导演式调教 GPT Image 2 人像:镜头、布光与「去 AI 脸」打法
模型的指令遵循足够好,能听你调度。把人像写成分层的拍摄 brief——主体、镜头、布光、构图——并用「视觉事实」代替夸奖词,摆脱塑料 AI 脸。
像写拍摄 brief 一样,分层写人像 prompt
GPT Image 2 在拍人这件事上的强项是高指令遵循:你的 prompt 越像一份真实的拍摄 brief,它执行得就越忠实。所以别再写一长串句子了,改成分层写、用换行分隔——模型对结构化输入的遵循更可靠。
Subject: a 38-year-old man, short dark hair, light stubble, calm confident expression, charcoal wool blazer.
Camera: 85mm lens, f/2.8, eye-level medium close-up, shallow depth of field.
Lighting: soft key from camera-left at 45 degrees, soft fill from the right, subtle rim light, slight catchlights.
Composition: light grey seamless backdrop, softly out of focus.
Constraints: natural skin texture, unretouched, no beauty filter, no plastic skin.最重要的习惯,是直接从摄影师那里偷来的:用「视觉事实」代替夸奖。「Beautiful、gorgeous、ultra-realistic(漂亮、惊艳、超写实)」会把模型拉向它那张过度修过的平均脸,给你一张塑料脸。「Window light from camera-left, 50mm, visible pores(机位左侧的窗光、50mm、可见毛孔)」则把它拉向真实照片。去描述相机会看到什么,而不是你有多喜欢它。
镜头与景深:暗示「真实」的相机语言
一个焦段就是一条单词级的指令,却承载着一整套观感,因为模型是从真实照片里学会每一个焦段的:
- 85mm:经典的头肩像,面部压缩讨喜,背景柔和分离。
- 50mm:自然、亲近的中景,带纪实感。
- 35mm film photograph(35mm 胶片照片):环境人像,街拍风格美学。
- Medium-format portrait(中画幅人像):高细节、商业级的渲染。
把镜头与一个光圈(f/1.4 到 f/2.8)和「shallow depth of field(浅景深)」搭配,让主体与背景分离。一条有记录在案的注意事项:堆砌技术参数会被「宽松地解读」。挑一个焦段、一个光圈、一套布光配方——而不是十个参数。

值得背下来的布光配方
命名好的布光样式,每一个都会触发模型早已理解的一整套灯位:
- Rembrandt(伦勃朗光):「distinctive triangle of light on the shadowed cheek, deep shadows on one side(阴影侧脸颊上有一块标志性的三角光区,一侧深阴影)」——经典、戏剧化,非常适合男性人像。
- Butterfly 或 Paramount(蝶光 / 派拉蒙光):「light from above and in front, small butterfly shadow under the nose, even cheeks(来自上方且偏前的光、鼻下一小块蝶形阴影、两颊均匀)」——讨喜,适合美妆与魅力风。
- Rim 或 backlight(边缘光 / 逆光):「bright outline of light around head and shoulders, hair glowing, dark background(头肩周围一圈明亮的光轮廓、发丝发光、深色背景)」——电影感的分离。
- Split 或 chiaroscuro(分割光 / 明暗对照法):「hard key from one side, half the face in shadow(一侧硬主光、半张脸没入阴影)」——张力最大。
然后加上那个能让死气沉沉的 AI 眼睛活过来的细节:「slight catchlights in both eyes(双眼中轻微的眼神光)」。虹膜里的一个反射光点,是区分「被导演过的人像」与「无生气的渲染图」最快的方式之一。

「去 AI 脸」打法
那张塑料感、蜡质感、对称得可疑的脸,是模型在向它那张修过的人像平均脸默认回退。你要从两条战线打败它。第一,删掉触发词——「beautiful、perfect skin、flawless、ethereal、ultra-realistic(漂亮、完美皮肤、无瑕、空灵、超写实)」——这些会把结果拉向滤镜化的均值。第二,开出真实皮肤的物理处方:
natural skin texture, visible pores, fine wrinkles, slight freckles, mild asymmetry, unretouched. No beauty filter, no plastic skin, no over-smoothing, no AI glow.强力的人像 prompt 都用同样的方式锚定真实感,用上「weathered skin with visible wrinkles, pores and sun texture(带可见皱纹、毛孔和晒痕的风霜皮肤)」这类短语。口诀很简单:不完美即真实。一点不对称、几处瑕疵会被读成照片;完美对称则读成渲染图。
三个可直接粘贴的人像模板
影棚商业证件照:
Photorealistic studio headshot of a 38-year-old man, short dark hair, light stubble, confident calm expression, charcoal wool blazer over a white shirt. 85mm lens at f/2.8, eye-level medium close-up, shallow depth of field, light grey seamless backdrop softly out of focus. Soft key light from camera-left at 45 degrees (loop lighting), large fill from the right, subtle rim light, slight catchlights in both eyes. Natural skin texture with visible pores and fine lines, unretouched, neutral color balance. No beauty filter, no plastic skin, no over-smoothing.户外自然光人像:
Candid photorealistic portrait of a 27-year-old woman with wavy auburn hair, light freckles, relaxed genuine half-smile, cream linen shirt, walking through a wildflower meadow. 50mm lens at f/1.8, three-quarter turn toward camera, eye-level medium shot, shallow depth of field with soft golden bokeh. Golden-hour backlight creating a warm rim light on her hair, soft reflector fill, gentle catchlights. Realistic skin texture, visible pores, slight asymmetry, subtle film grain, honest and unposed. No glamorization, no heavy retouching, no waxy skin.电影感情绪人像:
Cinematic photorealistic portrait of a weathered 55-year-old fisherman, deep facial lines, grey beard, contemplative gaze just off-lens, worn navy raincoat. 85mm lens at f/2.0, tight head-and-shoulders framing, eye level, dark moody harbor at dusk blurred behind. Chiaroscuro: single hard key from camera-right carving deep shadow across half the face (split lighting), cool rim light outlining the shoulders, strong catchlights. Teal-and-amber color grade, subtle film grain, weathered skin with visible pores and sun texture, every wrinkle visible. No stylization, no plastic skin, no symmetry, no AI glow.坑点:手、漂移,以及保持同一张脸
三种要提前规避的失败模式:
- 手画崩:构图取到头肩像以避开手;如果必须露手,加上「hands relaxed and anatomically correct, five fingers(手放松、解剖正确、五根手指)」。
- 参数过载:太多参数会被忽略——每个 prompt 保持一个镜头、一个光圈、一套布光配方。
- 脸在多次修改间漂移:GPT Image 2 能从参考图里守住一个身份,但你必须每一轮都重复这道锁定。
要在多张图里保持同一个人,上传一张参考图,按角色给每张输入打标签(「Image 1: base scene, Image 2: face reference(图 1:基础场景,图 2:人脸参考)」),并在每次迭代时重申不变量:「preserve face, facial features, skin tone, hair and proportions exactly(精确保留脸、面部特征、肤色、头发和比例)」。这套身份锁定工作流在我们关于角色与产品一致性的指南里有专门的深入讲解。
本站的 GPT Image 2 免费起步——注册送 20 积分,无需绑卡。去 gpt-img2.com 拍下你的第一批人像,并浏览人像 prompt。积分定价透明,远低于官方 API 的每张图价格。
Try these GPT Image 2 prompts





