举个真实案例：某国产护肤品牌想做一个"精华液在冰水中的清爽感"主视觉。初版提示词只写了"serum in ice water"，结果生成出来像泡面里的油包。后来我们改成："A clear glass serum bottle submerged in crystal-clear ice water, with small ice cubes floating around, splashes frozen in mid-air, bright aqua blue color palette, studio lighting, commercial photography style, ultra-sharp focus, --no label, no cap, no background text"，出来的效果直接让甲方当场拍板。记住，电商图的核心是"可读性+质感"，别让AI自由发挥。
比如你要设计一个"蒸汽朋克风格的女性机械师"，别只写"steampunk female mechanic"。你得拆解：发型（asymmetrical bob with copper wire braids）、服装（oil-stained leather apron over a corset, brass goggles on forehead）、配饰（spanner hanging from belt, mechanical spider pet on shoulder）、表情（confident smirk, one eye is a glowing blue prosthetic）。这样AI才能给你画出有记忆点的角色。
还有一个容易翻车的点：跨种族/跨文化角色。你要是写"African warrior queen"，AI大概率画成欧美审美里的"黑皮肤超模"。正确做法是加入文化指征词：比如"inspired by the Benin Kingdom royal court, intricate coral bead jewelry, woven raffia textiles, scarification patterns on cheeks"。这样出来的东西才不悬浮。
我常用的技巧是：在提示词里同时写视角和镜头（one-point perspective, wide-angle lens, 24mm focal length）以及材质清单（polished concrete floor, white oak veneer wall panels, brass metal accents, linen upholstery）。最重要的是，一定要加上"architectural digest style"或者"interior design magazine photo"这类风格锚点。
给大家看个翻车案例：一个设计师想表现"侘寂风客厅"，只写了"wabi-sabi living room"。结果AI画了个毛坯房，连个坐的地方都没有。后来我帮他改成："Minimalist wabi-sabi living room with low wooden platform sofa, hand-thrown ceramic vase with dried pampas grass, rough limewash walls with natural uneven texture, soft diffused afternoon light casting long shadows, tatami mat flooring, no decoration on walls, calm meditative atmosphere, photographed with Hasselblad camera, medium format look"——这才对了。记住，侘寂不等于简陋，是"精致的朴素"，提示词里要有质感细节。
比如你做咖啡品牌宣传，别写"a cup of coffee on a table"，试试："A ceramic coffee cup with a small chip on the rim, sitting on a rustic wooden table with visible grain and a few coffee stains, steam swirling in the air, a crumpled newspaper and a pair of reading glasses nearby, warm cozy lighting from a nearby window, candid lifestyle photography, film grain added"。你看，加入了"残缺"和"生活痕迹"，画面立刻有了故事感。
我的Midjourney最佳提示词策略是：明确受众年龄+风格类型+简化程度。比如："Children's book illustration style, a friendly cartoon heart character with a smiling face, showing the main parts (atrium, ventricle, aorta) with simple labels, bright primary colors, clean white background, educational poster design, rounded shapes, no realistic anatomy, no text (since we will add Chinese labels later)"。你看，把"不要什么"说清楚，AI就不容易跑偏。
另一个坑是版权问题。很多机构以为AI生成的图就百分百安全，其实Midjourney的训练数据里可能有受版权保护的风格。我的建议是：尽量选择"generic"风格，或者用"in the style of [年代/艺术运动]"，比如"1930s Japanese children's book illustration style"，而不是"in the style of [某个在世艺术家]"。这不仅是法律风险，也是职业道德问题。
