大多数 AI 影响者内容翻车,就栽在三件事上:人设每条帖子都长得不太一样、照片精致得不像 UGC、贴在图上的文字一团乱码。这三种翻车都不是模型的锅——是 prompt 的锅。
这篇指南给你能直接修好这三个问题的 prompt 公式。可复制粘贴的模板、人设锚点的写法、把人物锁住的具体语句。在 GPT-Image-2 和 Nano Banana Pro 上都能用。
AI 影响者 Prompt 公式
任何高效的 AI 影响者 prompt 都包含七个组件:
- 人设引用 —— 作为参考图传入的锚点肖像
- 人设不变量 —— 锁定身份的那句话("same character, same face, same skin tone, same hair, same eye shape")
- 姿态/动作 —— 用视觉语言描述人物在做什么
- 场景 —— 地点、时间、环境
- 服装 —— 精确的服装描述(或者写"same outfit as reference")
- 光线 + 风格线索 —— 窗边自然光、黄金时刻、刺眼的手机闪光、35mm film grain、随手拍的 iPhone 风格
- 输出规格 —— 比例、图上文字(逐字给出,不要替换)、构图
任何一项漏了,模型都会用自己的默认值补上。对 AI 影响者来说,默认值就是漂移——而漂移会破坏人设。
人设锚点(生成一次,永远引用)
锚点是这个影响者的母版肖像。后续每条帖子都引用它。第一次就要认真生成——它是真理之源。
锚点模板
Studio portrait of a [AGE]-year-old [GENDER], [ETHNICITY/FEATURES], [SKIN TONE], [EYE COLOR + SHAPE], [LIPS / SMILE DETAIL], [HAIR — color, length, texture, styling]. Wearing [BASE OUTFIT — neutral, recognizable]. [LIGHTING — soft natural window light from [DIRECTION]]. [BACKGROUND — neutral so persona dominates]. Slight 35mm film grain. Sharp focus on the eyes. Square 1:1 framing. No text on image.
填写示例
Studio portrait of a 26-year-old woman, mixed-Latina features, warm olive skin, almond-shaped dark brown eyes, full lips with a slight asymmetric smirk, shoulder-length wavy black hair with subtle copper highlights at the tips. Wearing a cream-colored ribbed turtleneck. Soft natural window light from the left, gentle shadow across the right side of the face. Neutral charcoal background. Slight 35mm film grain. Sharp focus on the eyes. Square 1:1 framing. No text on image.
把生成出来的图存好。后续每次生成都把它作为参考传进去。
五个内容模板
1. 生活方式贴(静态 Feed)
最基础的内容类型——咖啡馆抓拍、健身房自拍、早午餐桌面、OOTD。Instagram 用 4:5 portrait,X 用 1:1。
Reference image: persona anchor. Same character, same facial structure, same skin tone, same eye shape, same hair color and length. Now show her in [SETTING + TIME OF DAY], [ACTION/POSE], wearing [OUTFIT — describe in detail]. [LIGHTING — natural / golden hour / overcast / fluorescent]. Casual iPhone-style photo, slight motion blur, [DEVICE-AESTHETIC CUE — front-camera selfie / mirror selfie / friend-taking-photo angle]. 4:5 portrait orientation.
填写后:
Reference image: persona anchor. Same character, same facial structure, same skin tone, same eye shape, same hair color and length. Now show her at a sunlit Brooklyn coffee shop at golden hour, holding a ceramic latte cup with both hands, casual smile, looking slightly off-camera. Wearing an oversized vintage band t-shirt and high-waisted jeans. Warm window light from the left, soft shadows. Casual iPhone-style photo, slight motion blur, friend-taking-photo angle. 4:5 portrait orientation.
2. UGC 视频帧(动画前的静帧)
这张图会喂给 UGC 视频流水线。帧的质量直接决定视频的质量。"乱拍的手机照"那种线索特别关键——太干净,最后的视频一眼就被识破是 AI。
Reference image: persona anchor. Same character, same face, same hair. [POSE — mid-action, dynamic, not posed]. [SETTING]. [HARSH or IMPERFECT LIGHTING — on-camera flash, mixed color temperature, blue refrigerator light, neon sign cast]. iPhone 14-style photo, mild motion blur, slight grain, lens flare from a single bright source. 9:16 vertical orientation.
填写后:
Reference image: persona anchor. Same character, same face, same hair. Mid-laugh, head slightly thrown back, hand reaching toward something off-frame. Standing in a crowded house party hallway. Harsh on-camera flash, dim ambient light behind, blue color cast from a TV in the background, slight lens flare. iPhone 14-style photo, mild motion blur, slight grain. 9:16 vertical orientation.
3. 商单/产品植入贴
人设 + 产品 + 场景。传三个参考图:锚点、产品,可选场景参考。
Reference images: 1) persona anchor, 2) product reference. Same character, same facial structure, same skin tone. Show her [HOLDING / USING / WEARING] the product naturally in [SETTING]. [LIGHTING]. Product clearly visible with [LABEL / LOGO READABLE]. Bold sans-serif caption at bottom reading "[BRAND CTA]" verbatim — no extra characters, no substitutions. Casual lifestyle photo aesthetic. [ASPECT RATIO].
填写后:
Reference images: 1) persona anchor, 2) sneaker product shot. Same character, same facial structure, same skin tone. Show her sitting on a sun-drenched concrete stoop in a city neighborhood, wearing the white-and-cream sneakers from the reference, jeans rolled at the ankle, looking down at her shoes with a small smile. Late afternoon golden light. Sneaker logo on the side panel clearly visible. Bold sans-serif caption at bottom reading "OUT NOW // LINK IN BIO" verbatim — no extra characters, no substitutions. Casual lifestyle photo aesthetic. 4:5 portrait orientation.
