Practical guide · 实用指南
Character consistency in AI short drama: identity, costume and shot continuity
AI 短剧怎样保持角色跨镜一致:人物、造型与镜头状态
Source checked 2026-09-12 · 源码核对日期
Separate the person’s persistent identity from the current costume and each shot’s action. Keep only genuinely unchanged visual facts in a continuity lock; track hands, props and gaze as shot state. An original three-shot archive scene below shows how to connect these decisions. A character sheet or an image-prompt ID is not a generated reference image, and written constraints do not guarantee a model’s results. For the wider adaptation workflow, start with the novel-to-short-drama guide.
先把不随换装改变的人物身份、当前造型与每镜动作分开。只把真正跨镜不变的外观写入连续性锁,手持物、视线和姿势则按镜头记录。下面用原创的档案馆三镜示例,说明怎样把这些决定连起来。有角色设定或图片提示词编号,不等于已有真实参考图;文字约束也不保证模型最终效果。完整改编流程可从小说改短剧指南开始。
Before you start · 开始之前
- An accepted scene or episode script. It should establish who appears, which visible state matters and what changes during each shot; resolve story meaning in writing rather than hiding it inside an appearance prompt.
- Creator-approved identity facts and scene styling. Do not invent age, ethnicity, facial traits, branded clothing or other details merely to fill a template. Mark unresolved choices instead.
- If you plan to use the documented skills, a configured supported agent host and project folder. This guide does not execute image or video generation and its example contains no generated reference asset.
- 准备已接受的场次或单集剧本,写清谁出现、哪些可见状态重要、每镜发生什么变化。知识、目标和关系属于编剧,不要把故事语义藏进外观提示词。
- 准备创作者确认过的身份事实与本场造型。不要为了填模板擅自补年龄、族裔、五官、品牌服装等细节;未决定的内容继续标为待定。
- 如要使用文档中的技能,还需已配置的受支持 Agent 宿主和项目目录。本指南不执行图片或视频生成,示例也没有生成参考素材。
Steps · 操作步骤
- Write Character as the persistent identity Use observable, comparable anchors that survive a costume change and work beyond a front-facing close-up. In the example, Lin Lan has a V-shaped nape hairline and a right shoulder that sits slightly lower than the left. These belong to Character; the rain jacket, hand-held key, gaze and pose do not. Ask
short-drama-assetsto record the accepted facts in剧集/<EP>/视觉设定.md. - Add one compatible Look for this scene Reference the Character, then record only clothing, hair arrangement, makeup, injury or weathering that can coexist now. Lin Lan’s scene Look is a low-tied hairstyle and a dark-teal matte hooded work jacket with two rows of silver snaps on the left cuff. Keep the brass key as a Prop and its current holder as shot state; neither is part of the Look.
- Lock only the invariant visible surface A LOCK is justified when the fact appears across at least two shots, viewers could notice a mismatch and the story does not change it. Use the minimal phrase “dark-teal matte hooded work jacket, hood down, two rows of silver snaps on the left cuff” across the affected
IMG-*body, frozen keyframe andMOTION-*body. Do not lock the key’s position, hand pose or gaze because those change. - Chain three shots with explicit states For every shot write
start → one primary action → end, including each hand, held prop and gaze. The next shot must start from the previous shot’s actual end unless the screenplay establishes a cut or state change. A frozen keyframe normally describes the start only; it must not show an action or prop state that first appears at the end. - Keep prompt IDs separate from media evidence Use
short-drama-image-promptsonly to author entries such asIMG-CHAR-LINLANandIMG-LOOK-LINLAN-RAIN; those IDs do not prove files exist. Inshort-drama-storyboard, bind an inspected project image asREF-*. If the creator will supply it later, keep aPLAN-*slot with order, intended file, purpose, control scope and exclusions. Missing files must not be silently downgraded to text-to-video. - Stop at a production-honest handoff A
PLAN-*shot may describe the intended image-to-video path, but it cannot entershort-drama-produceuntil the real file is present and the slot becomesREF-*. For a continuing shot chain, later segments also need the preceding actual video and actual tail frame; planned frames or prose cannot substitute. Text constraints improve the handoff, but do not guarantee a model will preserve face, clothing, space, props or motion.
