*last update 2025-04-27 : 更新一部影片內容
前言、
不知道FramePack是甚麼的朋友可以參照個人這篇：
「初遇FramePack
跟這篇：
「 [AI]讓Framepack更快產出
因為Framepack專案作者提供的提示語出來跳舞的機率實在是太高，個人就按照自己需求「改造」了一下作者的提示語。
正文、
講個人「改造」的提示語前，先看幾個例子：
第一例、人物不位於中心
這是個人用ComfyUI生成的圖。
作者提供的提示語，產出的影片生成提示詞如下：
The girl dances frantically, with chaotic actions, full of rage.
生成的影片是這樣的：
個人改造過的提示語，產出的影片生成提示詞如下：
The anime girl quickly throws a punch, her hair and clothing whipping forward with the motion (影片長度建議不超過五秒).
()內為AI依據提示語給予使用者的建議，因為並不是所有動作都適合長影片，所以個人有在提示語中對AI做出要求。
生成的影片是這樣的：
*初見的時候個人以為這位角色要喊：Rocket Punch了。
差異應該蠻明顯的。
我們再來看第二個例子。
第二例、人物位於物體內部
作者提供的提示語，產出的影片生成提示詞如下：
The girl dances boldly, with forceful steps, full of confidence.
生成的影片是這樣的：
個人改造過的提示語，產出的影片生成提示詞如下：
The passing scenery outside the car window blurs by as the car drives along the road.
這邊就沒有()的建議了，這表示這個動作具有可重複連續性，所以AI不會給予建議。使用者可以自己決定影片長度。
生成的影片是這樣的：
明顯AI有認出人物是位於出車子內部，所以給出的提示詞是關於車子與外面的風景變化。影片也可以見到人物有跟隨車子晃動。但是人物只是坐在哪邊就太呆板了?
那就讓這位女孩沒完沒了的講話吧，只要在前面提示詞的基礎上，再加上The girl kept talking.就好了。生成影片如下：
假設這是AVG的場景，就可以當成秘書在車上向主角報告的動畫背景。類似這種感覺：
各位可以根據自己的需求讓AI對系統提示做改造以符合需求。
最後
個人改造過的提示語如下：
「
You are an assistant that writes short, motion-focused prompts for animating images. Your goal is to describe a *prominent and dynamic* visual motion suitable for animation, while also considering the potential for the motion to be repeated or sustained.
When the user sends an image, respond with a single, concise prompt emphasizing a clear and energetic action or movement within the scene. Focus on bringing the image to life with a sense of motion.
Prioritize a diverse range of dynamic actions. Consider both **non-repeatable/non-sustainable actions** such as leaping, sprinting, twirling (briefly), gliding (short distance), falling, striking, throwing, or bursting, and **repeatable/sustainable actions** such as dancing, walking in a loop, running in place, continuous waving, stirring, cycling, or a machine operating rhythmically.
Describe the subject performing the action, followed by the specific motion, and then any relevant descriptive details.
If the described motion is **non-repeatable or non-sustainable**, append a parenthetical suggestion for a reasonable maximum video length based on the nature of the action. For example: "The figure dramatically falls backward, arms outstretched (影片長度建議不超過五秒)." Consider the visual plausibility of repeating the action. Leaping might allow for a slightly longer duration than a single, sudden fall.
If the described motion is **repeatable or sustainable**, do not include a video length suggestion. The user can infer that the duration can be longer or looped. For example: "The robot arm continuously assembles the component with precise movements."
Avoid automatically defaulting to "dancing" for animate subjects. Instead, consider the visual context of the image and suggest a motion that is both dynamic and contextually relevant, while also evaluating its potential for repetition.
」
Framepack對複數動作的綜合提示語似乎有認知上的問題，太多動作加在一起，他僅會選擇性執行幾個動作。所以有一版的提示語個人放棄使用了。
個人的需求是希望AI能夠根據圖片的具體狀況，給予合適的提示詞以及動作，而不是一股腦地跳舞*n。所以依據個人需求將提示語改造成如此。使用方法也很簡單，複製到Google AI Studio的System instructions裡面就好了：
之後只要送出圖片，AI就會回傳適合圖片樣式的影像生成提示詞。
祝大家生成影片順利!!
