imagine a world where you can create not just any character but one that lives and breeds in a multitude of environments expressing a range of emotions and actions now what if I told you that this world isn't just a figment of your imagination but a reality we can dive into right now welcome back to the channel everyone today I'm thrilled to introduce you to saraphina and her friend Marina these aren't just any characters they're embodiments of Versatility and creativity each with their own uniq
ue presence across various settings activities and styles but here's the twist both saraphina and Marina weren't made with Coya or other Laura creation tools that you might have heard of instead their creation was made extremely easy by an amazing tool called scenario who also happen to be the proud sponsors of this video I'm thrilled to partner with them and it's led to a special bonus for all of you a unique discount code that you can find in the description below so what's the magic behind sc
enario and why is it revolutionizing character creation well scenario is actually intended to be a generative AI tool to help game developers artists and Studios to easily craft unique and consistent game assets such as characters using custom trained AI models and let me tell you this it is an absolute game Cher not just for game developers but for anyone looking to create consistent characters with depth and personality from visual novels to Comics to creating AI influencers scenario opens a w
orld of possibilities today I'm going to walk you through the character creation process and together we'll unravel how to bring characters to life once you're signed up and logged in you'll be presented with this screen now it might seem overwhelming at first glance but don't worry I've got you guys if you're eager to start experimenting you can click on one of the Start From Here templates which offer a guided experience with preset options however these are quite limited so we'll actually ski
p them for now instead let's focus on the side panel here on the left where we have several options the main feature we'll be using today is models and images however I'll also touch on canvas as it contains some important features worth exploring we'll also cover the rest of these features in a later video jumping into the model section you'll find a range of models curated by the scenario these include a mix of stable diffusion XL and SD 1.5 checkpoint models picking one of these will be our s
tarting point you've also got some Community models which offer even more variety however I do want to highlight that not only are these models carefully curated but you'll also find a selection of models thoughtfully chosen specifically for game applications these include models based on isometric Landscapes icon models and even even model specifically for creating backgrounds honestly the more I explore scenario the more I'm tempted to create a series about developing a game from scratch start
to finish using just AI tools if that's something that you'd actually be interested in please let me know in the comment section below especially since they have unity and go. Export options the first step in creating our custom character is to actually develop a foundation for us to work with to achieve this we need to create a set of images with a base model in an art style that we like then we'll throw in a few prompts and generate images until we find a look we love so to get started we fir
st have to select the model now if we observe most of the models here they're all designed with game elements animation cartoon Styles in mind for this video we're actually going to try and create a realistic looking character now there are a couple of realistic models in the community section however I find that because they are Laura based the designs or the outputs while consistent tend to revolve a lot around the character that the Laura is based around so we're going to go ahead into Founda
tion models and start from scratch and you can actually see here some attempts that I had created earlier so we're going to go ahead and generate with this model so to start off with since we want it to be realistic I'm going to go ahead and add in realistic photography high quality you can also add in stuff like 4K 8K detail to make sure that it comes out well in negative we can add things like deformed animation cartoon typical stable diffusion stuff then going back to the positive prompt we j
ust want to put a simple description of the character that we're looking for in this case let's type in brownhair girl and let's go ahead and generate about 16 images the maximum that it lets us so we've got some variety to choose from so after hitting the generate button a few times more than a few times we now have a vast library of images to choose from so our next objective is to go through the images and pick out faces that look similar that share facial features that look like they could b
e the same person to start to build the data set for our Laura we want to aim to get a minimum of five images and a maximum of 20 you can go higher but for the purposes of this video we're going to stick to the 5 to 20 rule so we can already see here from the first two images while not identical they do share a lot of similarities we can see here the lip structure looks kind of similar the face shape shape looks kind of similar this is a slightly different angle the noses are not quite there but
not too far apart and the forehead distance also looks about the same so we could consider these two winds one of the things I like to do is to go ahead and tag the images to make them easy to find later so let's call her Lucy and we'll just tag the images that we like as Lucy so I would have normally picked this one uh again it's not the same person you can tell but the facial features are similar enough however her eye color is slightly different so I'm just going to actually add this with a
