Master Character Design: Create Consistent Faces with Stable Diffusion!

[Music] hey everyone welcome back to alchem zero phase this is Eric and in this tutorial we're going to show you how to use a character template sheet that will allow you to get different facial angles um and even expressions and stuff like all while maintaining the same character consistency as you can see in this I just used a bear with this sheet and uh was able to get lots of different angles with the bear looking down and kind of looking up and looking all over the place so we can do this with animals we can do with people and uh this tutorial will show you how to do this so you can create consistent characters whether it's for story books or other illustrations or just getting the face you want and then using that to create other faces so let's move into this what we're going to do I'm using reality's Edge right now it seems to work pretty good for what I'm doing uh but honestly this will work with pretty much any uh checkpoint okay so the prompt I have set up I am currently using uh is photo grid and then in this one here I used anime girl because what I'm describing is a picture that um I'm taking the character and I'll show you here real quick so so this image that I generated quite a while ago actually and um converting that over to a more realistic photo version of the person with the hair you get the kind of wisps of hair and uh yeah even the eyes let's pull all this up so you can kind of see so you can kind of see the how the character was carried over and made more realistic it's a very cool feature um something that I know a lot of people are looking for to be able to do storybook illustrations that seems to be pretty popular uh or in other things I don't know um videos and whatnot okay so I showed you the checkpoint the uh prompt is pretty simple you can either say character design sheet I I I like Photo Grid um if you use character design sheet it tends to create like these characters which is fine maybe we'll go into that describe who it is or what it is you want to see in there um because you're describing a grid it should maintain this I'll show you the other settings that help maintain this as well with a white background emphasized so you can grab that prompt if you want seems to work pretty good I'm using automatic 11-11 Forge so there's a few things that might be a little different uh with a Samplers I'm using the turbo editions of the Samplers that are pretty common uh so UL a DPM Plus Plus+ 2m and uh DPM Plus+ 2m SD each one is a turbo they're designed for lightning models and turbo models and for right now we're just using the DPM Plus+ 2 MD turbo works pretty good we got our sampling steps set to eight so depending on which model you want to use you're going to need to read up on what the sampling steps are going to be if it's just a regular sdxl model or even a 1.5 model uh you're going to want to have this up to above 20 probably and then the config scale for this particular model works really well at config scale two uh there's a couple of lightning models that work really well at one or 1.5 but two seems to work really well for this one the character sheet I'm using has a bit of an odd aspect ratio I'll come down here show you it's this right here it's actually like a 1024 by 614 pixels and um I so basically you had to figure that out program in the pixels here and then I lock it using aspect ratio extens as aspect ratio helper extension and then I can drag it up and I like to have it a little higher on the resolution so 1 12200 by 720 that gives me a little bit more pixels to work with and makes for a higher quality image and allows the faces to come out a little bit better okay these you don't have to worry about if you want to do multiple batches you can a detailer I do turn on when working with uh human faces whether they be coming from anime or going from Human to anime um don't mess with any of the other settings in that just turn it on down here we're enabling the first control net if you only see one control net in your automatic 1111 interface and you're not using Forge you'll need to go into your settings go under the the control net settings and increase the number of control net um Windows there should be a setting in there I typically like having it at three so what we're doing is we're just dropping this character Grid in here and I'll make this available on my Google share you can go there and grab it and uh it seems to work pretty well and I'm using it with canny so we're enabling it we're doing I'm allowing a preview not that that's necessary but doing Pixel Perfect so you can least see that it's recognizing all the uh different aspects of the images pretty well well uh you know I do notice that it's cutting off the hair in some of these uh so it's not doing a ponytail and a lot of that is kind of gauged by how much control weight you're giving it so cany leaving these defaults I'm setting the control weight down to 045 for what we're doing here now I've gone as low as like3 the problem you run into with that is it doesn't the coherence of the grid starts to fall apart meaning uh some grid squares will be bigger than the others it like focuses on the center too much I find that right around 04 to 045 maintains that integrity and then on the ending control step I'm dropping that down to 55% or 0.55 of the way through the generation so about halfway through I'm giving the AI full freedom to incorporate whatever details it wants so it works really well maintaining the consistency but adding that finer detail in like around the hair and facial features now the other thing I'm doing here is I'm enabling a second control net bringing in the picture of the character or person that I want to uh transfer I guess you could say the face to the character sheet and in this case