AI Character Expression Sheets: One Sheet or One Expression at a Time?
Tested with Tsubaki.3 on PixAI · Test date: 30th October, 2026
Disclosure: this article is part of a paid collaboration with PixAI. It also contains my personal referral link, and I receive credits if someone qualifies through it.
Character expression sheet: the quick answer
A character expression sheet is the page that sits next to a character design and shows the same face reacting to things: happy, angry, sad, surprised. If you write a visual novel or draw a webcomic about an OC, it stops her from quietly turning into somebody else by chapter two. Drawing six different faces is the easy part. Keeping them recognizably the same person is the hard one.
I wanted to know which route in PixAI costs less effort. Route A asks for the whole AI expression sheet as one image. Route B edits one expression at a time, going back to the original reference each time. Both used the same original character and the same six emotions: neutral, happy, angry, sad, surprised and shy.
Short answer: make the sheet when you want to judge a character’s emotional range, and make separate images when you need files you will actually use. The sheet gave me all six emotions in one run and she stayed recognizable. But two panels read weakly, each cell holds about a sixth of the pixels of a full image, and my one correction regenerated the whole page and shifted panels I never asked it to touch. Route B took more runs, yet every result was a full-size file, and fixing the shy expression left the other five exactly as they were. In credits, the sheet route came to 9,200 and Route B to 29,700, not counting the 3,200-credit reference both routes share.
Figure 1. The master reference, the complete sheet, and the six separate Route B images. Layout is labeled; the images inside are unedited.
Start with a reference worth editing
Both routes depend on the starting image. I wanted a front-facing portrait with the eyebrows, eyes and mouth clearly visible, one distinctive accessory, and a plain outfit and background so nothing competes with the face. I generated it from text with this prompt:
Reference prompt
Create an original anime-style portrait of a young woman with shoulder-length dark auburn hair, warm amber eyes, a small beauty mark beneath her left eye, and a distinctive silver crescent-moon hair clip. She wears a simple dark green outfit. Give her a neutral, relaxed expression with naturally positioned eyebrows and a gently closed mouth. Front-facing, head-and-shoulders composition, clearly visible facial features, clean soft anime illustration style, plain light beige background.
I asked for dark auburn and got something closer to deep burgundy. I kept it. Every later edit has to match the image, not the prompt that made it.
Before testing, I wrote down what must not move and what is allowed to move.
Stays fixed: face shape, shoulder-length hair, amber eyes, the beauty mark under her left eye, the silver crescent clip, the dark green top, the beige background and the head-and-shoulders framing.
Free to change: eyebrow angle, how open the eyes are, the mouth, blush, and the general tension in the face. Several problems I hit later, like a changed crop and neighboring panels drifting, were just items from the first list moving when they should not have.
Model: Tsubaki.3
Make your first expression variation in PixAI
Before the full comparison, here is the smallest useful version of this whole article: one reference, one new expression. If you have an OC, you can try this with your own character in a few minutes.
- Open the PixAI image generator and choose Tsubaki.3.
- Add your neutral reference image so that it becomes @image1 in the prompt. In the prompt box, click the image icon, pick your reference, and refer to it as @image1 in the text.
- Paste the prompt below.
- Swap the emotion phrase for your own, and the list of preserved features for your character’s.
- Run the generation, then put the result next to your reference.
Screenshot: the PixAI generator with Tsubaki.3 selected, the neutral reference attached as @image1, the happy prompt entered, and the 4,100-credit cost shown beside the generate button.
Tested prompt (happy)
Use @image1 to create an anime-style portrait of the exact same character. Give her a genuinely happy expression with slightly raised eyebrows, bright eyes, and a warm, open smile. Preserve her face shape, shoulder-length dark auburn hair, amber eyes, beauty mark beneath her left eye, silver crescent-moon hair clip, and dark green outfit.
Maintain the original illustration style, head-and-shoulders framing, and plain light beige background. Change only her facial expression while keeping her identity recognizable.
To make it yours, keep the structure and change two things: the emotion phrase (the part after “Give her”) and the preserved-features sentence. Leave the last two sentences alone; they are doing the work of keeping her the same person.
