One Reference Image, Nine Sheets: Where an AI Character Sheet Generator Holds and Where It Guesses

You have one good picture of your character. One picture does not tell you much, though.

You cannot see the back. You cannot see the side. And you do not know what the face does when the character laughs or shouts. A character sheet is what shows you all of that.

So here is the question I wanted answered. Can an AI character sheet generator take that one picture and build the rest, the turnaround, the expressions, the poses, and keep it looking like the same person?

I ran nine tests on Tsubaki.3, split into two experiments. The first five use one swordsman I designed, run through the standard sheets. The last four use a single image I grabbed from PixAI's public gallery, to see what one found picture can turn into.

The real question underneath all nine is the same: can it tell apart what the reference shows from what it has to invent, and keep the invented parts steady.

The scorecard first

Before the walk-through, here is how all nine landed. Two of them needed a second run, and I have listed both.

Test	What it demanded	Score
Base reference	One clean front-facing design	Pass
Turnaround	Front, 3/4, side, back from a front-only image	6, then 8.5 on re-run
Expression sheet	Nine emotions, identity held	6, then 7.5 on re-run
Pose set	Five full-body poses	7.5
Expressions in context	Nine emotions in nine real scenes	8
Action pose sheet	Six combat poses with motion	8
Storyboard	Six manga panels from one frame	7
Three roles	One face across three worlds	8
Color script	Six lighting states of one room	7.5
Key-visual poster	A finished, title-ready poster	8.5

The pattern in that column is the whole story: solid outputs, held back by panel counts and small identity marks. More on that after the tests.

What counts as a real sheet

A character sheet is not just several pictures of the same person on one page. Each part should tell you something useful about the design.

A useful anime character sheet usually covers a front, side, and back turnaround, an expression sheet, some poses, and sometimes outfit variations. You do not need every one every time.

The point is that each panel should give you real information: the shape of the coat from behind, how the face moves, how the body holds a pose, not just another nice illustration. That is what turns a set of images into a real AI character reference sheet.

So the bar I judged against was not "does it look good." It was "would this help me draw or reuse the character."

Experiment one: standard sheets from a character I designed

I built one original swordsman as the base and used him for the first five tests.

He is a weathered older man with tied grey-streaked hair, a black eyepatch over his right eye, a long cross-style scar on that cheek, a fur-lined coat, a sword sheathed on his back, a carved wolf-head pauldron on his left shoulder, and prayer beads on his wrist. Those details are the anchors I tracked, specific enough that any drift shows up fast.

One thing matters before we start. The reference is front-facing, so the back of the coat, the far side, and the full sword are things the single image never shows. The model has to invent them. That is the real test, telling apart what the model kept from what it made up.

The reference itself came out strong. Every anchor landed in the right place, and the outfit worked as a coherent design. A clear pass.

Turnaround

Using this exact character, create a clean turnaround model sheet on a plain
background, four views at the same height: front, three-quarter, side profile,
and back. Keep his tied grey-streaked hair, eyepatch and scar, fur-lined coat,
wolf-head pauldron, back-sheathed sword, and wrist beads identical across every
view. Show the back of the coat and the full sheathed sword clearly in the rear
view. Neutral A-pose, consistent model-sheet line work.

The first run showed real problems. The front view stayed close to the reference, but the three-quarter view drifted almost fully into a side profile, so I basically lost that angle.

Worse, the model gave me two back views instead of a proper set of four, and the two backs did not agree. The wolf pauldron was missing in one and present in the other. Since that pauldron is a main identity piece, that is a real miss. First run, a 6.

So I ran it again, and the second run was much better. This time I got all four angles, a proper three-quarter view, and the wolf pauldron stayed consistent across every view, including the back. The hair, eyepatch, coat, sword, and beads all held together. The sword shifted shape slightly between views, but that is minor next to the first attempt. Second run, an 8.5.

Left: the first run. Right: the second run.

The honest takeaway on the AI character turnaround: it can build a believable back from a front-only image, but the consistency is not guaranteed on the first try. A re-run got me there.

Expression sheet

Using this exact character, an expression sheet of nine head-and-shoulders
portraits, changing only the expression: hard glare, faint smile, open laugh,
grief, cold anger, weariness, surprise, suspicion, and calm. Keep his tied
grey-streaked hair, eyepatch over the right eye, cheek scar, and one visible
eye identical in every panel. Plain background, consistent line work.

The expressions were great, but the identity slipped. All nine emotions came through clearly and the face structure held, so as an AI expression sheet the range is strong.

The problem was the anchors. In the first run, two panels lost the eyepatch entirely and showed both eyes, and the scar came and went in those same spots. For a character whose whole identity leans on that eyepatch, that is a real failure. First run, a 6.

