Most people think AI character design means getting one good picture. It does not. The hard part is building a character once, then getting her back again in a new outfit, a new expression, and a new scene, still looking like the same person.
That is what I set out to test. I took one original character through ten generations on Tsubaki.3, PixAI's newest model, starting with rough concept sheets and ending with full cinematic scenes. Rather than march through them one to ten, I have grouped them into the four stages the work really moves through: building the design, developing it, pushing it into scenes, and combining it with a second character. Every prompt and score is below.
How I approached it
Good character design AI work follows a path, not a single prompt. You start wide with concept exploration, lock the direction you like into one clean reference, then use that reference to build everything else. That path is the core of AI character design, not one lucky prompt.
Tsubaki.3 handles natural language, so I wrote full descriptive prompts instead of tag lists, and I repeated every defining feature in every prompt. The model drops details you stop mentioning, so the two-tone braid, the gold eyes, the red rune, the brass arm, and the potion vials appear in almost every prompt below. If you want the structure I leaned on, the PixAI prompt formula guide covers it.
A note on setup that applies throughout: I ran every test on Tsubaki.3 in Ultra mode, generating a batch of four each time. In the image pairs, the left image is the one I reviewed, so my notes always describe that one. The right image is a second pick from the same batch, there for extra visual appeal. Test 7 is the one exception, and I flag it there.
Stage one: building the design
Concept exploration
The first job of any AI character generator is to give you options. I asked for five takes on one idea so I could pick a direction.
A character concept exploration sheet, five distinct full-body design
directions for a young female monster hunter, each visually different: a
rugged wilderness ranger in furs with bone charms, an elegant aristocratic
slayer in a tailored dark coat, a punk street hunter with piercings and a
spiked jacket, an alchemist hunter with a brass mechanical arm and a
bandolier of glowing potion vials, and a shrine-maiden exorcist with a staff
and long paper talismans. Varied silhouettes and color palettes, clean
concept art, neutral poses, plain background, labeled panels.
This was a strong start. Tsubaki.3 gave me all five directions as separate characters instead of blending them into one, and each had its own silhouette and personality.
The alchemist came out the most defined, with the brass arm and potion bandolier showing clearly. That is the direction I carried forward.
The one real weakness was the labels. The model wrote text-like labels under each figure, but they are not real words. That is a common limit with generated images, and it does not hurt the actual AI character concept art.
Left the base run. Right the Clear Style LoRA run.
I ran it twice more. With the Clear Style LoRA, the rendering got cleaner and sharper, but the five designs drifted toward the same face and body, and potion-tube details leaked onto characters that were not the alchemist. A third run with the Glassy Anime style preset came out worst. The plain base run stayed the strongest and most varied, so that is the one I kept.
Locking the design
Once I had the direction, I generated one clean full-body reference. This image becomes the anchor for every test after it.
A full-body character design of a young monster-hunting alchemist, modern
anime style. Athletic, agile build, standing confidently with a faint knowing
smile. Her hair is two-toned, snow-white on top blending into black
underneath, worn in a messy side braid over her right shoulder. Sharp
molten-gold eyes. A small red alchemical rune tattooed just under her left
eye. Her left forearm is a polished brass mechanical prosthetic with visible
gears at the elbow and wrist. A worn leather bandolier crosses her chest,
lined with glowing green and amber potion vials. She wears fitted dark travel
leathers with buckled straps. Neutral standing pose, plain soft-grey studio
background, clean detailed anime character design, crisp linework, soft even
lighting.
Most of the important features landed. The two-tone hair, side braid, gold eyes, brass mechanical arm, and glowing vials all came through, and the arm in particular looks convincing, with visible gears. The rune is the weak spot.
It rests under the eye as asked, but it came out small and a little messy, so it lacks the clarity I would want from a signature feature. The outfit also turned out more revealing than the prompt described.
Even so, this is a solid AI character sheet to build on, and it scored a 9 out of 10 for me. The left image is the reference I used from here forward.
Test 2, Left the reviewed reference. Right an alternate from the batch.
Stage two: developing the character
With the reference set, the next three tests all pull from it: new outfits, new expressions, and a full turnaround.
Outfit variations
I tested it as an AI outfit generator, using the locked reference as a Character Reference.
Using this exact character, show her in four different full-body outfits lined
up side by side, the same young alchemist in each. Keep every defining feature
identical: the snow-white-to-black two-tone messy side braid, the molten-gold
eyes, the small red rune tattoo under her left eye, the brass mechanical left
forearm, and the leather bandolier of glowing potion vials across her chest.
