chibi-sticker-sheet

name: chibi-sticker-sheet description: Use when generating a LINE/WeChat-style chibi sticker sheet (4x8 grid, 32 expressions) from an anime character reference image via Gemini, including transparent PNG output and individual cell slicing.

Chibi Sticker Sheet

Generate a 4×8 chibi sticker sheet from a character reference image with Gemini, then key out the white background and slice into 32 individual transparent PNGs.

Scope

• Input: 1 character reference PNG + a list of 32 expressions
• Output: sheet_white.png, sheet_transparent.png, cells/*.png (512×512 each)
• Requires: any one of GOOGLE_AI_STUDIO_API_KEY / GEMINI_API_KEY / GOOGLE_API_KEY (AI Studio) or VERTEX_AI_KEY (Vertex Express) or GOOGLE_GENAI_USE_VERTEXAI=true + GOOGLE_CLOUD_PROJECT/LOCATION (Vertex ADC); Python + uv; model gemini-3.1-flash-image-preview
• Not covered: per-cell text overlays; use PIL ImageDraw post-hoc for captions

Key Finding: Double-Matte Fails

The natural "white bg + black bg → α extraction" requires pixel-aligned foregrounds. Gemini ignores bg-change instructions in image-to-image edits and returns the same image. Use edge flood-fill keying instead (see scripts/key_alpha.py).

Inputs

Workflow

[1] Generate white-bg 4×8 sheet (two Gemini calls, each 4×4)
      -> gemini-3.1-flash-image-preview, image-to-image with CHAR_REF
      -> Call A: expressions 1–16  -> sheet_white_a.png
      -> Call B: expressions 17–32 -> sheet_white_b.png
      -> Stitch A and B vertically -> sheet_white.png

[2] Key out background
      -> scripts/key_alpha.py: edge flood-fill via scipy.ndimage.label
      -> save sheet_transparent.png

[3] Slice 4×8 grid
      -> scripts/key_alpha.py: min(cw, ch) square crop, centered per cell
      -> save cells/01_*.png … cells/32_*.png (512×512)

Why Two Calls

Gemini's image output resolution caps around 1024×1024. A single 4×8 canvas would squash each cell to ~128×256px — too small for clean chibi art. Generating two 4×4 sheets and stitching keeps per-cell resolution at the same quality as 16-sticker sets.

Prompt Structure

Two calls (Call A: expressions 1–16, Call B: expressions 17–32), each image-to-image with the same character reference.

Use identical art-style and character-lock text for both calls to ensure visual consistency.

Generate a 4x4 grid sticker sheet of 16 chibi stickers of the same character
from the reference image. Seamless pure white (#ffffff) background; no grid
lines, no cell borders, no text, no captions. Stickers evenly spaced.

Art style: LINE / WeChat Japanese chibi sticker, extreme super-deformed
2-head body ratio, oversized round head, tiny stubby body, thick uniform
bold black ink outline, flat cel shading, warm creamy pastel palette, two
large round pink cheek blush dots on every face, large round eyes with a
single bright white highlight, mochi chibi aesthetic.

Character lock (must match in every single cell):
[list each attribute: hair color/style, accessories, eye color, outfit details]

16 expressions, left-to-right top-to-bottom:
1. <action-first description>
...
16. <action-first description>

Run the same prompt template a second time with expressions 17–32 substituted in. Then stitch:

from PIL import Image

a = Image.open("sheet_white_a.png")
b = Image.open("sheet_white_b.png")
# Resize b to match a's width if they differ (Gemini output dims can vary)
if b.width != a.width:
    b = b.resize((a.width, int(b.height * a.width / b.width)), Image.LANCZOS)
combined = Image.new("RGB", (a.width, a.height + b.height), (255, 255, 255))
combined.paste(a, (0, 0))
combined.paste(b, (0, a.height))
combined.save("sheet_white.png")

