how you doing hi Cliff is it Cliff yes I think so Hi how are you doing menath hi thanks for joining pretty cool hi ailia thanks for joining so yeah let's see by the way I haven't likeed I haven't massively used it myself so it's not like I'm an expert in comu ey but it's not it's not that hard to use there's some very nice examples online we can check that out a little bit and I want to show you some things that um you can do with it that are just like to get you in the in the thoughts in the in the mindset right of of using com youi hi sa how you doing thanks uh for joining from Brazil that's pretty amazing uh Neo furo hi thanks for joining hi Gregor from Swiss I guess and then Gabrielle from Santo Domingo wow nice hi to nice to see you hi uh watch so many videos you just said my my band name oh okay cool for all that help yeah that's amazing yeah you're welcome your band name is control all delete that's pretty cool Hi Fi Fi fan hi thank you very much Hi how are you doing hi Andress how you doing is automatic 111 dead no I'm using it as the main UI still and also most in my tutorials but a lot of newer things just take some time until they come to uh automatic 1111 because it's kind of I would say like centralized in the way that there is like a developer or team of developer I don't actually know uh but they decide um uh the kind of extensions that work with automatic 1111 and they are kind of like slow upgrading the um the UI while with comi you can just create your own notes on the fly if you want to if you know the coding a little bit you can also have CBT help you with the coding and this is why if something new comes out you can first like most of the time first check it out on comi and also comu ey can be a lot a lot more low weight because you only need to use what you need at that moment and not load all of the other stuff like in inside of automatic 1111 you have all of these extensions active of course you can dis deactivate them and stuff like that if you want to um but with coni you can build like precisely the workflow that you need for that process and you can do a lot more like you can for example load uh multiple models at the same time and work with them at the same time which you can't even do in automatic 11 11 so we're going to check that out um who else is here hi Ai Dan how you doing hi roll hi um Adriano from Brazil nice to see you hman how you doing we we kind of got lost with this kind of tradition of everybody telling where they are from so we should we should pick that back up maybe hi Axel how you doing thanks for joining hi CIT Citrus is it is it like that hi um thank you making all these great videos yeah you're you're very welcome my friend um Neo furo hi thanks for uh joining the live stream hi Robert hi Vera thanks for joining um seem to grab setting IM what seen some grab of a settings image and then loaded the notes is that a thing or am I M no no no if you if you any image that you generate with com youi has the complete workflow built as metad data inside of the image that is another huge benefit of using com youri is that you don't have to save the different like workflow builds that you create as separate Chas and files they are part of the image and you can also send them to other people and then they can use the exactly same workflow that you have so it's very very useful right um how many people from Brazil yeah apparently several hi Kevin from basing basyc soak what I can't I can't speak that bay syn soak in England hi hi Cliff how you doing um have you looked at the focus yes I have already made a video about Focus it's very nice I really like it uh grer garer sorry today my reading is bad hi from Germany that's nice greetings from Germany I'm also chairman hi Ali how you doing hi Bor Kua flow hi from Puerto Rico that's pretty cool hi F hi s sorus hi uh red OCTA we should actually like uh from Portugal hi we should maybe start the video Daniel is also from Portugal no from Portland sorry for that uh CK is from Croatia FC Lee hi how you doing uh sorus I already said that one um okay um hymon is also from Brazil Andre is from Romania drag is from South AF Africa Hi how are you doing that's so nice kopan um this is one of the most International streams I've seen that's good David is from Sweden hi um Dwight from Cornwall England hi that's also good Andres from Colombia a lot of amazing uh countries I want to visit aen is from Bangladesh hi NE says image generated on comu ey yeah have the whole workflow embedded right in the metad data right right right I iush is from India hi nice to see you I have been in North India but not uh South India so far very nice in North India very beautiful country but also very large like a lot of your smaller cities are bigger than most cities we have in Europe like the major cities so that that's crazy you come to like a town with your bus and you think like maybe a couple of thousand people here and then it's like Millions it's really really big it's crazy hi chip from the UK Ren is also from Germany hi how you doing okay I guess we should get started uh let me switch over here uh and then switch the screen so I can see something I hope it works yes there we go okay cool um what is this oh this is the this is the image preview this is the High um the high resolution um yeah upscaling all right uh I guess you can't see everything so I need to I mean I can scroll here and move it over so you can see it a little bit better this is what I showed you last time in the video with this kind of impromptu um Laura that you can create so you can do some really crazy stuff whoa this is jumping around like crazy one one second let's put it here and then let me move this or let's put this over here here and then we can like just balance this until you're out of view there we go Okay cool so now we're good and the cool thing here is like you load six different images any kind of thing you want uh should be similar style and then this is like converting them into an image based on that style right so when I let's turn this off over here this is also something like you can turn things on and off on the Fly uh let's click on Q prompt here and then we will render this you can also see with the things that are getting green those are the things that are active at that moment so the sampler is running right now and we should have an image in seconds there we go you see so this is the output um based on these kind of six input images I already made a video about that instant Laura um and we got whoa it jumped sorry one second we got this as an output it's one thing I'm still struggling with is that it's not completely centered that is one thing you see the ears are cut off stuff like that that might have to do with the uh with the checkpoint I'm using stuff like that right um yeah for beginners it's too much I just had this still loaded so this is why I showed it right now so let's make a new um let's clear this clear yes okay so we're going to start very simple uh one thing uh you can over here on the right side you can't see that my big head is in the way let's put this up here so here is Click uh clear and then you can load and save different workflows if you want to if you have files for that and clear just like clears everything um hi antara from uh Algeria um there we go and um when you double click you have a search so you can actually search for every kind of um of note you want to have here's a full list of all the notes I have so far because you can also easily install other nodes from other people there's a lot of nodes out there that you can use and then you just put them together so for example if you want to load you see you you type load and then you have your checkpoint loader simple right you can load an image you can load a image mask you can load a Laura clip loader um and so on right there's a lot of different stuff here but what I would suggest to you is that you can go to Let's close this one we don't need that anymore you can go to this page the official page um comu examples and there you find very easy um nice uh uh examples to follow you can also download them so for example here we have a very no this is Laura um this is a very easy build for imageo image examples right so here you can see um we load a checkpoint so we need to load checkpoint so we go uh one second a yeah let me let me put put this here so we double click load checkpoint no we don't have that okay checkpoint simple loader that looks good enough and then we can load a model here that we want to use you can also type here let's use ref animated for that my kind of still my favorite model and we can look here again what do we need we need a clip text encoder for the positive prompt and then a clip text encoder for the negative prompt so that is pretty easy to do so you click here clip text encoder first thing that comes up like that and then you can simply copy that with contrl c contrl v so you have another one and there you go now we have two clip text encoders for the prompt if you want to um you can rename them or not I thought you can I'm not quite sure right now you can give it some description or not properties ah yeah here you can rename it positive prompt negative prompt if you want to in the properties here uh but you don't have to usually you see it right so each model has an output for a model for a clip and for a vae in in case that the vae is baked into the um into the into the model right so you just connect this here and connect this here and easy right super easy easy to understand it has it has color codings for the ins and outs you know where they go it's pretty easy right title do we have a problem with the title oh for the live stream no conf ey Basics live stream no probably some something else all right I also had to download comi the other day uh I think it's for beginners just join IP adapter only takes a Center Square yeah with I you can actually go away you can actually get away with less images for the uh the thing I showed you before with with the instant uh Laura it's the IP adapter yeah yeah yeah all right um I can also by the way I can also uh uh scroll in and out so this is also good for people um with yeah I don't know with who have a small screen or not so good eyes or stuff like that so this is very very flexible as you can see okay what do we need next next we need a case sampler this is all of the kind of settings you have in automatic 11 and even more Choice than that so we go back here we double click there's already the case sampler there we have it right so here you can see positive negative and it has an orange button so that means this is what we need the conditioning from the clip text encoder right super simple right so this is our positive prompt and this is our negative prompt so we simply connect that and boom there we go right okay so in here you can see you have a seat number you have the way the seat number is created so in this case randomized or it's fixed increment or decrement so increment is basically count up decrement is count down so we leave it at randomized you can have here the steps the CFG scale um the sampler you have a lot a long list of sampler you can also have a search here through the sampler list and you can see these are only the sampler types because you have also below that whoops one second below that you have also here a scheduler right so here you have for example normal caras exponential SG uniform simple DD IM uniform right so when you go into automatic 1111 they're usually automatically combined for you so for example that you have um a DDM this down here DPM Plus+ uh 2 m caras for example right uh but here you can choose it in any way you want to use it right so you have more uh flexibility in playing around with that right talked about the Box title oh yeah you can change the title yes there we go Ren nameing can be done in title yes yes yes right right right all right so that's also very basic like very very similar to what you see inside of um automatic 1111 the cool thing is here you can click and then move your mouse left to right to um basically how do you say scroll roll through the values so you don't have to type them you can also use these kind of arrows here if you want to so it's all set up pretty easy you can double click and then I can just type a number in here hit enter it's now in there so yeah it's it's it's it's very like easy to use basically right so let's see what we do we need next we need a vae decoder and this is also the benefit here is that you learn a little bit more about how does AI image generation actually work because uh first this is so basically your text is using the clip encoder to convert it into the information that the sampler needs and then you create this latent image as you can see here it says latent the latent image but the latent image not a pixel image you can't see that it's latent information that only the AI can understand basically you can understand it like the ones and zeros on your hard drive and then it's converted into pixels when you want to look at an image right so the V the vae decoder is doing that for you right so for that we need a Vie decoder let's go here vae decode so we have that and then we need a vae that we can load separately unless it's already part of our checkpoint so that's uh uh easy and then here we have an image output uh so for that we simple Type image um preview image for example which does not say save the image it only prev it if you want to save the image you need to have a save image um uh file here we can actually do that let's delete that uh let's go image again save image so now upon the generation the image is automatically saved so we connect this here we connect this here and then what else do we have we have here the model this goes here to the model obviously the latent image is the resolution of the image you want to have right so for that we need kind of an image size which I'm not quite sure get latent size image get image not H one second no I want to set that I think we need a number for that I I kind of forgot how that works um one second we need a number here sampling loaders empty oh empty latent that's that's a good point empty yes empty latent image there we go thank you very much for that thank you very much I was stuck a little bit so here we put that in here and then you can set the value here for the width for the height let's set this here to 765 or is it 56 765 uh let's let's just keep it like that batch size is one and then we also need to load a vae vae uh loader over here so we go with this one here 84,000 steps uh which I kind of prefer and then you just connect this over here you can also bring the note over here to make the line a little bit shorter and then you can simply write um the text you want to have right um beautiful woman in a red dress let's see what we get from that I don't have any negative um uh or let's let's write a little bit more Masterpiece and then we go back down here and write bad quality deformed ugly uh nude also whoops that didn't work nude and then not save for work let's hope that works let's click here on Q prompt and see what I get from that you can see here it's load the model first and then it's already uh like rendering here and boom we done there we have the image you see how quick that was also I took me um it took me one second to get an image one second there we go uh on a 4080 so let's go here make eight images let's see how long that takes boom rendering and encoding and and we have eight images there we go boom shakala with a with a resolution of 512 by 768 we have eight different images rendered each at one second that is a lot faster than what you have also in automatic 1111 uh because this is lightweight we don't have all of the extra crap right we don't have control net and rub and all kinds of extra extensions loaded we only have loaded what we need loaded at that moment right so that makes it very very easy and lightweight right so yeah that that's the the easy thing here by the way I I guess I want to go to um civit Civ a i there we go and then let me find ref animated because I want to get a better negative prompt because mine is kind of not so great and I haven't um saved it anywhere so we take this one and then let's do some magic here put it in here let's set this back to one and now it should look a lot better let's see boom we are done hello look how good that looks right and then of course you can upscale that also so this is by the way this is just like the low resolution right now which also looks pretty good I would say um sorry I should also look at the 768 yeah I guess it's so right uh you can drag from the latent image also you can drag what um a little tip for you want to make this part easier is to use the efficiency notes extension it has most of these notes combined into one or two yes right there's a lot of different um there's a lot of different uh like collections out there of different notes that combine notes together make the workflow simpler you also can download workflows uh from uh from simit AI so when you type here uh comy UI uh workflows you did it work there's a