Learn how an AI character design workflow turns one rough idea into outfit variations, expression sheets, and turnarounds you can actually reuse.
Designing an anime character can be one of the most fun parts of a creative project, but it can also be harder than it looks.
Let's say you generate a character with an AI character generator, the image comes out better than you expected, and you save it immediately. Then you try to make a second image of the same character, and you find that you can't.
The tool didn't fail you. The problem is that you never actually designed anything. You have one picture, and it only holds so much information. You may not know what the character looks like from behind, or which outfit truly suits them, or what their face does when they are angry.
This is the gap that AI character design has to close. Artists have always worked through concept art, outfit passes, expression sheets, and turnarounds, because each stage answers a question the previous one left open.
AI doesn't skip those stages. It lets you move through all of them in an afternoon, generating five design directions side by side in the time it used to take to sketch one.
In this article, we'll see how we can take one original character through a complete AI character design workflow, from concept to design direction to outfits to expressions to a reusable character sheet you can work from later.
We'll generate every image with Tsubaki.3, PixAI's latest image model, built around control, consistency, and practical creative workflows rather than just polished one-off illustrations.
It's designed to produce the materials that sit between an initial idea and a finished piece, including concept variations, outfit sheets, expression sheets, and character turnarounds, which means the whole workflow below runs on one model without switching tools between stages.
Meet Suzu; she's the anime character we'll be testing with. She's a slightly anxious, over-prepared witch with a smoke-grey cat. She began as three lines of description.
Suzu generated in a single pass with Tsubaki.3, the kind of result that looks finished until you try to draw her again.
What an AI Character Design Workflow Actually Includes
Typically, a character design AI tool will show you the finished illustration. It rarely shows you the stages that produce it, so let's start there.
A complete AI character design workflow moves through seven stages. It starts with early concept exploration, where the idea is still loose, then settles into a design direction once you commit to one look.
From there come outfit exploration, expression development, and pose exploration, each testing the design under different pressure. The last two, a character turnaround and a set of reference sheets, are where the design stops being a series of images and becomes something you can hand to someone else.
However, not every project needs all seven. A one-off illustration may stop at design direction. A VTuber persona leans on expressions and rarely needs a turnaround. A comic needs the turnaround most of all.
These outputs feed each other rather than sitting side by side as unrelated features. Your concept produces the design direction, the design direction is what outfits are built on, and the outfit you settle on is what appears in the expression sheet and the turnaround.
To put this into action, before generating anything, we locked four traits for our character. Suzu has a dark plum bob with a single blunt-cut silver streak on the left side, round amber eyes, a small crescent moon pin at her collar, and one gold earring in her right ear only.
The four fixed traits marked on Suzu's first generation, the details that have to survive every stage that follows.
Everything else stayed open. The hat, the robe, the silhouette, the color accents, and the cat's role in the composition were all treated as variables.
That split gives you something specific to check for later. Without it, "does she still look like herself" is just a feeling.
Start With AI Character Concept Art, Not a Finished Design
The first prompt is where most people go wrong, because they try to describe the final character. That gets you one specific result and no information about the alternatives you never saw.
A better starting prompt defines the character's core and leaves the visual details open. For Suzu, that meant naming the archetype and the personality, an anxious, over-prepared witch, along with a rough age range and the four fixed traits. Everything visual beyond those four was left unstated.
Here is what we started with:
"1girl, solo, witch, teenage, anxious and over-prepared personality,
dark plum hair, short bob, single pale silver streak on left side,
round amber eyes, crescent moon pin at collar,
single gold earring in right ear only,
smoke-grey cat companion,
anime style, masterpiece, best quality, absurdres"
Notice how little that prompt actually specifies. There is nothing about clothing, pose, background, or composition, which means Tsubaki.3 has to make those decisions itself.
The Tsubaki.3 model selected in the Generate tab, with the opening Suzu prompt and the four variations it returned.
That is the point at this stage. You want to see what the model reaches for by default, because those defaults tell you which parts of your concept are doing real work and which parts you assumed but never wrote down.
But, to be clear, that first result isn't the design. It is a reading of the concept, and the useful thing is what it reveals about the gaps. In our case, the robe, the wide-brim hat, and the warm interior setting were all invented by the model. None of them were in the prompt, which means none of them are load-bearing yet. Any of them can change.
The reason to resist locking things down early is that the first result is almost never the best available direction. It is just the first one you saw. The next section is where that gets tested properly.
Comparing Design Directions in an AI Character Generator
One concept image tells you what the model reaches for by default. Four tells you what the character could be. This is the stage where an AI character generator comes in handy, because generating four full design directions is quick.
We used the first Suzu image as a reference and prompted only the outfit direction, leaving the face and hair to carry through on their own. Hair is usually one of the strongest levers at this stage, since a different cut or color can reshape a character entirely, but the bob and the streak were already fixed traits, so we left them alone and let the clothing carry all the variation.
