想让AI生图看起来像真实时尚摄影，关键不在堆形容词，而在把人物、服装、光线、镜头一次写死。这条GPT Image 2.5的提示词就是个好例子：先传一张人物参考图锁定身份，剩下的交给画面描述。
提示词原文，直接照着用：
Use uploaded image as the character reference. Preserve her facial identity, sculpted mauve spiral updo, beauty marks and crescent forehead detail. Create an intimate fashion photograph of her wearing an oversized oatmeal cashmere sweater and an ivory silk skirt. She sits sideways on a pale oak stool, sleeves loosely covering her wrists, hands relaxed on her lap. Her expression is calm, gaze directed toward a large window. Minimal cream studio with soft morning light and subtle shadow gradients. Three quarter seated composition, 85mm lens at f/2.8. Clearly visible knit fibers, delicate silk folds, natural pores and soft tonal contrast. Understated luxury clothing campaign, real photography, no illustration or text.
这条提示词分四层写。第一层锁人物：用上传的图做人物参考，保留面部特征、淡紫色螺旋发髻、美人痣和额前月牙细节，换人只换参考图，描述不用动。第二层定服装和姿态：超大号燕麦色羊绒毛衣配象牙白真丝长裙，侧坐在浅色橡木凳上，袖口盖住手腕，双手放松放在腿上，表情平静，望向大窗。第三层给场景和镜头：极简奶油色影棚、柔和晨光与微妙的阴影渐变，四分之三坐姿构图，85mm镜头、f/2.8。第四层逼质感：清晰可见的针织纤维、精致的丝绸褶皱、自然毛孔、柔和的色调对比，最后一句定调——低调奢华的服装广告，真实摄影，不要插画、不要文字。
照着改的时候注意两点：参考图和提示词里的人物描述要对得上，参考图换了人，发髻、痣这些细节也要跟着改；结尾那句no illustration or text别删，丢了它很容易出插画风。