Principles
- Asset = descriptor + reference image; the descriptor appearsverbatimin related prompts.
- State sets
@hero/@hero_wet/@hero_blood; same for day/night/rain locations. - Model has no memory: unlocked assets = later junk frames.
Three character design images
- Face close-up (the only source of the face)
- Full-body front — head cropped out(prevents wide-angle from sneaking in a blurry face)
- Full-body back
The sheet must be boring: gray background, flat lighting, pores visible;cinematic feel does not go on the sheet. Prioritize close-ups 3/4. Choose a believable face over fake beauty;pupil catch-light。
Change state
Local edit → mask and paste back.Do not run the whole image fully through the model twice(plastic dead texture).
Stress test: recognizable 10/10 across many poses and lighting + no collapse in the same frame; if it collapses, change the descriptor first.
States / Props
| Wrong | Right |
|---|---|
| One sheet cramming open hatch + bloodstained clothes + holstered gun | One asset per state, reference it when used |
| Reference image shows the hatch open | The model will force-play the hatch opening → delete unneeded states from the image |
| Small caption text as an instruction manual | The video prompt must still spell out the operation process |
Weapon condition: separate sheet for close-ups.
Character comes before appearance; secondary assets (mecha, etc.) align with personality. With multiple protagonists, first make them distinguishable by silhouette/clothing, then play with costume changes.
Voice line (fixed sentence)
Voice: [音色]; [节奏]; [口音]; [压力变化]
Attach only if there is dialogue; consistent with the performance section, unchanged between scenes.
Performance
Axiom: behavior under pressure, not emotion labels stuck on.
Master, one English paragraph (about 150–220 words): posture history · psychological engine · voice line usage · triggered tic · mask + However when X · eye life · optional sole softening target · do not write clothing/camera position/color grading.
Scene changes: write only those present in this shot; the core stays unchanged; actions that cannot be doneconvert into exitsdo not delete;state mid-swing, not a long process.
Five pillars: goal (verb directed at a person) · obstacle cost · tactics (must change on failure) · 2–4 visible beats · subtext leaking out
Aliveness: react before the other person's sentence ends · micro-pause in thought · interrupting the current task = punctuation · stillness = tension · one micro-event every 1–2s · group scenes ripple out of sync
Forbidden: indicating face · single tactic · dead eyes · emotion instantly reset to zero · slow-motion threat telegraphing · facial features overacting in close-up
Character design prompt skeleton
Film character sheet, three studio photographs, same person, mid-grey flat light:
headless full-body front | full-body back | large 3/4 portrait with catch-lights.
[role, build, hair, marks]. Outfit A. Natural pores, no beauty retouch, no cinematic grade on sheet.
In-shot anchor:@Name: [状态] + 本镜关键可见点. 100% matches reference.(if there is a reference, write less long-form facial description)