manager here that's also pretty good um there you have different packs here custom nodes and so on workflows ah there you you see you can download different kind of workflows that have been created by the community for you that do different kind of things um and you can just like play around with them you can learn from them so all of that can help you a lot right now um okay let's have some fun here right so what I want to do here is I want to use the same negative prom but I want to use different positive promts at the same time let's let's just use this as an example right so I'm going to copy uh basically one second I'm going to take these parts here and then I'm simply going to copy that okay that didn't work let's try this individually doesn't matter go like this so now we have it twice all of it now I go uh sorry no I go down here I go over here just to give you a a custom workflow that you could build as as an idea it's the same resolution it's the same model we go to latent image in here we load the same vae we have the image output here so now we have down here um uh what what what are we going to do um beautiful man beautiful man in a black dress no not black dress black suit there we go okay so now when I hit this button it renders both of them you see Boom the woman is ready the man is ready there we go we have both of them I clicked only one button right and so with that basically you can you can build like crazy machines that you want to have for your process for the things you want to do you can resize that I just can't there we go um you can also use the one image as the input for the next image there also no problem right so I can use this image as an input um for imageo image render so that's also possible Right most workflows are too complex uh yeah in that case just like um go backwards uh through the process and just look how they are connected because most of them use kind of the same stuff and also you can go on Discord ask people if there is like notes you don't understand uh just ask them what do these notes do why are they used why are they used in this and that way but with this as you can see you can just like really adapt it to the kind of thing you want to create with that let's make another um combination here there you you go another guy another woman boom super easy just one click it's not upscaled yet but um yeah you you can do that kind of thing right or we can go in here and say we have a be she's standing in a park standing in a park and then um let's put the same thing here and then we put this standing in a street right so you can be experimental with that and with that you can just build it um for your for your for the workflow you want to do now this of course is not the same woman if you want to have the exact same woman you can have it like with a Laura for example stuff like that um yeah but you can like you you can do like 10 others of these and then just have an output and with these kind of simplified notes that that uh one of you mentioned before you don't even have to do the full process by copying all of these kind of things because some of the notes have multiple of these combined and then it's it's a lot it gets a lot easier right efficiency notes yes right the efficiency noes right right right we can actually let me copy that over we can have a look here onto that on second um there we go you see so that's that installing is also very easy um I think you just like download it and put it into the extensions folder and then you restart comi and then um you have that so you can see in this case there um goes the um here is the input already combined um for uh the model you need the vae you need the clip skip the Lowa the Lura strength um no the Lowa model strength the Lowa clip strength you have a prompt here negative prompt you have the width the height and the batch size and all of that is then going into uh like this uh bigger case sampler it also has this kind of like image preview in it right so a lot of that is happening at the same time with these kind of notes right um so yeah you can do a lot of things with that right you can actually um we could try that but I want to show kind of the basic comu ey thing that might look a little bit more complex at the moment but um just to give you like get you started right let's let's play a little bit more with the basics here and then we can we can use custom nodes right how about that use groups and you can bypass group notes group notes what do you mean by that is missing or a model is the problem is when a note for a workflow is missing or a model is missing you start with the ex exception and nothing will work um I would I would uh I would uh suggest to you to install the manager which you can also find here on comi uh one second um comi manager does it work does it not work ah one second conf few IM manager there we go confi manager so with this when you have something from from GitHub the only thing you need to do is to go into the extensions folder um for sorry the custom nodes folder inside of comi and then you type CMD and then git clone and use the web address from the GitHub and then all of this is copied over and what this does is if you open up um if you open up an install uh if you open up a workflow that does have nodes that you don't have you click here on the manager and then you say install missing C some notes you click on that and it will show you what's missing uh and then it has an install button next to that so most of the time this gives you the notes that you don't have and then you close com UI you restart com UI and then basically that should work right um Alternatives of automatically Lear I haven't clicked on that maybe there's even some workflows in here I don't know um so yeah that's pretty cool uh what else do I want to show you um [Laughter] um um let's go back here so here we have image to image here we have a VI encoder we don't even need that and then goes in here and then goes in here oh yeah so we can do that so if you look at this kind of thing and you understand it a little bit it's easy to just like um uh get your own ideas from that so here for example you have an input image that then is code it into a latent image but because we already have a latent image I can just use that right so we go back here uh what do I want to do next okay instead of using the empty latent image from here I'm now using the latent image oh there's a second one sorry for that uh is there oh no okay I'm using the latent images coming out of here to put it in here which means I have basic basically um now image to image render all right so let me set the D noise to 0.5 like that and then we should have now when I render it we should have some pretty similar images let's let's check this out aha they are they are basically um the same image the face doesn't look so good right now um okay we need to set this a little bit higher here let's go 7.5 yeah oh no well the woman changed now okay so I was almost there let's go a little bit lower here try it again H maybe the change I'm trying to do is a little bit too extreme let's go here and say a blue dress instead of a red dress let's go like that yeah look at that so you have kind of the same pose here and now the the dress is changing right or um let's go with the red dress and say um here we say blonde woman blonde woman maybe make this with double round brackets to make it a little bit stronger there we go let's try that yeah look at that that looks pretty similar wait do I still have City over here oh s okay Park all right let's try this again yeah look at that so we have two similar very similar images again they are not upscaled yet but um we have oh we yeah we should like use the same seat that would also help of course right that's true like if I use the same seat here and there uh that would make uh let's make this a fixed seat and then just copy this over to here and make this also a fixed seat and then let's see what that looks like there we go well the background is still very different probably because my D noise is a little bit High let's set this to 0.5 H didn't let's go a little bit lower here why is it am I crazy right now one second let's go here but higher H interesting so the background is getting sharper oh I know what's going on um case sampler CFG scale let's make this CFG scale here a little bit lower set this to three let's try this again H interesting well this is the usual like experimentation you do with AI just to figure out what's going on uh I'm kind of confused why the background is getting so sharp for some reason in the second picture but you can say they have the same Poe so usually this should work just fine I don't know what's going on right now um set this a little bit higher here shed carus exponential now normal should be okay set this to three the hell set it to four I'm clearly missing something here um light is coming out of there steps maybe I guess it's because this is fixed so let's set this to randomized yeah there we go the problem was that it was fixed it uses the same seat on both images and that's just not giving you a good result and it the the faces look kind of shitty because because I have the original resolution I didn't upscale it yet right so we can also add that to it right um and then it's fixed but using the same seed on both images just gives you not so not not a goodlook result right um yeah so there we go that's pretty easy and you can see also that this is using smart um it's using in a smart way because this seat is fixed it doesn't render this image at all it only renders the second image because in the second image the seed is changing so it checks what kind of resources do I need per render to only use the resources required um while on the other hand with automatic 1111 if you have the same seed it's rendering the same image 10 times if you want to right um yeah where is Monkey when you need him who is monkey who is that how to download comu ey um just like Google comu ey I guess right com youi and there you have it it's the first like it's the first link and then you scroll down and um here it says installing comi click on that and then here you have a direct download this is a portable Standalone build for Windows right so this should work on its own but you need a Nvidia C GPU as always um and you need an extractor for a 7even zip file window R can do for that for you or anything else you want um so that should actually like work really well just download this file and run it and that should actually work right that should be anything everything you need to do it's super super easy to install right uh we can do for these images for example let's add here a hus upscaling let's go to the example highis fix there we have it boom so we check that out we already have built most of the stuff but we need this part which I can't see right now because it's so small uh we need an upscale latent and a case sampler and I guess that's it that's that's it like two things okay upscale upscale what upscale latent Okay cool so let's try that do we don't we need an upscale model oh the upscale model is loaded with this right upscale method okay cool uh let me go back here here and then we type in here whoa hello my friends a little bit big what what's going on right now oh I'm I'm I'm zoomed in I'm sorry that's my um up scale Laten upscale that that looks good yes oh it's a method uh should be good enough okay and then we need another case sampler we already have that over here bring it here so for that uh we're going to take the output of this latent image and then we're going to put um no sorry we take the output in here and then put it in here and then no this is not it this is not a high-risk fix is it you can download it yeah but I want to show people how to build it right the model for the case [Music] sampler we can try that but it's kind of like something is missing here highest fix no latent upscaling more examples can we have this ah ESR gun es ESR upscale this is what we need so what we do here uh decode we have an image we send it to load upscale model upscale image using model and then an UPS using up scale okay okay okay good so what do we need we have already decoded it we not we need to load the apps scaler so let's delete this and let's delete this um load where is it load latent upscale model loader there we go I want to use I want to use this one let's go with this one that's good enough okay what else what else do we need where is it there we go okay next One upscale image using model all right okay let's do that upscale image upscale with model I think yeah they they the nodes have different names than uh the um anyways okay and then what do we need upscale down scale you can technically down scale it afterwards also if you want to but actually uh let's have here image preview let's put this in here um so what is happening inside of out automatic 1111 is that this is upscaled by the value you see here so if you ever wondered why does it say 4X because the model is actually upscaling to 4X and then you downscale it to the size you actually want to have right so in this case it's a 4X upscale and it looks like so we did something wrong okay this didn't work um I'm missing something here where's the D noise also for that blinear upscaling method upscale image am I stupid today I didn't use it for a long time to be honest wait is there more steps included here we have to encode it again the model using this model H one second is this is not the the complete process I thought it is let me see preview image open image no this is this is like really upscaled it's just not it didn't improve anything of the image H you can use upscale with model node encode but it's already like an image so what are we going to encode here if it's already an image that's kind of confusing but okay let's just build what they built so we have here upscale image which is then setting the size let's do that um is that what it is no pixel no one second let me see here sampling image upscaling that sounds good upscale image upscale image by nearest Bel linear way let me check this out here two times oh but this okay this is something else one second one second upscaling upscale image this is something different man sorry one second this is what we need Okay cool so now we're going to put this in here and then what are they using for that be linear Bel linear let's set this to uh 1,24 1024 by um 1536 right if I'm correct crop disabled let's see what we get from that let's render this real quick downscale what's happening decode yeah this is this is we don't need that part down here uhuh okay let's go on so we have that down scale 05 you need down scale after the four upscale 05 five [Music] um okay we a decode so end code then we have here Cas sampler I don't I don't understand why they encode it first and then decode it again that's kind of my question right now but I guess they have their reasons so one second for this I need my vae which is here and then I need my prompt which is here wait let me zoom out a little bit and my negative prompt and then also my model and then looks like this is everything oh let's set this to 0.3 because K sampler needs latent note right oh I missed one vae one second there we go we can delete this part here so let's see ah that already looks a lot better look at that beautiful very nice Okay cool so now we have a nice image wow that looks actually pretty good look at the nice legs and everything wow so that's good also the render time was pretty fast uh for an image that's upscaled uh how long did it take 6 seconds 1 second for the image 6C for the upscale uh I guess let me let me change the um let me change this here to randomized again so we get a randomized image and then we go back here oops sorry no this did the same one twice one second randomize this is going in here yes now we have a new image two new images and we have an upscale running boom there's the upscale so this is our low resolution image down here with the ugly face and this is the high resolution image here with the nicer face so that's pretty cool very nice details on the hands and everything wow pretty impressive let's look at the render times so we have 1 second for the first image 1 second for the second image because I'm rendering A variation to and then I have 6 seconds for the upscale right so for some reason it says here like almost like 12 seconds for everything um yeah look at that nice arms yeah pretty good right that's ref animated for you ref animat is really good with arms right um so you're recoding you're decoding an image upscaling it then encoding it into latent space again and then let the case sampler work with that upscale image yeah this is I kind of like I'm not quite sure why this is happening with the upscaling and down scaling and stuff and why we need to case sample it again when it already ran through this but no matter it's okay if that what needs to happen that's okay um by the way I kind of want to see what happens if I don't use this let's just connect the image here without the upscaling model and see what happens then so we don't use the