How to Make Consistent AI Characters in Higgsfield AI (Step by Step) Transcript, AI Summary & Key Points

00:00 One of the biggest challenges in AI video creation is keeping a character consistent from one scene to the next. And after years of creating AI videos, I found a workflow that locks in a character's face, body, and outfit from the start, so you can reuse them across as many videos as you want. So, in this video, I'll show you the entire process from creating a perfectly consistent character to changing their outfit or even their art style without ever losing their identity.

00:24 And by the end, you'll be able to keep your characters looking the same in every single project. The tool I'm using through this entire video is Higsfield, which is an all-in-one platform that gives me access to everything I need to keep a character consistent. If you want to follow along, I've left a link for it in the description below. Now, there's a specific reason why your characters keep changing between shots.

00:44 And the best way to show you why is to first build one the way most people would with nothing but a detailed description. Once you log in, you'll notice that everything is laid out along the top navigation bar. So, I'll head over to video and from the model selector, I'll pick Sedance 2.0. Now, by the time you're watching this, Cedance 2.5 is already out and it's raised the bar completely with way better character consistency, longer clips, and cleaner motion.

01:08 So, definitely give that one a try. I'll set it to 8 seconds, 16x9, and 1080p since this is just one short shot to see how well the character actually holds. Then I'll paste in my first prompt, which is an extremely detailed description of one specific woman, right down to a small beauty mark under her right eye and the cluster of freckles across her nose walking through a busy street market.

01:31 And you can find this exact prompt along with the whole workflow down in the description below completely for free, so you can follow along in real time. At the first glance, this first one comes out great, but that was expected. Seed dance is good enough so that from one specific prompt, it can create highquality results. The freckles are there, but the beauty mark is actually under her lips, not under her eye.

01:58 This is the first hint that a prompt alone isn't enough. But to actually see how she holds up in other generations, I'll keep that same description and only swap the scene to a rooftop at golden hour. >> [laughter] >> And already she is not quite the same person. Her jaw is softer and the beauty mark is now under her eye, but it doesn't match the woman in the previous scene.

02:27 If this was one continuous project, the character already looks different, even though I used the exact same description as the first one. For the third one, I'll do the same thing again and change the location to a dance studio. >> [sighs] [gasps] >> This time her face is rounder and the beauty mark is bigger. For standalone videos, these look good.

02:51 But for a project with actual continuity, I described the same woman and got three different faces. A written prompt can describe a scene, but it can never lock down an identity. So, the real fix is to stop making the model guess from just words and give it one solid reference instead. And that one reference is a character sheet, which is the single most important asset you'll need for maintaining perfect consistency.

03:12 So, I'll head over to image and from the model selector, I'll pick GPT image 2 since it's one of the best models for a photoreal face that actually holds every fine detail you give it. I'll set the quality to high and the resolution to 4K since a sharper reference gives the video model a cleaner character to work from later and keep the aspect ratio on 16x9.

03:34 For this one, I'll build a character called Elias, a late20s expedition ctographer. And I'll describe his identity once at the top along with his full outfit with lots of details. Then, instead of describing a scene, my prompt describes one image split into three panels. And each panel has a specific job. And this gives me one clearly consistent person from all three angles with the same build and the same outfit.

03:57 The left panel is a full body shot from the front [music] with the head removed. So, the outfit, the build, and the proportions are all easy to see with no face competing for attention. In the middle, it's that same body turned around with the head on, so you can see how everything looks from behind. And the panel on the right is a tight closeup of the face looking straight into the camera.

04:16 And this is the one that locks his facial identity. All three panels are on a flat gray background with no shadows. And that gray background is intentional because it produces much more consistent exposure. A white background naturally reflects more light, making your character appear slightly brighter, while a black background can make the character seem darker.

04:35 Those subtle changes can carry over into the videos you generate from the character sheet. A neutral gray avoids both. Keeping the exposure balanced so the model [music] focuses on the character itself rather than the background. Now, to actually turn this into something I can reuse, I'll save the sheet as an element, set the category to character, and name it Elias's row.

04:54 From now on, I can just tag at Elias Row in any prompt and Higsfield pulls in this exact character every time. So now, instead of spending time writing a detailed description, I can just tag my character in multiple videos and he'll come out the same. But saving him is one thing. The real test is whether he actually holds once I start dropping him into completely different scenes.

05:15 And to actually test it, I'll build five completely different scenes to see how our character will look in each one. So I'll head back over to video Pixie Dance 2 again and set it to 15 seconds. 1080p and 16x9 since these are full scenes and a few of them have dialogue. And instead of writing out a single word about what Elias looks like, I just tag at Elias Row in the prompt and only describe the scene around him.

05:38 For the first scene, I'll drop him onto a rocky mountain ridge at dawn with this prompt and generate. And he looks exactly like his character sheet now placed in the scene I described. And I never explained how he's supposed to look like in the prompt. Higsfield got everything from the saved element. Next, keeping the same settings, I'll put him inside a small tent in the middle of a mountainstorm at night.

