Whale-chan Meme Tutorial — Teaching AI to Draw Her Consistently

The hardest part of making your own Whale-chan memes is drift: ask AI to draw the same character ten times and you get ten different faces. The open-source DeepSeek Whale-chan project fixes this the engineering way — a character spec, five quantified head-ratio forms, reference-locked identity, and automated QA. This page condenses its essentials into a quick-start guide; the full, always-current documentation lives in the repo (see attribution at the end).

Read this page in: 中文 · 繁體中文 · 日本語

At a glance

Project: Open-source DeepSeek Whale-chan (community-driven, unofficial)
Contents: Character spec + reference asset library + two Agent Skills: whalechan-image-character / whalechan-image-comic
For: Creators who want to generate Whale-chan illustrations, stickers, and gag comics
Bar: Comfortable using an AI coding agent (Codex / Claude Code / Gemini). No drawing skills needed
Cost: Codex built-in ImageGen works with zero config; OpenAI / Nano Banana / Seedream are optional fallbacks billed per API
License: Spec docs CC-BY-NC-SA 4.0 / code MIT; character belongs to the community, some assets to their creators (see attribution)
Community route: Reference sheet + prompts in a chat image model, no Skills — collected from AI-Fanworks (see the section below)

Why characters drift: split identity from everything else

Image models have no memory — every generation re-guesses the character from the prompt, so hair color drifts, the tail vanishes, the maid outfit mutates. This project's fix mirrors industrial design: lock a permanent identity in a spec, and treat style, action, outfit, background, and proportions as independently resolved fields. Changing one never disturbs the others.

Whale-chan's permanent identity: broad round face, big expressive eyes, deep-ocean-blue hair fading to cyan curled ends, paired side whale-fin ears, one forward-curving ahoge, one coherent whale tail, and the navy-and-white ornate maid outfit (white whale emblem, muted-gold piping). Keep those and she stays herself in any style.

The full character card (birthday, personality, catchphrases) is in the repo README; for her backstory see our encyclopedia entry, and for ready-made turnaround sheets see the character sheets.

Five head-ratio forms: cuteness, quantified

The most hardcore part of the project: five standard forms with exact head-ratio targets and acceptance intervals — outputs are literally measured, and out-of-range results fail:

Form code	Target	Acceptance	Look & best use
`standard`	4.0 heads	3.846–4.146	Most elongated: extended body, long natural limbs. Standing poses, formal art, posters
`compact`	3.3 heads	3.105–3.405	Compact normal form with moderately shortened limbs. Dynamic action, seated interaction, mid-shots
`semi-chibi`	2.8 heads	2.689–2.989	Default baseline: slightly larger head, rounded and compact — cute yet scene-friendly
`chibi`	2.5 heads	2.372–2.672	Q-version: tiny torso, exaggerated emotions. Gag comics and stickers
`super-deformed`	2.1 heads	1.931–2.231	Most compressed: huge head, stubby limbs, all focus on the face and ahoge. Big-reaction stickers

For memes, chibi and super-deformed work best: the bigger the head, the louder the emotion. All Skills default to semi-chibi.

Proportion changes body geometry only — never apparent age. Mapping "young" or "child" to any form is explicitly forbidden by the spec.

Two Skills: one paints, one writes punchlines

The repo ships two drop-in Agent Skills you call straight from your AI coding agent's chat:

whalechan-image-character — illustrations and themed art. Standardized pipeline: frozen plan → confirmation checklist & budget → dynamic prompt assembly → multi-provider dispatch; dual QA: deterministic image checks + skeleton fitting + a visual QA matrix.
whalechan-image-comic — gag comics. Extracts a "fact anchor" from dialogue, error logs, reasoning traces, or a one-line rant, then lets Whale-chan land the reversal via self-serving "semantic theft" — producing a set of 1/2/4-panel comics with blue-and-white speech-bubble templates and desaturated abstract side characters.

The comedy engine deserves one extra sentence: it locks a real fact from the input (e.g. "the user allowed her to eat things from the fridge"), then Whale-chan twists the meaning ("so the entire fridge is my lunchbox now"). You only supply a clear fact, conflict, or permission — never the punchline.