Figure 2. Neutral reference and the edited happy result from the prompt above.
This first edit cost 4,100 credits and took 1 attempt. Face shape, hair, clip and beauty mark all carried over, with raised brows and an open smile. The framing came out a little wider than the reference, which matters later.
Try it with your own OC: upload a neutral portrait and create one new expression in PixAI here.
Here’s my personal invitation link. If you’re a new user, you have to join through it, generate at least once and log in on two different days. PixAI currently gives both of us 50,000 credits within 24 hours of qualifying, and I get that reward too. It arrives later and you don’t need it to start: the first expression above costs 4,100 credits on its own.
Route A: generate all six expressions on one sheet
The prompt has five jobs: anchor the identity to @image1 and name the features to keep, list the six emotions, set the 2 × 3 grid, ask for consistent framing and style, and keep the background plain with no extra characters. PixAI’s own Tsubaki.3 examples show a 3 × 3 grid from one reference; I asked for 2 × 3 because I had six emotions.
Get Rehan Jamshed’s stories in your inbox
Join Medium for free to get updates from this writer.
Route A prompt
Use @image1 to create a 2 × 3 character expression sheet featuring the exact same character in six distinct expressions: neutral, happy, angry, sad, surprised, and shy.
Maintain her original face shape, dark auburn hairstyle, amber eyes, beauty mark, silver crescent-moon hair clip, and outfit. Keep the illustration style, head size, and framing consistent across all six panels.
Make each emotion clearly recognizable through distinct eyebrow positions, eye shapes, mouth expressions, and facial tension. Use a clean, plain light beige background throughout. Keep each expression centered in its own equally sized panel with clear separation. Do not add extra characters or duplicate expressions.
Figure 3. The complete sheet: 2 × 3 panels, 1216 × 1216 px.
Figure 4. The same sheet with labels added outside the image. The six cells are cut apart and re-spaced so the labels fit; no cell was edited.
Coverage was good: six panels, six different emotions, no repeats, in the order I asked for. Identity held up too. Face shape, hair, clip and outfit stay steady, and the beauty mark shows in most panels, though I could not clearly find it under the blush in the last one. Because the faces are generated together, consistent character expressions seem to come more easily here than from six separate runs.
Readability was uneven. Angry and surprised are clear. Happy has a big toothy grin, but a slanted, tense brow turns it into a smirk. Sad is mostly heavy lids and a flat mouth, so it reads closer to tired. Shy has a strong blush and worried brows but no smile, which looks more embarrassed or upset than bashful.
The layout missed one request. I asked for clear separation, but the panels touch with no gutters and the hair reaches the cell edges. If you split the sheet evenly, each cell is roughly 405 × 608 px, about a sixth of the pixels in a 1216 × 1216 image.
Route B: edit one expression at a time
Every edit starts from the same master reference, never from the previous result, so one bad image cannot pass its mistakes along. I wrote one shared instruction and swapped only the emotion phrase. To rebuild any prompt, drop the phrase from the table into the template. The six images were generated from @image1 in the generator; only the two corrections used Edit with text.
Route B template
Use @image1 to create an anime-style portrait of the exact same character. Give her [EMOTION PHRASE]. Preserve her face shape, shoulder-length dark auburn hair, amber eyes, beauty mark beneath her left eye, silver crescent-moon hair clip, and dark green outfit.
Maintain the original illustration style, head-and-shoulders framing, and plain light beige background. Change only her facial expression while keeping her identity recognizable.
I edited the neutral version too instead of reusing the reference, so Route B counts six runs, not five.
Figure 5. The six separate Route B assets, labeled. Each image sits in the layout at its original 1216 × 1216 size, unedited.
Happy and angry are the clearest of the set. Sad and surprised are readable but quiet; I asked for “visibly sad” and still got something subdued. The first shy attempt was the weakest, and I come back to it below. All six are full 1216 × 1216 images, framed a bit wider than the master, with more shoulder and a smaller face. That is consistent across the set, which matters more for a group of assets than matching the reference exactly.