We generated a classic fantasy witch with robes and a staff, a modern witch in city clothing, a herbalist in practical layers with satchels and bundled herbs, and a stage magician in structured tailoring.
Four design directions generated from the same reference image, with only the outfit described each time.
Mostly, the directions came back genuinely separate. We expected the model to drift back toward the wide-brim hat and robe, since the witch archetype is one of the most heavily trained categories in anime models and the reference image was wearing both. Instead, "no hat" held cleanly across all three of the other directions, and the silhouettes stayed far apart.
The unexpected result was smaller and more useful. Look at her collar across the four images. The gold crescent moon pin, one of the four traits we fixed before starting, only appears in the fantasy robe version. It is gone in the city coat, gone under the herbalist apron, and gone on the magician's high collar.
That tells us something about how the model read the reference. It didn't treat the pin as part of the character. It treated it as part of the robe, so when the robe left, the pin left with it. Hair, eyes, and the single earring all carried through without being mentioned once, because the model reads those as belonging to the face. Anything sitting on clothing is at risk the moment the clothing changes.
The practical fix is to name accessories explicitly in every prompt, even when working from a reference. The broader lesson is that a fixed trait is only fixed if the model understands it as identity rather than costume.
The cat disappeared entirely from all four, which is expected since the prompts never mentioned it, but it makes the same point. A companion isn't a character trait unless you keep asking for it.
Choosing between the four came down to the character rather than the images. Anxious and over-prepared is a personality built on being ready for everything, and the herbalist direction is the only one that shows it, since the satchels and rolled sleeves are visible evidence of preparation.
The fantasy witch was a pretty result, and the magician had a sharp silhouette, but neither said anything about who she is. So, we picked the herbalist direction and carried it through the rest of the workflow.
Using an AI Outfit Generator to Build Variations
The herbalist direction gives us a character. But it doesn't yet give us a wardrobe, and a character who only exists in one outfit is still half-designed.
Outfit exploration is where you find out whether the design survives without its costume. If Suzu is only recognizable in a canvas apron, then the apron is doing the work, not the character. What should stay constant is her face, her hair, the silver streak, the amber eyes, and the single earring.
The crescent pin is off the list. The last section showed that the model reads it as part of the robe rather than part of her, and a trait that only survives in one outfit isn't a fixed trait. Rather than fight the model on it in every prompt from here on, we dropped it. That is a real design decision, and it is the kind you can only make after seeing the model run.
Tsubaki.3 can generate several outfit concepts in one pass while holding the same character design, which is a better test than four separate generations. A single sheet forces the model to keep her consistent within one image, where drift is immediately visible rather than something you have to hunt for by flipping between files.
Here is the prompt we used, with the herbalist result as the reference image:
"character outfit variation sheet, four full-body outfits of the same character,
same face and hairstyle in every outfit,
outfit 1 everyday work clothes, outfit 2 formal gathering clothes,
outfit 3 winter layers with a heavy coat and scarf,
outfit 4 field travel gear with a cloak and pack,
plain light background, clean reference sheet layout,
anime style, masterpiece, best quality, absurdres"
The four choices are deliberate rather than decorative. Everyday clothing establishes the baseline. Formal wear tests whether the design reads as the same person when the practical details are stripped out.
Winter layers bury the silhouette almost entirely, which is the hardest case for recognition. Field gear pushes the over-prepared personality further than the base design does, since packs and a cloak are the visual language of someone who has planned for everything.
This was our result.
The same character generated across four outfits in a single pass.
Across all four outfits, the bob, the bangs, and the round amber eyes are the same; the silver streak stayed on her left side, and the single earring appears in every panel without being mentioned in the prompt.
The proportions also settled down. In the design directions test, she ranged from a child to an adult across four images. Here she reads as the same age and build throughout, which suggests a single sheet holds a character together better than four separate generations do.
Each outfit also says something different about her. The travel cloak and pack push the over-prepared trait further than the base design does, and the formal dress in cream and gold is the real surprise, since it is the only one suggesting she comes from luxury. That was never in the brief, and it's now a decision we get to make about her background rather than a detail the model got wrong.
The only main issue is that the model hid her hands in every panel, tucking them behind fabric or out of frame, so the sheet says nothing about how she holds anything.
The personality read also weakened. The first Suzu image translated anxious and over-prepared into a hunched, clutching posture on its own. Here all four stances are flat and neutral. The sheet format trades expression for consistency, which is a fair trade at this stage, but it means an outfit sheet won't tell you much about who a character is.
Building an AI Expression Sheet for Your Character
The outfit sheet gave us consistency and took away personality. An expression sheet is where you get the personality back, because it is the only stage in the workflow where the face is allowed to change.