Outfit one, rugged field-hunting leathers under a hooded travel cloak. Outfit
two, an ornate alchemists' guild ceremonial robe with gold embroidery. Outfit
three, relaxed town clothes, an oversized knit sweater and a canvas satchel.
Outfit four, heavy snow-expedition gear with a fur-lined hood and frost
goggles. Same height and build in each, plain background, consistent modern
anime design, clean linework.
Character consistency was the standout. The hair, braid, gold eyes, rune, and face stayed steady across all four looks, and the rune came out cleaner here than when I locked the design. All four outfits are distinct, and the guild robe with its gold embroidery is the best of them.
Test 3, Left the reviewed outfits. Right an alternate from the batch.
The downside is that some accessories get hidden. In the snow gear especially, the mechanical arm and bandolier disappear under the clothing.
The model keeps her identity, but it does not always keep every signature piece visible. A couple of the outer figures also got cropped at the edges.
Expression sheet
Next I tested it as an AI expression sheet generator, six emotions on the same face.
Using this exact character, an expression sheet of six head-and-shoulders
portraits of the same young alchemist, changing only her expression. Keep her
two-tone white-and-black messy braid, molten-gold eyes, the small red rune
under her left eye, and her exact face structure identical in every panel. The
six expressions: a confident smirk, intense focus, worn-out exhaustion, a
bright delighted laugh, cold fury, and wide startled surprise. Plain
background, consistent modern anime style, clean linework.
This was the most consistent test of the whole set. The character stays clearly herself across all six panels, and the expressions are properly different from each other, not small tweaks on one face. The surprised look works especially well, and the rune holds its shape and position throughout.
Two small misses: the framing drops into the upper torso instead of staying head-and-shoulders, and the mechanical arm is not in frame, which makes sense given the crop. Neither hurts the result. I scored this one a 9.4.
Test 4, Left the reviewed expression sheet. Right an alternate from the batch.
Turnaround
Then the technical one, using Tsubaki.3 as a character turnaround generator: front, three-quarter, side, and back.
Using this exact character, a full character turnaround model sheet showing
four views at the same height: front, three-quarter, side profile, and back.
Keep every detail identical across all four, the white-to-black two-tone messy
side braid, molten-gold eyes, the red rune tattoo under her left eye, the brass
mechanical left forearm, the leather bandolier of glowing vials, and her fitted
dark travel leathers. Neutral A-pose, plain grey background, clean model-sheet
linework, even lighting.
This is where Tsubaki.3 is weaker. The four views are all there and the character stays recognizable, and the mechanical arm holds up impressively across every angle. But the side and back views drift.
The braid changes shape and position, and a few clothing and accessory details shift from view to view. The pose also looks more like a natural lineup than a strict A-pose. The straight takeaway: Tsubaki.3 is much better at holding identity and expression than at producing an exact technical turnaround.
The front view is excellent, and the side and back are where the small details wander. I scored it 8.5, still usable, just not precise.
Test 5, Left the reviewed expression sheet. Right an alternate from the batch.
Stage three: pushing into scenes
Sheets are the setup. The real question is whether the character survives once she leaves the plain background.
Key art
Using this exact character, dynamic full-body key art of the young alchemist
mid-battle, hurling a glowing potion that erupts into a burst of green
alchemical fire. Her two-tone braid and coat whip through the motion, her
brass mechanical left arm extended, the vials on her bandolier glowing bright,
the red rune under her gold eye catching the light. Behind her, the massive
shadowed silhouette of a monster looms in billowing smoke. Dramatic rim
lighting, cinematic modern anime illustration, rich detail, full body.
This was probably the best image of the whole run. The character moved from a design sheet into cinematic key art without losing her identity. The mechanical arm, hair, gold eyes, rune, and glowing vials all carried over, the green explosion has real impact, and the monster silhouette gives the scene scale without stealing focus. The one miss is the coat.
The prompt asked for it to whip through the motion, but the character wears it tied around her waist instead. Very dynamic poses like this make some clothing details harder to hold. Even so, her identity survived a big change in composition, which is the point. I scored it 9.2.
Test 6, Left the reviewed key art. Right an alternate from the batch.
A cinematic scene
After the sheets, I wanted to see how far a detailed, story-driven prompt could push it.
