PROMPT RADAR / VIDEO视频叙事

Seedance 2.5が凄いのは「本物みたいな動画」を作れる点です。 Seedance 2.0やKling 3.0でも、十分にリアルな動画は作

Seedance 2.5が凄いのは「本物みたいな動画」を作れる点です。 Seedance 2.0やKling 3.0でも、十分にリアルな動画は作れました。 ただ、AI動画を作っている人なら、 「これはAIだな」 とすぐに判別できました。 ところが、Seedance 2.5で作った動画は、 「これ、本物? AI? どっちだろう🤔」 と、AI動画を普段から見ている人でも迷うレベルのものが作れます。 これだけ進化すると、さまざまな分野で動画

@tanabe_fragm 1.2k 热度 10.7 万 浏览
编辑先看

它为什么值得进入灵感库

先别急着照抄整段 Prompt:真正有用的是主体、空间和镜头之间的约束关系。 把一段描述拆成起始状态、动作变化、镜头响应和结束画面,连续性比形容词更重要。

动作与镜头 Prompt

Seedance 2.5が凄いのは「本物みたいな動画」を作れる点です。 Seedance 2.0やKling 3.0でも、十分にリアルな動画は作れました。 ただ、AI動画を作っている人なら、 「これはAIだな」 とすぐに判別できました。 ところが、Seedance 2.5で作った動画は、 「これ、本物? AI? どっちだろう🤔」 と、AI動画を普段から見ている人でも迷うレベルのものが作れます。 これだけ進化すると、さまざまな分野で動画生成AIを活用できます。 とくにAIインフルエンサーやAI彼女・AI彼氏などのバーチャルキャラクターは、今以上に自然でリアルな存在として作りやすくなりそうですね😃 もちろんエンタメだけでなく、広告などのマーケティング分野でも、今以上にAI動画が活用されるようになるでしょう。 動画生成AIの進化は、まだまだ加速するはず。 今のうちから使い方を学んでおいて損をすることないと思います😃 今回の動画はSeedance 2.5で480pで作り(コスト削減のため)、その後Topazでアップスケールしました。 「Seedance 2.5」をうまく発音できなかったところ以外はまずまず😌 動画は3つの画像を参照して、1つのプロンプトで生成しています。 ご覧のようにとても長い(2万文字オーバー)ので、Claude CodeやCodexを使って生成するのがおすすめです。 プロンプト↓ ============================================================ REFERENCES ============================================================ @Image1 = the woman — a character sheet with three views on a plain grey background. Take from it ONLY her identity and her clothing: her face, her bone structure, her eyes and eyelid shape, her nose, her mouth, her jawline, her real skin texture and pores, her medium-length black hair, her green ribbed tank top and her oatmeal linen trousers. She must be recognisably the same person for the whole thirty seconds. She is NOT wearing glasses in @Image1; in this video she wears plain tortoiseshell glasses until 15.0s, as described below. Do not inherit the grey background, the flat studio lighting, the standing pose or the three-view layout. @Image1 is NOT a starting frame. @Image2 = the mascara. Keep its exact shape, proportions, matte black finish and the single fine silver ring. It carries NO brand name, NO logo and NO lettering of any kind — keep it completely unmarked and never add text to it. @Image3 = the lipstick. Keep its exact shape, proportions, matte ivory case, the fine gold band and the rose-pink bullet. It carries NO brand name, NO logo and NO lettering of any kind — keep it completely unmarked and never add text to it. ============================================================ ABOUT THE TIMESTAMPS BELOW ============================================================ The timestamps mark when things occur inside ONE continuous take. They are not shot boundaries, not clip divisions, and not edit points. Nothing changes at a timestamp except what she is doing. The camera keeps rolling straight through every one of them. ============================================================ NO DEAD AIR ============================================================ She is never still or neutral. She never stops, stares, or holds a blank expression between actions. Every movement overlaps the next — she is already beginning the following action before the current one finishes, and she talks through her movements rather than pausing to speak. There is no moment where she sits motionless waiting for a line, and no moment after a line where she holds a still expression. Speech, gesture and activity run continuously and overlap throughout. ============================================================ NO VISIBLE PHONE OR DEVICE ============================================================ The camera is her phone. No phone, screen, handheld device or recording equipment appears anywhere in frame at any point. During the opening she is carrying the camera — her right arm is extended toward the lens and heavily foreshortened at the frame edge, and her hand is out of frame. Her left hand is empty. She is never seen holding a phone while a phone films her. ============================================================ FORMAT ============================================================ VERTICAL 9:16 portrait. Front-facing phone camera held in portrait orientation. Mild wide-angle distortion. 24fps. ============================================================ CAMERA GEOMETRY ============================================================ THE ROOM — PHYSICAL LAYOUT. Build the space from this plan and keep every object in the same place for the whole video. Plan of the room, seen from above: A small rented Japanese one-room apartment, roughly 4.5 by 3 metres. Plain white vinyl wallpaper, light wood-grain laminate flooring, a visible plug socket with a cable running along the skirting board. A tall sliding balcony window fills most of one wall, plain lace curtains drawn across it, the balcony railing faintly visible through them. It is the only light source in the room. The low wooden desk stands against the wall next to that window, at right angles to it, so that a person sitting at the desk has the window on their LEFT. She sits at that desk on a wooden chair, facing the desk and the wall. A small standing mirror sits on the desk directly in front of her. The phone is stood upright in portrait orientation on the desk to her LEFT, propped against a short stack of books, between her and the window, offset roughly 35 degrees from the direction she is facing and angled back toward her and very slightly upward. It is never directly in front of her and it never blocks the mirror. The soft brown leather tote bag
查看原作者内容

不要一次改完所有变量

用 Seedance 复刻时,建议先单独生成稳定首帧,再把动作拆成“开始、变化、收束”三段;每段只保留一个主要运动。

  1. 主体:Seedance 2.5が凄いのは「本物みたいな動画」を作れる点です。 Seedance 2.0やKling 3.0でも、十分にリアルな動画は作れました。 ただ、AI動画を作っている人なら、 「これはAIだな」 とすぐに判別できました。 ところが、Seedance 2.5で作った動画は、 「これ、本物? AI? どっちだろう🤔」 と、AI動画を普段から見ている人でも迷うレベルのものが作れます。 これだけ進化すると、さまざまな分野で動画
  2. 视觉方向:实时 AI 视频案例;查看原帖媒体与作者说明
  3. 镜头与构图:按原帖镜头流程执行;查看原帖媒体与工作流
  4. 交付约束:auto,保留原作者方法与归属;复刻时避免水印、低清、畸变、主体漂移和无关文字。
负向约束

保留原作者方法与归属;复刻时避免水印、低清、畸变、主体漂移和无关文字。

模型建议

Seedance / Kling / Veo。先用原始 Prompt 建立基准结果,再根据模型对自然语言、镜头运动或文字排版的能力做局部改写。

来源与编辑原则

保留作者,也保留判断。

本页收录公开案例并提供中文索引、结构化拆解和复刻建议,不冒充原作者作品,也不替代原帖上下文。

作者:タナベ | AI動画 × マーケティング · 收录于 2026年9月6日

访问原始来源

沿着这条线继续看