upscale model we just set it to double the size and rerender it with the case sampler how does it look it looks good right did it did it change anything let me open this so this is what we got the hand is ugly I'll give you that um okay and now we reconnect it going to use the same was this the same the same um seat I guess right so let's try this again ah no it's making a new image okay let's try this again one moment so here we got this result now okay open and now I kind of not use the upscaler here and set it directly to this without the upscaling model let's try it again let's run it again and I mean okay I guess with the upscale model wait which is yeah with the upscale model it has more details so that's good to know you see like this is pretty nice but it's also kind of like Blurry and this is with the upscale model you can see like the face and the hair has more details so that is what that does you see also the clothing has a lot more details looks much nicer so this is what what this kind of in between step with the upscaling thing does interesting that's interesting to know okay cool by the way what time is it oh 52 minutes okay wow we covered a lot of stuff in that time huh it's crazy invoke do behind the curtain yeah that's yeah what automatic 1111 does behind the curtain right yeah yeah yeah but here you have full control you learn the process you can build it any way you want right so that's pretty interesting h i I really enjoy this kind of process of understanding what is AI doing at what step and how do I actually use it or what's what's the result at that point that I'm getting from that right so that is pretty insane right um and then of course you can have the same thing for this um to just put it um basically over here and another thing you can do here one moment I think we can combine all of these together like that one second and then I put here a group oh no like that and I drag out the group so now this is together you can see I can I can move this around now first of all right click I can bypass this so it's not used anymore I have turned it off if I render now you can see if I click here it's going to render the image down here but not up here uh it didn't render the other one because it was fixed but let's set this to randomized again you can see now it renders why is it always not render the the randomized here if I if I don't change the number also okay now it works you see boom we have two images but there's no upscaling here and I can turn it on again by set group to always and then it's It's On Again I think can I can I copy a group apparently not I think I I thought I can just like duplicate it bypass edit group color remove font size color is there no way to duplicate a group because that would be kind of nice byp pass huh apparently not I thought that's kind of that's kind of possible um yes you can C How can I copy it how can I copy a complete group with all of the notes inside of it because contrl C control V didn't work for me right now clone how do you clone it contrl C control V didn't didn't copy anything for me still here alt drag no also alt drag doesn't work control drag doesn't work um SD no no this also doesn't work it's the same as control um a clone how do I clone it there's no clone command in here shift click no also doesn't work shift click doesn't work so maybe it doesn't work I don't know hold the control key and then Mark with mouse and now oh no this only like it it yes it it copied the notes all of them now I have all these notes sticking here um yeah it it copied oh God copied all of the notes for me but not the group I mean kind of but not in the way I wanted to um control C doesn't work no doesn't work all right um yeah but from this you can basically just like for example have the second image come in here and then you have to connect all of these kind of other nodes by the way you also have this here uh one second utility and then you have here the reroot this little thing here right I can also copy that so for that I can for example put the vae in here and then I can put the model in here like so and now I can simply uh use that as a shortcut to not have to go up uh to the other place all the time right so I need also I can have um let me copy this multiple times here and then we take this as the positive this is the negative and then I can put this in here and here so you can see it's getting a little bit complex yes that's true but um if you build it yourself you still have a lot of control there we go so now we get an upscale from the second image should be ready in a second the first one is upscaling right now as you can see boom that's done beautiful image very nice photo or like drawing of a woman standing by a tree and then in a second the second one should also come unless I forgot something yes I forgot something of course I did why not um there we go let's try this again select multiple noes using control but to move them you use shift click uhuh so here we have upscale of the first version boom another beautiful image look how nice she holds the skirt wow is there another thing missing oh no it's not okay cool and here we have the blonde woman wow look at that very similar dress very similar setup because we're using image to image we upscale afterwards um crazy really nice I like that crazy like of course like small details are different you can also like play with that you you can go crazy detailed with these kind of things if you want to right but that's already like very very very nice look at the hand how it's holding this that's that's crazy I mean here it's a little bit like missing detail but wow that's just so beautiful crazy okay uh what else do we do Mark you can also Mark the note then RM and put it into a template you are much faster what is RM what do you mean by RM RM is what I can put it back down here and then I can for example put these here and then turn this into a second group right add group like so zoom out I take this I put it over everything I can go here I can then uh recolor the group let's set this to yellow and so I know that these are my two ups scalers I can put them on top of each other like that and so you can also get like stay organized if you want to oh RM right Mouse really right Mouse okay cool um interesting let me try that real quick what can you do like this this this this this and then right click and now I do what Mark the node RM into a template oh I can align them color them pin them shape there's different shapes apparently properties save image I don't see any template thing it was the right Mouse button H collapse which yeah you can also collapse them by the way way hey you picked up a 3090 wow nice very good I don't see that right now RM on the desktop on the desktop oh align oh save selected as template oh look at this Okay cool so let's do this and let's call this upscaling all right and now what do we do right click ah templates upscaling boom shakalaka there we have it very nice thank you very much for that they even have the same kind of like uh positioning as I as I had them before and by the way if you already know what's going on up here this gray button that you can see here this is for collapsing it so when I click there you can see it becomes this kind of teeny tiny note here so you you can for example say I don't need all of that kind of stuff because I know what's in there the the decoder is not important this is not so important you put all of that up here and then you just build yourself a nice little interface let's put this down here organize this a little bit like so there you go put this over here and just like that you have yourself a little like compact bit build you see much easier than this one here right much much nicer and of course then you can also have these kind of um what I showed you before what you showed me before better said um these notes here that are called efficiency notes right so you can also use that ah yeah the templates are actually a very nice trick I like that right so that is all of that is very very useful um I would say we download this now let me let me switch here to the big view real quick just to be able to prepare that for a second uh one moment okay that's already downloaded let's unpack this extract what do we have here I guess this simply goes over into the let me copy this um conf I let me find that real quick I have so many folders already custom notes yeah that looks like the same thing put it in here should work like that okay cool so here you can see it um I'm inside of my comu ey folder this is the content of the zip file that I just downloaded I just put it into the custom noes folder and then to load them I have to restart comi that's the only downside it doesn't like re like load new nodes on the Fly you have to close this down you go to the uh base folder and then you just restart run Nvidia GPU bet and this might download some extra stuff in the meantime uh for the for the new uh you see here does some installing and boom we already back and now I should have these kind of new notes in here there efficiency notes you can see down here and there we have these kind of like a lot of more noes now that we can use uh for doing all this kind of stuff we even have an X Epsilon input that is pretty cool right like in comi or XY input as you say I say epsilon because that is how we say it in German there we go look at that pretty cool sampling case sampler I look at this there we have it the model let me zoom in here model positive negative latent image optional vae script if you want to model output conditioning plus conditioning minus like positive negative latent output VA output image output boom all of that in one thing and then down here um you have the sample State I don't know what the sample state is you have the seat randomize it or last Cu uh seat the steps the CFG scale the sampler name the shuer the D noising the preview method and the vae whoops the vae decode the only thing I'm missing here is the preview of the image which was shown over here it might have been a different one case sampler efficient because this shows an image down here but maybe if we use it it also shows an image down there so I would say let's reconnect this um stuff here and then just see what we get from that so here we connect the model the positive and the negative um the empty latent image the vae script we don't need at that point what happens if I click Q prompt does it render anything it does holy what what what what's going on here how many oh it rented two images for me wait is there where's the batch size did I miss something why did it render two images for me uh and why are they different quality oh is this is this already split up in in uh uh like high frequency low frequency is that what I'm getting here that what I'm seeing right now this is high and low frequency for this kind of other other render thing that I completely forgot about but I made a video about it did I make a video no I didn't make a video about it it's for free you right with the low frequency um ah but we need to like have an image output here so let's go here um image preview image put it in here let's see what happens have you the sdxl sampler no I don't aha so now we get a normal image as an output and then we can also upscale it right uh so that's pretty cool wait a second so the manager doesn't work why doesn't it work it's a chasing file you open from load for example loading multiple luras why do you dislike loading multiple luras I don't understand like with these kind of simplify notes or compact notes uh you can also load multiple LS in in a single note right so that makes things a little bit easier uh let me connect this to here so that we have an upscale from that just render it again ah we have a render animation also I like this beautiful okay so now we get this here oh and this is also animated now what's happening right now magic oh look at this beautiful lady oo nice dress very nice Okay cool so this this worked that is that is amazing and and why does this suddenly have a preview did all K the other K Samplers had a preview before am I like stupid or what uh okay huh anyways uh what do we have here hris fix script oo hris fix script ladies and gentlemen one second Hest fix script are you kidding me right now upscale by two to nearest can we highis steps D noise really is that working right away one second let me DEC connect this and then maybe wait a second uh let me turn these things here off for the moment um let me put them in a group here no no go away like this bypass okay let me check this is this a highrisk fix without a upscaler no way oh no it isn't is it H it is it is high resolution so it does it does do upscaling it doesn't really fix anything H did I close it no it didn't okay cool uh set is to 25 steps uler schedule normal D noise auto prevum mode decode hus steps wait when we set them to zero does it mean it's the same no it's probably not the same right let's set them to 252 and then set this to 2.5 2.5 mm didn't like that okay let's go here set this to six one moment you should check out next time the comu Roy notes for Rock of fire and AA Suzi hey akasu isn't she in my uh like Discord um I think you using the upscale of lower and stack ah yes right okay let let me check this I mean it might be that this preview is already the upscaled version I'm not quite sure because this doesn't look any better at the moment uh let me let me note this down comfy Roy I will make a note of that one moment okay thank you very much ah triple head monkey Hi how are you doing I didn't even read the username thanks for joining there we go man when I click here it opens up this kind of like admin tool which I don't need I just want to copy the text there we go sorry for that okay cool you need to change the sampler state to script go to manager preview T SD what does that do preview method oh okay interesting what do I need to set to script sorry um sampler state to script ah ah H la la let's see what this does so it's it's doing its thing again now I set it to be linear I don't know if this the right choice or not so this is what we get over here which already looks a lot nicer from the face it just doesn't have a enough detail and this basically looks the same so I guess that the thing we get over here is already the upscaled preview let's open this up yeah that's already the upscaled preview this is why we don't see any difference between this image and this image all right but do I have the ability to use a like why don't I have the ability to use an upscaler over here upscaling model for that that would be nice right um let's go here and see on the efficiency notes sampling load stack M scripts Epsilon plot sampling case sampler case sampler Advanced optional vae script ah starts step end step this does have some extra stuff in here but strangely enough none of them have anything that uses an upscaling thing right loaders efficiency loader what does this do clip skip Laura positive negative batch size because most likely you're upscaling the latent image not the image uh that is that is true however they could have built inside of this note away where this is like because here you can see I have an output for latent image so why not have inside of here a way to put the um you you know like in automatic 1111 I can I can use for the highis fix an upscale model which I can't I can't select here confy roll notes have a multi upscale stack with models wow okay cool I guess that's a little bit too advanced for right now I just wanted to see what this uh these kind of notes do however you know what we're going to do right now let's go over here and just see what kind of stuff oh it it might be a little bit too much I thought Lowa stack control net stack apply control net stack MK unpack sdxl tupple I don't even know what a tupple is script except XY plot XY input image overlay uhhuh so I guess we still have to at least for this moment go with the method that um we are not using this as a script instead we outputting this and then putting it in here right so we have the upscaling model from this one which still reduces a lot of nodes basically with this so it it makes it oh didn't like that huh okay um I guess the reason for that might be it should actually work try it again um H I'm bit confused what's happening right now one moment randomize five they should all be okay right because it's the same thing we use before and then we have an image input here that comes from here and these things down here we have turned off so they shouldn't do anything we can actually delete all of that stuff to make everything a little bit simpler here uh edit group remove edit group remove oh okay one second just get rid of some stuff here because we don't need that anymore all right so that already looks a lot simpler um because it's 4X no no no we do the 4X and then we go here because this should output this is outputting a 512 by 768 image right oh maybe because of this let's set this to Auto oh and then set this here to sample sorry I was still set it to script so now it should work actually yeah it's working again so you see it was