This New AI Tool Does Consistent Characters (For Real)

hey everyone so consistent characters in AI imagery has long been a bit of a Holy Grail I know there's been a ton of videos with methods and processes to try to achieve them but let's face it the results have been janky at best there is a reason why I haven't made that video at least until today so let's dive in and check it out so in a video last week I briefly covered scenario which is a pretty interesting new AI image generator that actually has its roots in creating AI assets for games poking around I did think that scenario had some pretty interesting stuff going on and I said that we would probably Circle back to it pretty soon and boy here we are uh because yeah this is the solution that many of you have been looking for so first off credit where it's due I first ran across this workflow in a thread by Vlados AI who actually works at scenario as their customer success specialist and AI artist second I do want to mention that scenario is free kind of at least uh there are obviously paid tiers as well but everything everything that we're doing today you can do on the free starter plan the free starter is pretty generous I actually got through a ton of testing for this video just on that plan before I eventually jumped up to the essential plan to finish things off so once you've signed up for scenario that link is down below of course uh you will be greeted with the homepage now there is a ton of stuff in here today we're just going to be concentrating on creating a custom character and then prompting for that character so to start with you want to come over to this sidebar and hit the models tab you'll be presented with a number of different aesthetic Styles um I wouldn't get too hung up necessarily on trying to find the exact perfect thing just find something that's sort of in the vibe of what you're looking for considering that we'll be blending and remixing things as we move on once you've chosen a model just come over and hit generate with this model briefly going over our interface I did swap out our model to Illustrated anime and creepy Nordic fairy tale which is actually kind of funny are there non creep creepy Nordic fairy tales because this is a blended model we actually have sliders here that will control the overall amount of influence that each model has on our output for example if we were to crank up Illustrated anime obviously we would get a stronger anime style uh whereas if we crank up creepy Nordic fairy tale we just probably end up with lots more hungry wolves underneath that is our prompt box uh with two interesting little features um there's a prompt Builder here which will provide you with a number of keywords or ideas that maybe you want to think about say it's a retro art style in chalk art additionally there's a feature called prompt spark in which if we give it sort of a basic prompt let's say wolf in a dark forest and then by hitting the prompt spark button it will of course give you a number of different suggested prompts below that via some pull Downs are of course our negative prompt negative prompts do play a pretty large part in scenario so don't sleep on them uh and then your reference image is where our character creation now Begins the first thing you want to do is take a template of either a male character or a female character and no I cannot stop looking at the massive wedgie that the male template has either but listen that dude has clearly been putting in the Reps he deserves to wear that Speedo you can use a template of your own choosing if you'd like but both of these are available in a download along with a PDF that outlines all of the steps in this over on gumroad it is completely free but as always if you would like to leave a donation it is always super appreciated so for this character I'm going to aim for a female so uh just simply dragging her in here uh make sure that your mode is set to image to image and then you can experiment a little bit but I would initially Begin by putting the influence down to Zero from here we can begin prompting I wouldn't suggest using the prompt Builder or the prompt spark functions at this point uh just be sort of specific but General about your character including things like the age hair color clothing type uh that sort of thing the most important part here is to add impr prompts for front view side view and back view letting the model know that you know you want the same character in three different viewpoints as opposed to creating three different characters from here it really is just a matter of iterating uh changing your prompts and rrolling a bunch of times you know it's still AI it's still going to make weird choices you will end up with results that aren't necessarily coherent with the other viewpoints for example uh this like like I guess like a hair Shield that Freya here is wearing that she clearly does not have in the other viewpoints but if you keep playing around with it you will eventually end up with something that you like this was an interesting one where I turned the influence up to I think like 13 or 14 um and as you can see it kind of has this sort of washed out look that is very representative of our templates um kind of gray drab look to it I ended up landing on this model which I actually really liked interestingly by calling out Siri and witch in the prompt so uh I think that perhaps the model might know something about The Witcher considering that she did end up with white hair obviously not a onetoone of Siri from The Witcher but that's okay I think that it's always more interesting to create your own characters uh from here what we'll want to do is hit the upscale button in which you will have options to two exit or 4 exit uh you have face restoration photo optimize and enhanced up skill just within scenario I more or less recommend just sticking with enhanced upscale now that said what I did I took it over to magnific and did a creative upscale over there um magnific is just become a real go-to for me lately it's the results are really really pretty stellar yeah there's magnific has something special going on so anyhow taking this and downloading it so take your character image into Photoshop or really any photo editor will do this is very basic um and all we're going to do is just cut out uh the side view the front view and the back back view into separate views so basically we'll end up with this this and this additionally come back to your main template and just uh kind of do a square image of each side of the face so that we'll also end up with this this and this I guess it's kind of the Nordic mug shot here so I don't know if this necessarily mattered or not but one of the things that I did to buy another Viewpoint is essentially take that side profile and just flip the canvas that way in order to end up with you know her looking to the right as well anyhow once you have your character all sliced up it's time to head back over to the main menu and begin training once again we move over to the sidebar and hit the plus button next to models uh we'll be presented with two options train your own model or compose with Precision you'll want to start training with your own model drag in all of your