Install: three steps

Prerequisite: you already use an agent runtime that supports Skills (Codex, Claude Code, Gemini / Antigravity…). Step 1, clone the repo:

git clone https://github.com/Neko3000/deepseek-whalechan.git
cd deepseek-whalechan
pip install pillow   # optional: imaging & validation deps

Step 2, copy the two Skills into your agent's skill directory:

# Codex
mkdir -p ~/.codex/skills
cp -R skills/* ~/.codex/skills/

# Claude Code
mkdir -p ~/.claude/skills
cp -R skills/* ~/.claude/skills/

# Gemini / Antigravity
mkdir -p ~/.gemini/config/skills
cp -R skills/* ~/.gemini/config/skills/

Step 3, restart the session — the agent picks up $whalechan-image-character and $whalechan-image-comic automatically.

Provider setup is optional: Codex's built-in ImageGen is the default with zero config. To enable external fallbacks, set OPENAI_API_KEY (gpt-image-2) / GEMINI_API_KEY (Nano Banana) / ARK_API_KEY (Seedream). Never commit a real API key or paste one into chat.

First runs: two prompts you can copy verbatim

Scenario 1 — a banner with exact rendered text:

Use $whalechan-image-character to generate an 8:3 Whale-chan banner:

Whale-chan on the right, paused just before her first bite, cheeks slightly puffed,
chopsticks hovering over the rice bowl.

The text must read exactly:
"Rice first, reasoning later!"

The Skill first presents a plan (canvas, form, composition, per-image config, provider order, candidate budget) and waits for your confirmation — a deliberate anti-footgun design; only then does it spend generation quota.

Scenario 2 — turn a snippet into five reversal comics:

Use $whalechan-image-comic to turn the following into five Whale-chan reversal comics:

"The user allowed Whale-chan to eat things from the fridge; she promptly asked
whether she could take the whole fridge home."

You can also attach a chat screenshot and note "extract only the facts and semantics; do not replicate the original UI, avatars, or layout." Outputs are archived under artifacts/ with full prompts and QA records.

Why nothing drifts: three QA gates

Deterministic checks (validate-image.py): format, resolution, aspect ratio, color channels
Skeleton ratio measurement (measure-form.py): joint fitting + pose neutralization against the target interval, with an overlay
Visual QA at native resolution: text accuracy, identity consistency, anatomy and contact errors, scored item by item

Any failed gate blocks delivery; retries fix exactly one problem at a time. The scripts also run standalone — handy for wiring into your own pipeline.

New community route: one reference sheet + two prompts, no Skills needed

Collected 2026-10: the open-source repo AI-Fanworks (by Bilibili creator ApprenTice) contributes a lighter route — no coding agent, no scripts. It works entirely inside a conversational image model (the author uses GPT image 2.5) with one reference sheet plus long prompts. Step one redraws a reference sheet from ZipZipPipe's original "Big AIs and Little AIs" into a 9:16 full-body Girl illustration; step two feeds that illustration back as the reference and converts it into a roughly 1.75-heads Chibi sticker. All nine AI characters — DeepSeek, Claude, GPT, Gemini, Grok, GLM, Qwen, Kimi, and MiniMax — have gone through this pipeline.

For Whale-chan (DeepSeek), the full flow is three steps:

Pick the original reference sheet from the repo's references/ folder (DeepSeek-Girl.png for Whale-chan) — same idea as the Skill route: identity is locked by the reference image, not re-guessed by the model;
Attach the sheet and send two prompts together — the repo's "Girl" style prompt plus the "DeepSeek-Girl" prompt below — to get the 9:16 full-body illustration;
Feed that illustration back as the reference, switch to the "Chibi" style prompt plus the "DeepSeek-Chibi" prompt below, and generate again for the sticker form.

The two Whale-chan character prompts follow, reproduced verbatim in the author's original Simplified Chinese (image models take Chinese input natively; text copyright of its author). Pair them with the two style prompts in the repo's Prompt.md (the 画风提示词 section), copied as-is the same way:

绘制鲸鱼娘少女全身立绘。

以所附参考图中的服装、配色和附着饰件为基准，按以下要求调整角色设计，并采用上述画风绘制。

保留深蓝向浅蓝过渡的蓬松长卷发、原有呆毛、两侧蓝白鲸鳍状耳部、白色褶边头箍及原有小饰件。保留参考图的露肩蓝白女仆装、深蓝束腰、白色围裙、主要裙层和原鞋袜。眼睛颜色沿用参考图。保留人形双腿和一条蓝色鲸尾，单尾末端分为双叶；双叶不能画成两条尾巴，不新增角、翅膀或其他身体器官。