just like I set it to the wrong thing that was the problem okay so now let's have a comparison here one second bring it over here and see how they look yes this is what we need look at that the hands are not so great but that's okay look at that face and then look at this beautiful face hello hello hello all right very nice this is exactly what we need um so it kind of like the bundle is missing a highr kind of thing at least as far as I can see but I probably there is some other kind of um am I like crazy like here it says the model uh what wait a second script hris fix what hus fixed script latent upscale method SD latent upscale version one so why do I not get that choice this looks like an upscale model right or am I wrong SD lat upscale version one or does this maybe this needs some Stu inside of the folder I didn't up I didn't download that might be the case let me check real quick uh there's the repo currently a case sample efficientcy loader blah blah blah simple notes blah blah blah no um could combine the note with ultimate upscale right so why why can't I load this thing in here this is the future this is absolutely the future because like uh I mean if you if you go to other professional software let's go for example for blender notes this this is what the pros use right here this is this is a professional 3D software and then this is what you use to to to set up how the materials behave while rendering right so um yeah if if you want to be serious about that stuff I guess you have to kind of figure out how to use notes because you need all the kind of control in there right this is like for the 3D program to to show like Define the rate tracing and the material and textures and all this kind of stuff um with these kind of notes because it gives you so much more control and flexibility and you can like build it exactly to what you need so that's what's uh why this is so useful I just don't I don't understand why I don't get the choice here to put something of my own um kind of missing something here uh I use both we can do everything yeah you can use both too you can use both too yeah uh 3D Max is using it the Venture resolve is using it um Unreal Engine is using it right everybody's using like notes because it's just like so much more flexible for what you want to do and also because it allows everybody to like create other notes for you right um so is there like a wait a minut wait a minute uh comu I ah multi- upscale node that sounds good what do we have here oh multi upscale apply multi- upscale rounding mod modulus okay that's I don't know what that is uh load image preview image rescale this is like you can just go from this to this and it's just like like that really okay we we going to try that right now let me download it's only it's only 2 kilobytes wow that it reminds me of the good old age of like amga 2,600 uh wait a second we need to go let's let me keep this open let me go go here to the downloads We have this oh this this just like two Chas in files Let's do an extract here and then I guess we're going to copy this over there custom notes just I don't know let's just let's just drop it in here and then we restart and see what happens worst case it doesn't it doesn't load let me check this what are they called CR upscale uhhuh CR CR no it didn't it didn't load uh do I do something wrong here these CR nodes are made by aat Suzie yes how do I install them like do I put the Chasin file just in here without the folder or how do I use that custom noes [Music] um do I put these just in the custom node folder load missing notes yeah but there's no missing notes because I'm not using it right now right click the load button but then I have to load it manually that's that's no fun are these builds they are not oh these are builds they're not notes do I still have them in downloads let's extract this again wait a second load no let's not load this let me open up a new confu ey so we preserve the old one and then we load this downloads conf ey upscale uh template one template 2 ah load image batch oh even even batch loading okay let's go to the manager install missing ah install takes a little bit it's still installing so you can see down here what is this downloading it's like crazy is it that big I hope it's not like a six gigabyte file or something like that hello nothing happening let me click on this real quick it's not opening [Music] H H did it crash something is not right hit enter here doesn't do anything the was notes you is Big a okay ah not sure that was a good idea right now let's see if anything is arriving over here where is my folder the hell C youi ah there we go properties three megabytes that's it empty uhhuh H this might not have been a good idea to do that during the live stream comy roll custom notes Rock of fire maybe we close this being updated and update to custom notes are being checked for update only check for ex extension have been fetched uhhuh what do we do now I mean I can close it and just delete it or just like restart it this is the good thing about com VII because it's running in this kind of extra folder with these custom noes it's basically nothing um happening B if you if you just break it you just delete the folder uninstall should we uninstall that let's uninstall it for now there's a lot of Errors okay let's close this sorry one second give me one more try and then if it doesn't work we just going to do something else are you kidding me right now all right uh maybe because the browser windows are still open I'm not quite sure let me close all of that yeah so something is using is still using that folder for some reason I don't know even though I closed the um command window all right what's the time by the way 1 hour 29 minutes uh we can still go back to this doesn't work this doesn't work we can still this doesn't work yeah we can still go back to this one let's reload it and we still have that let's check it if it's still running boom boom boom rendering first image done second image done did I not have the oh okay but I can just load it oh there it is the other one okay cool let's try this also this is by the way this is also a benefit of com youi is you can have different work workf flows loaded in different windows and then they are just staying as they are unless you do something uh crazy uh or like close the tab and then you still drag the image into it right so you can still do that right I also can show you that right now one second so I go here I clear my canvas clear everything is gone completely empty canvas so now what I can do is I go to comu ey I go to my out put and I have here my images that I rendered last these apparently didn't work we can delete them and now I can just drag this image in here and as you can see it's loading my complete workflow from the image now this is the one the other one that that's not what we want uh the good question is where is the one that um one second let me for example open up this image here save that image on my download folder and then I can go here let's go to download folder that might be a little bit easier and I direct this in here and you can see it loads that build for me and you can also see like how much much simpler this one is uh with this like compressed um with this compressed efficient Cas sampler and we didn't even use the input for the prompt here so we can do that right now I right click go to notes I go to efficient notes and then I have here uh the case sampler the loader stacker oh Laura stacker we forgot about that MK uh unpacks scripts inputs um doesn't doesn't this have like a input for the prompt yeah it has efficient loader okay it's called efficient loader let me go back here so I go here and there and then the loader efficient loader so now I load here ref animated I can type this up here ref animated then um I'm using the vae it's all in here Laura we don't use at the moment here's the positive negative bront I can put the positive one here the negative one down here this is my latent size this is okay this is a little bit overlapping sorry for that one second um so here we go 768 batch size one okay so we have eliminated this one this one this one it's all gone this one is also gone and then we connect this over here model positive negative latent image vae uh is also in here we have also um like uh deleted basically this one uh let me make let me make a note over here real quick um reroot yes reroot put this in here put this in here connect this to here good so now I can also delete this one and we can also delete this one because that was just for the second um negative prompt right and when I find a note that has all of that stuff here combined into one note I only basically have three notes left for all of the process and another thing we can do here is we come we close these down here like that the preview we need this we don't need so I put all of that over here in this corner like so put that here put this here put that here make it bigger like so maybe and then I can put this here so we have a compacted group also and so now we have this right right and that I would say that looks super simple right this is not too complex compared to the kind of uh little bit monster we had before I can put this also in here if I want to let's put it here maybe so now when I click on render oh this is is missing something ah one second the model goes here the positive goes here the negative goes here um there we go connected and it's already rendering and boom shakalaka there we have it this is our low resolution image this is our high resolution image face could be a little bit nicer I would say but yeah pretty cool look at that look at these nice the nice dress and everything looks very very good very very nice Okay cool so we have everything in here um with just that build how how great is that huh 93 minutes ago what was 93 minutes ago when did the stream start no the stream started one and a half hours ago oh yeah that's that's that's 93 minutes yeah that's true that's okay yes yes yes yeah I would say that looks very tidy I guess I should like maybe include that in a video because I was thinking about making a video of like maybe like 10 reasons why you should use com UI and that might be more convincing than showing that kind of beem moth of like 50 different noes connected although when you start out I would still kind of argue to use the basic nodes at the start just to check out um what is what is like each node doing and then use the simplified version but maybe that's just old school talk and you don't really need that by the way another thing you can upscale this you can make it bigger so now you can actually see the positive and negative prompt in here right okay uh what is the convincing reason automatic 1111 but what about comu ey doesn't it also have its own challenges in terms of usability um well it in the sense of usability I would say with confu ey if I can if I answer that um I would say it has a steeper learning curve because of all the stuff you see here and because there is like a ton of different nodes you can use however once you have worked your yourself into the process um you can uh build some really interesting crazy things with that uh and like pretty like complex processes stuff like that and also things you can't really do with automatic 1111 like one thing I haven't showed you before but I do have a video about that on my channel which is called latent tricks uh where you can go and uh where can check out this thing that I'm showing you uh from also from here area composition this one here so there's two things the one is area composition the other one is basically similar to that so you can see here this is an image with a sky like galaxies then we have here a nice sky with a Sunset and clouds then we have here landscape right so these are basically three different images that are afterwards rendered together into one image and this is what is happening here with the process now of course you can imagine um if you use these simplified notes then you don't have to have this kind of uh like massive complexity in here that you have but also when you start out and you just look at the prompts you can see H at the notes sorry when you look at the notes you can simply see that these are the different prompt so this kind of easy to understand what's happening here and then you have here something that's called condition conditioning set area so this is basically the area of the image that's going to be rendered in that kind of um in that from that prompt right so that's also kind of easy to understand so here you can see that the image has a size of 74 by 384 but also this has a x y coordinate of 0 0 and then down here you can see that the XY coordinate is 0 320 and down you can see it is z for x and for y it's 512 right so from that you can easily imagine that this is rendering the first image the other one is on top of that and the next one is on top of that right um so different areas of that right and then you can see this is then combining these two images and then combining the combination of two images into this with the last one here right um and then afterwards this is then rendered um it's going down here into a sampler to render this and then afterwards it's it's rendering everything again over here with the complete prompt that describing everything um into the complete image rather than just the individual images right um yeah so just by looking at the prompts it's it's like not super hard to understand what they do there's some extra prompts down here so I didn't explain everything but um you can you can kind of see that the pros is not that difficult and another thing you can see here by the way is um we saw this one which is pretty cool with landscape only but you can also render a high quality high resolution character on top of the landscape afterwards right and the the um the way this is rendering or the way this is working is um if you have the lat image and it's rendering the first steps basically what is doing is it's it's rendering it's turning noise into an image so the the way that defusion works this is also where the name comes from defusion is um the AI learned to turn an image into noise and then try to turn a very similar image from the noise back into an image right so when you do the defusion you start with a noise that is the seat and then turn it into an image by arranging basically the noise kind of a easy way to understand that H so you sort the noise a little bit so when you preview that you can already see kind of the like a little bit of that girl and a little bit of the landscape uh so you prepare these different areas and then combine them afterwards and render all of that together as a prepared noise so that this already has a setup um and the reason why it's done in this way is because then you can have a different prompt for that part and different prompt for this part different prompt for that part so the AI is not getting confused and then on top of that a different prompt for the girl which is describing everything here and then afterwards everything is combined with the noise that is already prepared for the process right um might sound a little bit complex but it's it's not that complex actually right uh what does the what does the chat do right now I don't lot talk here let's see um there's R for comu ey uhhuh R IP adapter control net everything is avable yes like all of the things are available in comi U and they are often most like earlier available in comi for the reason that uh yeah the community can build the notes in any way they want because um comi is basically just a sand boox you can put in here whatever you want in this kind of environment um and and just use it right so if you find a new technology and you're good with coding or you know how to use or create nodes with ch GPT or things like that you can do custom things in here right um yeah and then just use them because this is completely open this is the kind of like like Advantage compared to to automatic 1111 um what else do we have here uh the answer for sufficient how strong is the video production o I guess maybe triple hat monkey knows about the video production because I have to admit I have never done a video with comu I but I'm pretty sure that is possible too right it's just a code method it's very easy to implement there you go okay right uh what else do we have here has visual conditioning area which makes it so much easier have a build of a seamless tile generator oh seamless tile generator that sounds pretty good um can I use multiple a details in comu ey yes basically Bally you can right you can use multiple ad tailers by the way that is that is kind of an interesting thing like um with the method that I showed you before by the way there's a different method in here that shows this even better uh latent uh noisy latent composition this is what I mean so um here you have these characters now this is not 100% last time I checked it uh because after all at the end you have to use a prompt uh to render the rest of the steps and now if you have here characters with completely different clothing