AI 短剧分镜与关键帧

references/keyframe-craft.md

目录

• Boundary and instant tests
• Purpose
• Ordered recipe
• 尾帧：什么时候可以有，以及它换来什么代价
• Start-only drafting discipline
• Static test

Boundary and instant tests

• SHT-05 — Project exactly one accepted shot boundary and bind the exact Character/Look, Location/View, and Prop/State variants visible there.
• SHT-06 — Describe one freezeable instant. Ordered action, expression arcs, camera movement, and transforming weather belong to motion, not the keyframe.

Purpose

Describe one frame that can exist at a single instant and accurately anchors the accepted shot. A keyframe is not a compressed video prompt.

Ordered recipe

1. Purpose: what the audience must notice now.
2. Focal subject: exact asset and variant IDs.
3. Frame: shot size, angle, lens intent, aspect-aware composition.
4. Geography: Location/View, fixed anchors, foreground/background zones.
5. Boundary projection: exact start position, body facing, pose, gaze, hands, held props. Give body facing for every person — it fixes which screen side they hold and which side the next reverse cuts from; gaze does not cover it. Add head facing only when it differs from body facing.
6. Performance instant: one visible expression/tension state, not an arc. Pick a channel the current shot size can actually read — gaze, breath, body set, object handling, or a held decision. A wide shot cannot carry an eyelid; a close-up cannot carry a full-body retreat. Naming only the emotion ("愤怒") leaves the translation to a downstream stage, and the keyframe is the first frame — a mistranslation there is wrong from frame one.
7. Light and atmosphere: inherit accepted direction/time/weather; add only frame-relevant detail.
8. Text policy: readable/symbolic/blank/postproduction for visible surfaces.
9. Exclusions: contradictions or drift likely for this frame.

Identity and boundary facts remain source references. The keyframe owns focal, composition, camera/lens, frame-only staging, and exclusions.

写完之后过一遍可渲染检查

判据是一句话：接手的人能不能照着画出来，而不用自己发明任何事实。逐条对照， 缺哪条补哪条——这些不是可选修辞，缺一条执行端就得替你决定一次：

必须写出来	缺了会发生什么
每个人在哪、彼此距离多少（"站在他正前方约两步"）	人物间距逐镜漂移，观众读成换了场地
前景、中景、背景各有什么	纵深塌成一层，主体贴在背景上
视线该落在哪（本镜的视觉重点）	画面平均用力，观众不知道该看谁
构图方式与景别	同一场每镜构图随机，剪不到一起
机位在轴线哪一侧、俯仰角度、焦段意图	越轴；正反打方向对不上
光从哪来、什么质感	同场光位逐镜跳变
每个在场者的姿态、朝向、手部、持物	下一镜接不住，道具凭空易手
哪些与上一镜相同	执行端把"没写"读成"可以变"

最后一条最容易漏，也最便宜：与上一镜一致的部分写一句"与上一镜相同"即可， 不必重述细节；真正变了的才展开写。这一句是连续性的承载点。

写紧：一件事只说一次，能并进一句就不另起一句。上表十二项写满大约五百字就够； 写到八百字往上，多出来的通常是连接词、铺垫句和把同一件事换个说法再说一遍——它们不增加 可渲染的事实，却按字数付费。检查办法：删掉任意一句，如果画面没有变得更不确定，那句就是 多的。

写出来是这样（合成材料，四百余字覆盖上表全部十二项）：

竖屏 9:16 中近景。机位在轴线左侧、高约一米二，平视略低约五度，中焦 50mm。前景是长凳
扶手与半张湿报纸（占下缘约六分之一）；中景是两人对峙；背景是虚焦的候车牌。三分构图，
她落在右三分线上。她：坐长凳右端，双手交叠压膝，身体朝向正前，视线落在对方鞋尖，
无持物。人物乙：站她正前方约两步、略偏画面左，双手插袋，身体朝向她，头略偏向出口。
光自画面右上顶棚缝隙斜下为主光（硬），左侧留暗，无补光。视觉重点在她压膝的双手——
本镜唯一不动的东西。位置与持物与上一镜相同，只有她的视线从水洼移到对方鞋尖。
排除：无动作过程，无雨丝拖影。

Every visible entry the finished frame carries is listed in the shot's required 视觉依据 field, resolving to a real 视觉设定.md entry; an IMG-... prompt item in 图片提示词.md is a separate axis and does not stand in for it, and 视觉设定.md itself does not define IMG-... IDs. Write that list from the finished frame, not from the shot's intent — the subject that gets missed is the one that entered the composition while the frame was being written. A frozen frame may determine whether the surface is legible in this composition; it may not replace the exact wording or policy with an untraceable prose instruction. If the policy is still undecided, say so instead of inventing a hidden candidate state.

