怎麼填變量：穿搭寫 3–4 件單品即可；背景可以換成"晴天午後的老街拐角""夜晚便利店門口"，但保留"路口 + 凸面鏡"這一核心。參考照最好是清晰的正臉或微側臉。

常見問題：

• 人物直接站在街上、看不出是鏡子裏：補一句"整個畫面就是鏡面裏的倒影，鏡子的紅色邊框在畫面邊緣清晰可見"。
• 臉被手機全擋住：改成"手機只擋住下巴"。
• 太精緻像海報：負面提示詞裏的"影棚光、乾淨完美的鏡子"一定保留。

迭代：換背景比換人物更穩，先固定人物，再逐個嘗試雨天、雪天、黃昏。

英文原版

text
Upload one portrait as the identity reference. Use it only for the subject’s facial identity, face shape, hairstyle, hair color, skin tone, and natural body proportions. Do not copy the original portrait’s clothes, background, pose, lighting, or image quality.

Create a hyper-realistic outdoor convex traffic mirror selfie, 9:16 vertical composition. The subject stands in front of a large round roadside safety mirror with a red rim, taking a flash selfie with a decorated phone case, phone partially covering the face. Use the uploaded portrait only for identity and hairstyle. Keep a casual youthful outfit: dark oversized zip hoodie, cropped white graphic tee, loose jeans, white sneakers. Rainy dusk street intersection background, wet asphalt, zebra crossing, trees, soft blue-gray sky, faint streetlights, one blurred passing scooter, dirty mirror surface with dust and water spots, strong camera flash glare, realistic fisheye distortion, candid smartphone snapshot, natural skin texture, real-life street photography, no watermark.

Negative Prompt:
anime, cartoon, CGI, 3D render, doll face, plastic skin, over-smoothed skin, fake eyes, bad anatomy, distorted hands, extra fingers, broken phone, duplicate person, messy face, unreadable facial features, overexposed face, low resolution, heavy beauty filter, studio lighting, clean perfect mirror, text, logo, watermark, UI elements, poster design, artificial background

改編自 @Shinning1010 發佈、EvoLinkAI/awesome-gpt-image-2-API-and-Prompts 收錄的提示詞，倉庫許可證 CC0 1.0；示例圖同樣來自該倉庫收錄的原帖出圖。