you might run into a problem a little bit but um you can you can use different ad tellers or I mean h or different luras on these different characters and then render the image finished that is one way you can do however another thing you can do inside of Auto of comu ey is you can also use masks do have an example here for masks uh in painting you can do you see you can like mask that area out and so you can just use as many different noes or luras or stuff you want to do and also they are running one after another they're not running at the same time um so yeah that you can you can do this kind of process and for example if you would use control net you would already know where the face is so you basically can prepare an alpha mask for that area where the face is going to be and then just replace it or or of course if you use a detailer a a detailer is going to find the phas for you and then replace it however um I'm not sure about automatic uh I'm not sure about com VII but inside of automatic 1111 um a detailer can identify different faces you can number the faces and then render different luras to the different numbers of the faces right so that way you can also like create different characters in one image with the same render method right with the same render process right um video production is currently behind on comu I but all the developers are just starting out this month or so releasing stuff oh that sounds good okay cool interesting if you ever want to take a break from imagery you can do a live stream or quick video on ugab Booga uh yeah I can do that I experimented a lot with ugab Booga in the past like now it's also very easy to install I'm might be interesting the only thing I have to admit is I don't I just don't know what to do with ugab Booga other than um like writing your like person personal kind of fictions like short short stories and stuff like that it's really good at that if you want to write any kind of story about any kind of character that works uh fairly well for other things like talking to it like um asking questions stuff like that I find that um jgpt is much much more powerful like there is not not any model out there that is even remotely as powerful as as um Chet GPT of course because CET GPT is also a much much bigger model and you would need a like a crazy computer or even a server to run that um so it's it's just not on the same level that's kind kind of the thing and the other thing is also the models you have for ugab Booga are mainly English models like I try to find one for chman not really good like they they give you just always like crazy output like most of them is very very focused on English language while on on CBT it's very easy to have um like all kinds of different languages and it works very well uh so yeah I would say I could make a video on that why not I I could do that but I have to talk probably also a little bit more about uh that with the community uh actually with animated diffusion you can make till 10 seconds of a video but you need a very powerful graphic card the an animated diffusion or any diffusion I wasn't really like happy they have a standalone version for that that apparently works very well I have never installed that I used the automatic 1111 extension and that didn't direct really good give good results but maybe the comu ey version is better for that I don't I don't quite know um yeah so these examples here are not great of course that you can see here but you can do masking and then render stuff into masks uh what else do we have should I go to the big screen maybe let's go like this for a little bit it's also like nearing the end of the stream anyways um samd what is Sam detector masks are created search workflow for civi yeah there's different workflows different notes you can download from on civi so Civ is a really good source for all like a lot of stuff for um uh comi of course also GitHub is pretty good are do you want to check out for that check out what stroubles is making in the share I art section of the Discord on my Discord or what do you mean is it on my Discord or another one stroubles segment anything oh yeah that's also a call from the past oh it's probably on my one okay I will I will note that down one second there we go I will check that out okay cool and then next week I hope next w Wednesday we're going to do the uh stage event again in my in my Discord so people can show their works and stuff like that that was also very very um very uh interesting and useful I still have a backlog from that by the way um on on stuff that I want to create on my channel I had I had so much back pain this week I couldn't do like I had a full night I couldn't sleep I was just lying on a on a hot um what do you say like water what do you use when you're cold like you fill it with this like rubber thing you fill with hot water I was just lying on that and then the whole day after that was also bad I still have back pain uh right now but yeah tomorrow going to go to the doctor get that checked out uh but I couldn't do like videos for two days it's just too crazy um he's a master of comu ey animate diffusion oh okay so I will check that out that sounds very good and that could be an interesting video I'm not quite sure if my main Community is into that but I kind of like want to push people to try a little bit more right because I mean one thing with AI is if you expect a a good artwork that is worthy of being an artwork from just putting a prompt you're not going to go to the like highest levels of what you can achieve with AI you just have to like and even if you use all of these kind of extensions with control net and so on and automat 111 you can do a lot of stuff with that so there's really good people who do like mainly their stuff with automatic 111 but you can also see from all of the other professional tools out there like I said with blender with uh denty solve with Unreal Engine and so on and so on that they all use nodes because you can just build much more complex processes that then also um render in in in one go because uh one thing that's important at that point is to understand the difference between doing it manually in comu ey where of course you can say well I can go render the image then I go to image to image then I go to extras afterwards there's multiple steps in between yes that's possible but if you see it as a workflow in front of you and how things influence the other things first of all it gives you a landscape an overview of what is happening uh but also it gives you a lot of hints of what you could do in between and the kind of connections that you can build and a lot of these connections are not even possible inside of automatic 1111 and that gives you a huge advantage in the sense of how can you actually play with AI so the stuff we looked at today is fairly simple but just like creating an image and upscan is nothing special you don't need uh comu ey for that but you have seen the flexibility of the things you can do and just leading for example latent images as an input into other stuff that you do like one thing um we could actually like we could do right now is uh to to go from one model to the next model right so you can see where's my mouse there it is okay um yeah discod is cool um hot water bottle there we go life can be so simple if you try if you don't try to overthink it uh where can I join uh that's a good question let me find that real quick I hope I I get a link that works this time I think last one it didn't really work or it like expired or something invite people addit link no it is on No Limit and never I don't know why that didn't work okay um one second uh whoops there you go this is my Discord server you can show join there if you want to uh this triple hat monkey is a mod there also by the way I I think you are right yes you're a mod there we go and an expert also there we go right so that's pretty cool yeah so we have very very awesome people in there very nice um okay as a last thing I guess we can check out the thing where we render the same image with different models uh just let me figure out how I do that real quick latent image optional should be fairly simple I give it a shot and we see if it works one second so the one thing I would do here is to just have a checkpoint loader simple and I load here for example realistic Vision which I don't know if this is the newest model but it doesn't matter right now um and then let me see we're going to copy this put it here and then I'm just going to put this model here and I'm going to use the latent image from here and then we're going to just connect these things let me put this over here so we is a little bit simpler put that here put that here and then afterwards the only thing we need to do we can just put this up uh well actually we don't need to do that right now that should be enough yeah I guess for the moment that's enough let's cue this up so this is now loading realistic Vision also and there we go you see already done so I have this image and now it's transferred into okay the D noising is very high let's set this a little bit lower set it to um let's go with 0.4 let's render this again there we go you see so here I have the style of ref animated and he have the style of realistic vision and of course now we could upscale that um okay the question here is upscale image I guess it comes in from here wait I was a little bit messy before yeah this should be the input so let's put this in here oh um wait sorry I was a little bit messy before now I don't know where what goes image image oh no it worked me open this up yeah that worked okay cool let's put this back in here try this again now it's rendering one [Music] second ah okay now now we're done no am I crazy right now what's going on here oh no this is putting it in here come on what's going on here this is the vae encode bring the picture from here to here oh oh oh oh I see my I see my mistake one second I put it in the model and then I put it back from the model again I need to have the model from here there we go okay that's the that's the thing that happened here you can also combine SD XL and SD 1.5 that's true that's another thing that doesn't work in automatic 111 so you see now we have this rendered in the style like from the posing colors everything in the style of ref animated but in the realistic style of uh realistic Vision right so this is one thing you can do right it's random Riz this here and randomize this here and then we do this again so we get another image so here you can see different image in the ref animated style different image in the realistic Vision style but the same composition and colors and now we do the upscaling and then there we go we have it um as a realistic kind of photo of course you can improve it you have to play a little bit around with the D noise setting and peo all these kind of things we can go here a little bit higher maybe and get a better result from that but like of course yes you can do that inside of automatic 111 also but there you would have to go in between the text to image and the image to image render afterwards and here you can see there one flow and that can say get some more ideas from that right you can for example also like inject here different kind of prompts from that um to maybe change the image from a day image to a night image you can create a whole series of images based on this kind of concept right or you can render it in different kinds of like um different kinds of models stuff like that so um seeing the process like the inspiration artistically is just that you see the process and this gives you inspiration on the workflow and what could be done differently or where where things could flow um creatively by reconnecting this in different ways right so that's basically the main point to take away from that right and that's like you can really you can really like um compare this for the way for example also modular synthesizer are used right um like if you have things like this of course it looks like a mess if you have this kind of stuff here let's let's make it a little bit simpler with this here but at the same time you can combine all kinds of modules you can connect them in any kind of way and then you come you you like you figure out in very short time things you never would have thought about um connecting it in that way right so that that makes it the interesting part here rather than you have a fixed situation that can also give you a lot of sounds but they also are limiting you in the way that um you can use it so you always think about it in the same kind of limited way you know if that Mak sense anyways um yeah that was kind of like pretty cool like you can you can do really interesting things with that right with this kind of model changing model swapping I call this model swapping this kind of process but um yeah you can you can do even more right like you can for example of course you can render an image you can use it as an input for com UI and then use the output from that as an input for something else for a different model that would also be a process right instead of using the image directly because then what you have is the composition that the model is creating but you don't need uh you don't use the colors or the clothing or stuff like that that the model is trained on and so you can switch it to another model and have the complete style of that model right there's a lot of different possibilities all right uh let me go back to the big screen here or should I keep that in the background is that nicer to look at that while I speak with you um there is modular synthesis note based UI called Native Instruments I love Native Instruments reactor I actually buil built multiple modules for that not modules sorry that's I built multiple um uh modular buildups for that we can actually look at that one second um let me see if we find that my music oh no these are the forums that's not what I mean M oh no wait um Native Instruments I I think I use a different username user Library let me check this real quick accept all how do I get to the user oh there we go Orbiter yeah baby look at that oh Olivia SAS there we have me um Orit 4 MXL uh this is um this is built on I think only on noise right based on noise generation right so this is built with these kind of reactor blocks the input is not like sine waves or anything else it's only um only noise as an input but the sound is pretty cool cool actually have to open this up in my second computer so you can actually hear that wait I'm going to play that back for you real quick and I going to mute my mic so there's no feedback loop one second so you can hear that sorry I forgot to put on the sound one moment [Music] [Music] [Music] [Music] all right there's even more stuff but never mind um I think it has a lot of presets in there from the community um I think like 128 prese sets or something like that I'm not quite sure 2,000 downloads not too bad okay cool six years ago wow time flies crazy VC VRE is also amazing but I haven't done much with that right VC VRE is really cool yeah it would be fun to do like a like a music live stream but I don't think my community would be liking that um because I really like all these kind of different like uh tools and synthesizers and stuff like that uh also like the Atura collection with all the classic synthesizers right so that's pretty cool um The Sounds might not be so crazy but it's interesting that doesn't have any oscillators in there so I thought like the the the presets have been a little bit like interesting like these are just like previews of the presets right it's not like supposed to be music it's just showing you different pres oh it say 200 presets right here I'm blind right and you can do a lot more with that right and the just the idea is not to use a classic synthesizer as everybody else with these kind of different waves but like just noise so you don't have to play around with the oscillator you play around with the settings that's the kind of idea right Music stream maybe on a different Channel it would be nice right yeah AI VSS are going to be a tight a thing I bet AI oh yeah AI bsts yeah I I am still waiting for something that's good right did you get woke over the keyboard yeah we could have done an actual like demo track that would might probably would have been nicer than this kind of thing right but the the intro music you hear every time when my uh when my stream starts it's actually from me um have to download yeah try it out like I have I have a lot of other stuff also on there oh I wanted to make a serious back then this was an idea for a Channel about modular synthesis and I built these kind of setups and then explain them in a video and then do the next one so it's going step by step but that didn't like really I think there wasn't too much interest for that um there we have one that's based on a another build that's is based on the Uno and then we made another like 150 presets and I did some adjustments to make it just sound fatter from The Sound let me actually check how does this sound wait a second if you have if you have good um