很多人用 AI 画写实人像,最后出来的效果总差点意思。要么皮肤像打了十层蜡,光溜溜却毫无质感;要么眼神空泛,看久了甚至有点 “恐怖谷” 那味儿。其实 AI 绘画没那么玄乎,尤其是写实人像,皮肤质感和眼神的刻画,80% 靠关键词精准度,20% 靠对光影逻辑的理解。今天就掰开揉碎了说,怎么用关键词让 AI 画出能 “以假乱真” 的写实人像。
✨ 先搞懂:AI 是怎么 “理解” 写实人像的?
别觉得 AI 多聪明,它本质上是个 “关键词翻译机”。你喂给它的描述越具体,它从数据库里调取的参考素材就越精准。但写实人像的难点在于,人类对 “真实” 的敏感度极高 —— 我们每天都在看真人,皮肤的细微纹理、眼神的微妙变化,哪怕差一点点,大脑都会立刻判定 “假”。
比如皮肤,真实的皮肤不是 “光滑” 这么简单。它有毛孔的凹凸,有皮脂腺分泌的淡淡油脂光泽,不同部位的厚度还不一样(脸颊薄、额头厚),光线照上去会有不同的透光感。AI 默认的 “smooth skin” 往往是均匀磨皮后的效果,这就是为什么很多人画出来像硅胶娃娃。
眼神更复杂。真实的眼睛有多层结构:角膜的反光、虹膜的纤维纹理、瞳孔的收缩状态,甚至眼白里若隐若现的红血丝,都是 “真实感” 的组成部分。AI 如果只收到 “big eyes” 这样的指令,画出来大概率是空洞的 “卡通眼”,和写实压根不沾边。
所以第一步,得扔掉那些笼统的形容词,换成 “解剖学级” 的细节描述。你对人体结构了解得越多,关键词就越能戳中 AI 的 “数据库痛点”。
🧴 皮肤质感关键词:从 “假滑” 到 “能摸到温度”
皮肤是写实人像的 “门面”,也是最容易露馅的地方。直接说 “真实的皮肤” 等于没说,得拆成几个维度来喂关键词。
先定基调:肤质和瑕疵。千万别写 “perfect skin”,这是大忌。真实的皮肤永远有小瑕疵 ——“skin with visible pores on the nose and cheeks”(鼻子和脸颊有可见毛孔)、“subtle freckles across the nose bridge”(鼻梁有淡淡的雀斑),甚至 “a tiny mole near the left jawline”(左下颌有个小痣),这些 “不完美” 反而让皮肤活起来。油性皮肤可以加 “oily sheen on T-zone under natural light”(自然光下 T 区有油光),干性皮肤则用 “slight dry patches on the forehead”(额头有轻微干纹)。
再讲光影:次表面散射是灵魂。这词听着专业,其实就是光线穿透皮肤的效果 —— 比如耳朵、鼻尖这些薄皮肤部位,会透出淡淡的红色。关键词里必须加上 “subsurface scattering effect”,后面最好补一句 “warm light penetration through earlobes and knuckles”(暖光透过耳垂和指节)。光线角度不同,质感也天差地别:侧光用 “side light casting soft shadows on the cheek, highlighting skin texture”(侧光在脸颊投下柔影,突出皮肤纹理);顶光则是 “overhead light creating subtle hollows under the eyes, enhancing skin depth”(顶光在眼下形成浅窝,增加皮肤层次感)。
最后加细节:肤色过渡和毛发。脖子和脸的肤色很少完全一致,“gradual color transition from neck to jawline, slightly warmer on the neck”(脖子到下颌肤色渐变,脖子稍暖)会更真实。还有那些细毛,“vellus hair on the cheeks catching light, visible at close range”(脸颊的绒毛反光,近距离可见),加这句,皮肤瞬间就有了 “呼吸感”。
试一组对比:只用 “realistic skin” 生成的,大概率是均匀的 “假面”;换成 “skin with natural pores, subsurface scattering, subtle freckles, vellus hair catching morning light”,出来的皮肤能看到细微的凹凸,光线照上去甚至有半透明的质感。
👁️ 眼神关键词:让 AI 画出 “有故事的眼”
眼神是人像的 “灵魂”,但 AI 最容易把眼神画得 “空洞”。关键在于别只描述 “眼睛大”“有神”,要拆解到眼球的每一层结构。
先抓虹膜:纹理决定真实度。虹膜不是一块纯色,里面有复杂的纤维和斑点。关键词得具体到 “iris with detailed radial fibers, golden flecks scattered near the pupil”(虹膜有清晰的放射状纤维,瞳孔附近散布金色斑点)。不同人种的虹膜质感也不同:亚洲人可以加 “dark brown iris with subtle amber undertones, dense fiber pattern”(深棕虹膜带淡琥珀底色,纤维密集);欧美人则是 “blue iris with thin, light-colored fibers, visible stromal patterns”(蓝色虹膜纤维细且浅,可见基质纹理)。
再控高光:位置和数量是关键。眼神光太多会像塑料,太少又显无神。“single catch light in the upper left of the pupil, soft and slightly blurred”(瞳孔左上方单一点光源,柔和微糊)是最自然的 —— 就像人在室内看窗户,光线只会从一个方向来。高光形状也要跟着光源走:窗户光用 “rectangular catch light matching window shape”(矩形高光呼应窗户形状);灯光则是 “circular catch light from ceiling lamp”(天花板灯的圆形高光)。