underscore we're just going to call it Lucy one to indicate that I like it as Lucy it will show up when I search for Lucy but the one tag will tell me that I'm not totally sure about it so this makes it easy as well to add images if we are desperate for them or discard IM Imes that we weren't sure about this next one very obviously looks like a completely different person and has completely different facial features so we're just going to skip over it the rest of this video is just going to be r
ushed through as you can see me pick out the images that we need so now that we've gone through picked out the images you might notice that there are a lot of images that don't look like the others and what might end up happening is as you go through the images you might actually find more than one set of faces that look similar to each other so what I have done is I've gone through and just tagged everybody is Lucy I filtered it so I'm down to 34 images and then I suggest you just go through an
d do a second pass through make the images a little bit bigger and once again go in and ret tag the images into the different variants that you've got so in this case I will have lucy1 lucy2 lucore 3 Etc and doing it this way allows you to see more easily which faces might look similar to others so let's go ahead and do that now so I finally set settled on these eight images here which more or less share several facial features they've all got the plump lips the relatively slim nose the same hai
r color relatively similar hair lengths more or less similar facial shapes some thinner than others high cheekbones so I've done my best to look for similarities between the characters and try and aim for facial traits that I want to have in the final Laura we also fortunately have different angles of the characters and different Zoom so we've got here a midbody shot we've got here a semi side angle which gives us enough variety for the model they're wearing different clothes so this should give
us a good place to start now one thing I will mention that you can do this is absolutely optional is if you are not 100% happy with any of the images before we move on to our next step you can go ahead and open them and scenario actually has a upscale feature so you can go ahead and click it and use a combination of face restoration and enhance upscale I suggest doing one or the other when you use both results are not great if you want to try and get a slightly more enhanced version of your ima
ge at this point I'm pretty happy with them I might put this one through just because it looks a little airbrush so you know what let's actually run it through and we can see it has a slightly lesser airbrush look the face looks a little bit more textured in my opinion it has mixed results if you have comfy UI you might want to run the images through a latent upscaler but for now we'll stick with what we've got so the next thing we want to do is we want to go ahead and download the images an imp
ortant thing to note is we've just used scenario to generate the sample images for our Laura we are downloading them and we will be re-uploading them into the Laura generator shortly because of that if you have a different model that you are happy with something that you've gotten off of civit AI and you've run the images through comfy UI or automatic 1111 and you've gotten a collection of images that you like feel free to use those as well by the way thanks for staying with me this far if you'r
e finding this video at all helpful this is a reminder to smash that like And subscribe button and if you really want to support the channel and get access to exclusive or early content content please check out the patreon it really helps the channel out and allows me to continue making videos thank you to everyone who's already a patreon supporter you guys make all the difference once we've saved the images that we want let's go ahead and come back over here to model and click the plus button o
ver here we have two options train your own model and compose with position we'll come back to this one later for now just start training give your model a name since we've called our character Lucy we're going to call this Lucy one keep your model type as stable diffusion Excel Laura and then set your preset to character and by the way if you'd like to see me do a style Laura please leave a comment in the comment section below I'd be more than happy to make one once we've done that we're going
to go ahead and grab the images that we have downloaded earlier make sure you drag them onto the box and not into the empty space as it won't register it scenario will then ask you if you want to crop the images since this is training data it will only except Square images 1ex One ratio so if you've created images from another platform such as mid Journey your own photos or on another stable diffusion UI please make sure that they can be cropped in a square format in my case because we generated
them using scenario they're already in the correct format so we're just going to go ahead and upload them one of the things I absolutely love about scenario is that it makes it super easy to train Aur in fact if you haven't trained Aura before this is a fantastic starting point because scenario does a lot of the heavy lifting for you making it extremely easy to Dive Right In that just by dropping in the images scenario has already suggested some learning parameters and even gone and used machin
e learning Vision to put some pre-written text en codes for the images it's so easy in fact that you can just click the start training button right now except for one little thing and that is that we have to give a name to each of the encodes now we could use the character's name Lucy but Lucy is a common name and that could cause confusion with the model what we want to do is add in a word that represents the character but is unique that the model might not already know so an easy thing to do i
s just remove the vowels so let's just add in lcy and just by having done that if you don't want to go any deeper you can just press a start training button however if we want to get the most out of our Laura training we need to put a little more time and effort into captioning our sample images what I just described earlier about choosing a unique word is called setting a trigger what that is is it's setting a unique word that the base model won't know and the Laura will now associate its train
