Principles
- Changing positions during the storyboard stage is the cheapest.
- Space before style.
- Cross-frame unity: face, clothing, injuries, props, light direction, screen direction.
Granularity
| Coverage | 4–8 frame overview, to probe rhythm |
| Pinned points | Frame-by-frame shot size + action, lock transitions/climax |
Unit = 4–8 consecutive beats in the same scene; for sequels useEnd-frame anchor, identity/clothing sentences unchanged.
GEO Use the script scene lock; this module only verifies whether it is readable and whether it matches the image.
Reference slots ≤4
Character · Location · Prop/Second person · Light mood (optional)
Bind person/place/object sources;Consistent aspect ratio; name the responsibilities (character reference / location reference).
Location reference:Only spatial materials, do not inherit composition/camera angle/color grading.
Location board
- 3/4 depth, not front-facing wallpaper
- ≥1 anchor per scene (pillar/lamp/table)
- Single-source light
- Reverse shot: output another corner of the same room, or capture a frame from a moving empty shot and then fix the texture
First frame and positions (verification checklist)
First frame already holds all required subjects in correct places.
No empty establishing unless designed.
Each person: screen L/R · front/middle/back · measurable distance to landmark · torso orientation and line of sight written separately · which hand holds the prop.
Dialogue: only the speaker's lips move; others are silent.
Timeline
0–1s 定场（可接上句尾）
1–Ns 每拍≤3 短句，动作不互斥
Default single take; hard cuts/reverse shots mustcut by cutbe written clearly. Multi-cut check: number of people, left/right, line of sight, light direction, injuries and props, no teleporting.
Frame sentence order
风格一句 → 主体 → 动作 → 环境 → 光/镜 → 与上帧一致
[Style] sequence. [@Cast] in [place].
(1)建立 (2)行动 (3)阻碍 (4)反应.
Same face, wardrobe, light, geography.
Project Style constants
Style Prefix + Constraints reused across the whole project;only SHOT/frame content is variable(STYLE / CONTINUITY / DIRECTION locked likewise).
Stop-motion/hand-drawn: write clearly on twos, what is smooth/stepped, no default smooth interpolation.
Before generation
- identity and clothing stable
- GEO/anchors readable
- first-frame placeholder correct
- line of sight and orientation not contradictory
- every frame advances
- can map to 5–15s video
finalized text and optical/physical density → prompt-engineering。