TS-16.md 13.3 KB

TS-16 交付件

一、最终提示词原文(① 型 · 全能参考模式六段)

subject_definitions: <Subject 1> is the male lead vocalist from <Video 1>, in his mid-twenties, with the short cropped hair, loose-fit bomber jacket, and baggy grey cargo trousers established in <Video 1>, and he moves with the same sharp, grounded hip-hop cadence as the dancer in <Video 1>.
<Video 1> is the street-dance reference clip, supplying <Subject 1>'s appearance and his movement style; its original location, its background, its edit rhythm, and the audio inside <Video 1> are not carried into the target video.
<Audio 1> is the voice-timbre reference for <Subject 1> (S1), providing his bright, slightly raspy mid-register rap timbre; the words sung inside <Audio 1> are not carried into the target video.
<Audio 2> is the voice-timbre reference for the female harmony vocalist (S2), providing her soft, airy upper-register harmony timbre; she never appears on screen, and the words sung inside <Audio 2> are not carried into the target video.

summary: [reference generation + audio reference] A male lead vocalist performs a ten-second street-dance themed music video on a wet night street: he raps the verse alone under a sodium street lamp, then the chorus arrives at 00:05.000, where a female harmony vocalist joins from off screen and never appears in frame while he keeps dancing through the hook.

retention_analysis: <Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - his short cropped hair, loose-fit bomber jacket, and baggy grey cargo trousers are retained across both shots.
<Video 1> (defines <Subject 1>'s look and movement): attribute_transfer - the dancer's appearance and hip-hop movement style are carried onto the male lead vocalist; the original location, background, edit rhythm, and audio of <Video 1> are not reproduced.
<Audio 1> (heard as <Subject 1>'s voice): reference - only the timbre is taken; the words sung inside <Audio 1> are not reused.
<Audio 2> (heard as the female harmony vocalist's voice): reference - only the timbre is taken; the words sung inside <Audio 2> are not reused.

detailed_description: Live-action music-video look, cinematic, with hard sodium street light as the key source and a shallow focus that keeps the male lead sharp while the wet street behind him smears into bokeh.
[Shot 1] The camera starts low behind the male lead's shoulder and pushes in to a medium shot as <Subject 1> plants his feet on the churned wet asphalt under a sodium street lamp, the loose-fit bomber jacket hanging heavy on his frame and the baggy grey cargo trousers pooling over his sneakers. He bounces once on the balls of his feet, drops his weight into his heel, and rolls his shoulders in the sharp, grounded hip-hop cadence carried over from <Video 1>. His breath shows briefly in the cold air each time he snaps into a step, and the zipper pull on his jacket swings a beat behind every shoulder roll. <Subject 1> (S1) raps: <d>[English] Wake up the block, we move at dawn, and we don't stop.</d> His lips and jaw drive the line while the lamp's amber light rakes across his cheek from camera left and throws a long thin shadow past his sneakers onto the wet ground, where the reflection of the lamp shivers each time his heel lands. A shallow puddle between his feet holds the lamp upside down, and the amber rim light rides the top of his shoulders and the crown of his short cropped hair while the storefront behind him stays two stops under.
[Shot 2] At 00:05.000, the camera cuts to a slightly wider three-quarter front framing and begins to truck left while panning right, holding <Subject 1> centred as he steps out of the lamp's amber pool and into a cooler storefront spill light that flattens the colour on his jacket. He snaps into the chorus pattern, one arm cutting across his chest, his shoulders still rolling on the same cadence, and a bead of sweat tracks down his temple through the cold light. The female harmony vocalist (S2), who never appears on screen, says in an off-screen voiceover: <d>[English] the city is ours.</d> while her lips remain completely closed, and her soft upper-register line sits a fifth above his. His sneakers squeak once on the wet asphalt, the hem of his jacket swings with the last shoulder roll, the puddle slides out of the bottom of the frame as the camera trucks, and the frame holds on him mid-step with the lamp still burning at his back.

overall_soundscape: Night street room tone runs under both shots, with distant tyres hissing on wet asphalt and a low ventilation hum from the storefront behind him. The sodium lamp carries a faint electrical buzz that sits just under the vocal and never changes level. No other on-screen sound source enters in either shot.

non_diegetic_music: A 92 BPM boom-bap instrumental carries the whole video: a sub-bass line and a dusty drum loop enter at 0.00 seconds, and a bright synth hook joins at 00:05.000 to lift the chorus under the harmony.