speakers bus speakers this might be something you like one second let me go to the second screen here real quick listen to that there you go that's some that's some pretty nice subbs for your tracks Sy G yeah I did some like several things and also these kind of like value sequencers six times um 16 values as a block right so you can just set the values and then you have an output here um so it's like these compacted notes stuff like that and such um different value faders yeah and things like that uh those kind of on and off switch for cables so you can like just have 16 inputs and then turn in on and off without having to disconnect the cable every single time right these kind of things right anyways could be that could be another life right different kind of um Ker anyways maybe that comes after my AI career let's see the music career or the photo career which I also do like more photography lately um yeah there we go one shop of AI art and synth yeah right and and other and other art forms right Max for life no I never tried Max for life never have used that one that one might have been nice but I I didn't use that by way let me let me uh I go here to the big screen real quick because I edited some more of my photos uh that I could maybe show you one second where do we have that this is the preview but that is not fixed um is there did I delete that file already um oh there we have first part one second you have already seen those yeah they a little bit mixed but okay so one second almost there just have to push this over here so this is not AI related now it's just like end of the stream talk there we go so here have some more of my photos um from last time I showed you a little bit this is like a fashion shooting uh kind of thing that I did uh a week ago I guess and then I also have versions of that where the background is black and white so you see the dress a little bit more because this is kind of a golden uh wallpaper and kind of mixes too much with the colors so this is without the colors pretty nice and there she's sitting on the ground just doing like uh more amateurish not like professional fashion shoot but still for a fashion store right she's a very nice model also very good at posing this one I like a lot uh the cropping is maybe a little bit too tight but I didn't have a a lens where I could go far away uh where I could have enough of the room because the room was kind of tight but my lens was also long Focus so I couldn't get like a wide angle enough to to include enough of the of the room so yeah this is this the color version I don't know why they are so like ah there we have it like different like okay yeah it's also my folder can could have moved the camera a little bit more to the side but yeah these are the kind of like uh yet another life that could could happen there's so many different things you can do right um there we go and I think for most of the part the only the only thing is to to just do them and try to see if they work or not right um especially with Arts there's not like any kind of requirement or limitation of what you can do or can't do so if you if you find people interested in the stuff you do you do them if not well I don't know right so um it's just like um yeah trying it there we are in oh there's a little bit of the um of the chair left I'm seeing right now this is in a barber shop I just asked them if I can take phot very nice wood here on the background this might be better with the color photo but I wanted to have this version also but I have to fix this part didn't see that um yeah so this it's interesting right with life you can do so many things but just like trying that especially if it's in a creative area you can do all kinds of things right and everybody has some needs everybody wants some and as long as you have the passion you can um kind of do them right and most creat Fields like they have a hand of like a handful of different creative rules that you should look into and a good way to get started is just look at what other people are doing right like with photography for example what I do is just I go to Pinterest and I I search for different poses I never do the same poses um because uh the the the it they have to fit to the model of course and the model has to feel good with them but it gives you some Inspirations what you can do with that right and it also makes you less nervous in that situation and it gives you better results because at least they're inspired by people who are much better than me as photographers so I already know like for example how high does the the lens have to be like do I have to kneel down on the ground or does the lens come down from the top and does she lean towards the camera or away from the camera stuff like that so it really you can do a lot of uh uh shortcuts like that hacker man could start a music stream that is a good idea actually right not so hot anymore under the Wick yeah that's true right right right when the um when the eii stuff is going down I I'm starting to do my my hacker man um cinti channel that would be nice right right right yeah and I I I think like people still going to do their own art kind of thing and then AI is just like supporting that right like uh think about the crazy potential that artists have right now with all these kind of Inspirations you get from Ai and then on top of that if you have the skill that you're really good at drawing or really good at editing or photo bashing stuff like that and you can just edit it the thei images into something that is much more perfect to what you want to have you can do crazy things and then you have all this kind of experience from the color harmonies the composition the storytelling all these kind of different elements it put in the the the Dynamics the positioning of the body all that kind of stuff that helps you so much right with all of the control net with all of the imageo image rendering and understanding of what is a good image what is a bad image stuff like that what works really really helpful right so you can go very far I don't think that people like just throw away things just because AI kind of can do that um you still yeah like you use your skill and this is then an extra tool inside of your bag right that's the that's the only difference right that's happening but but um knowing all of these kind of skills is still good and learning them is fun because it's basically just means you look at good art and analyze it that's the only thing you do like when you go to an art school most of the time what you do is you look at the works of other people you analyze them of course the professors uh tell you about like how to understand them correctly how to read them correctly of course there's not like a complete correct way to do like what it means but um you tell them okay I see this and that in the work and they say yeah but it's more like or or they let's say they know more about the background of the artist the biography the people that the artists know the kind of philosophers the artist work with stuff like that so they can like Notch you more into that kind of direction that helps a lot of course you understand that sorry I should look at the I should look more at the chat uh still many six fingers oh yeah the six fingers right is that would you think like that might become a fashion that people wear a six silicon finger wouldn't that be funny right if the youth picks that up and they just wear like extra limbs because of AI I think like that kind of for some reason you know at at some point like you have seen six fingers so often in AI that you kind of feel more at least not more comfortable with it but um it feels normal right at some point where the like extra fingers you just like yeah just like what it looks like right so that's kind of interesting kind of change um Vera says I think AI has re awakened a childlike love of Art and computers I thought was that absolutely right this kind of experimentation and exploration especially because this can do all kinds of things um like for example a friend of mine is a painter and then we uploaded some of his images to Mid Journey because we didn't have time enough to train aura and we created variations of his images so um he can use them as an inspiration as ideas to create his own paintings right because the AI doesn't have a mind it doesn't know what it's creating so it creates often a lot of goofy details in there but they can create connections um that are uh artistically very interesting right and like just image compositions that are that are just interesting we can actually look at that a little bit let me see I don't think that's a secret um because they are just like sketches so that shouldn't that shouldn't be much of a problem or I can just like show you some um wait a second uh what we do here let me let me quickly see if I find something that might be interesting to talk about there some crazy stuff in here um oh yeah there's a good series here let me switch over look at that here real quick so here let's let's maybe open this up a little bit bigger so this is not exactly the style he's painting in his his paintings look different but it they're kind of similar in this kind of like abstractness but they he doesn't have this kind of color shift in here these kind of gradients and stuff it's not that kind of painterly if you might but um because AI doesn't know what it's doing and we we based it on his artistic style and said we want to have a tower uh AI mchy just created Towers inspired by his image input but of course here a lot of interesting like crazy things are happening in the image that's kind of strange like for example here you have this kind of like letter or whatever it is it's going like it's ending in nothingness is Darkness down here and it's just coming up here with this kind of strange contract uh construct sorry so that's kind of strange and then the whole composition that you have this here in the middle then you have this shape here in the background that might look like a giant head so there's a lot of goofiness in there that you would never use in a painting um maybe you would use it in a painting about like a monster or something in like a comic illustration or something like that but like for a fine art it's kind of a little bit too on the nose but um then up here for example you have this kind of like floating figure here which might wear a skirt not quite sure it might be a bird head not quite sure uh but it's kind of an interesting concept and everything is so centered right the tower is completely centered the figure here is centered right and it's it's just a very interesting concept it doesn't mean anything but you can take hints from that that's a kind of interesting starting point here we have another one the Moon is kind of a little bit um uh strange uh or here we have this one with this kind of guy standing here it's also nice right there's often these kind of like strange details like this guy and then also the the idea that his either like his his upper body is a complete shape you can see like a pill but then the arm is kind of like weaved into that and then kind of like going into the pants down here but everything is one shape which is which is also strange I wouldn't like I don't think people would necessarily draw it or paint it like that and that's also good inspiration and kind of the way just like the way he stands up here it's just yeah you wouldn't necessarily think about doing it that way right there's a lot of like different little elements in here that are just like strange but at the same time interesting then you can use them as inspiration points right or this just like just like this this for no reason shape up here right and often like a interesting thing is that or this kind of thing down here an interesting thing with the human mind is that we find reason in things even if it doesn't have reason right so even though the II doesn't know what it's doing the interesting thing for you is that you can find reason in that and then turn it into something in your own artwork by using that kind of ele element maybe in a completely different painting and a different artwork but just like use this kind of shape in there for some reason that you figured out is artistically interesting for you right um yeah so that's that's just like um interesting interesting things that AI can do for you there's a lot more like we did dozens or hundreds I don't I'm not quite sure right and all of them have some kind of like yeah it's kind of kind of interesting elements in them right not everything is perfect but like for example here you have this guy sitting on a letter or St like crouching on a letter for some reason it's also kind of strange right and then up here you have this kind of like is it a horse is it a person there some for some reason it's kind of like this kind of bar situation up here with a window behind there so yeah it's just like interesting strange combinations that might inspire you to do something or what what is this this thing here you know is it an elevator is it like a you know for these kind of columns for advertisement is it for like a phone is it a toilet could all could be all kinds of things you know it's very strange okay anyways let's go back to the chat sorry there another very long chat but like using this for inspiration is pretty interesting I guess David cronenberg's last movie was sort about it I haven't seen that actually what was the last movie from David Cronenberg wasn't it the one with the with this kind of gas tower um this kind of gas storage wasn't that the last Cronenberg movie wait I amdb let me see David Cronenberg Star Trek he worked on Star Trek this covery what the hell oh actor okay um Crimes of the future interesting uhhuh I think the one that I mean is it maybe Nest the one I mean no that looks different I'm not quite sure to be honest is it just a short film no that's not what I mean I have to check that out I'm not sure okay uh where's the chat second um I already saw some silicon China six thing oh yeah really in in China they already created like Silicon six fingers that's fun um the idea behind the six finger prosthetic comes from the fact that a prosthetic is not just a replacement mud and an augmentation it's if someone is missing a limp replacing it is an augmentation that's also true right like um um I think often we think like we couldn't handle having extra limbs or maybe an extra eye or something like that because we are so tuned to have these kind of like five fingers on each hand and like two eyes but of course like if you have a person who is born with six or seven fingers on one hand they can just use all six and seven fingers no problem right so I guess if you for example would have a like an extra eye or an extra ear um technically you could like a adapt to that over time and just like have this kind of extra part of vision that you could take up extra kind of part of hearing and with hearing is pretty easy right you just put a microphone into a room and then you can hear that room at the same time and we already use that right by just people hearing into other locations while at the same time hearing what's going on around them so we have no we have zero problem uh about using like more ears than we have on our body so these kind of like augmentation or um uh how can I say additions to the body yeah that's there's a lot of potential for that right um extra finger People Are People Too okay Crimes of the all crimes of the future I will check that out yeah poly dualism that is a cool word wow okay um we are all missing a second thumb on each hand I could use two more arms to be honest like just to to do stuff um on the go additionally right or sometimes you just want to do like extra stuff like you want to type but then also drink coffee but not put the hand up or stuff like that you know um things like that repaint that yeah right repaint that or use it as inspiration I actually want to pick up painting again because I'm so like there's so much inspiration from Ai and I just think I want to do that too because it feels good to paint it with your own like tools with your own hands and then your own mix comes into that just it's it's just nice um you only need three fingers for your mouse some people have these mouses with a lot of buttons especially for these kind of roleplaying games and stuff like that World of Warcraft you can have a mouse with like 30 different buttons to have every kind of shortcut on there I never could use that I bought one but I never could use it it's just too much for me um yeah it h it hallucinates images right right right I'm a little bit behind on the chat sorry is there a hot key to generate the image in com eye when it's ready a hot key to generate is there I I'm not sure I don't think comi has the same thing automatic 1111 I think has shift enter right or control enter either of those you might be able to set up in the settings a shortcut for rendering pretty sure that should be possible to just have a render button let me see is there anything does anybody know if there's like a a key that you can put um