Animated sticker packs with GPT Image 2.6

On launch day people turned their own photos and characters into animated sticker sets: six chibi stickers from a profile photo (@goan999999), a character eating rice with one caption (@Gorden_Sun), "GIFs of this character's four emotions" for LINE (@tetumemo), custom sets sold on Xiaohongshu (@Leviwoyun), and a how-to for publishing sets with tips on WeChat's sticker platform, from an author who says AI stickers earned them over 1,000 RMB (@ama_huangama, 635 bookmarks).

The GIF mechanics (pixel budget, loop lock, failure modes) are in gpt-image-gif. This skill adds the pack plan, the per-platform specs and the export.

Our example pack, made on varg with gpt_image_2_6 (2026-11-04/05), from the mascot still Pip: LOL and THANKS! (6 frames of 768×768, 160 ms, transparent) and one failure, ON MY WAY! (three words garble between frames).

1. Plan the pack

• 6-8 emotions. LINE sets have 8, 16 or 24 stickers; WhatsApp packs 3-30; WeChat albums 8-24 (section 4). Eight covers every platform's minimum.
• One short word per sticker. One word held in every frame ("LOL", "THANKS!"); three words broke (measured). @goan999999's captions were 1-3 Chinese characters.
• One simple motion per emotion, one that reads at 160 px: a wave, a nod, a bow, a jump, a shake.
• The first frame is the whole sticker. WhatsApp "ends the animation on the first frame after looping", and LINE shows the first frame as the store image (section 4).

A starter set (word, motion):

@goan999999's set, for comparison: 嗨 "hi" (waves), 收到 "got it" (nods, thumbs up), 爱你 "love you" (finger heart), 哈哈哈 "hahaha" (laughs to tears), 加油 "you got this" (fist), 谢谢 "thanks" (palms together, bow).

2. Make it

1. The character still. From a photo: an edit with the photo in files, e.g. "A chibi sticker character of the person in the photo: big head, small body, same hairstyle, face shape, glasses and clothes. Full body, front view, neutral happy face, thick clean outline, flat colours. Transparent background." Settings: aspect_ratio: "1:1", background: "transparent", output_format: "png". Use a photo only of someone who agreed to it. Approve this still before anything else. 6-7 credits.
2. Upload the still (create_upload or POST /v2/files) to get its https://s3.varg.ai/… URL.
3. One GIF per emotion, always from the same still:

call_tool({
  "tool": "image",
  "input": {
    "model": "gpt_image_2_6",
    "prompt": "Animated sticker of the character from the reference image. Emotion: thankful. Motion: bows with palms together, then comes back up. 6 frames; the first and last frame show the same pose, so the loop has no jump. Same size and position in every frame, fixed camera. Caption above the head in bold rounded letters with a white outline: \"THANKS!\" The caption does not move or change. Transparent background. Flat colours, thick clean outline.",
    "files": [{"url": "https://s3.varg.ai/files/194b94d0-58ff-4e8b-91de-4f1959d04013/file_5ed17708b8d2.png"}],
    "provider_options": {"fal": {
      "output_format": "gif",
      "frame_count": 6,
      "image_size": {"width": 768, "height": 768},
      "frame_duration_ms": 160,
      "background": "transparent",
      "quality": "high"
    }}
  },
  "wait": true
})

6-7 credits each (measured 2026-11-05). For 2-8 frames fal always draws an eight-cell sheet (gpt-image-gif section 1), so 6-8 frames cost no more than 4. 4. Check every sticker: magick sticker.gif -coalesce +append strip.png. Reject a sticker whose caption changes letters, whose face changes, or whose last frame jumps to the first. 5. Export: bash scripts/export.sh sticker.gif all (section 5).

