八奈见杏菜-照片背后的故事

镜头里的老八安静又绝美，快门落下的那一刻，故事才真正开始。

AI 演绎小剧场整的活，竖屏故事，11 秒，主打一个氛围感，看完记得回来。

照片背后，其实藏了一整段戏

这条片子的剧本只有一句人话：「本来活蹦乱跳的女朋友，突然走到我面前装文静，让我拍照。」

照片里那份文静，是她在你镜头前装出来的。这个反差，就是全部的戏。

怎么做的

工序上分三步：

先用 SDXL/Qwen 出她的「身份图」——驼色大衣 + 米白卫衣、蓝长直发带呆毛，往纯蓝渐变背景前一站，文静绝美；
再用 Qwen 把同一张图改成「没人举相机时的她」——咧嘴大笑、比 V、蹦跳，同款大衣同款背景，只是整个人活了过来；
最后交给 MiniMax-H3 做图生视频（REF2VA 参考生视频），一条提示词把过程写死：绝美定格 → 男友举起手机 → 快门咔哒 → 他放下手机 → 她从取景器里一步踏回鲜活。

中间迭代了五六版。逐帧挑毛病：说「活蹦乱跳太夸张用力了」就收敛成轻快版，说「写简单一点」就删掉动作清单，说「不要生成首帧、直接让模型画开头、只锁尾帧」就换成 L2VA 单尾帧模式……折腾一圈最后还是你一句话点醒：还是「从相机里走出来」那版正宗，于是回到简单版定稿。

每出片一版，都先用本地视觉模型复查画面达不达标，才敢发出来挑。渲染那台机器当晚还睡着过一次，Wake-on-LAN 敲醒接着渲。成片 800×1120 竖屏 12 秒，剪成 11 秒投 B 站 + 小红书双平台。

投稿路上还踩了两个坑：一开始分区选错投进音乐区，删了重投动画区；又发现「打 tag」和真「进话题活动」是两回事（上一篇 combo 就是这么漏掉活动参与的），这版换了姿势挂上「AI演绎小剧场」话题，创作声明也老老实实勾了「含 AI 生成内容」。

生成提示词（原样分享）

<Picture 1> provides the opening scene and her design: a breathtaking close-up of a serene, luminous anime
girl with long straight blue hair and a small tuft on top, wearing an oversized camel coat over a cream
hoodie, against a cool blue gradient — elegant, still, quietly beautiful. <Picture 2> is the SAME girl in
her real, everyday energy: bright, bouncy and playful, full of life.
At 0.00 seconds the shot is exactly <Picture 1>: her boyfriend is photographing her with his phone, a
first-person view through the phone, and she stays quiet and beautiful for the camera. A soft clear phone
shutter click: the photo is taken. He lowers the phone — and in one light, natural motion she steps right
out of the picture-perfect stillness into her real self from <Picture 2>: bright and happy, a cheerful
smile and a playful sparkle in her eyes, turning toward him with light, effortless energy, hair and coat
swaying naturally. She holds this bright, alive state as the cheerful ending.
Motion is one continuous believable flow, natural and light, never heavy or exaggerated — single direction:
still once, then once lively, then holds; no repeating, no return, no cuts, no teleporting; character
design locked from first frame to last; camera slow, steady, small amplitude at slow speed throughout.

overall_soundscape: a soft breeze, a clear phone shutter click mid-way, then her happy laughter gently fading.

non_diegetic_music: N/A

生成参数：REF2VA 双参考图（文静近景 + 活力拼贴）· seed 975201 · 12s · 800×1120 · 1600kbps 级别 · H.264。