to rendering by the way it would also be interesting if maybe there is a timer note what what do you think about that like having a timer note where it renders an image let's say every 30 seconds so it can do some adjustments and then it starts to render on its own right so you just like stay in the flow do the adjustments and then once in a while it starts to render image not quite sure if that would be super useful but maybe uh but you can experiment with these kind of things right by just creating your own notes that might be interesting control enter works oh yeah with automatic 111 right multimodal eye that knows and understands the history evolution of Artistic Styles and can intervene NE Artistic Styles with reasoning behind and how and why those new styles were invented right yeah absolutely like even now you can go to chat GPT and you ask about different like for example story ideas or maybe sh ideas or stuff like that and it gives you different kind of combinations it can come up with and inspire you I I don't know if AI will ever be able to understand things because like one thing that I find interesting about the history of youi that we have seen over the last year is that it got a little bit better but it did not got in most cases it did not get dramatically better right and then um there's a difference between for example an AI that drives a car because it can be complex to drive a car but basically basically there is not much where you can go you can go backward forward left and right so even though you it has to take a lot of choices uh into consideration the output is limited to how fast you go and do you go backward forward left and right this is very very different from creating an image with reasoning behind that because reasoning is uh like incredibly complex and I don't know if AI will ever arrive that but and that's the interesting thing about an image is reasoning is required to create the image in the right way um for the right purpose and that is what's limiting a lot of the things we see right now with all of the errors we getting also with the hands because um the hand from the perspective can be so different from whatever you see because actually a hand I I can't turn it right now actually a hand sometimes only has one finger right right and sometimes it might be that you have hands that just have more fingers because it's two hands at that moment but it doesn't mean that you can see in the image of what's going on right so you need an AI that has sort of kind of a reasoning why this looks like this and can understand that right um I'm not sure if we will ever get there I think um I think right now there's a lot of optimism about what AI can or can't do uh but um I think a lot of that op optimism might be a little bit too uh too optimistic right like at the beginning when we had the internet right and it it was like the imagine the like remember the AI the sorry the ideas we had back then right people came up about this kind of amazing I mean they come up about like they re kind of dream that right now with this kind of virtual world where everybody's living and we connecting and like we only talking over the phone with video calls and everything is so interactive and blah blah blah all these kind of things but actually what it melted down to is a tool that helps us in the background with a lot of things that are happening while in the foreground we consume very simplified information that is already prepared for us but it's not a super extra complex Network that is um all involving for us right as we imagin it at the past being right we are not completely consumed by it it's just a site component that mostly most of the time lives inside of our smartphone and then sometimes we sit down on the computer because we need a bigger screen for working stuff like that so um and the other ways where it has melted into our everyday life is most of it is just in the background right like with online stores with online uh these kind of word libraries video libraries all these kind of of things but they they are just there so you can like ask for the information and get it but the whole process is automated in the background you don't need to touch that right um so this kind of like everybody owns his own online Universe never came true because nobody really needs that right and it might be the same with AI where at a certain point we realize it's it's pretty cool but we don't need more and we can't achieve more right it would be incredibly um difficult and expensive and resource hungry to get this kind of super uh uh super finalized um results like for example an AI that actually understands the image do we really need that or do we just need something um like for example the way that um 3D animation replaced a lot of um practical effects in movies right which is pretty cool but at the same time it's still just just a tool and you have to learn how to use it it's not just where you click and then you get an automatic like amazing explosion from that you have to do everything of that by hand still right sorry I have to look at the I I like to do my monologues actually we we made a background last time forgot about that I don't think I have it here right now though um book no that's something else okay never mind I can prove to you that it can actually understand in it's a trivial thing to do no it's not absolutely not it can also misunderstand but misunderstanding can be corrected no no no it's not a trivial thing like I I think what you mean by that is that AI can understand it in the limits uh in The Limited way of when you ask it for something and then it gives you something in return that fits what you asked for but that is not understanding has nothing to do with understanding right like for example if you give me um a box with different symbols and then you ask me for one of these symbols and I give you that symbol doesn't mean that I know what the symbol means it just knows it just means that I have an index of the symbols and this is why I can give you the symbol and if there is a combination of these symbols that make a perfect sentence it doesn't mean that I understand anything of that sentence it's just that I have a chart that tells me in what uh combination the symbols work so that the output is what you expect from me but I don't have to understand anything of that and this is what AI is doing right now with language the language AI models don't understand anything of what they are writing what they're doing is they calculate the most likely word that comes next in the sentence which sounds completely absurd um but yeah that that's what it's doing like based on all the training data it just figures out when this is the sentence so far this word is is most likely to be next doesn't mean it understands anything but just by the probability mathematically this is what should be there right doesn't mean any understanding right um uh where do we go from that let's see uh is a way to post the image generation certain I don't think uh confu I is any way to pause the image generation this is also something like I'm missing or like cancel the image generation I don't think it has that right I have never I I don't know I haven't seen it yet um clip vision is an is there in fact the various ways to do the believe knows the details of such things better I do have not done much with the art models um what what are we talking about is there a way to pause image generation ah okay that's the thing I think I get it image to image setup with com ey yeah check it out with comu ey uh triple hat monkey says true it might be better for now at least what that we have a visual fxu rather than a true word to image model at the moment there just won't be enough resources to run it um The Limited way I have to prove to quite a view like an example artificial intelligence models that they are capable of comprehension so yes providing it to human is quite trivial by comparison what the limited way I have had to prove Pro to quite a few like for example artificial intelligence models that they are capable of comprehension so yes proving it to human is quite trival by comparison I don't know what that sentence means I don't know what you mean by that you have to rephrase that um but I don't I don't think I agree like first of all even for humans most things to understand them is very complex um and we do have reasoning behind the things we are doing uh but then also the things that might look like the ai's understanding what you mean and talk to you is again just like the process of you um interacting with a machine who doesn't know what you say but gives you back what you want from the machine right um so I don't have a good example for that um maybe an example would be if you are in a video game and you do something with an NPC and you request something and it might be a complex dialogue with multiple choices and this and that um it might seem like the NPC is understanding what you want to do like uh for example as a simple very very simple example it's nothing to do with AI but you can say to your NPC companion in a video game wait here or follow me of course that has nothing to do with AI so it's not an AI sample but if you have a more complex task you might come to the conclusion that hey the NPC understands what I mean because he's doing what I'm saying he's giving me the stuff that I asked for but actually it's just like the function in the game right that the AI uh sorry not the AI the NPC is waiting there because the function weight means weight but it doesn't it doesn't like there's no purpose behind that for the a I right it's just like the way it's it's it's like it's it's coded you mean you know what I mean so I I would be very surprised if that is true there might be a very limited way of comprehension but comprehension is not easy to do if that was easy to do why is it why is it the the Holy Grail of AI to have ai that is smart that makes no sense right every like every AI scientist around the world who is like working on that kind of problem is absolutely obsessed by the idea of creating an AI that has human intelligence and actual comprehension and understanding of topics and we are super far away from that so if that was so super easy to create ah yeah I I just I don't think so right then why isn't it here right and of course this guy from Google he thought it was intelligent and stuff like that yes you can be like you can think you have an intell conversation with an AI but it's just because um you get back what you what you think you want from that right like for example even with humans like a uh if you want to have if you want to have your uh the op your like if you meet someone and you want them to be impressed by how good of a conversation partner you are um the best way to do that to convince them you're a really good conversation partner is to say nothing nothing and they will after the like with a with a super high you can try that with a super high um percentage of likeliness if you say nothing they will end the conversation by saying wow you're really nice conversation partner I guarantee you you have to I mean you have to say some things the only thing you have to say is ah okay that's interesting tell me more about that these kind of three combinations in different variations just Bist listening you don't say anything they talk all of the time but they have the illusion that you said amazing things you didn't say amazing things you just listened intently to what they said right so that's the thing um I wouldn't say that I could prove it if I couldn't back up such a statement I think you're convinced you can prove it but I if that was possible where is it why do you have that ability ility and Google and meta from Facebook and um stability Ai and open AI which have millions and upon millions of funding don't have that ability yeah it's the it's a Chinese room problem right it's the you think it might have understanding but doesn't really have like even chat GPT doesn't have it doesn't have any kind of understanding of self- reference it tells you oh sorry I made a mistake but that means nothing it will repeat that mistake exactly in the same in the same like answer um because it doesn't it doesn't get it right like one thing like I will I will um wait a second uh I will show you right now one thing two things actually oh wait a second let me check my check Che cheap beforehand um can you hide oh you can hide the sidebar I didn't know that okay cool okay I want to show you two things the first thing we're going to ask jet GPT is how many ns are in the word pause how many n oh how many letter N are in the word pause there's no n in there we can agree on that no n in the word pause okay CH gbt the word pause has two letter NS in it anything else you'd like to know or chat about are you sure oops my bad I apologize for my mistake there is only one letter N in the word pause check again I think you are wrong let's go you're absolutely right I apologize for any confusion earlier I double checked and there is no letter N in the word pause okay so now we have a yeah it goes with a no okay but let's let's try again H uh check again please let's see check again oh now it is with no letters okay so pretty cool it kind of stuck with that good so now I want to have a another thing I want to have um please please tell me the title of the uh Tom Scott YouTube video about the moon um with the link and a summary all right there we go does it do something no it doesn't do something um wait is he called Tom Scott because he comes up with fantasy uh videos Tom Scott yes that's that's the guy one second uh do you no Tom Scott he is a popular YouTuber okay um what is the title and summary of His YouTube video about the shortest river let's see ah why doesn't it work right now Tom Scott the shortest river um do you know past titles of Tom Scott YouTube videos it doesn't work right now ah shortest river the signs off ah here it starts the short location of the shortest [Music] river let's see is it does it does it is it called the shortest river no is not called the shortest river weird places um in this video Tom Scott might explore oh it might explore I had in the past where it tell me like a completely like untrue oh wait a second um in 2018 Tom Scott made a video about the shortest river um give me a summary let's see if it falls for that ah there we go there we go I apologized for the confusion earlier in 2018 Tom Scott indeed made a video uh titled the shortest river in this video Tom Scott explores the Row River located in Montana USA which is often considered the shortest river in the world the video is likely discussed the unique what likely discusses the unique features of the Road River its length and its significance um and so on blah blah blah um and the thing is like if you now ask chpt are you sure it's not a real video then like at least in the past jgpt uh told me know this is absolutely a real video they might have fixed that by now I'm not quite sure but uh again here you can you can get jbt to just tell you yeah of course like he made in 2018 a video and first of all you can see here um the video has a different name and it's 4 years ago when is that uh 2019 actually right no wait a second oh yeah this 2019 right so it's it's also not from that year like I could also like write any kind of different name in there um wait a second do you have a YouTube link to that video let me see ah in the past they gave links fake links not real links fake links um yeah anyways anyways but you can see it's like not really like knowing what it's saying anyways um so that's that's kind of the that's kind of the problem of these kind of things of um like um uh it it it might be very like knowledgeable and give you very good explanation but doesn't mean really mean it has like understanding of course understanding or comprehension can be defined in a lot of different ways right um something looks like is only one of the first layer even humans can possess all layers of comprehensions or everyday of everyday items okay right right right yeah I mean that you have to also like take into consideration at that point that of course AI if it analyzes an image it does it in a different way right uh than than we do doesn't again it's like it's it's not so easy to to to actually prove understanding in for example in comparison with categorization right like you can have a machine like for example U one thing you could do let me switch over here again um one thing you could do is like I make a board not a board let's let's say I I have a um um what is it called um yeah let's say I make a board I make a board and I drill different holes into the board and then I I put different stones on the board and I have a uh like I have different amounts of boards with different sizes of holes and I shake it and of course every board then at the end only has this kind of stone size on the board it doesn't really mean that the board understands what the stone stone size is right it only means that the filter of the board is set in a way where this Stone size doesn't fit but the rest can go through