it was an anime girl that I had generated and as you can see it transferred it really well now you do want to make sure that in your description you're describing the character that you want to have so in this case I did describe uh anime girl red hair black shirt you know I specified black shirt because I wanted something a little different and blue eyes those were the main characteristic features that I wanted to maintain and then the IP adapter so uh we bring this image and then we're selecting IP adapter and Pixel Perfect but the IP adapter is going to take the rest of it you know kind of the general shape of the face and characteristics that way and bringing them over even the like some of the hair you can see that uh it brought over um that wisp of hair in a lot of these images not completely but you know that's okay all the other settings I I turned the control weight up at one point I don't think that's entirely necessary I did have great SU ESS at one I didn't see much of a difference going to 1.3 or or above one honestly now here is the big difference in playing around with this there's three different pre-processor models and what I found is that the that's interesting I thought it was a second one but the Insight face clip 4 which is the default one that it goes to seems to work the best at maintaining the character consistency without too much random Ness so we're going to actually bring in a different image here and see what what we can do with it so let's uh grab this other image where you go give me one second okay let's grab this one bring it in okay um so we got characteristics I mean this is pretty generic you know this is another generation I did we got black long black wavy hair and beautiful woman with dark complexion we'll just say that so we're going to go here we're going to grab that we're going to say beautiful woman long black wavy hair dark complexion maybe maybe not dark complexion we're go tan see if that works leave everything else the same we've already gone through all the settings we don't have to modify anything else and let's just hit generate and see what it does I'll let you watch it through this first generation here so you can kind of see if it maintains the consistency of the grid that's all brought in through the Grid on here and it translates over into the uh cany preview which is nice looking good so far hero fix just kicked in a detailer should kick in find all the faces there it goes yeah it actually looks a lot like the character in the picture and it looks like it maintained the angles of the faes pretty well sometimes depending on where you have the control net weight at down here if you go too low you lose consistency on that and the character will be kind of looking off in random directions and uh doesn't work well all right let's blow this up and we have a very consistent character that looks a lot like the original let's bring that over here so you can kind of see it so we have same eyes eyebrows look the same same cheeks nose looks the same so let's kind of go over this different angles look good side profile looks great like it is amazing how well this does this now you can take these and run them through the background removal and create a transparent PNG and place these characters in anything you want now this doesn't just work for human figures we can be doing this for a lot of different things it all depends on what you're putting in your description so let's try one other thing let's try something a little different we're going to do Photo Grid let's [Music] do character grid and we're going to do um I don't know what's this will be a little bit of a trial run here I want to see what it does we're going to do um purple alien woman with I don't know tentacle tacle hair I'm just going to leave it at that and give it a try we're just experimenting right now and I'm going to show you how to do things like the bear and other stuff here um cuz those are pretty pretty easy too yeah it's looking pretty good all right so now we've got a purple alien I guess uh oh it is woman so did I did specify a woman so it is a woman but she's purple she's got tentacle hair which I thought I think is awesome look at this but what's cool about this is that the character maintains the same structure so it carried over characteristics from the image that I I put in the IP adapter and brought that over so you can have a lot of fun with this creating all sorts of unique characters but having consistent characters character um positioning facial features we could even specify smiling uh we can say angry and let's say um biomechanical okay so it's not changing the girl a whole lot we're still getting the same consistent female face and and then the reason is because we're using that second IP adapter we wanted something completely different we'd shut that one off but this is crazy so get almost something that looks like something from Dragon Ball Z uh you could have a lot of fun with that too change the hair color whatever okay so in fact let's do that what we're going to do now is we're going to come down here and shut off this second IP adapter you don't need I'm just show using that to show you that you can carry a face whatever face you want to put in there across illustrations photography cartoons whatever you can just put that in there adjust the settings whatever character comes up here is going to have a similar face if not the exact same face uh you obviously depends on your description up here so now that we've disabled that we're going to leave this description the same we're to see what we get now with out the oh you know let's just get rid of woman too let's run that this should be interesting now we definitely got something that looks a lot different specify alien you're still going to get something that's fairly humanoid but these look great