looks um as a quick note it does crop to one one for each of the training images so um we can just accept this and when you run across say these horizontal images where they don't fit into a one one you can just actually just scroll the image down um so that it does fit with in the bounding box um it doesn't seem to have too much of an effect with the white space on either side as a quick note because it took me stupidly longer to figure this out than it should have uh you do need to actually take this scroll bar down in order to hit the accept button from here what you want to do is go into the description box and change that around uh here you can actually name your character so I named her Siri uh and then let it know here's a turnaround here's her head facing left here's her head from the back um it just kind of gives the model just a little bit of an extra nudge and probably reinforces who that character is to the model after you have all of that done just come up and name your model uh in terms of the advanced settings here I'd be a little bit wary about fussing with that too much namely because in an earlier experiment I FedEd a ton of um these stock photos of kind of like the this you know cop show kind of thing um yeah a bunch of them and then I cranked up all of the training knobs not all the way but you know pretty high I am a guy that likes to turn things to 11 but the results kind of end up speaking for themselves so yeah uh maybe just leave the default settings on so now we have our model the thing is that you don't want to just start using your model right away uh if you do you do more or less end up with these extremely Bland kind of things like this is just her holding a sword it doesn't have the sword in her hand she's just kind of standing there it's essentially the default pose um mostly because this model doesn't really know what a sword is the solution to this is that we will now need to blend our model with another model uh to do so we head back over to the main menu again uh come back up to models hit our plus button once again and this time we want to hit compose with Precision head over to your models and then add in your model the easiest thing to do from here is just to click public models once again and then choose another model model that you think is kind of Representative once again of the style that you're looking for so we'll go back to creepy Nordic fairy tale again but you're not just limited to one so uh let's just add in cell shaded illustrations as well um again you will be able to control the amount um that you would like for each to influence once you've added them in adjusted the sliders to your taste um you can title it now here's what's interesting is that you can actually come down to this test button and run a couple of prom just to see if you like it or not overall it's kind of cool it's definitely not the style that I'm necessarily going with which is kind of what's neat about having the ability to test you can continue to play around with the influence sliders or choose different models ultimately I just Blended my Siri model with creepy Nordic and ended up with some pretty decent results here she is you know standing in front of a very snowy Castle uh holding her sword uh a couple of other quick uh variations on that you can see obviously it is still very much our character now one thing on the importance of negative prompts I did tend to end up with a lot of like elf ears um so that is something probably influenced a lot by her overall character and the fact that we're sort of in a fantasy setting so um that was easily taken care of by negative prompting elf ears now if you're looking for specific poses for your character you can do so by either image referencing or using something called control net so by coming down to the reference image and then adding in I'm just going to actually use a picture of the real Siri so taking this image of Siri and then bringing her in as an image reference we can either choose to use image to image in which it will take sort of that mid Journey thing of like I'm looking at this whole thing and I'm kind of getting the overall idea of what you're looking for but I'm going to go ahead and do my own thing or we can take this tab and come down to control net um and this offers a few more options so using control net will end up with a secondary pull down uh with character landscape structure pose depth Etc ET um character obviously being character landscap being landscape so an interesting one is if you come down to pose that will open up this mode map here now if we click inside here you can actually see sort of an outline of what it thinks the character is doing in this kind of skeleton shape while it's possible that you end up with pretty much a one: one um it may take a little bit to get there like this isn't quite on it looks fairly acceptable um this one has an elbow that's completely jacked backwards this one's not bad except of course she's holding her sword in the wrong direction unless she's about to do that like cool hero thing where she stabs somebody from behind just keeps on walking straight without even looking at him I mean I didn't prompt for that but she clearly does what she wants but again I just want you to know you will still end up with stuff like this it is still AI art you still do need to roll numerous times but I do think that that will all sort of end up working out as we move along for now the real win here is the fact that you know you you have this consistent character even when I add another model in like I took our classic 3D realism and then added that into our Siri model and our creepy Nordic fantasy and ended up with this image which is still clearly her just in a new style now briefly and I don't want to get too crazy here CU like scenario is an entire Rabbit Hole uh I do want to point out that when you do end up with an image say like this which I really did like this was Siri hanging out with a barbar in in some tavern having an ale uh and then obviously I ended up with like two Siri heads here you can uh use the canvas function in scenario to uh edit that out so admittedly I haven't fully delved into all of the canvas features in scenario but it is interesting in that it works in you know uh sort of your standard uh in painting out painting style you know you make a square or a lasso or a brush you prompt for that uh run the prompt and then you have four options to choose from but what is interesting about scenario and I haven't really played with this too much is the fact that you actually have this layers tab over here in which you can try out different Generations so uh yeah that's kind of cool um is that James Franco looks like James Franco so yeah as you can see scario is a pretty interesting platform and for such a young platform I'm actually really excited to see how it ends up developing again you can check out scenario for free so you know nothing lost there the link is down below is the link to the PDF guide on creating consistent characters I guess that's it for me I thank you for watching my name is Tim

Channel: Theoretically Media

Views: 156,673

Keywords:

Id: v_FXC0iq1Sk

Length: 13min 36sec (816 seconds)

Published: Thu Jan 04 2024