This is also the harder consistency test. Outfits change everything below the neck while leaving the face untouched, so holding a character together is comparatively easy. An expression sheet does the opposite. It deforms the exact features that make her recognizable, then asks whether she is still the same person.
We used the herbalist image as the reference again rather than the outfit sheet, since a single clean face reads better as an anchor than one of four small panels.
This was the prompt we ran:
"character expression sheet, six head and shoulders portraits of the same character,
same face, hairstyle, and outfit in every panel,
expressions in order, neutral, happy, angry, sad, surprised, embarrassed,
plain light background, clean reference sheet layout,
anime style, masterpiece, best quality, absurdres"
And this was our result.
Six expressions of the same character in one generation, with the face, hairstyle, streak, and earring consistent across all six panels.
The bob, the bangs, the jaw, and the round amber eyes are the same in all six panels; the silver streak sits on her left every time, and the single earring appears in every panel on the correct side. The outfit collar carried through too, which keeps the sheet usable as a matched set.
Following the prompt exactly is where it slipped a little. We are missing the happy option and have two angry options. If we still wanted a happy expression example, we can generate it separately rather than asking for the whole list again.
A happy expression generated separately after the sheet returned two angry panels instead.
The three that landed really well are the three that deform the face in genuinely different directions. Anger opens the mouth and lifts the brows inward. Surprise widens the eyes and drops the jaw. Embarrassment brings a heavy blush across the nose and cheeks, with the eyes pulled down and away, and it is easily the best panel on the sheet.
Creating a Character Turnaround for Reference
Everything so far has shown Suzu from the front. A turnaround fixes that by asking for front, side, and back views of the same design.
It is the point where a character stops being a collection of images and becomes something another person could work from, since it answers what a front view leaves open:
- How the apron fastens at the back
- Whether the bob is longer at the nape than it looks
- Where the satchels sit when you can't see them from the front
It is also the strictest consistency test in the workflow. The outfit sheet changed clothing while keeping the face. The expression sheet changed the face while keeping the angle.
A turnaround changes the angle, which means the model has to invent information that was never in the reference, and inventing is where drift comes from.
Here's the prompt we used:
"character turnaround sheet, same character shown three times,
front view, side view, and back view, full body,
identical outfit and hairstyle in all three views,
plain light background, clean reference sheet layout, neutral standing pose,
anime style, masterpiece, best quality, absurdres"
We got this output from our character turnaround generator.
Front, side, and back views generated in one pass, with the outfit and proportions holding across all three and the silver streak disappearing in the back view.
The outfit reconstruction is the strongest part of this result. The back view had to invent everything from scratch, and what it produced makes sense.
The apron straps cross in an X across her back, the apron opens at the back the way a real one does, and the belt pouches, boots, and bundled herbs all carry through. The bob holds its shape from every angle too.
The one thing the back view lost was the silver streak. It reads clearly in the front and side views and vanishes entirely from behind, which is the same asymmetry problem the single earring created earlier. An off-center detail only exists from the angle that shows it, so a turnaround is where you find out whether it survives the rotation.
Meanwhile, the side view answers a question we have been carrying since the first image. Her left ear has been hidden behind hair in every generation so far, so "one earring only" was never actually confirmed. The profile shows that ear clearly, and it is bare. The asymmetry survived.
Her proportions also held across all three views, which continues the pattern from the outfit and expression sheets. That is what makes a turnaround the most reusable asset in the set. Every earlier stage produced images of Suzu, but this one describes her, which means you can come back to it months later, or hand it to someone else, and rebuild the character without guessing.
Adding Poses and Other AI Character Sheet Assets
At this point Suzu has a design direction, four outfits, six expressions, and a three-view turnaround. That is enough to work from, which means the question stops being what else you could generate and becomes what would actually help.
Plenty of things could go on the pile, including pose sheets, lighting studies, additional outfit sets, character design boards, and prop breakdowns. But most creators don't need most of them.
With respect to Suzu, the clear question is her hands. The outfit sheet hid them in every panel, tucking them behind fabric or out of frame. The turnaround kept them straight down at her sides, holding bundles of herbs.
Across the entire workflow, there isn't one image showing how she holds anything, which is a real gap for a character whose defining trait is being over-prepared. Preparation is something you do with your hands.
So we generated a pose sheet using this prompt:
"character pose sheet, four full-body poses of the same character,
same face, hairstyle, and outfit in every pose,
pose 1 crouching and sorting herbs on the ground,
pose 2 reaching up to a high shelf,
pose 3 walking while carrying a full basket,
pose 4 kneeling and writing in a notebook,
plain light background, clean reference sheet layout,
anime style, masterpiece, best quality, absurdres"
Each pose is doing a job rather than filling a slot. Crouching and kneeling test whether the model can bend her without losing the proportions that held so well through the last three stages.