别忘了细节:瞳孔、眼白和眼睑。瞳孔会随光线收缩,“pupil slightly dilated in dim light, sharp edge with subtle irregularities”(暗光下瞳孔微扩,边缘清晰带细微不规则)更真实。眼白也不是全白,“sclera with faint red veins near the corners, slight yellowish tint at the edges”(眼白角落有淡红血丝,边缘带浅黄调)才对。眼睑的褶皱和睫毛也很重要,“upper eyelid with natural crease, casting soft shadow on the eyeball”(上眼睑自然褶皱,在眼球投下柔影)、“eyelashes with varying lengths, individual strands visible, some crossing slightly”(睫毛长短不一,单根可见,有些轻微交叉),这些细节能让眼睛瞬间 “立体” 起来。
试过一个案例:用 “eyes looking forward, realistic” 生成的眼神呆滞;换成 “iris with radial fibers, golden flecks, single rectangular catch light, sclera with faint red veins, eyelashes with varying lengths”,出来的眼睛不仅有焦点,还能看出人物像是在看窗外,有了 “故事感”。
🛠️ 工具不同?关键词得跟着变
不是所有 AI 绘画工具都吃一套关键词。Midjourney、Stable Diffusion、DALL・E 这几个主流工具,对关键词的 “敏感度” 差得很远。
Midjourney V6 更吃 “场景化描述”。它擅长捕捉氛围,所以描述皮肤和眼神时,最好带上环境。比如皮肤:“skin with natural pores, lit by morning sunlight through sheer curtains, subsurface scattering making the cheeks glow softly”(皮肤有自然毛孔,晨光透过薄窗帘,次表面散射让脸颊微微发光);眼神:“eyes focused on a book, iris with amber flecks, catch light matching the lamp on the desk”(眼睛盯着书,虹膜带琥珀斑点,高光和桌上台灯呼应)。它不喜欢太生硬的技术词,用 “glow softly” 比 “subsurface scattering effect” 效果更好。
Stable Diffusion(尤其是 RealVisXL 模型)则要 “精准到苛刻”。它对细节描述更敏感,关键词可以加参数。皮肤可以写 “skin texture, 8k, ultra-detailed, pores:1.2, subsurface scattering:1.1, vellus hair:1.0”(皮肤纹理,8k,超细节,毛孔强度 1.2,次表面散射 1.1,绒毛 1.0);眼神则是 “eye detail, iris fibers:1.3, catch light:1.0, red veins:0.8, 8k render”(眼部细节,虹膜纤维 1.3,高光 1.0,红血丝 0.8,8k 渲染)。加这些数值,能让模型更聚焦你想突出的部分。
DALL・E 3 适合 “平衡描述”。它介于两者之间,太技术的词会懵,太笼统又出不来效果。可以试试 “skin with visible pores and a little oil on the nose, light from the window making the ears look slightly red; eyes with small lines in the colored part, a tiny white spot from the lamp, and thin eyelashes”(皮肤有可见毛孔,鼻子带点油光,窗户光让耳朵微红;眼睛的有色部分有细纹,带一点台灯的白点,睫毛细细的)。用简单的话把关键细节说清楚,反而更出效果。
🚫 避坑指南:这些关键词组合千万别用
踩过的坑多了,就知道哪些关键词是 “雷区”。避开这些,至少能少走 80% 的弯路。
别用 “完美” 相关的词。“flawless skin”“perfect eyes” 会让 AI 直接调取 “理想化” 素材,结果就是皮肤像打了十层粉底,眼神像玩偶一样规整。换成 “skin with natural imperfections”“eyes with subtle irregularities”,真实度立刻提升。
别堆砌太多 “光效词”。“glowing skin, shiny eyes, bright light, reflection” 堆在一起,皮肤会像涂了油,眼睛像装了灯泡。光线描述要 “专一”,定好一个光源方向,再细化效果。
眼神别只说 “looking at camera”。这样 AI 容易画成 “瞪镜头”,很僵硬。换成 “looking slightly to the left of the camera, eyes softening as if seeing something familiar”(视线略偏镜头左侧,眼神柔和像看到熟悉事物),瞬间有了互动感。
皮肤别忽略 “动态感”。人不会一动不动,“skin stretched slightly at the corner of the smile, pores compressed near the cheekbone”(笑时眼角皮肤微拉,颧骨附近毛孔受压)这种动态细节,能让静态人像有 “瞬间感”。
🎭 实战案例:3 组关键词生成不同风格写实人像
光说不练假把式,放三个实战案例,看看关键词怎么影响最终效果。
案例 1:自然光下的少女(通透感皮肤)
关键词:“Portrait of a 20-year-old girl, morning light through window falling on left side of face, skin with visible pores and subtle freckles, subsurface scattering making earlobes translucent, vellus hair catching light; eyes looking down at hands, iris with brown fibers and golden flecks, single small catch light from window, sclera with faint red veins, eyelashes with varying lengths. 8k, realistic texture, soft shadow.”