ing with that trigger in addition to that when captioning images we can think about other descriptive words in the form of classes and descriptors classes are ways to categorize key elements common in most of our image sets such as woman man sword Etc when we include these words we are linking our trigger to these broad categories that the model already understands then we also have descriptors these are smaller more unique items that are not in all images by including these we help the model un
derstand that they are not part of the character but can occasionally be present these would be things such as colors clothing accessories hair lengths Etc these are things that we can change from image to image and we want the model to know that so now that I'm done with the captioning you can see not only do we have the trigger word I've gone in and tweaked the descriptions to cover things that the character might be wearing how their hair might be done you'll also notice that all of them have
a woman to represent the category we've got the trigger word and then the rest of them are the descriptors that we mentioned earlier ultimately when deciding the text encoder we want to keep in mind that the more we identify a name across all images that is the same the more we will need that word to keep things consistent now that we've captioned the images moving on to advanced settings as I mentioned earlier it's enough to leave them as is if you don't want to get into that start training ho
wever if you want to understand what's going on here the total training steps is the number of times that the model or the trainer will go over the images in an attempt to learn who the character is by default the trainer will multiply the number of images that you have by 175 and this formula is true for image sets between 5 and 20 images this will give you a quick character training now if you want to do a longer character training I've put some parameters up on the screen for you these are pr
ovided by scenario in case you want to experiment the next parameter is unet learning rate this is the rate at which the model learns from the images so meaning each training step each pass it will put more or less of an effort to understand what's happening in the image in this case it is 5 to the power of minus5 the important thing to remember is that the higher the number the faster the model will learn that means that it will go through each step faster picking up less details now you might
think why don't we Crank that number all the way down so that we get the maximum information out of the images the reason for that is something called overtraining or overfitting can happen where the model will go through the images too much that it won't be able to create variety when you try prompting it so the goal is to find a balance between the model understanding the character enough but not so much that it can't generate variety from what it knows on the other hand we don't want it to le
arn so little that it can't keep consistency in the Laura the default settings usually work fine you can always come back and tweak it if you're not happy with the output of the model next we have the text encoder training ratio like the total training steps this is the parameter that involves connecting what the model understands from the image to the words that we've provided the ratio is the percentage of steps that are used to connect visually what the model understands with words again you
can leave it at default I have experimented between 0.2 and 0.35 and the results are usually fine again you can come back and tweak this if you feel that the text prompts you're giving don't line up with the images that you're getting maybe you want to let it train with a slightly higher Ratio or if you need to use more words that you'd like you can lower the training ratio and finally the text encoder learning rate similar to the unit learning rate is how long or how much of an impact each trai
ning step of encoding will put on the image so use this information to experiment with the settings if you're not satisfied with the training results as I showed earlier here are some alternative settings that you can try from Scenario for longer training Cycles and in fact so that you can understand the difference in the quality of the images the encoding the text and the learning rates here are some images from one of scenario's artists who showcases how important text encoding can be by compa
ring bad text or average text en coding with low images average text en coding with a high number of images and high quality text en coding with a small number of images and a large number of images and you can see that the best situation is obviously where you have a large data set and fantastic text encoding but in the event that you don't have a lot of images great text encoding can really make a difference in helping the model understand your character so that you can get expressions and mor
e Nuance out of it so if we're happy with the settings let's go ahead and click start training now the training time varies depending on your settings and the number of images that you're using it can take anywhere between 20 to 30 minutes to 2 hours while we wait I want to use the opportunity to tell you a little more about scenario aside from having a really powerful and simple Laura generator scenario has great tools to help you build out characters including their canvas tool which you can u
se to produce a plethora of different Expressions change your character's outfits fix things with regional in Painting and More as well as tools to upscale as we saw earlier pixelate which allows you to convert your images into a more pixelated format to f F in with whatever art style you're going for a vectorized tool which similar to pixel8 converts your images to a vector format and isolate which helps separate your character from the background in case you want to use them in different backd