Using this exact character, a cinematic modern anime scene on the rooftop
platform of a speeding elevated train at sunset. She wears a stylish modern
black techwear jacket, fitted top, high-waisted cargo pants, and sleek combat
boots. A massive horned monster is leaping between train cars behind her while
she grabs a metal railing with her brass mechanical left hand, twisting around
to look back with a confident grin. Her long two-tone braid whips violently in
the wind, potion vials swinging from her bandolier. A second train races
alongside them, its windows glowing orange, while passengers inside stare in
shock. Below, a sprawling futuristic city rushes past: glowing billboards,
elevated roads, colorful storefronts, distant skyscrapers, and hundreds of
tiny lights. Sunset clouds fill the sky behind the monster, with sparks, loose
papers, and broken glass flying through the air. Vibrant modern anime film
frame, dramatic perspective, extreme sense of speed, rich environment, detailed
modern fashion, cinematic lighting, dynamic composition, vivid colors, polished
anime keyframe, highly detailed.
This test needed two attempts, and that is the useful part. The first generation ignored the story completely and gave me a neon rainy street scene, no train, no monster, no passengers. So I refreshed and ran it again. The second attempt came together.
The elevated train, the speeding city below, the passengers, the sunset, and the monster behind her all showed up, and the whipping braid and flying papers sell the speed. Her whole identity carried over, and the modern techwear outfit looked more convincing this time.
A few things still drifted: she is not grabbing the railing exactly as described, and the second train looks like one track rather than two. The lesson is that a specific, dramatic scene prompt can push Tsubaki.3 into genuine film territory, but you should expect to re-roll it. The revised version scored 9.2.
Test 7, Left the reviewed image. Right an alternate from the batch.
A candid moment
Every test so far was a posed shot. I wanted to see if it could do the opposite, a quiet, unposed moment.
Using this exact character, a candid slice-of-life moment, not posed. The
young alchemist sprawled sideways in a worn armchair on a lazy afternoon, one
leg hooked over the armrest, half-asleep with a book about to slip from her
real hand while her brass mechanical hand dangles a half-eaten pastry.
Sunlight pools across her through a window, dust drifting in the beam. Her
two-tone braid is loose and messy, molten-gold eyes barely open, the red rune
under her eye catching the warm light. Cozy cluttered room softly out of focus
behind her. Natural, unguarded, warm, cinematic modern anime illustration,
off-center composition.
The pose is the win here. She looks comfortable and half-asleep, sunk into the armchair with the book in her hand, and the warm light, dust, books, and worn furniture make the room feel lived-in. Her identity holds up cleanly, and the rune is one of the cleanest versions across all ten tests.
The miss is a small object detail. The pastry rests below the mechanical hand instead of being held by it, so the model included both things but not the exact interaction. The framing is also a little more centered than the prompt asked.
The result feels like a real quiet moment from her life, which is what I was after. I scored it 9.1.
Test 8, Left the reviewed image. Right an alternate from the batch.
One thing I have to tell you: the small red alchemical rune never stayed fully consistent throughout this guide. It is always there, but you will notice it drifts slightly from the base image each time.
Stage four: two characters in one scene
The last two tests do something the design sheets cannot: combine two separate character references into a single scene.
A two-character rescue
I fed it the alchemist and a second character I uploaded.
Using the exact alchemist from @image1 and the exact girl from @image2. A
dramatic but beautiful anime film scene on an old moss-covered stone bridge
surrounded by lush green mountains and wildflowers. The girl from @image2 has
slipped over the edge of the bridge and is falling, reaching upward in
surprise. The alchemist grabs her wrist with her brass mechanical left hand and
braces herself against the stone railing, desperately pulling her back to
safety. Her two-tone braid and potion vials swing with the motion. A small
fluffy cat sits safely on the bridge watching the rescue, while colorful birds
burst from nearby trees. Ferns, vines, flowers, moss-covered stones,
butterflies, and tiny glowing particles fill the foreground. A clear turquoise
river winds through the valley far below, with waterfalls cascading down the
cliffs. In the background, a young boy stands on the bridge, startled and
running toward them. Bright morning sunlight filters through the trees. Strong
sense of height, movement, danger, and friendship, warm beautiful atmosphere.
Vibrant modern anime film scene, cinematic composition, expressive faces,
dynamic pose, rich green environment, natural character interaction, dramatic
perspective, atmospheric depth, highly detailed anime illustration.