效果:少女左侧脸颊被晨光照亮,皮肤能看到细密的毛孔,鼻梁的雀斑若隐若现,耳朵透着淡淡的粉。低头时,眼神向下,瞳孔里有个小而清晰的窗形高光,眼白角落有极淡的红血丝,整体透着温柔的灵气。
关键词:“Portrait of a 20-year-old girl, morning light through window falling on left side of face, skin with visible pores and subtle freckles, subsurface scattering making earlobes translucent, vellus hair catching light; eyes looking down at hands, iris with brown fibers and golden flecks, single small catch light from window, sclera with faint red veins, eyelashes with varying lengths. 8k, realistic texture, soft shadow.”
效果:少女左侧脸颊被晨光照亮,皮肤能看到细密的毛孔,鼻梁的雀斑若隐若现,耳朵透着淡淡的粉。低头时,眼神向下,瞳孔里有个小而清晰的窗形高光,眼白角落有极淡的红血丝,整体透着温柔的灵气。
案例 2:室内光下的男性(硬朗质感皮肤)
关键词:“Portrait of a 35-year-old man, warm desk lamp lighting right side of face, skin with rough texture on jawline, slight stubble shadow, T-zone with oily sheen; eyes looking forward, slightly squinted, iris with dark brown dense fibers, catch light matching lamp shape, upper eyelid fold casting shadow, eyelashes short and thick. Ultra-detailed, 8k, sharp focus.”
效果:男性右侧脸被台灯照亮,下颌线皮肤粗糙,带着淡淡的胡茬阴影,额头和鼻子有自然的油光。直视镜头时眼睛微眯,虹膜纤维密集,高光呈圆形(呼应台灯),上眼睑的褶皱在眼球投下阴影,眼神锐利又带点疲惫,很有故事感。
关键词:“Portrait of a 35-year-old man, warm desk lamp lighting right side of face, skin with rough texture on jawline, slight stubble shadow, T-zone with oily sheen; eyes looking forward, slightly squinted, iris with dark brown dense fibers, catch light matching lamp shape, upper eyelid fold casting shadow, eyelashes short and thick. Ultra-detailed, 8k, sharp focus.”
效果:男性右侧脸被台灯照亮,下颌线皮肤粗糙,带着淡淡的胡茬阴影,额头和鼻子有自然的油光。直视镜头时眼睛微眯,虹膜纤维密集,高光呈圆形(呼应台灯),上眼睑的褶皱在眼球投下阴影,眼神锐利又带点疲惫,很有故事感。
案例 3:逆光下的老人(岁月感皮肤)
关键词:“Portrait of a 70-year-old woman, backlight from sunset, skin with wrinkles around eyes and mouth, deep pores on cheeks, subsurface scattering making the edge of face glow; eyes half-closed, looking at the sunset, iris with grayish-blue fibers, faded color, catch light dim and soft, sclera with more visible veins, eyelashes thin and sparse. Realistic, 8k, natural shadow.”
效果:夕阳逆光下,老人的脸边缘泛着暖光,眼角和嘴角的皱纹清晰,脸颊毛孔深陷。眼睛半闭着看向夕阳,虹膜呈灰蓝色,颜色有些淡化,高光微弱柔和,眼白的血丝更明显,睫毛稀疏,满满都是岁月沉淀的温柔。
关键词:“Portrait of a 70-year-old woman, backlight from sunset, skin with wrinkles around eyes and mouth, deep pores on cheeks, subsurface scattering making the edge of face glow; eyes half-closed, looking at the sunset, iris with grayish-blue fibers, faded color, catch light dim and soft, sclera with more visible veins, eyelashes thin and sparse. Realistic, 8k, natural shadow.”
效果:夕阳逆光下,老人的脸边缘泛着暖光,眼角和嘴角的皱纹清晰,脸颊毛孔深陷。眼睛半闭着看向夕阳,虹膜呈灰蓝色,颜色有些淡化,高光微弱柔和,眼白的血丝更明显,睫毛稀疏,满满都是岁月沉淀的温柔。
看出来了吧?AI 绘画画写实人像,真不是靠 “运气”。皮肤要拆成毛孔、光影、瑕疵,眼神要细化到虹膜、高光、睫毛,再根据工具调整描述方式 —— 做到这些,你生成的写实人像,大概率能让别人问一句 “这是真人照片吧?”
【该文章由diwuai.com第五 ai 创作,第五 AI - 高质量公众号、头条号等自媒体文章创作平台 | 降 AI 味 + AI 检测 + 全网热搜爆文库
🔗立即免费注册 开始体验工具箱 - 朱雀 AI 味降低到 0%- 降 AI 去 AI 味】
🔗立即免费注册 开始体验工具箱 - 朱雀 AI 味降低到 0%- 降 AI 去 AI 味】