@goan999999's prompt (ChatGPT, published in the reply), original:

用我上传的照片做一组 Q 版微信表情包，保留本人的发型、脸型、眼镜和服装特点，所有表情都要像同一个人

做6款：嗨（挥手）、收到（点头点赞）、爱你（比心）、哈哈哈（笑到流泪）、加油（握拳）、谢谢（合掌鞠躬）

每款单独画一张2×2的连续动作图，共4帧，按左上、右上、左下、右下播放。固定人物大小和位置，字幕不要移动，白色背景，不要边框和序号。动作要有变化，最后一帧能接回第一帧

画好后，分别合成320×320的GIF，每帧200毫秒，无限循环。给我6个独立文件，再打一个压缩包。

English translation:

Use the photo I uploaded to make a set of chibi WeChat stickers. Keep my hairstyle, face shape,
glasses and clothes; every sticker must look like the same person.

Make 6: 嗨 "hi" (waving), 收到 "got it" (nod and thumbs up), 爱你 "love you" (finger heart),
哈哈哈 "hahaha" (laughing to tears), 加油 "you got this" (clenched fist), 谢谢 "thanks" (palms
together, bow).

Draw each one as its own 2×2 sheet of consecutive poses, 4 frames, played top-left, top-right,
bottom-left, bottom-right. Keep the figure's size and position fixed, don't move the caption, white
background, no borders or numbers. The motion should change, and the last frame should lead back to
the first.

When they are drawn, combine each into a 320×320 GIF, 200 ms per frame, looping forever. Give me 6
separate files, and also a zip.

On varg you skip the hand-made 2×2 sheet: output_format: "gif" draws and assembles the frames. Keep the rest: same person, fixed size and position, a caption that does not move, last frame back to the first. The author's warning: in ChatGPT, asking for "a GIF" sometimes returned one still with motion lines, and in their set "the hair and the text jump a little".

@Gorden_Sun's one-line version: 「使用这个角色，生成一张GIF图，内容是她吃大白米饭，图片里有文字：真香！」 "Use this character and make a GIF of her eating a big bowl of white rice, with the text in the picture: 真香!" ("so good!").

3. Failure → fix

4. Platform specs (checked 2026-11-05 on the official docs)

How each was checked, and what we could not verify:

• Telegram, WhatsApp, LINE, Slack: read on the linked page with WebFetch on 2026-11-05. Telegram "animated stickers" (.TGS) are Lottie vectors, a different format a GIF cannot become; this skill makes video stickers.
• WeChat: the guide page is built by JavaScript and returned empty to WebFetch, so we read the specification text out of the page's own script bundle on sticker.weixin.qq.com (2026-11-05). The same page says single stickers (表情单品) "暂不支持包含推广形象、推广信息，或人物肖像（如人像照片、肖像画等）" ("do not yet support promotional figures, promotional information, or portraits of people, such as photos or painted portraits"). A chibi drawn from your photo may count as a portrait; we could not verify how reviewers treat it.
• Discord: the help-center sticker FAQ returned HTTP 403, so the numbers come from the developer docs' "Create Guild Sticker" section. How many sticker slots a server has was not verified.
• Slack names no pixel size; the script uses 128×128.
• KB: the script treats 1 KB as 1,000 bytes, so files pass under either reading.

5. Export: scripts/export.sh

bash scripts/export.sh thanks.gif all            # every platform -> thanks_stickers/
bash scripts/export.sh thanks.gif line out/      # one platform, chosen folder

What it writes per platform, and what it does to fit:

Every output is re-read and checked (size, pixels, length, frames, fps, audio); the script prints PASS or FAIL per line and exits 1 if anything fails. It strips metadata, so the C2PA content credentials the model embeds are not carried over.

Tested 2026-11-05 on the LOL sticker (768×768, 6 frames, 960 ms loop, 444 KB), all six platforms passed:

The LINE frame rules were also tested on made-up GIFs: 32 frames → 16, 3 frames → 6, an 8 s loop → 4 s. The mascot wave GIF (white background) fails the LINE check as it should.

Then upload with each platform's own tool: Telegram's @Stickers bot (named on core.telegram.org/stickers), an app built on the WhatsApp/stickers sample, LINE Creators Market, sticker.weixin.qq.com, the Discord server's sticker settings, Slack's custom emoji page.

Related skills

gpt-image-gif (GIF rules, loop lock, WebP/MP4 for the web), gpt-image-2-6 (model guide, prices), gpt-image-markup-edit (fix the base still before you animate it), varg-character-design (designing the character itself).