PROMPT RADAR / VIDEO视频叙事
THE AGENT THAT BUILT A FULL ANIMATED AD FROM ONE INSTRUCTION IN SEEDANCE
THE AGENT THAT BUILT A FULL ANIMATED AD FROM ONE INSTRUCTION IN SEEDANCE 2.0 × GPT IMAGE 2.0 One prompt. One AI agent. A full 15-second chase-to-punchline commercial, from character sheets to final cut, with no manual sh
@Primee32 503 热度 8.3k 浏览
编辑先看
它为什么值得进入灵感库
它最有价值的部分,是把一个模糊想法推进到了可以被模型执行的程度。 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。
01 / 原始方法
动作与镜头 Prompt
THE AGENT THAT BUILT A FULL ANIMATED AD FROM ONE INSTRUCTION IN SEEDANCE 2.0 × GPT IMAGE 2.0 One prompt. One AI agent. A full 15-second chase-to-punchline commercial, from character sheets to final cut, with no manual shot-by-shot work. No re-prompting between shots. No building character references one by one. No stitching the sequence together by hand. Most people still treat AI video like a one-shot-at-a-time job — generate a frame, check it, generate the next, repeat until it's done. The real shift happens when you hand the whole pipeline to an agent that turns one instruction into tracked tasks and runs them at the same time. 1. Give the agent the full brief once — the characters you want, the tool stack, how many shots, the format and resolution 2. The agent builds a character sheet for each character in GPT Image 2.0 first — full body views, poses, close-ups, locked before any video gets made 3. The agent plans the shots itself — chase, standoff, reveal, sharing moment, each with its own timing 4. Every shot becomes its own task, generated on its own in Seedance 2.0 5. The agent checks in as each task finishes — accepting what's ready, flagging what's done, running the rest in parallel instead of waiting one by one 6. The last step stitches every finished shot into one sequence, product shot included, ready to post Why this works - Treating each shot as its own task means one bad generation doesn't hold up the rest - Locking the character sheets first is what keeps the officer, the thief, and the product looking the same across five different shots - Running shots in parallel cuts the total time way down - The agent handles the planning and checking, so your job becomes approving, not generating Use cases: ⁃ Branded or product-placement animated shorts ⁃ Multi-character sequences with dialogue and a clear story beat ⁃ Small teams wanting agency-level output without an agency ⁃ Testing a full commercial idea before spending a real budget on it One shot needed a redo because the officer's expression looked off in the standoff. Still one instruction in, one finished ad out — the agent did the part that used to take a whole team.
02 / 复刻路径
不要一次改完所有变量
用 GPT Image 2 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。
- 主体:THE AGENT THAT BUILT A FULL ANIMATED AD FROM ONE INSTRUCTION IN SEEDANCE 2.0 × GPT IMAGE 2.0 One prompt. One AI agent. A full 15-second chase-to-punchline commercial, from character sheets to final cut, with no manual sh
- 视觉方向:实时 AI 视频案例;查看原帖媒体与作者说明
- 镜头与构图:按原帖镜头流程执行;查看原帖媒体与工作流
- 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
GPT Image 2 / Seedance / Veo / Kling。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