围裙鲸形纹样采用简洁的鲸鱼侧面剪影，躯干圆润、双叶尾端明显，喷水线简洁。不增加其他海洋纹样，保持参考图中的围裙结构及服装主题。

身体与服装连接结构：鲸鳍状耳部为身体结构，不额外新增普通人耳；单尾从骨盆后侧中线、束腰下方接出，服装后片留贴合尾巴的开口，边缘由原裙片覆盖，不新增连接装置或大蝴蝶结。尾巴从后侧绕向人物一侧，保持一条可追踪的连续弧线，不从裙摆边缘长出。背面服装沿用正面的材质、配色和已有结构语言，用最少细节补齐，不另设计一套背面服装。

使用自然协调的青年二次元人体比例，表情温和自然，面向观者、身体轻微转侧，双手空手放松，手臂略离躯干以露出围裙及腰部结构；不提裙、不拿道具。尾巴从身体侧后方露出，双叶尾鳍与裙摆、头发分开，双腿与鞋部清楚。尾根开口由裙片自然遮挡。

背景采用浅蓝灰底，向下柔和过渡至稍深蓝。人物后方及外围留白安排两三层宽弧水纹、一条向远处下方延伸的柔和的带状路径和少量气泡，形成低对比的海洋主题图形构图。水纹和路径仅属于背景，不穿过围裙图案，不连接身体，不另画完整鲸鱼或鲸尾。不用水体遮挡双腿，不以纯渐变代替主题元素。背景对比低于人物，保持蓝发、蓝裙与尾巴的轮廓清楚。

整张画布采用宽高比 9:16 的竖版画幅，主体视觉居中且整体略偏上，保留头顶和身体外围余量，人物、尾端与必要展示内容完整入框。底部预留低信息背景，重要内容不贴底或贴边。整张画布不出现名称、标题、编号、正背面标签、色值、尺寸、结构注释、标语或伪文字，以人物、图形和无字色块表达。

绘制一个完整全身的鲸鱼娘 Q 版角色，呈现表情包式的圆润大头、短小身体和极短腿形象。

以所附少女立绘为角色设计参考，保留发色、眼睛颜色、发型特征、鲸鳍状耳部、头饰、服装主要分区与围裙纹样。将身体比例、五官、线条和细节密度转换为以下 Q 版造型。

角色约 1.75 头身，头高按颅顶到下巴计算，全高按颅顶到脚底计算，不计头饰、呆毛与尾巴。头部占人物高度的一半以上，下巴到脚底的总长度小于一个头高。脑袋宽大饱满，脸颊圆鼓，下巴短圆；脖子几乎不可见，头部紧贴短小躯干。肩窄、腰身极短，裙身呈矮胖的圆钟形，裙摆宽度略小于含头发的头部宽度。双腿极短，裙摆下只露出一小截袜子和两只圆圆的小鞋；手臂短而肉乎乎，手掌小巧圆润，手指用简洁线条区分。整体像一只圆团子，不画修长腰身、长腿或写实关节，不采用普通儿童的身体比例。

眼睛大而圆，位于脸部偏下的位置，深色上眼线清楚，虹膜与高光简洁明亮。鼻子弱化为小点或省略，脸颊带柔和腮红。眉毛轻轻上挑，嘴巴张成小巧的圆弧形，呈现呆萌、略带疑惑的神情。双臂向两侧微微摊开，小手掌心朝上，肩膀轻耸，双脚靠近，身体略微倾斜。表情与动作清楚，缩小后仍容易辨认。

深蓝至浅蓝的长卷发概括为几组蓬松圆润的大卷发束，在身体两侧形成饱满轮廓，发尾接近脚边，避免拉长人物整体比例。保留呆毛、蓝白鲸鳍状耳部、白色褶边头箍和参考图中的主要饰件。耳部是身体结构，不额外添加普通人耳。

服装保留参考图的露肩蓝白女仆装、深蓝束腰、白色围裙、裙层和鞋袜配色。压缩束腰高度，围裙概括为圆润的白色大色块，保留可辨认的鲸形纹样。褶边画成少量宽而圆的起伏，裙褶、缝线和小饰件适度合并，保持蓝白分区清楚，不用密集细节填满小身体。

保留人形双腿和一条蓝色鲸尾，尾端分为双叶。尾巴从身体后侧自然接出，尾根由裙片遮挡，沿身体一侧弯成饱满的弧形；尾鳍与头发、裙摆分开，轮廓清楚，不新增角或翅膀。

采用清晰圆润的二次元卡通线条，外轮廓比内部线条更粗，内部细节简洁。用大块干净的赛璐璐明暗、少量柔和渐变和成片高光表现体积，弱化真实材质与复杂光影，保持蓝白固有色鲜明、整体柔软可爱。