and this is the same thing with like analyzing an image that you have a filter in there that says oh this is what a face is like saved in my database or in the in the I don't know in this kind of system on how I I recognize a face so this is um fitting the pattern of what I'm looking for so it's looking for a pattern this is the pattern right so this must be a face and so it categorizes as a face doesn't mean it understands what a face is so it also can have something like a shadow that looks like a face or there a print of a phas on a wall where we would say that's a picture of a face it's not a face it's a shadow looks like a face it's not a face it might be a dock that is from a certain perspective looks or a monkey or something looks like a face but it's not like a human face that we looking for so we understand that the machine doesn't know it because it looks for patterns but that's mean this doesn't mean there is an understanding of that right because this this is the like this very complex um process of understanding is to um how can I say not just categorize things but actually have the understanding between all of these different steps and cultural references and putting it into the situation and what does it mean in the situation there's a lot of information that goes into understanding rather than just having knowledge or having an archive or having a pattern or having a certain system or Ruster or any kind of thing you're looking for and then going by an index or a system or a software to give back that kind of data none of that requires any kind of understanding right you might have a book like a uh how do you say like a like a index of all the words in the English language it does not mean that the book can speak English but it it has a knowledge in that sense of all the worlds on the English language so it knows more about English than you do but it still can't write a single sentence from that right so there's a there's a huge difference between understanding something and just preparing information in a certain way sorry again like my monologues they are always coming in yeah um neither do stable diffusion understand what an image is as jpd doesn't even know what the words are it's all just code wait yeah right it's all just code right it's just because Chet like stable diffusion makes a beautiful image that looks like artistic images what it does is it takes noise and then arranges the noise basically simple explanation into the way it is most likely to what what is trained on the image to look like but stable diffusion will do absolutely everything for you that you train it on any kind of art style so um uh if you tell if you tell um uh uh if you tell stable diffusion this is an image of a dock and this is abstract art even though it's an image of a banana and it's a photography if you train it like that it's probably going to do it like that because you Associated the the the the the image of a banana with the word face it doesn't know it's not a face right doesn't mean anything just to be clear I wasn't speaking hypothetical when I say I can prove it yes I know when you can say you can prove it you can prove it then please if you can do that I back you write an email to Facebook meta to Google to open a i and stability AI because if you can prove it that AI at this point has comprehension and understanding of actual human knowledge in a way that humans do or even in a simpler version of that they will pay you Millions to be able to do that because nobody has that right now this is what everybody's looking for doesn't it's just not there right even if you see these like cool uh videos that you have from these kind of labs with the with the robot dancing that is fake it's not real they they can't do that right now you you don't have a robot just like doing back flips willy-nilly they take hours upon hours and this stupid robot is crashing a million times until they have one take where it does back flip correctly and then they cut all of that together to one like take that looks like the robot did all of that at the same time but it does not do that at the same time they have a video showing that that it's not real it's fake the robot isn't enh like able to do that right now I can't just walk around and pick up stuff and then dance with the stuff it's just not possible right it's it's super super complicated one of the biggest complications with that is also because the the the sensors that we can build are just not precise enough for that and if they are very precise they're at the same time extremely not only extremely expensive but also extremely um uh what do you say um can be extremely affected by anything around them the temp temperature the light situation small damages on it all kinds of things to just go wrong right so the stuff that we have in our skin all these kind of senses that is like incredible magic we can't do anything like that from this kind of like how robust it is and at the same time how good the sensor feedback is and also um a robot only has a bunch of sensors in the complete robot while we have millions of sensors all over our body and this is while for for us it's so easy to just like balance and do all these kind of things because we get like tons of feedback and at the same time our Motors are organic soft flesh muscle while the robot has kind of Hy hydralic or whatever kind of motor in it which is very unprecise and rigid at the same time it can't adapt to the situation right it's just it's just not working right um he's saying he has trained his own llm which actually understands things if you want to believe that believe that right um send it to to the big companies and let this let them test them but I think you're just falling for the same fcy like the guy on Google that you think you're talking to as sension Ai and it's just everybody's just saying nah bro it's just you got fooled by it's just sounding very human but it's not it's not real right it's like muscle memory it's not understanding some point uh when it comes comes to practice a sport yeah muscle memory is also an interesting thing that the the robot doesn't have it's actually a very good point because a lot of the stuff that our body does is not um this kind of communication between the the brain and the the rest of the body the body is doing stuff on its own that is happening without the brain in between because the time would just be too short even with the with the speed of the of the energy going up into your brain and back would not be enough time um to uh to do that kind of action to react so fast to that so this kind of muscle memory that we train over time uh to be able to like for example walk without falling over stuff like that or when you there's some unevenness on the ground and you just even it out a lot of the body can do stuff on its own at that moment right if it's alive of course right so yeah um robot playing ping pong yeah that yeah is also cool right there's a cool there's a lot of cool examples but they they they are made look they are made to look much more advanced than they actually are uh in in the real world right I'm actually very surprised that these kind of electric self-driving cars are as good as they are um so far even though they do a lot of mistakes right if you watch actual videos of people testing them you will see that they do um still you have to like steer a lot like intervene where it it might do some some uh stuff that is not so is a little bit sketchy right they still very very good though at what they do um also planning of publishing a formula is something I don't know much about doing and I'm not sure what I will publish first but I'm planning on doing so yes do that that's good you might mix a cons on appro than being control you mix you mix to be conscience on a process in being control I wasn't really trying to convince anyone here I was only stating a it's it's okay if you believe that uh send it send it to some experts and and and let them see what they think about that right um it just from my perspective it doesn't make sense that you trained a model on an AI structure that doesn't have the ability and you say the model that you trained also with your limited like uh uh uh uh uh funding that you have like the computer have and stuff like that and suddenly this kind of training that you did outranks the the much much bigger training that the people did to develop all of that kind of AI and and the system behind that right because they like training these kind of models they they cost like I don't know what was it like um stability AI the model didn't it cost like a100 million do to train it it's it's a really like intense process it's not just like and the open AI also it costs like at least I don't know what it cost exactly but let's say 50 60 70 million doar to train the language model of of open AI jet GPT well we can actually look that up to train CET gbd4 uh the report claimed that the competition Sam Altman stated that the cost of training CET GPT was more than $100 million there you go so and CPT is not even close to being conscious or smart or understanding of anything it's just like it's a nice tool to to help you with things so I don't know but maybe it might be I don't know just send it to expert and see what they say right um there's always a chance right there's always a chance um anyways we are at oh my God again I talked for one hour after the end of the stream holy I always go to my monologues I'm sorry about that but I think it was still kind of interesting and fun um yeah I I mean I hope that AI is going as far as it can um I just like I see that recently in the last year um the gains have become smaller and the the hardware that is required has become much much more right um and if and and still like for example if you if you think about sdxl and what it's bringing to the table and then at the same time how much more GPU time and how much better Hardware it needs and it's not even that much better if it would be much much better you would say cool okay I can deal with that but it's not so if you have something that is really a lot better and has really a lot more resolution and really lot more understanding and a lot more parameters is drained on imagine how crazy the hardware might be like that and if the hardware requirements for example if the hardware requirements and the parameters that the model needs to have to be better at the thing that it needs to do is exponential holy right if that is the thing who wow then you need like an extra an actual super computer to to use the models of the future I don't know you know I mean if you think about 3D Graphics you have pixel movies where one frame a single frame can take like 30 hours to render right and that's with like modern rendering Hardware that is specifically built and configured to render 3D images so we didn't like go to a to a situation where we can render these kind of 3D movie quality images in in like real time we have very good computer graphics for video games but the thing about that is we didn't go down the route of actually making the technology better for the graphics instead the studios learned how to trick us into making the lights and the shadows and the reflections and everything look real and feel real even though they're not real right and now of course we have rate tracing it's kind of cool but at the same time you see uh when you compare R tracing to the fakery that they did before rate tracing is actually not improving it that much and you need much much better hardware and the gain is like minuscule the only thing where it's actually really nice is when you have a reflection on the ground but to tell you the truth when I play Call of Duty or some other game I couldn't give less of a crap about what the the the reflection of the puddle on the ground looks like I'm not looking on the ground when I play the video game I'm looking on the on the competition right so it doesn't even matter if it looks nice or not the fake light in most cases even look better because that's artistic light applied for dramatic reason realistic light is following nature which is cool for some games but not for all games right so it's it has some benefits but is it really is it really worth getting a 4080 to have a Graphic that's like a little bit better than what you can render with a 1080 that's much cheaper I don't know you know um anyway doesn't do llm it's not that I don't do L llm stuff it's that there is no development in AI that is showing any signs of Consciousness yet right that's just the point right or or comprehension that's just the thing like if if would be nice if it was but there isn't right rate there's a there's improvement with rate tracing but it's like it's you can take it you can leave it right wait a second [Music] um like the the the only the only times you really see much difference is when they show water Reflections most of the other time it's like not sure if I need it one second oh there might be a feedback loop now with the sound and also this is ah shut the hell up man with your stupid advertisements here one second uh want to sure if this is a good example because it's very dark okay I don't know there's some comparison one second okay there's a scene can you see that one second not sure how much you can see that how well you can see that so this is like rate tracing off and then you have rate tracing on looks like the same like don't tell me this is magically better and you have to like spend $33,000 for a new computer because this is what you see before and this is what you see after like the the sign is like glowing a little bit more and the Shadows are like of course real time but who cares about that right look at that this is without rate tracing this is with rate tracing did you see the difference other than the glowing up here I didn't really see any difference and then there's path tracing and yeah of course but it it's brighter and it's a little bit more saturated I can do that with a overlay filter on the on the video right I don't have to have a super good hardware for that this is just like a filter that you can like uh what is it called the filter you can put on video games uh where you can set your own settings so the game looks nicer and sharper and stuff like that it doesn't do any kind of special rendering or stuff um FX something I can't remember it but like this is darker and then the I mean also like if if if this is the artistic vision of the game and it's dark because it's Blade Runner idea dystopian and then you have this fancy Pence version here that looks nice and bright and pink and it's like this is not the artistic version of that right the artistic vision is kind of like not reflected in that also this is why where I said this is uh sorry this is the artistic light how they imagined it to be because the light is set to give you the exact impression of what the artists want to have and this might be what the real light sources would do but who gives a crap about what the real light sources would do when you want to watch the when you want to have the actual thing that's like saying you go to an action movie and then complain about oh but the weapon doesn't work like that yes but it's an action movie you want to see the action the car doesn't explode in a in a ball of fire yeah but I want to see it because it looks cool and it's action I don't want to see a car just crashing down the mountain with nothing happening at all I want to see the artistic version of it exploding into a fireball right with crazy sound or with hand grenades that explode in a fireball they don't they make a little puff but nobody cares about the little puff you don't want to you want to see that right you want to see the fireball anyways too much talk there we go um let me see I will I will research the topic of of comprehension in AI but I don't think there's anything like that um maybe in I don't know 30 years or something but so far I don't think we have seen anything like that okay anyways W thank you very much for that the the talk was way too long afterwards but I think the first two hours were pretty interesting about comu ey so that was fun um yeah so thanks for that and thanks for the interesting talk um and uh yeah see you soon and next time I will I will keep my monologues a little bit shorter anyways okay my friends where's the where's the Q button for the music one moment there we go okay cool see you soon my friends thanks for watching [Applause] [Music] bye [Music] [Music] [Music] n [Music] is [Music] no

Channel: Olivio Sarikas

Views: 16,381

Keywords: oliviosarikas, olivio sarikas, olivio affinity, olivio tutorials, stable diffusion, stable diffusion tutorial, stable diffusion ai, stable diffusion ai art, how to use stable diffusion, stable diffusion 2, 2.0, automatic 1111, install, setup, local, SD local, SD automatic 1111, SD 2.0, Midjourney, adobe competition, Adobe Alternative, stable diffusion gui, midjourney ai, stable diffusion install, stable diffusion browser, stable diffusion installation, ChatGPT

Id: REdk5UuvTyE

Length: 190min 33sec (11433 seconds)

Published: Mon Oct 09 2023