The story is clear at a glance. The girl is falling, the alchemist is reaching to save her, and the cat and boy react from the bridge. You understand the moment without the prompt. Both characters stayed recognizable too, the alchemist with her hair, gold eyes, braid, rune, and brass arm, and the second character with her pink hair, blue eyes, white dress, and bow.
The environment does a lot of the work, with the mossy bridge, the valley, the waterfalls, and the butterflies giving you plenty to look at.
The miss is the contact point. The alchemist reaches toward the falling girl, but whether she is gripping the wrist at all is ambiguous in the left image, and the boy looks more like a background extra than part of the moment.
The right image handles it better. Still, combining two references into one narrative scene is a real step up from a character sheet. I scored it 9.3.
Test 9, Left the reviewed scene. Right an alternate from the batch.
Modern fashion transformation
For the last test, I kept both characters but changed everything else, swapping their fantasy roles for modern fashion.
Using @image1 and @image2 as character references. Keep both girls recognizable
with their original facial features, hair colors, eye colors, and identities,
but completely replace their fantasy outfits and roles with modern fashion. A
cinematic close-up of both girls relaxing together inside a stylish contemporary
creative lounge, dressed like modern Japanese fashion models. @image1 wears an
oversized charcoal blazer over a fitted cream top, a short pleated skirt, sheer
black tights, silver earrings, and sleek loafers. @image2 wears a soft blue
cropped cardigan, a white pleated skirt, delicate jewelry, and fashionable
sneakers. They sit together on a curved designer sofa, laughing naturally over
colorful iced drinks, leaning slightly toward each other as they look at
something on a phone. Around them are lush indoor plants, chrome furniture, art
books, vinyl records, a small neon sign, abstract artwork, hanging lights, and
large glass windows overlooking a distant sunset. Warm sunset light mixes with
soft colorful interior lighting, natural expressions, fashionable layered
clothing, detailed fabric textures, intimate friendship moment, close-up
cinematic framing, vibrant modern anime film aesthetic, polished anime
illustration.
Identity preservation carried the test. The alchemist kept her white-to-black hair, gold eyes, rune, and face, and the second character kept her pink hair, blue eyes, and bow, while both moved convincingly into modern outfits.
The lounge feels contemporary and social, and the phone gives them a natural reason to sit together rather than just pose. Two things to note. The close-up became more of a medium shot showing most of their bodies, and the alchemist's mechanical arm is gone entirely.
That last one makes sense, since I asked to change her role and outfit, but it means her identity now rests on her face and hair rather than her strongest design feature. I scored it 9.2.
Test 10, Left the reviewed image. Right an alternate from the batch.
What held, and what drifted
A few patterns ran through all ten generations. Consistency is the top strength. Tsubaki.3 held the character's face, hair, and signature details across outfits, expressions, and full cinematic scenes. It also combines two references well, keeping both characters recognizable in a single scene, which is hard for most tools.
The weak points are just as clear. Turnarounds are the soft spot, since side and back views drift, so this is better for identity and expression than for an exact multi-view model sheet.
It follows the big picture, not the fine print, so object interactions, exact framing, and specific poses often come out approximate, like the pastry, the coat, and the wrist grab. And tiny details drift the most. The small red rune under her eye was the least consistent feature of the whole run.
It is in almost every image, but its shape and clarity shift a little each time, so a small marking like that is the first thing to wander from the base design.
Two habits handle most of this. Name every feature, every time, since the model drops details you stop repeating. And when a scene prompt is complex, expect a re-roll, because a specific, dramatic prompt gives cinematic results but does not always land on the first try.
How to keep a character consistent
If you plan to reuse a character, the reference workflow above will carry you a long way. Set your locked design as a Character Reference and build from it, exactly as I did from the second test onward. PixAI's Reference Pro guide covers that in depth. For a character you will use constantly, the stronger move is to train a character LoRA on PixAI. That turns your design into a model you can call any time instead of referencing an image on every generation.
Your next step
Anime character design with AI is a process. You explore directions, lock one, then develop it through outfits, expressions, and scenes, checking consistency at each step.
Tsubaki.3 handles most of that well, and it is strongest exactly where it counts, keeping your character recognizable as you move her around.
If you want to build your own character this way, try it on PixAI. Start with a concept sheet, lock the one you like, and reference it forward. You will have a usable character in a handful of generations.
Top comments (0)