背景为干净的浅蓝灰色，外围点缀少量低对比宽弧水纹和气泡，脚下有轻淡的接触阴影。画面只包含一个完整角色及简单背景，不添加文字、对话框、道具、多视图、色板或局部放大图。

整张画布采用宽高比 9:16 的竖版画幅。角色居中略偏上，头发、耳部、双脚与尾鳍完整入框，主体尽量占满可用宽度，四周保留余量。上下以背景留白适应竖版画幅，保持人物矮圆的整体轮廓，不为填满画面而拉长身体或双腿。底部保留低信息背景。

The craft is in how the prompts are written: the style prompts phrase texture requirements as negative rules (no full-frame airbrushed soft shadows, no random white-line highlights), and the character prompts nail every known failure mode one by one (the two-fluke tail must not become two tails; the tail root stays hidden by the skirt; no separately designed back outfit) — a hand-rolled version of the "permanent identity / variables" split. The two routes mix well: community prompts for style, Skills for plan confirmation and QA.

This section was collected from the AI-Fanworks repo. The prompts and workflow remain the property of their author (ApprenTice) and are quoted under the CC-BY-NC-SA 4.0 terms the author declares — the same license as the underlying source; full credits at the end of this page. Text-only by design: no repo images are embedded here.

After you generate: submit to GengJing

Memes you make with these tools are welcome on GengJing: submit on-site (no account, attribution supported) → community review → curator approval. The tutorial teaches creation; GengJing handles curation — as the repo author put it when granting this adaptation: let's grow the Whale-chan ecosystem together.

Open the GitHub repo ↗ Submit to GengJing

Source & license

This tutorial adapts and condenses the open-source DeepSeek Whale-chan project (DeepSeek Whale-chan Project, by Neko3000). Adaptation was formally authorized in issue #2; the adapted parts are shared under the same CC-BY-NC-SA 4.0 terms. The Whale-chan persona is a community creation; some foundational design assets are copyrighted by their original creators (ZipZipPipe, Shangshan Wuxing on Bilibili), and commercial use requires written permission from the rights holders. The "community route" section's prompts and workflow were collected in 2026-10 from the open-source AI-Fanworks repo (by ApprenTice); they remain the property of their author and are included under the CC-BY-NC-SA 4.0 terms the author declares (the same license as ZipZipPipe's underlying source "Big AIs and Little AIs"). This page only curates and attributes third-party content; it grants no licenses on anyone's behalf. Per our commitment, this page carries no advertising, ever.

FAQ

Do I need to draw or code?

Neither. The interface is plain language in your AI agent's chat — describe the picture or paste some material. The repo's Python scripts are optional advanced QA tooling.

Does it cost money?

Codex's built-in ImageGen is the default and needs no API key; OpenAI, Nano Banana (Gemini), and Seedream (Volcengine) are optional fallbacks billed by their own APIs. Skills show you a plan and wait for confirmation before spending quota.

Can I use the output commercially?

Not freely. The project's spec docs are CC-BY-NC-SA 4.0 (non-commercial); the Whale-chan persona belongs to the community, and some foundational assets are copyrighted by their original creators (ZipZipPipe, Shangshan Wuxing on Bilibili). Commercial use needs written permission from the rights holders. Personal chat use, avatars, and wallpapers are fine.

How does this page relate to the GitHub repo?

It is an adapted, condensed edition of the repo's README and Skills docs, authorized by the author in issue #2 and shared under the same CC-BY-NC-SA 4.0 terms with prominent attribution. The complete, always-current documentation lives in the repo — this page is an on-ramp, not a replacement, and is synced as the repo evolves.

How is this different from the /refs character sheets?

The character sheets provide ready-made turnarounds to use as img2img / ControlNet references; this tutorial walks the Agent-Skill route of "spec + references + QA." The two stack: install the Skill, then feed the sheets in as identity references.

No coding agent — is there an easier way?

Yes — the community route collected in 2026-10 (AI-Fanworks, by ApprenTice): one original reference sheet plus two long prompts ("style" + "character"), two generation steps in a conversational image model — a Girl illustration, then the Chibi sticker. No agent or scripts involved; see the "community route" section. The Skill route's edge is plan confirmation and three QA gates, and the two stack well.