the only one that didn't do the aail on was this last one it didn't quite recognize that as a face but the other ones it worked phenomenal on look at that now we could get into animals um if you want do is come in here we can leave angry in there that's kind of fun let's just say eagle head just something simple see what it does yeah I can already tell it's not going to do it quite right so we're going to interrupt that I think what we're going to need to do now is I have it up at4 five in order to give this the freedom it needs we need to come down to and before I think I went down to35 I think helped me maintain some of that let's try that in fact we're going to shut off high-risk fix we're not going to worry about that right now we're just going to quickly go through these see what setting it was I'm pretty sure it was .35 we may need to go a little bit lower but I don't want the Grid it's still doing this is interesting okay sorry I just had to do a little bit of experimenting the reason why it was still giving me characters is because I had up here character grid so once I switch this over to photo grid and we started getting a much different result so what we're going to do now we're going to turn on highes fix and we're going to get rid of angry because honestly an angry Eagle kind of looks a little funny it did maintain the character grid pretty solidly so we're going to go ahead and drop that down just to three oops this should be a little more consistent and give us what we're looking for I think there we go yep perfect that last one there is not right but you know we could fix that later or just rerender it honestly does the highres fix so that it fills in all the extra details I think a detailer did we shut a detailer off yeah good okay there you go Eagle heads in different directions not exactly looking up and down working with animal faces a little different got two commas in there let's do this let's do chip Monk head again we're just going to quickly go through some of these that one worked out beautiful down kind of up and a little bit more up this one are you know kind of about the same level as these ones here but these are looking down worked out nice now let's do something a little more uh illustration grid chipmunk head let's do cartune chipmunk head this model does pretty good reality Edge is pretty nice and there you go you got yourself a kind of a cartoon chipmunk let's do a cartoon fairy head yeah kind of looking up on these ones I'd use the fa the ad tailer on this to fix the faces but there you go now again you can use the second control net if you have a face in particular you'd like to apply to these and uh you would apply it let's see is there anything else that would be good to show you on this I think that was pretty much it um I mean there's a lot of experimenting you're going to do with this I hope this really helps uh I did have quite a few people asking me about consistent characters I had a lot of people asking me about doing a video with a character sheet so you could do uh so they could understand better how do you create a consist in character style and getting those different facial positions um works out really well let's do one more dark wizard oh it's interesting it's kind of doing it in um almost like drawn I would actually bring up the so bring this back up to like let's TR4 five I said dark wizard so it might just assume black and white honestly yeah you get up above 04 and what you end up what ends up happening is it's adhering to the shape here so I'm going to show you one more thing uh we're going to use open pose so this character sheet does the reason I use this character sheet is it works really well with open pose um I've got a couple other character sheets that are just drawings they're they're like uh learn to draw a drawing so it's got the character looking in different directions but it's just like the outline of a head with a cross across the face so open pose has a hard time figuring that out with this one though because they're actual faces it's able to pick up all the faces and by using open pose it's not restricting you to the outline of a human face so though the um cany works really well this actually works really well when uh working at I think like lower numbers or yeah no we're going to have to increase the number sorry you can actually work at higher control weight numbers with this um just because there's no pre I don't know what you call it pre- determined uh face you see in just a second here I think you do lose a little bit of consistency with I'm not really specifying clothes do get a fairly similar character across the board let's just do this wearing hooded cloak red hooded cloak see what that brings across here so it's pretty similar looks good what I would do though with this I think you can get much better consistency if you do bring in a character face so let's do this let's bring over this guy here this was an image generated by somebody on my Discord I don't remember who was trying to out paint the rest of the body um I think we got them taken care of too so we're going to enable that Pixel Perfect just leave it as is we're going to say um we're going to leave it as hooded CL let's just see how that carries across here yeah not quite what I was thinking one of the reasons uh I like to use canny is because when you're using the um open pose you're losing the grid as opposed to if you have cany selected it actually maintains that grid like that so if we bring this character over here even just for consistency's sake it might be good just to have something here um I think even if it's just a colored picture of nothing I get the idea the the feeling that that would actually work really well we're going to bring this down to three4 five no four come on 2.35 I think that should maintain the grid it should