These three fields form a layered synergy mechanism and are the core guarantee for cross-shot character consistency.
Describe key props (handheld items / tabletop items / environmental decor). Use a three-part location prefix + item description:
| Prefix | Meaning | Standard Example |
|---|---|---|
/held | Direct hand control | /held: The fingers of the left hand are holding a transparent glass with amber liquid in it. |
/table | Stable support surface | /table: Stainless steel thermos cup placed flat on marble countertop, lid tightened |
/bg | Distance / background | /bg: There are green potted plants in the corner of the conference room with wide leaves. |
Full example: /held: right hand grips a black leather briefcase with silver metal clasp; /table: open laptop lying flat on a solid wood desktop; /bg: city night skyline visible through floor-to-ceiling window
The single source of visual truth for the character consistency system. Stores reference photo paths to provide a direct visual anchor across shots.
Path format: /characters/{character_english_name}/{reference_photo_file}, named portrait_{version}.{extension}
Example: portraits: ["/characters/chen_zhuo/portrait_v1.png"]
Update mechanism: version increments, old versions retained, full-level cascade — new reference photos are injected into all related shots, triggering prompt re-fusion.
Text-level anchor words that complement reference photos (visual reference). Use a three-part format: facial features + costume + body type.
Example: square face shape, black short hair parted on the side, thick straight eyebrows, single-fold eyelids; dark gray fitted suit jacket over white shirt, black straight-leg trousers; medium build around 175 cm, well-proportioned
Must include character constraint clause: Character constraint: only Chen Zhuo appears in the frame, no other people allowed, no repeated appearances of the same character, no character clones or duplicates