尾帧：什么时候可以有，以及它换来什么代价（SHT-17）

默认每镜一张关键帧，冻结的是 start。但首尾帧接续是 AI 视频里最常用的工作流之一：把首帧 和尾帧一起交给执行端，中间由它补。套件此前没有尾帧的位置，于是这条路要么走不通，要么被 私自绕过——后者更糟，因为绕过时尾帧往往是独立画出来的，它就成了第二个终点权威，与 镜头 end_boundary 各说各话。

structural_invariant：关键帧记录必须声明 boundary_role（start 或 end）， boundary_ref 指向同一镜头对应的那个边界字段。尾帧是 end_boundary 的投影， 不是新的终点事实——它与首帧对 start_boundary 的关系完全一样：可以决定这一帧怎么构图、 用什么景别镜头、光怎么落，不能决定人在哪、手里有什么、看着谁。尾帧与镜头终点不一致时， 错的是尾帧。

每镜的关键帧数量不是固定的。默认一张首帧；只有当交付工作流真的要消费尾帧时才加一张， 不为凑齐而画。补尾帧不改变 SHT-10：首帧仍然只能写 start 事实，尾帧只能写 end 事实， 两张各自守自己的边界，不互相借用。

代价要说清楚：交出一对首尾帧，等于把两帧之间的运动交给了执行端插值。 而运动路径本来 是运动规格拥有的东西。所以选了首尾帧接续的镜头，其运动规格不是"照旧再写一遍"，而是被 两端夹住了：它仍然要写清中间必须发生什么（动作顺序、对白落点、摄影机行为），但要意识到 执行端会优先满足两端的画面一致性。因此——

• 中间必须被看到的动作不要只靠插值兑现。如果一个动作是本镜存在的理由，两端之间没有 任何东西保证它会发生，就该拆镜，或者让该动作落在其中一端。
• 两端差异越大，插值越自由，中间越不可控。首尾帧接续适合"状态改变清楚、路径无所谓"的 镜头（转身、递交完成、坐下），不适合"路径本身是戏"的镜头。
• 终点报告仍然对照镜头 end_boundary，不对照尾帧。尾帧只是它的投影，不能自证到达。

Start-only drafting discipline

SHT-10 · reviewed_invariant: rendered keyframe prose may contain only facts from the boundary that frame declares. 首帧写进任何由运动或终点首次产生的事实是漂移； 尾帧写进在它之前就已经消耗掉的事实同样是漂移。两帧各守各的边界，不互相借用。

Keyframe 默认是 shot start，不是“本镜最有戏的时刻”。为避免把 end 提前：

1. 先明确本帧冻结 SHOT-... 的起点，草拟正文时暂不读终点与运动描述；
2. 只填 start 已成立的 position/pose/gaze/hands/held props/visible state；
3. 再与 end 做“新出现事实”差集；差集中的事实不得出现在 keyframe prompt；
4. 写完 Markdown 后从自然语言反向提取手位、持物、目光和可见状态，与镜头起点再比一次； 不能只看条目说明而忽略真正交付的正文。

反例：start 是“右手空置、看对方”，end 是“右手握铃绳、看门”。冻结帧写 “手已握铃绳”即使很好看，也属于 boundary drift。

Static test

Ask: could a still photographer capture every described fact at once?

Move these to motion:

• ordered verbs (先、再、随后、最终);
• expression changing from A to B;
• camera push/pan/track over time;
• entering/leaving/turning/reaching sequences;
• dialogue delivery arc or sound progression;
• light/weather transforming during the shot.

A single held pose may imply tension, but it must not require several moments.

Prompt economy

Do not restate full character or location 设定集s. Bind accepted variants and repeat only facts the frame needs to prevent ambiguity: distinguishing anchor, current Look, crucial spatial anchor, held prop, light direction, text state.

Generic “cinematic, 8K, masterpiece” language cannot replace subject identity, geography, composition, or continuity.

Failure examples

• incompatible Looks appear in the same frame;
• “turns, runs, then looks back” appears in a still;
• the start hand/prop differs from the shot boundary;
• background changes location identity or orientation;
• light direction resets without source change;
• readable evidence is paired with no-text;
• prompt lists subjects but not their spatial relationship;
• framing has no focal hierarchy.
• structured projection matches start but rendered prompt describes a fact that first appears in the shot end or motion.