- 先写长期不变的 Character 身份 选换装后仍成立、可观察、可比较,并能在侧面或背面识别的锚点。示例中的林岚有 V 形后颈发际,右肩略低于左肩;这些属于 Character。雨衣、手里的钥匙、视线和姿势都不属于身份。可用
short-drama-assets把已确认事实写入剧集/<EP>/视觉设定.md。 - 再写本场相容的 Look 先引用 Character,只补本场能同时存在的服装、发型、妆伤与天气化。林岚本场是低束发,穿墨绿色哑光连帽工作夹克,左袖有双排银色按扣。黄铜钥匙仍是 Prop,由谁拿着属于镜头状态,不能吞进 Look。
- 只锁真正不变的可见表面 同一事实跨至少两镜可见、观众能发现不一致、剧情又不会改变它,才写 LOCK。示例锁面是“墨绿色哑光连帽工作夹克,帽子放下,左袖双排银色按扣”,在受影响的
IMG-*正文、冻结关键帧和MOTION-*正文中原样携带。钥匙位置、手势和视线会变化,不上锁。 - 用状态链写三镜 每镜都按“起点 → 唯一主要动作 → 终点”记录双手、持物与视线;除非剧本明确跳时空,下一镜起点应继承上一镜真实终点。冻结关键帧默认只写起点,不能提前出现终点才成立的动作或道具状态。
- 把提示词编号与真实素材分开
short-drama-image-prompts可写IMG-CHAR-LINLAN、IMG-LOOK-LINLAN-RAIN,但编号不证明图片存在。到short-drama-storyboard,检查过且项目内可读的图片才写REF-*;创作者计划稍后提供时保留PLAN-*,注明顺序、拟用文件、用途、控制范围与不得控制范围。缺图不能静默降级为文生视频。 - 以诚实的投产边界收尾
PLAN-*可以记录预期的图生视频路径,但真实文件入项并改为REF-*前,不能交给short-drama-produce。连续时空的后续段还需要上一段实际视频和实际尾帧;文字或计划帧不能冒充。文本约束能改善交接,却不保证模型最终稳定人脸、服装、空间、道具或动作。
Example · 示例
ORIGINAL EDITORIAL EXAMPLE — written for this guide, not skill output, generated media or a production test.
Scene: At a rain-soaked archive loading bay, Lin Lan must open locker 17 while keeping a red evidence folder dry. The steel bench stands directly beside locker 17, so the three shots require no off-shot repositioning. Her jacket hood stays down, leaving the nape anchor visible.
Character — Lin Lan: V-shaped nape hairline; right shoulder slightly lower than left.
Look — LOOK-LINLAN-RAIN: references Lin Lan; low-tied hair; dark-teal matte hooded work jacket; hood down; two rows of silver snaps on the left cuff.
LOCK — LOCK-LINLAN-JACKET: “dark-teal matte hooded work jacket, hood down, two rows of silver snaps on the left cuff”; active in SHOT-01 through SHOT-03.
Prop — brass locker key and red evidence folder remain separate assets. Throughout these three shots, the left arm pins the folder and the left hand supports its lower edge; only the right hand handles the key.
SHOT-01 — Start: folder tucked under left arm, key on the steel bench, right hand empty, gaze on the key. Action: Lin Lan picks up the key with her right hand. End: folder under left arm, key in right hand, gaze shifts to locker 17.
SHOT-02 — Start: inherit the preceding end. Action: she inserts the key into locker 17. End: folder under left arm, key remains in the lock, right hand stays on it, gaze on the keyhole.
SHOT-03 — Start: inherit the preceding end. Action: she turns the key once. End: folder still under left arm, key turned in the lock, right hand on the key, gaze moves to the door seam.
Prompt/reference record: `IMG-CHAR-LINLAN` and `IMG-LOOK-LINLAN-RAIN` name prompt entries only. `PLAN-LINLAN-RAIN-01` points to the explicitly non-existent placeholder `剧集/EP001/待提供/林岚雨夜造型.png` below, first in reference order, for identity and compatible Look only; it must not control hand pose, key position, gaze or camera. No real file has been provided, inspected or bound, so all three shots remain PLAN and are not production-ready. No image, video, face-consistency or motion result is claimed.