Prompting Rules

• Lock first, change second. Open with "Character lock" before anything changes.
• Action-first expressions. "waterfall tears streaming" beats "sad". Include a visible prop or gesture.
• Be concise. Overlong prompts (>300 words) trigger MALFORMED_FUNCTION_CALL. No negative-list clauses (must NOT).
• No in-image text. Gemini cannot reliably render Chinese/Japanese — add captions via PIL post-hoc.
• Style vocabulary that works: super-deformed 2-head ratio · thick uniform bold black ink outline · flat cel shading · mochi chibi · large round pink cheek blush dots
• Character attributes to lock: hair color, hair length/style, ALL accessories (花饰 must name species: pink cherry blossom sakura flower hair ornament), eye color, every garment piece.

Expression Diversity Rules

Gemini tends to reuse the same pose for cells that share a column, especially cells 9 and 13 (column-1 of rows 3–4) and the equivalent pairs in the second sheet (25 and 29). Prevent duplicates:

1. Span all six emotional axes across the 32 slots — no axis should appear more than 6 times:

   Axis	Example actions
   Joy / excitement	raised fist, jumping, sparkle eyes
   Sadness / crying	waterfall tears, wilting head, tissues
   Anger / frustration	puffed cheeks, steam from head, finger-point
   Surprise / shock	wide-O mouth, hands-on-cheeks, dropped jaw
   Shy / embarrassed	hands over red face, hiding behind sleeves
   Calm / smug	arms crossed, side-eye, tea-sipping

2. Assign a distinct body verb to every cell. If two cells share the same verb (e.g., both "crying"), Gemini collapses them. Make verbs orthogonal: waterfall-tears arms-spread ≠ single-teardrop hands-clasped.

3. Critical column-1 pairs: 9 & 13, and 25 & 29 must each differ in both emotion axis AND body posture. Write each pair side by side before finalising and check they are visually distinguishable.

4. Vary facing direction and limb action across rows. Cells in the same column naturally echo each other; counteract by alternating pose direction (facing left vs. right) or prop presence.

5. No expression from Call A (1–16) should be repeated in Call B (17–32). Cross-check verbs before submitting the second prompt.

Reference 32-slot layout (copy and customize):

Call A (sheet 1, expressions 1–16):

1.  arms raised, jumping for joy, sparkle eyes
2.  waterfall tears streaming, arms limp at sides
3.  puffed cheeks, steam wisps from head, fists clenched
4.  wide-O mouth shock, both hands on cheeks Home-Alone pose
5.  smug smile, arms crossed, eyes half-lidded
6.  shy embarrassment, hands pressed together, deep blush
7.  thumbs-up grin, winking one eye
8.  single teardrop rolling, trembling lip, hands clasped
9.  hyper-excited wave, leaning forward, mouth wide open    ← must differ from #13
10. exhausted slumped, sweat drop, drooping eyes
11. index finger raised, lecture pose, tiny smile
12. heart eyes, both hands framing face, rosy cheeks
13. angry stomp, foot raised, fist shaking at sky           ← must differ from #9
14. sleeping ZZZ, head tilted, eyes closed
15. nervous laugh, hand behind head, eye twitch
16. victory peace-sign, tongue out, confetti burst

Call B (sheet 2, expressions 17–32):

17. cheering both fists raised, eyes glittering, mouth wide
18. sobbing face buried in hands, shoulders shaking
19. furious vein-pop, pointing finger, leaning in
20. startled leap, arms flailing outward, pupils tiny
21. lovesick float, dreamy spiral eyes, hearts around head
22. pouty sulk, arms crossed, cheeks puffed, eyes averted
23. excited run, legs spinning wheel, sweat drops flying
24. relieved sigh, hand on chest, eyes closed in relief
25. cheerful skip, one leg up, waving hello              ← must differ from #29
26. defeated head-desk slump, tiny sweat rivers
27. determined fist-pump, one eye closed, cape flutter
28. panicked sprint, eyes wide, papers scattering
29. grumpy arms-crossed side-eye, tapping foot impatiently ← must differ from #25
30. yawning stretch, arms out, eyes teary from yawn
31. embarrassed covering ears, blushing furiously, hunched
32. triumphant pose, foot on imaginary podium, sparkle burst

