人物肖像提示词实战:五官、表情与姿态的精准表达
人像是最常见的生成需求,也是崩坏重灾区:手部畸形、五官漂移、表情僵硬。稳定出人像的框架是把人物信息拆成四段,依次填写。
第一段基础特征:性别年龄加发色发型加脸型特点,a young woman with long black hair, soft round face, warm brown eyes。第二段表情:表情词要选具体的,faint smile(浅笑)/ bright smile(开朗笑)/ pensive expression(若有所思)/ serious gaze / laughing with eyes closed,避免只写 beautiful——它不产生任何表情信息。第三段姿态:用日常可信的动作,looking over her shoulder(回眸)/ hands in pockets / holding a coffee cup / leaning on the wall,复杂动作崩坏率飙升,能用静态姿势就不用大幅动作。第四段镜头与光:close-up portrait, 85mm lens, shallow depth of field, soft window light 或 golden hour backlight。
完整示例:a young woman with long black hair and warm brown eyes, faint smile, looking over her shoulder, wearing a cream knit sweater, close-up portrait, 85mm lens, shallow depth of field, soft window light, muted colors, film grain。
三条降崩经验:一,五官描述点到为止,写两三个最强特征即可,特征太多模型会强制全部呈现导致五官挤压;二,手部入画的画面先检查负面词是否带上 bad hands, extra fingers,仍崩就改为手插口袋、手持物体等遮挡式姿势;三,需要系列一致的角色时,把第一段特征描述原样复制到每一张图,只变表情姿态与场景,一致性会明显高于凭记忆重写。

