this is the best way for creating consistent characters with AI whether you want to create cinematic AI movies children's books or AI influences you'll need to have your characters stay the same throughout multiple shots but this is actually quite the challenge even with my consistent and posible character workflow that I showed you in one of my last videos you would still need to try out multiple prompts and seats to get it right and details especially on complex clothes still might change a little bit from image to image but that has changed now thanks to a
new image model called flux I updated my workflow to it and look at these amazing results I can now generate images where the characters look exactly the same whether they are animated or liveaction oh and this also works for multiple characters in the same image which is especially important if you want to tell a story so let me show you how it works and how you can set it up on your own computer for free the trick is that we first generate a character sheet that's depicting the character from multiple angles and with different emotions
and for that I create this pose sheet that shows the characters bones in the open pose format and we can now use flux stable diffusion XL or even stable diffusion 1.5 to generate our character sheet you could use this post sheet with most AI interfaces but I created a workflow for comi which automates all of the steps and as usual I've created a free step-by-step guide on how to install comi and where to download and put all the models for this to work and let's start with the flux version once you have everything set up
you can just drag and drop my workflow into the comi interface this workflow consists of four steps but the last three will happen automatically and we don't need to use them yet so we can deactivate them using this fast mutter here let's focus on the first group the only thing that you need to do here is double check if the correct models are loaded and import the post sheet here and you can just drag and drop that in there for the flux version you can use a prompt format like this and the weather is getting
quite cold out there so let's create like a character for like a fashion magazine showing Autumn fashion you basically just leave all these keywords here and add short descriptions of the character that you want to have and that's all you have to do in this first group if you want you can also add like a name for your character here but then you can just click Q prompt and wait for the first step to finish so the prompt is working but this image is a bit dark so what you can do is just change the
seed a little bit and try out a new one so when you found a character sheet that you like and let's just say even though the hair is bit red instead of brunette and the Cod is more like gray I would say a little bit Brown let's say you found an image that you like a character that you like and you want to continue with it then you can just activate the second group and click Q prompt and this will then upscale that image this already looks a lot better than this first image here you
can see the faces and everything is really broken here it already looks a lot better and then then in the second step in that group it will actually go through all the individual phases and just add more detail and make them a bit cleaner for the next step it will be important that we have images of all the individual phes and we can just simply save them out by activating this node here using contrl B if we now click Q prompt it will actually save out all the individual faces next we can activate the third
group and click Q prompt and this should only take a second or so this is another group that will only save out the individual poses and now we can already activate the last group and click Q prompt and this will quickly create different emotions for our character I just set some example emotions here but you can customize them you have all these sliders and you can just change them and find the emotions that you like maybe we can make her even more happy in in this one by just raising the eyebrows no that's not happy
but um it doesn't matter too much which emotions exactly you create it's uh more about having a a wide variety and what you could also do is like rotate the head so you have like different angles so let's just change the rotation here into the other direction change the rotate pitch and just see what that creates now we have her looking to the right we have a looking to the left and this will be very good for the next step if you support me on patreon you can also get your hands on the advanced version
of the workflow which will add upscales to different parts of the workflow for example it will upscale the individual face images for some better quality and it will also upscale the emotions at the end here sometimes it can happen for like stylized characters like pixel characters or anime characters that there are some weird glitches around the edges of your face and using this upscaler here you can fix that now it's important to know that the emotion setup will not always work perfectly for any type of character like for example here I created this lizard person
that I showed you in the beginning and I really love how well that came out but for the emotions the face was just a bit too far away from a typical human phe so it was not able to add the emotions correctly but at least we got some different angles out of it another important thing to note with the flux workflow is that it can be a bit tedious to work with for one it's pretty slow even on better computers and the control net is just not as robust as for example the sdxl control net
so when it works it typically works really well like for this character for example but you probably get a lot of weird and broken results as well and I don't want you to get frustrated so I also created an SXL version of this workflow which is not only a lot faster but it will also give you way more consistent results so you can see it looks pretty much the exact same except the beginning here is a bit different we don't need as much stuff I'm using the Wild Card turbo model which I've been using a