Grid Slicing: Auto-Detect Boundaries

Gemini does not divide the canvas into equal cells. Row/column heights vary (e.g., top row 550px vs. bottom row 480px). Hard image_size / 4 cuts cause sticker overflow into adjacent cells.

After stitching the two 4×4 sheets into one 4×8 canvas, run _find_cuts() on the combined image. Pass n_rows=8, n_cols=4.

Fix: compute per-row and per-column dark-outline occupancy profiles instead of white-fraction profiles. True inter-cell gaps have zero or near-zero dark pixels because the black outline disappears entirely in the white gutter, while white hair or clothing inside a sticker still leaves some outline pixels somewhere on that same row/column.

dist = np.max(np.abs(rgb.astype(np.int16) - 255), axis=2)  # 0 = white
dark = binary_dilation(dist > OUTLINE_THRESH, iterations=1)
row_profile = dark.sum(axis=1)          # dark outline count per row
col_profile = dark.sum(axis=0)          # dark outline count per col
# find contiguous near-zero-dark runs, then select the deepest valleys
# under broad cell-size sanity bounds instead of equal-spacing fallback

If a gutter is noisy and no exact zero-dark run survives, fall back to the lowest dark-count valleys in the smoothed profile. Cells are saved as cells/01_*.png … cells/32_*.png (512×512). See scripts/key_alpha.py: _find_cuts().

Alpha Keying: Edge Flood-Fill

See scripts/key_alpha.py. Core algorithm:

# dist[y,x] = max channel delta from pure white (0 = white, 255 = farthest)
dist = np.max(np.abs(rgb.astype(np.int16) - 255), axis=2)
near_white = dist < WHITE_TOL          # WHITE_TOL = 28

lbl, _ = label(near_white)            # connected components
edge_ids = {lbl[0,:], lbl[-1,:], lbl[:,0], lbl[:,-1]} - {0}  # border-touching
bg_mask = np.isin(lbl, list(edge_ids))

alpha = np.where(bg_mask, 0, 255).astype(np.uint8)
# feather boundary: GaussianBlur(radius=1.2) on alpha channel

Interior white pixels (shirt, skin highlights) are enclosed by foreground and never reach the border — they stay opaque.

White clothing leak: if the outline has 1–2 px gaps, the flood leaks through into white fabric. Fix: dilate dark outline pixels (dist > OUTLINE_THRESH=150) by OUTLINE_DILATE=2 iterations before flood-fill to plug gaps.

Retry Pattern

Gemini 3 image models have three transient failure modes:

See scripts/generate.py for the full retry wrapper.

Cross-Sheet Consistency (32-sticker specific)

Two separate Gemini calls will often produce slightly different rendering — brightness, line weight, palette temperature, or shading style may drift. Control it with the following steps:

1. Lock the style prompt byte-for-byte

Copy the exact art-style paragraph and character-lock paragraph from Call A into Call B without any edits. Even synonym substitutions (circular vs round) can shift Gemini's style.