Copyable text-only handoff request:
For EP001, use `short-drama-assets`, `short-drama-image-prompts` and `short-drama-storyboard` to organize Lin Lan and these three shots. Keep Character, Look, LOCK and shot state separate. Stop at text files; do not generate images, video, voices or music. `剧集/EP001/待提供/林岚雨夜造型.png` is a placeholder path, not an existing file. Keep it as PLAN and do not prepare a production submission.原创编辑示例——为本指南编写,不是技能执行输出、生成媒体或投产测试。
场景:雨夜的档案馆装卸区,林岚要打开 17 号柜,同时护住一份红色证物夹。钢台紧挨 17 号柜,三镜之间无需补走位;夹克帽子保持放下,让后颈锚点可见。
Character|林岚:V 形后颈发际;右肩略低于左肩。
Look|LOOK-LINLAN-RAIN:引用林岚;低束发;墨绿色哑光连帽工作夹克;帽子放下;左袖双排银色按扣。
LOCK|LOCK-LINLAN-JACKET:“墨绿色哑光连帽工作夹克,帽子放下,左袖双排银色按扣”;SHOT-01 至 SHOT-03 生效。
Prop|黄铜柜钥匙与红色证物夹分别建资产。三镜中,左臂始终夹住文件夹,左手压住其下缘;只有右手操作钥匙。
SHOT-01|起点:证物夹在左腋下,钥匙在钢台上,右手空,视线看钥匙。动作:林岚用右手拿起钥匙。终点:证物夹仍在左腋下,钥匙在右手,视线转向 17 号柜。
SHOT-02|起点:继承上镜终点。动作:她把钥匙插入 17 号柜。终点:证物夹仍在左腋下,钥匙留在锁孔,右手扶住钥匙,视线看锁孔。
SHOT-03|起点:继承上镜终点。动作:她把钥匙转动一次。终点:证物夹仍在左腋下,钥匙已在锁孔内转过,右手仍扶钥匙,视线移到柜门缝。
提示词与参考记录:`IMG-CHAR-LINLAN`、`IMG-LOOK-LINLAN-RAIN` 只命名提示词条目。`PLAN-LINLAN-RAIN-01` 指向下方明确标为非现存文件的占位路径 `剧集/EP001/待提供/林岚雨夜造型.png`,在参考顺序中排第一,只控制身份与相容 Look,不控制手势、钥匙位置、视线或机位。现在没有真实文件被提供、检查或绑定,所以三镜都保持 PLAN,不可投产;本例不声称已获得图片、视频、人脸一致或动作效果。
可复制的纯文本交接请求:
请为 EP001 使用 `short-drama-assets`、`short-drama-image-prompts` 和 `short-drama-storyboard` 整理林岚与以上三镜;分开 Character、Look、LOCK 和镜头状态。只交文字,不生成图片、视频、配音或音乐。`剧集/EP001/待提供/林岚雨夜造型.png` 是占位路径,不是现存文件;保持 PLAN,不准备投产提交。
Expected files · 预期文件
剧集/<EP>/视觉设定.md: the Character entry, one compatible Look, separate Props, scoped states and only the necessary continuity LOCK.剧集/<EP>/图片提示词.md: copyableIMG-*prompt entries carrying the accepted anchors and LOCK text. They are text deliverables, not evidence of generated assets.剧集/<EP>/分镜.md:SHOT-*state chains, frozen start frames, visual-basis references and honestREF-*,PLAN-*or missing-reference status.视频提示词.mdis added only when the reference path is ready enough for that stage.
剧集/<EP>/视觉设定.md:Character 条目、一个相容 Look、独立 Prop、有范围的状态,以及必要的连续性 LOCK。剧集/<EP>/图片提示词.md:携带身份锚点与 LOCK 锁面的可复制IMG-*提示词;它是文本交付,不证明图片已生成。剧集/<EP>/分镜.md:SHOT-*状态链、起点冻结帧、视觉依据,以及真实的REF-*、PLAN-*或缺图状态。参考路径满足该阶段要求后,才按需添加视频提示词.md。
Verify the result · 验证结果与边界
- Editorially compare all three shots: Character anchors remain the same, the Look is compatible, the exact LOCK surface is carried wherever required, and every hand, prop and gaze at a shot start matches the preceding end.
- Audit evidence labels before production: every
REF-*resolves to an inspected real file with a stated purpose and boundary; unresolvedPLAN-*slots stay non-production. Passing this document review does not validate or guarantee generated-media quality or continuity.
- 逐镜做编辑核对:Character 锚点不变,Look 内部相容,LOCK 锁面在所需位置原样出现;每镜起点的双手、持物与视线都接得上前镜终点。
- 投产前核对证据等级:每个
REF-*都能找到已检查的真实文件,并写明用途和边界;未解决的PLAN-*继续保持不可投产。文档审阅通过不等于生成媒体质量或连续性已经验证。
Sources and version notes · 来源与版本说明
- Asset boundaries: identity, variants, shot transients and continuity locks
- Character anchors and compatible Look definitions
- When a visible fact merits a continuity lock
- Where continuity-lock text must be carried
- Shot states, frozen keyframes and reference-slot status
- Image-prompt structure and scoped reference control
- Real REF images, planned inputs and non-production states
- Actual video and tail-frame requirements for continuation