Animated Sticker Studio · 动态表情包工坊

Pipeline: plan → reference → atlases → cut → build → look → fix → deliver. Scripts live in scripts/ (Python 3 + Pillow; ffmpeg for Telegram/MP4). Run every script with --help before first use.

1. Plan (manifest.json)

Ask only what you cannot default. Defaults: 16 stickers, Simplified Chinese captions, generic GIF; add wechat if they mention 微信表情开放平台/上架, telegram/line/discord/slack/whatsapp if named. WeChat albums take any 8–24 stickers; LINE packs must be 8, 16 or 24.

Write manifest.json next to the work (format: references/manifest.md). Pick captions from references/phrases.md unless the user gave their own, and give every entry:

• pose — what the body does, in English, concrete enough to draw (arms, props, expression). Two entries may never share the same pose with only new text.
• motion + icon + optional caption_fx — chosen to match the line (table in phrases.md).
• For the 4–8 stickers where the gesture IS the joke (挥手, 鼓掌, 打字, 干饭, 点赞, 哭, 打哈欠, 比心…) put them in a group with "keys": 4 and write four key_poses. These become real frame-by-frame animation. Single poses get procedural motion only.

Show the user the plan as a short table (编号 / 文案 / 动作) before spending on generation when the pack is ≥24 or the user seems picky; otherwise go.

2. Reference

• User image → copy into ref/, set character.reference. Write character.anchors: one English sentence listing what must never change (shape, colours, markings, accessories).
• 3D model → render a front/three-quarter view with the face visible, use that.
• Only a description → gen_atlas.py manifest.json --provider hiapi --make-reference "<description>", show the result, get a yes.
• Use only art the user owns or may adapt. Do not recreate a third party's character from memory.

3. Atlases

One sheet per group (≤16 cells). Either:

• Your own image tool (Codex image_gen, etc.): gen_atlas.py manifest.json --print-prompts, generate each sheet with the reference attached, save as the printed path, then set the group's atlas and grid.
• Bundled providers: gen_atlas.py manifest.json --provider hiapi|openai (keys only from env: HIAPI_API_KEY, or OPENAI_API_KEY + optional OPENAI_BASE_URL/OPENAI_IMAGE_MODEL). Skips groups that already have an atlas; --group N redraws one. Never print or echo a key.

Open every sheet at full size. Reject and redraw (--group N) when the identity drifts, a cell is missing or doubled, limbs are clipped, or two poses are the same.

4. Cut

cut_atlas.py manifest.json — removes solid or fake-checkerboard backgrounds, finds each sprite by its pixels (not by a rigid grid), writes sprites/ and fills every entry's sprites list. Open each atlases/*.index.png and confirm number N shows the pose written for entry N. If the model swapped cells, fix by editing sprites paths in the manifest — never by renaming captions to fit wrong art.

5. Build

python3 scripts/build_pack.py manifest.json out --mp4          # platforms from manifest
python3 scripts/build_pack.py manifest.json out --only 3,7-9   # quick re-render of a few while iterating

Fonts are found automatically (PingFang/Hiragino/Noto CJK/YaHei); --font overrides. Each platform file is re-encoded until it fits that platform's byte limit; report.json records the setting used.

6. Look (required — you cannot watch a GIF, so look at these)

• out/review/motion-XX.png: every sticker as 4 frames side by side on alternating white/dark rows. Check: the motion reads as the caption's emotion, keyframes don't jitter sideways, nothing clipped, no white halo on dark rows, captions legible, accents not covering the face.
• out/review/contact.png: whole pack at once — check variety and that no two look alike.
• The build's JSON summary: failing (per-platform spec failures) and pack_issues (near-duplicate poses, wrong pack size, stickers that barely move). Fix all of them or tell the user exactly which remain and why.

Fix loop: wrong motion/icon/caption → edit manifest, --only; bad art → redraw that group or cell; then rebuild everything once at the end.

7. Deliver

Give the user: the ZIP path, 预览.html (opens offline; light/dark/checker backgrounds and a chat mock-up), the MP4 if made, and a one-line pass/fail per platform from the summary. Mention platform limits that still need a human: WeChat needs a 750×400 banner and copyright paperwork, LINE/Telegram need account-side upload. Specs are in references/platforms.md; say they are as of 2026-11 and the platform console wins.

Keep the original reference and atlases out of the ZIP unless asked.