The practical upside is control. Each prompt carries its own intensity wording, like “slightly raised” or “unmistakably angry,” and I can redo one without touching the others. The limit is that the wording does not guarantee the result.
What happens when one expression needs fixing?
On the sheet, I picked the happy panel, the smirk. I used a plain-language instruction using Edit with Text.
Sheet correction prompt
Fix only the happy expression. Give her a genuinely warm, cheerful smile with softly raised eyebrows, bright, friendly eyes, and slightly lifted cheeks. Make her look happy and approachable. Keep her facial features, hairstyle, accessories, and all other expressions unchanged.
The happy panel improved: softer brows, a warm smile, a light blush. But the sheet came with more changes than just a single-cell edit, and “all other expressions unchanged” did not hold. On my read, the neutral face picked up a worried brow, angry softened toward flat and weary, and hair details shifted.
Figure 6. Sheet route: the six panels before (top) and after (bottom) the happy fix, cropped to faces for comparison.
With separate images, I corrected shy. The first attempt had the blush but read tense, with her head angled a little. The instruction names what to keep:
Route B correction prompt (shy)
Change only her facial expression to look shy and hesitant rather than sad. Keep her head position, angle, direction, and gaze exactly as in the original image. Do not turn or tilt her head. Give her a small, nervous smile, subtly blushed cheeks, and slightly raised inner eyebrows. Her expression should feel bashful and pleasantly self-conscious, not sorrowful or worried. Preserve all other facial features, hairstyle, accessories, outfit, and framing.
Figure 7. Route B: the first shy attempt and the corrected version, both at their original size.
Now it reads as bashful: lowered gaze, a small smile, blush, inner brows lifted. The other five files could not drift, because they are separate files. The unintended change is that this result came back framed tighter than the rest, so the face is larger in the frame and would need a re-crop to match the set.
Counting runs from my prompt file, Route A was two runs (the sheet and the correction), and Route B was seven (six expressions and the correction). I did not re-roll any of them, so there are no failed attempts to add. Credits: 9,200 against 29,700, not counting the 3,200-credit reference both routes share (4,100 for the sheet and for each Route B image, 5,100 for each Edit with text run). The sheet is cheaper to start, but after its fix the set was less settled than before. Route B took more runs, and each result could be judged, kept or redone on its own.
Which workflow should you choose?
A character emotion sheet is a visual overview; a folder of images is a set of assets, and the two are not the same thing. Cropping a sheet does not hand you transparent backgrounds, matching canvas sizes or engine-ready sprites. My Route B files are better starting material than sheet cells, but they still have the beige background and are square portraits, not sprites. I did not test background removal, automatic splitting or engine export, so I cannot promise any of them.
My rule after this test: sheet first, to see whether the design holds together, then separate edits for the expressions I will really use. If you want to take the same character further, with turnarounds and outfit variations, the OC / VTuber Character Sheet Generation Guide covers a wider reference package, and PixAI’s Tsubaki.3 prompt and reference examples show expression editing and other grid layouts.
Short FAQ
How do I make a character expression sheet from an existing character?
Use a clear neutral portrait as @image1, ask for a 2 × 3 grid, list the features that must stay fixed, describe each emotion through eyebrows, eyes and mouth, and ask for equal panels with no extra characters. Add labels yourself afterward.
Why do several expressions on an AI expression sheet look almost identical?
Usually the prompt names the emotion without describing the face. “Sad” alone lets the model change only the mouth. Spell out eyebrows, eyes and mouth for each one, and compare the panels side by side. On my sheet, the sad panel sat closest to neutral.
How do I correct only one expression?
On the sheet, an instruction-only fix regenerated everything and moved other panels. For tighter control, redo that expression as a separate edit from the original reference, say what to keep (head position, gaze, framing), and check the result’s crop against the rest of the set, so the faces still read as consistent character expressions.
Does a character expression sheet become a visual novel sprite pack automatically?
No. A character emotion sheet is an overview. Sprites need cutting, matching canvas sizes, clean or transparent backgrounds and usually more, and I did not test those steps. Expression artwork is also not Live2D: these are flat images, not a rigged model with eyebrow and mouth parameters.