rops different situations or even as a logo it's an extremely powerful tool and if you want to try it out don't forget we have a discount code for you as well as a referral Link in the description below if you are interested in using scenario for game development and are part of a studio please feel free to reach out for me and I can connect you with the sales team coming back to the Laura once it's done you'll be presented with a screen like this showing you that the model is done the name and
telling you that you haven't created any images so let's go ahead and try and create some images of Lucy so here we're just going to put in a prompt of Lucy doing something let's say Lucy in a cafe drinking coffee so even though I made an error here props if you can uh spot it before I tell you put it in the description section below if you figur it out what the problem is you can see that we are getting a relatively consistent character in these images it's not perfect some of the face widths a
re not consistent and that could be because when we generated the text encodes we did not call the trigger Lucy but instead lcy and I've put here Lucy so let's go ahead and try that out again so we're going to actually use the trigger to try and call her in let's see if we get slightly more consistent images and would you look at that the faces are considerably more consistent after using the lcy trigger word now because this is a stable diffusion based model it runs into many of the common issu
es that we might experience with stable diffusion in this case the hands are not the best and even though we have specified that she's in a cafe drinking coffee stable diffusion insists on giving us multiple cups of coffee so let's clean that up slightly let's add in here bad hands let's put in here one cup bad face disfigured and let's see what we get so still not perfect but we do have one image here that could potentially be a winner so one of the really cool things that we can do with scenar
io is if we found an image we more or less like we can take it into the canvas and this allows us to do slight modifications to our image in this case I'm going to try and change her facial expression and again we may have mixed results just because of the limited data set that we created in our Laura but just to show you how versatile this platform is I think this is going to come out relatively unnatural simply because of the nature of the pose but let's see so as you can see the model has suc
cessfully redrawn the face unfortunately we've lost a lot of the facial features of Lucy you can see see the face changes somewhat but with a little bit of fine-tuning I'm sure we could get that working the main thing I wanted to show you was how powerful this is and how cleanly it's replaced the face a few other things that are worth noting in the generate page before we move on is you'll see here that there is a section called Lowa components you can actually adjust the strength of the Laura s
o the images that we've created so far were the Laura set at 80% let's crank it up to one and see what that does this time let's take Lucy Into The Streets of London okay we're going to put in some symol prompts Lucy walking down Oxford Street we're saying it's winter and that she's wearing a sweater and we can see here Lucy walking down streets that could very much look like they're London you can even see a little red bus over here like with the hands one of the issues with stable diffusion is
when you make the subject smaller the image quality drops slightly we can try and fix this as usual with prompts and you can see here just by adjusting the prompts the image quality does get better one final thing I want to show you before we finish up that is one of the absolutely coolest features in scenario is the ability to blend models so we've created our character Lucy now we've created a few images of her in natural situations but what if we wanted to turn her into a cartoon character l
et's go ahead and go back to models and this time we're going to click on compose I quite like this viven Laura so we're going to grab that one and we're going to grab Lucy and we're going to set them both to a strength of 075 let's call this Vivian Lucy and we're actually ready to test it out and this is a really cool thing is before you save that Laura composition you can go ahead and test out a few images so let's put here Lucy sitting by the cide and see what kind of images we generate when
we combine the two luras so not only do we have a new character in a completely new art style that character now remains consistent across all the frames we can see here the new character has a style that blends the two luras together kind of this semi- painted look and she remains consistent in all the images if we add line art to the promt we can see how the art style continues to evolve while still remaining consistent if we're happy with the output we can go ahead and save it and here are a
few other examples that I've tried I introduced you to Marina earlier and this is what some of her images look like as a pirate Queen when combined with another Laura and here is a grittier version of saraphina so as you can see the ability to combine luras is incredibly powerful there's even one that gave Marina Asian features and you can see the results continue to be relatively consistent in fact across two anime models it gave us Asian Marina and that is just the tip of the iceberg of what y
ou can achieve with scenario if you want to keep track of the girls saraphina and Marina you can check out their socials in the description below as well as my own socials on Instagram and Twitter go ahead and follow if you want to see how I improve the quality of the luras as well as tips and tricks to get the most out of your image generation if you found this video helpful please don't forget to like And subscribe and if you want to support the channel as usual please go ahead and check out t
he patreon thanks again to all of the supporters that make this happen it makes a tremendous difference I hope you all had a good time and I'll catch you on the next one
Comments
I'd like to see a style modal.