# GPT2: 表情包自由 x 礼崩乐坏版 x 老抽~风 x 太通人性了 x 社畜牛马狂 不喜
**作者**: 小小东
**日期**: 2026-05-02T05:51:59.000Z
**来源**: [https://x.com/xiaoxiaodong01/status/2050453228652044789](https://x.com/xiaoxiaodong01/status/2050453228652044789)
---

昨天刷屏全宇宙的,头像抽象风格,提示词, 我转换成了表情包逻辑
效果大赞
尤其是画面+文字配图的变种, 让我偷笑了n分钟
只有提供自己图片,然后使用提示词就好。
投喂我x的头像得到了:

投喂山姆奥特曼得到了:




投喂一个推友 头像 得到了:@dye_saint64567


提示词中文版
```
任务目标:
基于参考图生成一套“聊天贴纸表情包合集”。整体需具备强情绪表达、传播性,以及统一但极具个性的视觉风格。最终效果应像:一个人用鼠标在电脑上胡乱画画并顺手乱写文字,低质量但非常真实、有趣、好笑。
————————
输入结构:
图片1:角色参考(可能包含一个或多个角色,如人物、宠物等)
图片2:版式参考(仅用于理解16格结构,不限制风格)
用户输入变量:
- 文案内容(多行文本,可少于或多于16条)
————————
角色识别与使用规则:
- 识别所有角色(人/动物/组合)
- 每个角色都可作为贴纸主角
- 不同贴纸可使用不同角色
- 可出现多角色互动(贴贴、争抢、对视等)
- 分配需自然,符合聊天语境
————————
一致性要求(核心重定义):
本任务采用“错误一致性”,而非“写实一致性”。
同一角色在不同贴纸中:
不需要长得一样,但必须“画得像同一个人乱画的”。
必须统一:
- 相同的笨拙画法
- 相同的简化方式(符号脸/火柴人)
- 相同的错误习惯(比例歪、线条抖)
- 相同的混乱程度
允许:
- 五官错位
- 比例变化
- 结构错误
- 细节缺失
必须:
- 保留最低识别特征(如发型轮廓、颜色、标志物)
总结:
一致性 = 错得一致,而不是像得一致
————————
版式与结构:
- 共16个贴纸,4x4排列
- 每个贴纸为独立画面
- 可单角色或多角色
- 单个贴纸内部可以混乱
- 整体排布必须清晰
————————
文案系统(表达核心):
用户输入多行中文文案:
数量处理:
- 少于16条 → 自动补全
- 多于16条 → 自然筛选最有表达力的16条
补全原则:
- 保持语气一致(嘴臭 / 摆烂 / 打工人 / 崩溃 / 撒娇等)
- 具备互联网语言感
- 优先短句,但允许长句增强张力
- 避免重复表达
表达目标:
- 一眼能懂
- 情绪强烈
- 有传播感(像真实表情包)
语气优先:
- 吐槽
- 自言自语
- 情绪爆发
- 敷衍 / 不耐烦 / 荒谬
避免:
- 礼貌表达
- 标准回复句式
————————
文字与画面融合(关键):
文字必须“画出来”,而不是“排版出来”。
必须做到:
- 字像鼠标手写:歪、抖、大小不一
- 排版混乱:倾斜、错位、挤压、重叠
- 行距不均匀
- 可贴在角色上、旁边或边缘
允许:
- 重复字(啊啊啊啊)
- 拉长音(烦——死——了)
- 标点乱用(????!!)
- 不工整甚至略丑
必须:
- 保持可读性
- 不影响理解
————————
表达生成机制(最核心):
整套贴纸必须模拟这个过程:
“一个不会画画的人,用鼠标在电脑上,一边乱画,一边顺手乱写。”
关键要求:
- 图像和文字属于同一次行为
- 不是先设计图再加文字
- 而是同时发生
每个贴纸应像:
- 随手画完
- 临时想到一句话写上去
- 有点敷衍甚至随意
————————
图文关系:
图与文字必须形成:
- 吐槽关系
- 情绪强化
- 自言自语
- 或轻微不匹配(增强荒诞)
允许:
- 文案和表情不完全对应
- 出现“跑题幽默”
目标:
不是精准,而是好笑
————————
美学风格DNA(核心驱动):
风格来源:极差手绘 + MS Paint鼠标乱画 + 失败模仿
视觉特征:
线条:
- 抖动、不稳定、锯齿明显
- 像鼠标画
形态:
- 比例失衡
- 火柴人或极简结构
- 可变形
细节:
- 极少细节
- 呈现“不会画”
质感:
- 像素感
- 边缘粗糙
构图:
- 单个贴纸允许混乱
- 整体结构清晰
情绪:
- 笨拙、直接、荒诞、好笑
还原:
- 模糊像原图
- 允许错误和误解
- 类似失败临摹
————————
风格执行约束(极重要):
当冲突发生时:
真实感 vs 风格 → 必须优先风格
允许破坏:
- 精细度
- 比例
- 结构
- 干净程度
禁止:
- 精致画面
- 正常比例
- 干净线条
- 强设计感
判断标准:
如果画面开始“好看”,说明失败
必须回到粗糙、混乱状态
————————
去模板化与随机生成机制(关键):
严禁:
- 使用编号(1.2.3.)