4. 品牌图(文字主导贴)
金句图、公告、新品发布、"关注我"置顶——这类帖子文字本身就是视觉。
[BACKGROUND DESCRIPTION — out-of-focus persona / abstract gradient / cityscape]. [PERSONA ANCHOR REFERENCE if persona is in shot]. [LARGE-FORMAT CAPTION — exact copy, font weight, color, placement]. Bold sans-serif caption reading "[EXACT TEXT]" verbatim — no extra characters, no substitutions. [SECONDARY CAPTION if any]. [ASPECT RATIO].
填写后:
Out-of-focus persona looking at the camera, magenta and cyan neon city light cast, deep shadows. Reference image: persona anchor. Same character, same face. Bold geometric sans-serif caption centered in white reading "NEW DROP // FRIDAY 8PM" verbatim — no extra characters, no substitutions. Smaller caption below in neon magenta reading "@personahandle" verbatim. 1:1 square orientation.
5. 多语言本地化贴
同人设、同场景,本地化文案。模型能处理非拉丁字符。
Reference image: persona anchor. Same character, same face, same hair. [SETTING + POSE]. [LIGHTING + AESTHETIC CUE]. Bold sans-serif caption reading "[EXACT TEXT IN TARGET LANGUAGE]" verbatim — no extra characters, no substitutions. [ASPECT RATIO].
填写后(俄语):
Reference image: persona anchor. Same character, same face, same hair. Sitting on a Moscow café terrace at dusk, wrapped in a wool coat, holding an espresso, looking thoughtfully off-camera. Warm string lights overhead, cool blue evening light behind. Casual phone photo aesthetic. Bold sans-serif caption reading "Новый пост в профиле" verbatim — no extra characters, no substitutions. 4:5 portrait orientation.
让人设保持稳定的几条铁律
- 永远把锚点当参考图传进去 —— 纯文字 prompt 哪怕描述得再细,也会漂移
- 每次都把不变量重新写一遍 ——
same character, same facial structure, same skin tone, same eye shape, same hair color and length是承重的那一句 - 服装状态要写清楚 —— 要么"same outfit as reference",要么把新服装完整描述出来;不要留模糊地带
- 系列内容要锁光线 —— 拍"day in the life"系列时,每一帧都要重申光线条件
- 锚点一次到位 —— 锚点糊了,后面所有帧都会跟着糊
让图上文字正确渲染的几条铁律
- 逐字引用要写的字 ——
reading "DROP FRIDAY" verbatim - 指定大小写 —— ALL CAPS、Title Case 还是 lowercase
- 指定字体风格 ——
bold sans-serif、geometric sans、serif italic、handwritten marker - 指定位置 —— 顶部居中、下三分之一、右下
- 加上免责语 ——
verbatim — no extra characters, no substitutions
不加这些,连 GPT-Image-2 都会自己造字。加了,准确率高到可以直接发,不用人工返工。
三个要避开的坑
- 形容词堆砌 —— "stunning, beautiful, professional, gorgeous"是噪音。"Sharp focus on eyes, soft window light, 35mm film grain, slight motion blur"才是信号。
- UGC 用了过度精致的美学 —— "studio lighting, glossy retouch, magazine quality"产出一眼 AI。UGC 需要瑕疵:手机闪光、motion blur、混合色温。
- edit 时忘了带锚点 —— 不重申人设不变量就做像素级编辑,每次都会让脸微漂;十次之后人设已经肉眼可见地变了。
系列内容的迭代策略
锚点定下来之后,做系列内容(一周的内容、一次活动节奏)时,每次只改一个变量:
- 锁住锚点 + 服装,换场景 —— "same persona, same cream turtleneck, now in Tokyo street market / now at home / now in studio"
- 锁住锚点 + 场景,换服装 —— "same persona, same coffee shop, now in oversized denim jacket / now in workout gear"
- 全部锁住,只换文案 —— A/B 测试帖子文案,不重新生成视觉
这种"每次只改一个"的纪律,是你能搭出连贯 feed 而不是 30 张只共用一个 handle 的散图的关键。
FAQ
为什么我的 AI 影响者每条帖子的脸都不太一样?
大概率是没把锚点作为参考图传进去,或者没在 prompt 里重申人设不变量。两个都是必须的。
哪个模型在人设一致性上做得最好?
GPT-Image-2 配合"锚点 + 引用"的 workflow,在单人设内容上很可靠。Nano Banana Pro 在多角色场景上更强。详见 GPT-Image-2 vs Nano Banana Pro。
AI 影响者的 prompt 应该写多长?
60–120 个英文词是甜点区——足够锁住人设和场景,又不会长到模型丢细节。
这些都得我自己手写吗?
不需要。OmniGems AI 自动处理"锚点 + 引用"的 workflow——创作者只需要把人设简介写一次,平台后续每条帖子都会自动构造结构化 prompt。
这些模板能用在非影响者类的 AI 内容上吗?
可以——"锚点 + 不变量"这套模式适用于任何角色一致性场景(NPC、品牌吉祥物、连载角色)。模板风格偏影响者,但结构是通用的。
内容要连贯
AI 影响者内容能涨粉还是被取关,差别就在一致性上。同样的脸,同样的氛围,跨上百条帖子。模板和"锚点 + 引用"的纪律是你抵达那里的方法。
或者你可以完全跳过 prompt engineering,直接用 OmniGems AI Studio——把人设简介写一遍,连贯内容自动产出。