maintain the character positioning while still carrying over that character face possibly did I select something wrong here enabled Mak perfect IP adapter Insight face yeah contr weight is one that's fine yeah it could be that I'm just not describing the character very well let's do this dark elf wizard you can see this is giving you a lot to work with you'll be able to do a lot of really cool things I said dark so it's doing darker skin we got the elf I don't know if it's carrying the character face over as well as I'd like it to and maybe that's because I'm uh not using ad detailer but ad detailer doesn't utilize this let see here wait a minute yeah there we go so three maybe let's bring this up I mean at this point you get the idea I hope you have a lot of fun with I'm just kind of experimenting here to see what I can get out of it do Photo Grid yeah I think I might have clicked something wrong somewhere but if you go back to the beginning of the video you definitely get a better view of what's going on here to try this one right here the clip vit big G well yeah no that definitely didn't work that's probably because of this yeah that's what I thought okay so um the where is it here so it's the second one here this clip viit Big G the other one worked for some of the other faces but I think this one uh definitely works for these and you can tell it brought over that face quite a bit quite a bit quite well you can tell my voice is getting a little rough I've been sick the last week and a half and today's like the first day I've had any kind of a normal voice we're going to turn on high res fix and a detailer on this one I want to see this thing full res and we'll call it good maybe we'll use this one as the title picture okay so you can see it's just kind of finishing up here kind of wanted you to see that it's creating that consistent character face in a very similar fashion to the original picture there you go cool stuff it's interesting the uh the a detailer sometimes will end up like missing part of the face like as it renders the face so the a detailer you are losing it's not utilizing this face though there is supposedly ways to go into a detailer because it has the ability to set up a control net here but the only one that uh um would work with is these T2i adapters I'm not sure how to use those necessarily so I'll have to experiment with that and see I hope you enjoyed the video I hope it wasn't too long for you if you're watching great um just for those of you who are using my online prompt Forge we are implementing some updates here uh that will create a better user experience um just so I can show you here real quick bring it up so I've condensed it down moved some of the tools and everything into a control panel this dragable we're going to be adding some more tools to this but you have your standard picking the Gen prompt generator from the list of generators there selecting your number of prompts um doing a picture analysis some some of the features of my core prompt generator and then you know modific modification requests now some of the things we're going to be changing in this is adding the ability to list your luras a lot of people like using luras I've honestly I never use them um but I did get a request the other day um that got me thinking about redoing the interface to make it look nicer and have a lot more functional so you now have two buttons for a control panel and then the artist options and references so you can come in here and you can leave these open if you close them they maintain their settings but just as kind of a sneak peek here let's see if I can show this so this is um kind of a rough inlay of the functional components so as you can see it looks like the other one except down here you have the ability to add Lura so you name your Laura you know kind of maybe like a description of what it is um and then you actually put in the Laura itself you know whatever that is you know sometimes I guess they look like this most of the time and what that does it puts it in a drop- down box that you can select it and you have the ability of adding multiple luras to it so it refreshes it so you can come in here and select these and the whole point is that when you hit generate on your prompt it will add these to the end of whatever prompts it generates so if we say two prompts and we're not going to put anything in there we're just going to have it generate some random prompts it would put it at the end of The Prompt here so triggering your lauras when you go to render them in automatic 1111 or I'm assuming I don't know come if comy UI has that but it's not quite there yet I should have this finished hopefully in the next few days uh this is going to be an amazing function that will save people a ton of time that's my whole goal with with creating this this what I call my prom forges the ability to save people time to help them find inspiration and I found that there are a lot of people out there that use luras and in fact this guy he was adding something like four luras to the end four or five luras to the end of each of his prompts and he was saying is there a way we can get this put in because honestly I hate you know having to go find each one and select each one when you can have all of them listed here and you just pick each one you want and they stay there and then it just automatically adds it to the end of each of your prompts you will have the ability to save and load so when you save this it'll save it as a CSV file and then if you refresh the page it obviously doesn't hold those settings it doesn't save those settings so you'd have to come back in here and just click this or you just drag and drop the CSV file onto this field right here it'll load it in and you have all your prompts um just as a quick example here let me grab some that I had so let's