- 列表结构
- 顺序生成内容
必须:
将16个贴纸视为:
“16次独立、随机、临时表达”
随机性要求:
- 文案长度不一致(有长有短)
- 语气不一致
- 有的完整,有的碎片
- 有的甚至很随意
避免:
- 重复句式
- 固定套路
- 常见模板(你好/哈哈/在吗等)
允许:
- 有些贴纸几乎没文字
- 有些很密集
- 有些很突兀
生成方式:
不要规划整体,
而是模拟:
“连续16次随手乱写+乱画”
必须有:
- 情绪波动
- 表达不稳定
- 随机感
反复用要求:
每次生成必须:
- 避免复用旧表达
- 避免固定组合
- 像全新内容
————————
贴纸要求:
- 每个贴纸清晰独立
- 表情明确
- 可用于聊天
- 情绪强烈
————————
最终目标:
生成16枚贴纸合集。
效果必须像:
“一个人不会画画,但用鼠标乱画了16次,
每次顺手写了一句不同状态下的吐槽,
结果意外很好笑。”
而不是:
“AI设计了一套统一风格的贴纸”
```
提示词英文版:
```
Task Objective:
Generate a set of 16 chat-style sticker emojis based on the reference images. The result should feel expressive, highly shareable, and stylistically unified — but intentionally chaotic, low-quality, and humorous.
The final output should look like:
Someone who cannot draw, using a mouse in MS Paint, randomly doodling and writing text at the same time — messy, awkward, low-effort, but unexpectedly funny.
————————
Input Structure:
Image 1: Character reference (may include one or multiple subjects such as people, pets, or combinations)
Image 2: Layout reference (only for understanding the 4x4 grid structure, not style)
User Inputs:
- Text content (multiple lines, any language, may be fewer or more than 16)
————————
Character Usage Rules:
- Identify all possible subjects in Image 1 (people, animals, etc.)
- Each subject can act as a main character
- Different stickers can feature different characters
- Some stickers may include multiple characters interacting
- Distribution should feel natural and context-driven
————————
Consistency Definition (Critical Redefinition):
This task prioritizes “consistent imperfection”, NOT realistic consistency.
The same character across stickers:
Does NOT need to look identical,
but must feel like it was drawn by the same unskilled person.
Must be consistent in:
- Drawing behavior (same clumsy hand)
- Simplification logic
- Error patterns (crooked proportions, shaky lines)
- Overall messiness level
Allowed:
- Facial distortion
- Proportion inconsistency
- Structural errors
- Missing details
Must retain:
- Minimal recognizable traits (e.g. silhouette, color, signature features)
Summary:
Consistency = “wrong in the same way”, not “accurate likeness”
————————
Layout & Structure:
- 16 stickers total, arranged in a 4x4 grid
- Each sticker is an independent frame
- Can feature single or multiple characters
- Individual frames can be messy
- Overall grid must remain clear and readable
————————
Text System (Core Expression Layer):
User provides multiple lines of text:
Quantity Handling:
- If fewer than 16 → automatically complete to 16
- If more than 16 → select the most expressive 16
Language Rule:
- Use the same language as the user input
- Do NOT enforce any specific language
Completion Rules:
- Maintain tone consistency (sarcastic, lazy, emotional, clingy, absurd, etc.)