06:11 And this time, I'll even give him dialogue. This is a real test because the second a character starts talking, you need accurate lips sync. Every chart of this valley is wrong. Good. Wrong means nobody's been in it. And most models can even alter facial consistency, but Cedance handles this one perfectly. The lip sync is accurate, the consistency still holds up, and the lighting and audio make everything feel cinematic.

06:45 For the third one, I'll drop him into a snowedin cabin while he's preparing to go out to the blizzard. >> All survey teams are ordered back below the pass. The route is closed. What makes this one hard is that it contains a lot of movement and big changes between him, the blizzard, and the lighting change, but everything holds up, especially the transition when he opens the door.

07:14 For the fourth scene, I'll put him in a dim tavern at night sitting across the table from an older guy. So, this time there are two people in the same shot. >> They'll never print your name, Ro. The mountains don't care who draws them. And normally a second phase is exactly when the model starts combining them together. But Elias looks exactly like he's supposed to while interacting with the other character.

07:46 And for the last one, I'll drop him into a cramped map shop at dusk, standing across the counter from an old map dealer inspecting his work. You drew this yourself. Men get buried over maps like this. This is a totally different look again with a second character right next to him. These are five different scenes with completely different environments, movement, and lighting.

08:16 And no matter what changed, Elias still looked exactly the same from his outfit to his face, even from multiple angles. But we can take this a lot further. Right now, he's stuck to that same outfit. So, it's only a match if I put him in a setting where he's meant to dress like that. But what if I want to put this same character in a completely different place and still have him look the same?

08:36 With Higsfield, you can do just that because his whole identity is saved into that one sheet. I can change everything around it and still keep the same person starting with his clothes. So, I'll head back into image with GPT image 2 on the same 4K settings. But this time, I'll upload the base sheet as a reference and prompt for a brand new outfit while telling it to keep his face and body exactly the same.

08:59 First, I'll put him in a heavy winter setup. Now, he's wearing a thick wax jacket with a sheerling collar and a beanie, and it's clearly still Elias, just dressed for the cold. Then, I'll switch him into a clean, smart, casual look with a white tea under a navy blazer. And the face is exactly the same, even though the whole feel of him just changed.

09:16 Next is a desert field outfit with a light linen shirt and a widebrim hat. And again, it's the same guy, just ready for a completely different climate. After that, I'll give him a rainy coastal look with a dark rain jacket with a hood down. And last, I'll put him in a full formal outfit, a charcoal suit with a white shirt, which is about as far from his field gear as you can get, and it's still unmistakably him.

09:37 So that's five completely different outfits, each one suited to a different kind of scene, all worn by the same character. And you can save each one of these as an element the same way, so you can drop them in any video you make. Now, changing his clothes is one thing, but I can push this a lot further and change the entire art style he's rendered in.

09:55 I'll use that same base sheet as the reference again, except this time I'll ask for him in a totally different visual style while keeping his key features so he still looks like the same person. First, I'll turn him into a clean anime character, and even as a cartoon, you can immediately tell it's still him. Then, I'll render him as a 3D animated character, the kind you'd see in a Pixar style film.

10:14 Next, I'll drop him into pixel art, and even at that low resolution, he's still recognizable. After that, I want a heavy comic book ink style with a thick black lines of a graphic novel. And last, I'll make him a claymation figure, like a handmade stop-motion puppet. And even as sculpted clay, it's unmistakably the same character. So whether I change his clothes, his setting, or even the entire art style, the identity always holds because it's locked to that one saved sheet while everything else around it changes.

10:42 Now, everything we've done this far has been with a completely madeup character. And the real question is whether this same system works on an actual person like me. So to actually test this, I'll start by making a character sheet of myself. I'll head back into image with GPT image 2. And this time, instead of describing what I look like, I'll just upload a reference from one of my videos.

11:02 You can do this with your own picture. Just make sure it's one where your face is clearly visible. And it gives me a clean character sheet of myself with the same front, back, and close-up layout built off that one frame. Then I'll save it as an element. So now I've got a locked version of myself I can tag into any prompt I want. Now to actually put myself in a scene, I'll head over to video with Cedense 2.

11:21 Set it to 15 seconds in 1080p, but this time on a wider 21x9 for a more cinematic frame. For the first one, I'll tag both myself and Elias Row in the same [music] prompt in a scene where I meet him inside a cave. Finally, I was looking for you. I want to talk to you about your maps. And this is the part that really proves it because there are two separate saved characters in a single generation and both of us come out looking exactly like our own sheets.

11:58 Then for the next one, I'll just drop myself into a completely different scene, sprinting through a metro station to catch the last train. Intense movement like this is [music] what most models struggle with, but I still maintain my consistency from every single angle. So whether it's a character I completely made up or a real version of myself, I can lock anyone in once and reuse them across multiple videos, and they'll look like the exact same person in every single scene.

12:36 This way, your characters will always look the same throughout your projects, and your videos will look professional. So, if you want to make AI videos with consistent characters yourself, use the link in the description to sign up to Higsfield.