当所有人都在让 AI 出更复杂、更逼真、更炸裂的图时——试试反过来：只给它三种颜色。 黑线、一抹蓝、一点红。一个物件、一个动词、一句小字。没了。 结果？越少，越像「某个人画的」。因为克制本身就是风格——它在说「我知道可以加更多，但我选择不加」。这种「选择不加」的感觉，就是 AI 图最难伪造的东西：审美判断。 Prompt 的高阶操作不是往里塞更多描述，是知道该删掉什么。 而这 4 张还是同一套模板：风格锁死（奶白纸底 + 手绘黑线框 + 蓝色填充 + 红色手写标签 + 单个日常物件 + 右下角小字注释 + 版画颗粒感），换的只是「画什么物件、配哪个动词」。模板抄走直接用—— { "style_name": "Loose Scribble Riso Print Style", "style_slug": "loose-scribble-riso-print-style", "style_version": "1.0.0", "style_summary": "A sparse handmade riso or screenprint poster style with one large simplified subject, wavering black contour drawing, rough off-white paper, flat blue and coral-red overprint accents, handwritten margin text, and visible print grain.", "environment_variables": { "SUBJECT": "The new main subject or simplified object for the print.", "SUBJECT_ACTION": "The simple pose, placement, or action of the subject.", "PRODUCT_OR_PROP": "A small supporting prop or object detail that is not the original source prop.", "LOCATION": "A minimal setting or surface that grounds the subject.", "BACKGROUND_ELEMENTS": "Sparse background details such as a border, ground line, paper speckles, or one small accent mark.", "MAIN_TEXT": "A short handmade phrase to place near the margin.", "SECONDARY_TEXT": "Tiny supporting handwritten text or edition-like microcopy that does not mimic the source identity marks.", "ACCENT_SYMBOL": "One small decorative mark, scribble, star, arrow, dot, or stamp-like accent.", "WARDROBE_STYLE": "Styling direction for any character or object treatment, kept flat and handmade.", "STYLE_FIDELITY_ANCHORS": "Observable style traits that must remain visible in the generated sample.", "SOURCE_CONTENT_TO_AVOID": "Literal source content that generated samples must not recreate.", "ASPECT_RATIO": "9:16 or 16:9" }, 雨后的伞、深夜的灯、桌上的收音机、一包种子——哪个最像你此刻的状态？