drag this over here we'll drop it right there refresh the list and there's all the tests I put in there and yes you can select as many as you want in this adding luras to your heart's content okay and again just as if they're in here they will be added I didn't feel like there was any reason to enable or disable um maybe we'll put a clear button so you're not having to click off each one of these to clear them but that's what I'm saying this kind of in the rough stages trying to figure out what the layout looks like like and what buttons to get in there that makes it functional so I'm looking forward to this uh and I think this is going to greatly improve for those of you aren't aware or aren't familiar with this this is uh my online prompt generator um that gives you the ability to generate prompts for a variety of image generation platforms um and giving you vastly powerful control over those prompts um so let's put in something in here troll with sword that's all we're going to give it hit submit it's going to generate two prompts here when that submit button comes back we have our prompts so we got a wide angle oil painting it just it it just kind of picks the medium uh unless you specify so if I come up here and put uh let's say water color painting submit and it'll come back with oh put medium in there it does that once in a while medium watercolor painting so it's like the shot um and then adds a kind menacing troll wielding a sword so I kind of expanded on that a little bit Dynamic composition sharp blade intense eyes Fierce expression impending attack Lush Force background yeah so I've I've got this thing trained to intuitively based on what you put in there it will expand on and incorporate things you can put as much as you want up here and it'll incorporate those features the artists and options panel gives you access to a variety of artist samples and names and a Color Picker and this is something that uh I find to be incredibly powerful if you want to actually control the theming of the of the image you know the color theme you can come in here and select the color and it will intuitively grab the color name it it uses AI to interpret a hex code on the back end it's really cool the way I got this figured out and then you any color you give it select and got sky blue we'll come over here to a deep deep blue is like navy blue and then the prompts when you generate these will incorporate those colors into various aspects of the promp it's not really you're not really controlling it necessarily so you come in here and we got uh raging expression deep maroon skin sky blue eyes navy blue sword and we could add more colors in there and you render that and you get some great images and then we got a ton of drop- down menus with all sorts of different um words for different uh features visual features that if you're looking for inspiration this is where you want to find it at so um you select a name it'll drop a name in there and enable this over here and you'll try to incorporate that particular artist's name into the prompt watercolor techniques of AJ KAS so most these are there's about 1,200 artists here these are all artists that have been to some degree or another trained in the stable diffusion models and um it's great to have that there you can scroll through that and see great examples of of what they are you know two different hum samples structural sample and a kind of a natural nature sample fun to look at fun to see the different artist Styles I find this much better than the inspiration extension in automatic 1111 uh it seems to be more consistent easier to use you don't have to worry about generating the images yourself I've already done that and uh you need actually do a search too there's Giger do real time search pretty cool stuff anyway if you hung on this long I hope you uh see something of value uh we do offer a free version out on the website if you go to shop. zeras decom go under zeren go to zeren light and you have a free version you can work with this one um you can do you generate one prompt every 30 seconds it just has you wait for 30 seconds uh so most people aren going to need more than that um this may change really depends on how I got a lot of people now using it and I know open AI on their API has some kind of limitations on like the prompts or what is it tokens per minute ran into that the other day and uh so far I haven't run into it again but as more and more people use this uh the number of tokens being utilized will go up uh this my prom generator with the request typically takes up about 3,200 tokens for each request so uh it's a pretty extensive uh set of instructions to get the results we're looking for anyway and if you're interested in working with a full Platinum Edition uh we actually have three different levels uh we have a uh um silver gold and platinum and uh they oops let's go over here and just show you real quick this kind of gives you a rundown of uh what they do each one or the features each one I need to update the picture on this one since we're changing the interface pretty considerably and there's a video kind of a video tutorial walk through so you can kind of see the functionality a little bit more in depth okay like the video subscribe um I should have said that earlier on the video but uh we'll talk to you later

Channel: AIchemy with Xerophayze

Views: 17,501

Keywords: Character Design, Stable Diffusion Tutorial, Automatic 1111 Forge, Consistent Character Faces, Art Tutorial, Digital Art, Character Template Sheet, Character Angles, Step-by-Step Guide, XeroGen Prompt Generator, Free Art Resources, Character Consistency, Comic Book Creation, Storytelling, Character Creation, Digital Artists, Art Tips, Creative Process, Detailed Character Design

Id: 82bkNE8BFJA

Length: 39min 23sec (2363 seconds)

Published: Tue Apr 09 2024