Reaching up exposes the torso and the apron straps from an angle the turnaround didn't cover. Walking with a basket is the only one that shows her carrying weight, and all four put her hands in the frame doing something specific.
Four poses generated in one pass, with her hands finally visible and doing something specific in every panel.
All four poses landed, and the model followed each instruction rather than substituting easier alternatives. The crouch and the kneel hold her proportions without the collapsing that difficult poses usually cause.
The hands problem is solved. This one sheet says more about how she handles things than every earlier stage combined.
How Consistent Is the Character Across the Whole Workflow?
Now, let's put it all together. We evaluated each stage on its own as we went, but the harder question is whether the outputs work as a set, because a character design AI is only useful if someone can look at everything and see one person.
Her face never broke. Her plum bob, blunt bangs, round amber eyes, and jaw shape are the same in the first concept image and the last pose sheet, across roughly twenty generations and six stages. Since identity rests on the face more than anything else, this is the most encouraging result.
Proportions held from the outfit sheet onward. They collapsed only at the design directions stage, where she ranged from a child in the fantasy panel to an adult in the magician panel.
Every stage after that used a single-sheet format, and each one kept her the same age and build. The art style held throughout as well, because nothing in the workflow required switching models, so all of the outputs sit together as one set without any cleanup.
The failures clustered more tightly than we expected. We lost the crescent pin at the first stage that changed her clothing, since the model read it as part of the robe rather than part of her. The silver streak held everywhere it was visible and dropped out of the back view, and the cat disappeared after the concept stage because no later prompt asked for it.
In other words, the pattern is that the model holds whatever it reads as face and loses whatever it reads as attachment. It treats hair, eyes, and bone structure as the character, while a collar pin and a companion animal are all treated as scene content that exists only while the surrounding context supports it.
So can these outputs be used as design materials for the same character? Yes, definitely!
Where AI Helps and Where You Still Make the Decisions
Everything in this AI character design workflow came out of one model, but almost none of the decisions did.
Tsubaki.3 was fast at the things that are slow by hand. It generated four full design directions in the time it would take to sketch one, produced six expressions in a single pass, and reconstructed a back view of an outfit it had only ever seen from the front. When exploring is that cheap, you can look at more options before committing.
What it couldn't do was tell us which option was right. The fantasy witch looked great, but we discarded it because the herbalist was the only one that showed the personality we started with. That decision came from knowing who Suzu was supposed to be, and nothing in the model had access to that.
The same applies to what we kept. Fixing four traits before generating anything is why the consistency section had something to measure. Dropping the crescent pin once the model showed it wouldn't survive a change of clothes was a judgment call.
So the split is that AI handled the generating and comparing, while every decision about what the character actually is stayed with us. It produced the options. Choosing between them, naming what was permanent, and knowing which failures needed fixing was the part that made this a design rather than a folder of pictures.
Tips for a Better Anime Character Design Workflow
From taking Suzu through every stage, we learned a few things about what an AI character design workflow rewards and what it punishes.
A checklist of what we learned from taking Suzu through the full design workflow.
Here's an overview of our insights:
- Fix three or four traits before generating anything: This turns "does she still look like herself" from a feeling into something checkable. Keep the list short, since every extra trait is another thing to verify at every stage.
- Don't attach identity to clothing: The crescent pin disappeared the moment the robe changed, because the model read it as costume rather than character. Traits on the face survive almost anything. Traits on an outfit survive only that outfit.
- Ask for variations in one sheet rather than four separate generations: Proportions collapsed across the four design directions and held in every single-sheet stage that followed. One image forces the model to keep the character consistent within itself.
- Prompt only what changes when working from a character reference: The face, hair, streak, and earring carried through without being described once. Re-describing them fights the anchor and pulls the result toward a generic version.
- Return to the original reference each time: Chaining generations lets small drift compound. The apron simplified slightly by the pose sheet stage, which is minor on a project this size but would accumulate over a longer one.
Designing Characters With AI and Tsubaki.3
Suzu started as three lines of description. She now has a design direction chosen from four, four outfits, six expressions, a three-view turnaround, and a pose sheet, which is enough for someone else to draw her without asking a single question.
None of that came from one perfect generation. It came from treating AI character design as a sequence where each stage answers a question the last one left open, then checking what survived between them. The gap between a good image and a usable character is exactly that process.
The failures are part of it too. Knowing that the model holds a face and drops a collar pin, or that an asymmetric detail vanishes the moment a character turns, is what lets you plan around it instead of regenerating and hoping.
So pick a character you have been meaning to develop, write down the handful of traits that make them themselves, and run them through the stages. Try it with Tsubaki.3 on PixAI and see how far one idea gets.
Top comments (0)