The second run improved. The eyepatch and scar held in seven of the nine panels, so I will call it a 7.5. Better, but one identity-losing panel out of nine still is not a clean pass. This was a heavy ask, nine faces in one go, and that is likely why it stumbled.

First expression run left and the second run right

Pose set

Using this exact character, a pose reference sheet of five full-body poses on a
plain background: standing at rest, drawing the sword from his back, a low ready
stance, kneeling with the sword planted, and walking with the coat trailing.
Keep his proportions, tied hair, eyepatch, fur-lined coat, wolf-head pauldron,
and beads consistent across all five. Clean model-sheet style, consistent scale.

Four of the five poses were solid, and the character held well across them. Standing, drawing the sword, the low ready stance, and the kneeling pose all came through clearly, and the eyepatch, hair, coat, pauldron, and proportions stayed steady. A good result for an AI pose sheet.

The fifth pose missed. It looks like another standing pose rather than walking, and the coat shows none of the trailing movement I asked for. The other weak spot is the face, which goes soft and muddy at this full-body size compared to the closer shots. A 7.5.

The five-pose set.

Expressions in context

For this one I pushed past the plain grid and asked for nine emotions inside nine real scenes.

Using this exact character, a nine-panel grid of cinematic close-up moments,
each showing a different emotion in a different setting, same man throughout:
laughing by a campfire, glaring in the rain, grieving at a grave, calm in a
snowfall, angry in a tavern brawl, weary on a mountain road, surprised in
torchlight, suspicious in a market crowd, at peace under cherry blossoms. Keep
his tied grey-streaked hair, eyepatch, and cheek scar identical in every panel.
Cinematic modern anime illustration, varied lighting.

This was harder, and the character held better than in the plain expression sheet. The eyepatch, hair, beard, and face stayed consistent across all the different lighting and settings, and this time the patch did not vanish. He also fit each scene naturally instead of looking pasted onto a background. The campfire, rain, graveyard, tavern, and cherry blossoms all came out well.

The one real failure is the count. I asked for nine panels and got eight. The "surprised in torchlight" moment dropped out entirely. An 8, strong consistency under changing light, held back by the missing panel.

The in-context expression sheet

Action pose sheet

Using this exact character, a dynamic action pose sheet of six full-body combat
poses on a soft neutral background: mid-swing slash, blocking overhead, spinning
parry, lunging thrust, sheathing the sword, and landing from a jump. Coat and
hair moving with each motion. Keep his proportions, eyepatch, scar, fur-lined
coat, wolf-head pauldron, and back sword consistent across all six. Clean,
energetic model-sheet style.

The motion here is a clear step up from the static poses. The model understood this needed real combat energy, and the coat, hair, and sword all move convincingly while the character stays recognizable. The eyepatch, pauldron, coat, and beads held across the set.

The miss is precision. The six specific actions are not all clear. Several poses look like slashing or parrying variations, and the sheathing pose is missing or hard to spot. The face is still a little soft in the smaller renders, though cleaner than the pose set. An 8.

Base image (left). The action pose sheet (Right).

Experiment two: what a stranger's single image can become

The standard sheets answer the brief. Then I wanted to see what else one image can become, so I switched references entirely.

I grabbed a single moody close-up from PixAI's public gallery, a blue-haired student resting his head on a classroom desk. It is a found community image, not my own design, and I used it purely as a test input.

It is a tight, angled shot, so almost everything below his shoulders is unknown. That makes these four tests even harder, and more like real life, since most people start from one good picture, not a clean model sheet.

The anchors I tracked: dark blue layered hair, grey eyes, a cross-shaped cheek scar, and the row of ear piercings.

The found reference image

A storyboard from one frame

Using this exact character as the starting frame, create a six-panel
black-and-white manga storyboard of the scene this moment belongs to. Panel 1,
a wide shot of the quiet classroom. Panel 2, him walking in alone. Panel 3, this
exact moment, him resting his head on the desk, tired. Panel 4, another student
speaks to him from the doorway. Panel 5, he lifts his head and looks over. Panel
6, he stands to leave, bag over his shoulder. Keep his dark blue layered hair,
grey eyes, cross cheek scar, and ear piercings consistent in every panel.
Cinematic manga composition.

The visuals are good, but it missed two story beats. I asked for six panels and got five. The model merged the wide classroom shot and him walking in into one panel, and because of that, the student speaking from the doorway never showed up.

The rest of the sequence works. The desk moment matches the reference, the head-lift is clear, and the final shot comes across as him leaving. His hair, scar, and piercings stay recognizable, and the black-and-white manga treatment fits well. A 7. The character holds up better than the storyboard instructions do.