lot in my videos because it's like super fast so the settings are fine-tuned to this model if you want to use another other sdxl model make sure to change the K sampler settings here but let's quickly try out a prompt like for example let's use this one to create an anime character I'm just activating all the groups here and just let that run through and this only took a couple of minutes it did an amazing job of creating the sheet the faces look cool and even the emotions look really good but you can see that
for example this one is a bit broken so you could try to fix that by just changing it a little bit or you can use the upscaling in the advanced workflow to fix it so now why did we do all this well we now have a perfect data set to create a Lura Aura is basically a way to give an AI model more context for an object or a subject and it's definitely not a new technique but it has never been so easy let's say we want to create a Laura for this woman here that
we created earlier let's first just create a folder with all the good images of her think of a unique name for her maybe maybe she's called a i something that's not a typical word and let's just look through the images and copy over the images that have a high quality and show her from multiple angles I want the different emotions and this is honestly probably already enough so I usually use between 10 and like 15 images to train the Laura let's just look through some of these additional face poses maybe we can use some of
these extreme side views here that's cropped in a weird way so I just open that with paint and just focus on the face here and I would say that's a pretty good data set to train the Lura we need a tool called flux gy and they call flux gym a dead simple web UI for training flux lauras and I can only confirm it it really is that simple the easiest way to install it is using Pinocchio Pinocchio is this super easy oneclick installer for many AI tools so just download that go to discover and search
for flux gym and install it from here once it's done you can just start it put in the name of our Laura and I'll just call that Tina Ai and I also want to use this as the trigger word as well so when we put that in the prompt and we have the Laura activated it will create this character you can choose which model you want to train I want to train flux Dev and I have more than 20 GB of vram so I choose that but I also tried this tool on one with 12
GB and it worked pretty well as well just takes longer that's all we need to do here and then I just take my images and drag and drop them here it had already added the caption with our trigger word and now we have two options we can caption the images manually or we can use Florence 2 to tag our image so this is what it would look like so here you can see for example Tina AI a woman with long red hair wearing brown turtleneck sweater and a gray coat she is standing in front of
a wall and these are actually quite good I usually prefer tagging the images myself and keeping it a bit simpler for example let's just drag and drop these images in here again so what I like to do is gray background and then the pose and emotion and um shot type something like this this should be enough and that's pretty much it now we can just click start training and it will download the model which will probably take a long time the first time you run this because the flux Dev model is huge now I just
let that run okay it's Now 37 minutes later and the Laura is finished so we can go to the Pinocchio folder API flux gym uh outputs and it created a folder with the name for our Lura we have different variations so it's doing like snapshots at different stages of training but let's just try out the last one go to my comi directory comi models luras and I created an extra folder for flux and my own flux gym models and I'm just going to put that here so this is my custom flux workflow we have the
option to input two luras here but let's just deactivate one and only use this one here and we can put the strength pretty high I sometimes lower it to give it a bit more freedom but something between 0.9 and one uh will work so let's just use this prompt click Q prompt and this is just the first image and it already looks really good let's just try out another seat and this is also looking pretty nice it's definitely her let me quickly explain this custom flux workflow that I'm using here and you can also get
that for free on my patreon basically the beginning here the beginning stages is just the typical comu iflex workflow and then in for the second image here I added a little hack that I learned from materal from latent Vision basically during somewhere in the middle of image generation you add some extra noise and and this will make the textures of the image look less like plastic and more realistic and it will also add some Minor Details after that I added an upscaler you can also deactivate this part if you don't need that and then finally
I add some film grain using this node just for this analog type of um photography style that I'm going for here and generating images with this workflow and the luras is so much fun because pretty much every single image looks cool no cherry picking these are just all the images that I created the same goes for this character here this is just a character that I trained on this character sheet to try out if this workflow works for other Styles as well it looks really good like these are all the images that I created but
as I said this workflow also works for multiple characters and it's actually really easy you just need to load in both luras and I like to decrease the strength a little bit just to give mlux a little bit more flexibility and then I'm going to use a prompt like this first we describe the general composition of the shot mentioning both keywords of the Laura and how they relate to each other so they are standing next to each other to keep them from from mixing because that's usually what happens when we stack luras on top of
each other it creates these sort of hybrid Styles and creatures so to actually fix that I just add character descriptions for both of these characters so as the image is generated the luras actually have a chance to fall into place in the correct regions and and create images that just look really good I hope you enjoyed this video and give Laura training a try if you would like to support my work and gain access to the advanced workflows and exclusive example files like all the data sets character sheets luras and prompts I used and created
for this video consider supporting me on patreon your support makes this channel possible so thank you very much and see you next time
