One Reference Image, Eight Character Sheet Runs, and What the Model Made Up

Character sheets are a reference format. Front, side, and back views, a range of expressions, a set of poses, enough detail that another person could work from the sheet without asking you questions.

The interesting engineering question is what happens when the input is underspecified. One image of a character does not contain the back of the coat. It does not contain the object hanging off a strap that runs behind the hip. Ask a model to expand that image into a full sheet, and it has to produce values for fields that were never set.

So I ran it as a controlled test on Tsubaki.3, the model PixAI currently has in early access, and measured what came back.

Test setup

One original character, generated once from text, then used as the only reference for every run after it. Eight generations: a turnaround, the same turnaround again on a different seed, a third with one sentence added, a six-panel expression sheet, a nine-panel expression sheet, a four-pose sheet, a single pose, and an outfit sheet.

Every setting held constant across runs: style preset off, Pro mode, Prompt Helper off, default negative prompt untouched, square format, one reference image attached.

The character design is the instrument, and it splits into three groups.

NAMED IN EVERY PROMPT      hair length and color, eye color,
                           red scarf, olive coat over grey sweater

REFERENCE ONLY             which side the hair tucks behind the ear,
                           which side the scarf tail hangs on,
                           strap across the chest with a metal slider,
                           notebook in the left chest pocket

ABSENT FROM THE REFERENCE  the back of the coat
                           the object on the end of the strap

The first group is the control. If those drift, nothing else is worth reading. They did not drift in any of the eight runs.

Result 1: silence is deterministic, promises are not

The turnaround came back with three genuine views at a matched scale: 983, 978, and 978 pixels tall in a 1024 pixel frame. The back view had to be invented, since the reference never showed it. It came out as a plain quilted panel with a center seam.

Then I ran the identical prompt, byte for byte, on a different seed. Same invented back. Same center seam.

That is the useful finding. An unset field gets a default, and the default is stable enough to rely on.

The strap behaves differently. The reference shows it crossing the chest and running to something behind the hip. Both turnaround runs dropped it entirely from the back view rather than resolving it, on two different seeds.

For the third run, I held the first run's seed and appended one sentence naming the bag. The bag appeared, correct material, correct hip. Two things came with it: a second strap crossing the first on the back, which no single shoulder bag produces, and the disappearance of the strap from the front and side views of that same sheet.

One added constraint, satisfied at the point of the constraint, paid for elsewhere in the same output.

Result 2: panel count is a real budget

Six expressions in a two-by-three grid was the strongest output in the set. All six visually distinct, the face measuring between 243 and 260 pixels wide across the panels, and a metal slider a few pixels across surviving at roughly 341 by 512 pixels per portrait.

Nine expressions in a three by three grid did something I did not predict. The faces stayed on model. The framing simplified: every panel squared up to the camera, the hair fell symmetrically, and a reference-only detail that held through all six panels of the smaller sheet vanished from all nine of the larger one. Two emotion pairs also converged, so nine slots produced seven clearly separate faces.

Result 3: the sheet format itself costs resolution

The pose sheet held proportions and broke layout. Four requested poses produced three distinct ones, and the crouching figure landed in the gap between two others at a different camera height instead of on the row.

More useful was running the same crouch as a single figure in the full frame. The notebook in the chest pocket renders at 71 pixels wide there against 32 on the sheet, with the cover, page block and spine thickness all readable. The reaching hand resolved into five separated fingers.

A quarter of a 1024 pixel frame has no room for that. The sheet was not failing at drawing; it was failing at budget.

What survived, ranked

Across every figure, the reference-only details came out in a clear order.

notebook in the chest pocket   25 of 26 figures wearing the coat
strap across the chest         23 of 26 front and side views
scarf tail on one side         10 of 17 full body figures
hair tucked behind one ear     14 of 26 views that face the viewer

The ordering is not random. The notebook is a high-contrast object in a fixed place on the garment, and every format keeps that place in frame. The strap depends on which side of the body is visible. The hair parting has no object holding it, so it goes first.

One caveat that cuts across all of it. The notebook and the strap are missing from the outfit sheet completely, because the coat they live on was replaced. Identity carried by an object lasts exactly as long as the object does.

Takeaways if you try this

Design for the format. Put the identifying details on objects, keep those objects on a garment you plan to keep, and treat anything that lives purely in the silhouette as the first thing you will lose.

Run the turnaround twice before you trust the back. The second run tells you whether the invented answer is a stable default or a coin flip.

If a detail matters, generate it as a single image rather than a panel. The overview and the detail are different jobs.

One last thing. Treat the output as a working reference rather than a production model sheet. Everything on it that your original image never showed is a plausible guess, which is a different thing from correct.

If you have one good picture of your OC sitting in a folder, the whole experiment costs you a few generations. Try it on PixAI and see which of your character's details are carried by an object and which are carried by luck.