Sticker Pack Maker (微信表情包制作)

Turn digital-human / reference sheets into a complete, upload-ready WeChat
sticker-album package. The pipeline below is proven end-to-end; the bundled
scripts are deterministic and reusable.

When to use

• "做个微信表情包" / "用数字人做表情" / "把参考图做成表情专辑"
• Need looping GIF emojis from AI frames
• Need to clean ImageGen artifacts, make transparent covers/icons, package for 微信表情开放平台

Pipeline (5 steps)

1. Decide scope — count × actions. Example: 16 stickers × 2 actions (blink +
   nod) = 32 AI generations (dropping a 3rd action like "shrug" simplifies cost).
2. Generate frames — ImageGen img2img from a reference grid. Expect ~32–48
   raw PNGs at 1024×1024.
3. Clean frames — run scripts/clean_frames.py to strip the right-edge ghost
   artifact and chroma-key the background to transparency.
4. Build GIFs — run scripts/build_gifs.py with a sticker→frames JSON map;
   produces 5-frame loops (睁→眨→睁→点头→正) at 240×240.
5. Make covers + package — scripts/make_cover.py for a solid-bg 240 cover and
   scripts/crop_avatar.py for the 640 real-person avatar; assemble the upload
   folder + zip.

Bundled scripts (run with the project's Pillow venv)

• scripts/clean_frames.py [input_dir] [--out cleaned_dir]
  Detects and removes the ImageGen right-edge ghost (cols ~959–1023) via
  column-run gap detection, re-centers, then edge-floods near-background pixels
  to transparent. See reference pitfalls #1, #2, #3.
• scripts/build_gifs.py --cleaned [dir] --map stickers.json --out [dir] [--static [dir]]
  chroma-keys static bases, bbox-normalizes every frame (crop alpha bbox → scale
  to 82% of 240 → center) to stop scale/position jitter, assembles the 5-frame
  loop with loop=0, disposal=2, transparency=0, optimize=True. Keeps each GIF
  under 500KB.
• scripts/make_cover.py [in] [out] --size 240 --text "专辑名" [--mode character|photo]
  Composites a transparent character onto a solid engineering-blue gradient (or
  square-crops a photo) with an opaque bottom text band. Avoids the
  "floating transparent avatar" ugliness.
• scripts/crop_avatar.py [in] [out] --size 640
  Center-crops any photo to a square and resizes for the platform's real-person
  640×640 artist avatar.

Critical constraints (read before uploading)

Load references/wechat_sticker_spec.md for the full field table, upload steps,
and the 8 hard-won pitfalls. Highlights:

• Name ≤ 8 chars; intro ≤ 80 chars (the "≤11 char intro" rule circulating
  online is wrong — verify against the live form).
• Two different "avatars": 240×240 album cover (use a solid-bg image) AND
  640×640 real-person artist avatar (AI art will likely be rejected).
• Chat icon must be 50×50 and transparent; do NOT put a background on it.
• WeChat allows 16 main GIFs max, each under 500KB; "add works" binds to one
  表情形象 with only 1 later change — upload all at once.
• ImageGen background:transparent does not work → chroma-key instead.
• You (the agent) cannot see the images — rely on numpy pixel stats (density
  grids, bbox, column runs) and always ask the user to eyeball sample GIFs.

Typical file layout produced

AI表情帧/cleaned/   →  transparent RGBA frames (from clean_frames.py)
GIF动效/            →  16 *_loop.gif (from build_gifs.py)
上架包/              →  16 gifs + 封面.png + 聊天图标.png + 详情页横幅.png
                         + 赞赏引导图.png + 致谢图.png + 头像_640.png + 上架清单.txt
上架包.zip          →  upload-ready archive