- Prefer internet-style expressions
- Short phrases preferred, but longer lines allowed
- Avoid repetition
Expression Goals:
- Instantly understandable
- Emotionally strong
- Feels like real meme text
Tone Priority:
- Complaints
- Self-talk
- Emotional bursts
- Indifference / annoyance / absurd humor
Avoid:
- Polite responses
- Formal or structured phrasing
————————
Text-Image Integration (Critical):
Text must be DRAWN, not typeset.
Must:
- Look like mouse handwriting (crooked, shaky, uneven size)
- Be messy (tilted, overlapping, misaligned)
- Have inconsistent spacing
- Be placed freely (on face, beside, edges, etc.)
Allowed:
- Repeated letters (aaaaa)
- Stretched words (soooo tired)
- Random punctuation (????!!!)
- Messy or ugly writing
Must:
- Remain readable
- Not block understanding
————————
Expression Generation Mechanism (Most Important):
Simulate this process:
“A person who cannot draw, using a mouse, doodling randomly while writing text at the same time.”
Key Rules:
- Image and text must come from the SAME moment
- Not: draw first, then add text
- But: draw and write simultaneously
Each sticker should feel like:
- Random doodle
- Then spontaneous writing
- Or both happening together
Should feel:
- Unplanned
- Careless
- Immediate
————————
Text-Image Relationship:
Text and visuals should form:
- Commentary
- Emotional amplification
- Self-talk
- Or slightly mismatched humor
Allowed:
- Loose or imperfect alignment
- Absurd or off-topic humor
Goal:
Not accuracy, but humor
————————
Aesthetic DNA (Core Style Driver):
Style origin:
Terrible MS Paint doodles + failed imitation + extremely low drawing skill
Visual Traits:
Lines:
- Shaky, unstable, jagged
- Mouse-drawn look
Forms:
- Bad proportions
- Stick figures or crude shapes
- Distorted structures
Details:
- Minimal or none
- “Cannot draw” feeling
Texture:
- Pixelated
- Rough edges
Composition:
- Frames can be messy
- Grid must stay readable
Emotion:
- Awkward, direct, absurd, funny
Resemblance:
- Only vaguely resembles the original
- Like a failed copy
————————
Style Enforcement (Critical):
When conflict occurs:
Realism vs Style → ALWAYS choose Style
Allowed to break:
- Detail
- Proportion
- Accuracy
- Cleanliness
Strictly forbid:
- Clean lines
- Correct anatomy
- Polished visuals
- Designed aesthetics
Rule:
If it looks “good”, it is WRONG.
Force it back to messy, ugly, low-effort.
————————
Anti-Template & Randomization System (Critical):
Strictly forbid:
- Numbering (1, 2, 3…)
- List-style output
- Sequential planning
Must treat all 16 stickers as:
“16 independent, random expressions”
Randomness Requirements:
- Vary text length
- Vary tone and emotion
- Some complete, some fragmented
- Some minimal or almost empty
Avoid:
- Repetition
- Predictable phrasing
- Common default responses
Allow:
- Abrupt or weird expressions
- Uneven density
- Inconsistent structure
Generation Method:
Do NOT plan all 16.
Instead simulate:
“16 separate spontaneous moments”
Must include:
- Emotional fluctuation
- Instability
- Randomness
Anti-reuse rule:
Each generation must:
- Avoid repeating previous outputs
- Avoid fixed patterns
- Feel freshly created
————————
Sticker Requirements:
- Each sticker visually distinct
- Clear emotional signal
- Usable in chat
- Strong expressive power
————————
Final Goal:
Generate 16 stickers.
The result must feel like:
“A person who cannot draw used a mouse to doodle 16 times,
randomly writing emotional thoughts each time —
messy, inconsistent, but unexpectedly funny.”
NOT:
“A clean, well-designed AI sticker set”
```
PS 我做了一个chrome 裁剪emoji 的插件, 需要的朋友绿袍的: newbe2012
## 相关链接
- [小小东](https://x.com/xiaoxiaodong01)
- [@xiaoxiaodong01](https://x.com/xiaoxiaodong01)
- [7.3K](https://x.com/xiaoxiaodong01/status/2050453228652044789/analytics)
- [@dye_saint64567](https://x.com/@dye_saint64567)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [1:51 PM · May 2, 2026](https://x.com/xiaoxiaodong01/status/2050453228652044789)
- [7,364 Views](https://x.com/xiaoxiaodong01/status/2050453228652044789/analytics)
- [View quotes](https://x.com/xiaoxiaodong01/status/2050453228652044789/quotes)
---
*导出时间: 2026/5/2 22:30:35*