2. Submit Call B with the sheet_white_a.png as additional reference

Pass sheet_white_a.png alongside CHAR_REF as the image input for Call B. Gemini will use it as a visual anchor. Example (google-generativeai SDK):

parts = [
    PIL_to_part(char_ref_img),
    PIL_to_part(sheet_a_img),  # style anchor
    types.Part(text=prompt_b),
]

3. Post-process for brightness/contrast parity

After stitching, compute the mean luminance of the top half (sheet A cells) and bottom half (sheet B cells). If they differ by more than 8 luma units, apply a PIL.ImageEnhance.Brightness correction to the dimmer half before writing sheet_white.png.

from PIL import Image, ImageEnhance
import numpy as np

def mean_luma(img_crop):
    arr = np.asarray(img_crop.convert("L")).astype(float)
    return arr.mean()

combined = Image.open("sheet_white.png")
h = combined.height // 2
luma_a = mean_luma(combined.crop((0, 0, combined.width, h)))
luma_b = mean_luma(combined.crop((0, h, combined.width, combined.height)))
if abs(luma_a - luma_b) > 8:
    # brighten or darken the bottom half
    factor = luma_a / luma_b
    bottom = combined.crop((0, h, combined.width, combined.height))
    bottom = ImageEnhance.Brightness(bottom).enhance(factor)
    combined.paste(bottom, (0, h))
    combined.save("sheet_white.png")

4. Visual QC before alpha keying

Open sheet_white.png and scan for:

• Line weight mismatch (bottom half lines thinner/thicker)
• Palette temperature shift (warmer/cooler tones)
• Hair color drift across the boundary row

If drift is visible, re-run Call B with a slightly adjusted prompt (e.g., add "same warm creamy pastel palette as reference sheet") and a fresh seed. Repeat QC.

5. Common cross-sheet failure modes

Common Mistakes

WeChat Submission Extras

Each sticker set needs three additional assets. Generate with scripts/generate_extras.py:

uv run generate_extras.py <sticker_dir> <char_ref_image> "<theme hint>"

Theme hints by set:

• Snow/winter → "snowy winter wonderland"
• Autumn ginkgo → "golden autumn ginkgo forest"
• Summer sailor → "sunny summer beach ocean waves"
• Spring school → "cherry blossom spring school campus"
• Kimono/plum → "red plum blossom Japanese garden"

WeChat rules:

• Banner: colorful background only — no white, no transparent; no text; story-rich scene
• Cover: transparent bg; no white outline; avoid over-cropping (half/full body preferred)
• Icon: transparent bg; head-only, no square border; must differ across sets

Multi-Platform Banner

scripts/generate_banner.py produces a standalone banner for any registered platform — independent of the WeChat extras flow. WeChat itself is one preset; Twitter/X is another. New platforms are added by appending to the PLATFORMS dict.

uv run generate_banner.py <char_ref> "<theme>" <out_path> --platform twitter

Each preset is a Platform(width, height, aspect_ratio, composition_hint). The hint is appended to the prompt and is the place to encode platform-specific safe-zone layout (avatar overlay, mobile crop, letterbox bars). WeChat needs no hint; Twitter needs explicit letterbox awareness — see below.

Letterbox Lesson (Extreme Aspect Ratios)

Gemini 3 image preview only supports a fixed aspect-ratio menu (1:1, 4:3, 3:4, 16:9, 9:16). For ratios more extreme than 16:9 (Twitter 3:1, LinkedIn 4:1) we generate at 16:9 then center-crop the middle band. The naive prompt strategy fails in two opposite ways:

The fix is to tell Gemini that the top/bottom 22% are off-screen letterbox bars (showing only background overflow), and place subjects in the middle 25%–75% band of the source canvas. This produces full vertical fill in the final visible frame plus a 3% safety margin against the crop line on each side.

The Twitter preset's composition_hint encodes this verbatim. When adding a new extreme-aspect platform (e.g. LinkedIn 4:1 → middle 44% of 16:9 source), compute the letterbox percentages from the math and follow the same structure.

Verification

# α channel: expect ~30-40% transparent pixels for a sticker sheet
python -c "
from PIL import Image; import numpy as np
a = np.asarray(Image.open('sheet_transparent.png').convert('RGBA'))[...,3]
print('min', a.min(), 'max', a.max(), 'transparent%', (a<10).mean()*100)
"

# Cell sizes must all be square
ls -la cells/*.png | awk '{print $5, $9}' | head