The five-panel storyboard

The same face, three lives

Using this exact character, keep his face, dark blue layered hair, grey eyes,
cross cheek scar, and ear piercings exactly the same, but reimagine him in three
completely different roles side by side: a cyberpunk street mercenary in a neon
alley, a medieval knight in worn armor, and a modern rockstar on stage. Same
face and identity in all three, only the world, outfit, and role change.
Detailed modern anime illustration.

This is where identity preservation really showed. The model turned him into three completely different characters, and all three still look like the same guy. The blue hair, grey eyes, face structure, and ear piercings carried across every version, and the three roles feel distinct instead of the same outfit recolored.

The cross scar is the weak point. It is there in all three, but its exact shape and placement drift a little, and the knight's face is slightly different in build. An 8. Keeping one face across three worlds is a strong result.

Test 7 - Base image (left). The three-roles image (right).

A color script

Using the mood and color palette of this exact image, create a horizontal film
color-script strip of six small thumbnail frames of the same classroom scene at
different times: early dawn, bright noon, golden afternoon, blue dusk, night with
lights off, and stormy grey. Each thumbnail keeps the boy resting at the desk,
small in frame, but the light and color of the room change completely across the
strip. Cinematic color-script layout.

The lighting nailed it, the layout did not. The classroom stays recognizable while the light shifts hard across the frames, and the blue dawn, warm gold, purple dusk, deep night, and stormy grey all come through clearly. The mood control is excellent.

Two things missed. I asked for a horizontal filmstrip and got a 2x3 grid, and the boy is far too large. He was supposed to stay small so the room and light were the focus. A 7.5. Great color, wrong format.

The color-script grid. I’ll go with the right one for this review.

A key-visual poster

Using this exact character, design a cinematic anime key-visual poster with him
as the centerpiece. Keep his face, dark blue layered hair, grey eyes, cross cheek
scar, and ear piercings exact. Build a moody, atmospheric composition around him
with dramatic lighting, a color palette matching the original, space at the top
for a title, and a lonely, introspective tone.

This one came out the strongest for mood and finish. The model understood the lonely, introspective direction and built a dark classroom, cool blue palette, and window light around him. His face, hair, eyes, and cross scar are all present and closer to the reference than in some earlier tests, the scar especially.

The miss is the poster part. I asked for empty space at the top for a title, but his hair runs high into the frame and leaves almost none. It comes across more as a cinematic illustration than a designed poster. An 8.5.

The key-visual poster

The pattern across both experiments

Put the two experiments together and the same three-part pattern shows up every time.

Big features hold, tiny ones drift. His hair, face shape, eyes, and general build stayed steady almost everywhere. What slipped were the small identity marks, the swordsman's eyepatch and the student's cross scar. Those dropped out or changed shape most often, especially in the busy nine-panel sheets and under strong expressions.

Scale hurts the face. In the full-body poses, the face goes soft and muddy. It is still recognizable, but it loses the sharpness you get in the close-up shots.

Exact counts and layouts are shaky. More than once, I asked for a set number of panels or a specific layout and the model quietly dropped one or reorganized it: the missing ninth expression, the five-panel storyboard, the grid instead of a filmstrip.

And the invented parts, the back of the coat, the full body from a close-up, are believable but not guaranteed to stay consistent, which is exactly why the turnaround needed a second run. The first turnaround gave me two different backs in one sheet, which tells you the model is guessing, not remembering.

The verdict: a strong starting point, not a finished reference

For developing an OC, planning a comic, working out a mood, or sharing a design with someone, these sheets give you plenty to work from. The in-context expressions, the action poses, the three-roles test, and the poster were all useful outputs. As a quick OC character sheet or a reference to draw over, it does the job.

What it is not yet is a clean production model sheet you can hand to an animator without checking. The turnaround needed a re-run, panels go missing, and the small anchors are not reliable enough to trust blindly.

A re-run usually helps, and naming your key features in every prompt helps more. But the more your character leans on small, specific marks, the more you will want to check each panel. If you want to go deeper on the reference side, PixAI's Reference Pro guide covers that workflow.

So, can Tsubaki.3 build a character sheet from one reference image? Mostly, and better than I expected, as long as you go in knowing where it slips. It is strongest at expressions in context, poses, and the creative expansions. As a character turnaround generator it is not fully reliable yet, and it is weakest at precise panel counts and at holding tiny identity marks across a big sheet.

If you have one good picture of your OC in a folder, this is a fast way to turn it into something you can build on. Try it on PixAI with your own character, start with the expression sheet or the poses, and you will know within a couple of generations how well it holds your design.