2024-05-08 / Workflow
Prompt Structure and Scene Rhythm
How character assets become shot prompts without overloading the model.
- character asset
- prompt
- workflow
Back to Lab
Question
角色资产里有很多信息,但镜头 prompt 不是资产表的复制。复制得太完整,模型会同时处理身份、服装、情绪、场景、动作、镜头和风格,结果反而稀释了这一镜头真正要完成的叙事任务。
从资产到 prompt,关键是选择,而不是搬运。
Observation
一个角色资产通常包含:
face
hair
costume
accessories
body posture
emotional range
social status
backstory
relationship to other characters
但单个镜头只需要其中一部分。reaction shot 需要脸和情绪;wide shot 需要 silhouette 和 costume;prop interaction 需要手、道具和袖口;dialogue shot 需要眼神方向和相对站位。
Method
我会先写 shot goal,再选 anchors:
SHOT GOAL:
She watches the auctioneer and hides anxiety.
KEEP:
black shoulder-length hair / ivory suit / pearl earrings / defensive stillness
DROP:
full backstory / family history / unrelated costume details / generic luxury terms
然后再写成 prompt:
Young heiress in an ivory tailored suit, black shoulder-length hair tucked behind one ear,
single pearl earrings. She stands still at the edge of the auction room, watching the
auctioneer while hiding anxiety. Slow push-in, medium close-up.
这个 prompt 的重点不是写满,而是让身份锚点和镜头目标互相支持。
Boundary
提示词太短会丢连续性,太长会丢节奏。判断标准不是字数,而是每个信息是否服务本镜头。
identity anchor: 保证她还是她
scene anchor: 保证她还在同一个空间
camera behavior: 保证观看方式稳定
event: 保证镜头有动作或信息变化
如果某个词不属于这四类,大概率可以删。
Lab Note
我会把 prompt 当作 shot contract:
model must preserve: identity, costume, key prop
model must change: expression, attention, camera distance
model may improvise: background texture, minor gesture, atmospheric detail
当 prompt 能区分 must preserve、must change、may improvise,它才真正变成可控的镜头指令。