二、资产层清单

数组顺序 = 输入菜单原始顺序(order 1→3)= 模型自动编号顺序。未含任何 schema 外字段;url 字段只放 url 原文。

{
  "references": [
    {
      "id": "street_dance_clip",
      "type": "video",
      "role": "舞者形象与动作来源",
      "url": "<OSS>/h3_v3_ts16/dancer.mp4"
    },
    {
      "id": "rapper_voice",
      "type": "audio",
      "role": "男声歌手音色来源,对应目标视频中的男主唱",
      "url": "<OSS>/h3_v3_ts16/rapper_voice.mp3"
    },
    {
      "id": "chorus_voice",
      "type": "audio",
      "role": "女声和声音色来源,对应目标视频中的女和声(不出现在画面中)",
      "url": "<OSS>/h3_v3_ts16/chorus.mp3"
    }
  ]
}

素材落位说明(说明层,不进 JSON):street_dance_clip → <Video 1>(取舞者形象 + 动作风格,attribute_transfer);rapper_voice → <Audio 1>(对应 <Subject 1> (S1) 音色);chorus_voice → <Audio 2>(对应稳定声音描述 + (S2),该说话人全程不在画面)。三条素材均无 note 限制,全部可用、全部被认领,无 excluded。

三、交付说明

  • 形态判定(§0 第 0 步):需求是「MV 十秒 + 三件参考素材 + 各素材职责」,没有逐窗显示内容要求,也没有要分镜脚本交付物 → 单件 ① 型,未拆链路四层。正文语言英文(① 型正文英文硬判据),资产层与说明层中文(说明层语言自由)。
  • 模式判定:多素材身份/风格锁定(视频定形象与动作 + 两段音色参考)→ 全能参考,六段顺序 subject_definitions → summary → retention_analysis → detailed_description → overall_soundscape → non_diegetic_music。
  • summary 任务类型标记:[reference generation + audio reference]。reference generation 来自视频与音频参考锁定主体;+ audio reference 来自两条音色类参考素材(§1.7 第 21 项:音色/风格/节奏类参考 → 组合标记)。本单无任何帧锚点输入,故不标 keyframe completion。
  • duration_seconds 取值:10(raw_brief 给定「10 秒」,落在 4-15s 内,未超 H2,无需拆分方案)。
  • H13 素材总闸:视频 1(≤3)+ 音频 2(≤3)= 3 个文件(≤12),组合合法——音频不是单独输入,有视频素材同单,未触发 doc1:35 组合条款,无需阻塞。
  • 音色参考不带入原台词(doc2:577):<Audio 1> / <Audio 2> 在定义句与 retention_analysis 都显式声明「只取音色,不带入其中的词」。两条音频各自显式绑定到目标说话人,绑定关系写在 <Audio N> 定义句内(不是只放在资产层或说明层);<Audio 1> 复用 <Subject 1> 的全局说话人 ID 写成 <Subject 1> (S1);<Audio 2> 的说话人不在画面、不立为可见内容 Subject,按官方允许的「稳定声音描述 + (Sx)」形式绑定,不另造 <Subject 2>。
  • 说话人 ID:(S1) 男主唱、(S2) 女和声,按目标视频实际发声事件顺序各分配一次,跨镜不重号;(Sx) 只出现在 detailed_description,retention_analysis 零 (Sx)。
  • 画外音:女和声「不出现她的画面」→ 用官方固定短语 says in an off-screen voiceover(逐字相邻、未插语打断),紧跟声明她唇部保持闭合;她全程不在画面内,该声明与画面无冲突。
  • 台词口径:本次输入未给逐字歌词,两段 <d> 文本为按「主唱开口唱歌 / 和声加入」这一意图所作的创作补位(低风险可拍摄细节),已在正文写成英文歌词并在此声明为可替换项;口径按 §1.5 ③ 默认 V(无参考音频复用——两条音频只取音色),故 <d> 内词与标点保持我所写原样、未做规范化增删。若你方有定稿歌词,请整句替换,我按 V 口径逐字重挂。
  • 素材角色不裸奔:三件素材各自声明职责(形象/动作来源、男声音色、女声音色),无堆叠。
  • 运镜:[Shot 1] 单一主导运镜 = push in;[Shot 2] 单一主导运镜 = truck left + pan right 组合(官方指导层的环绕写法,Arc Shot 未用)。
  • C 级构造声明(非官方条款,如实标注):幽灵素材禁令、<Audio N> 与说话人全局 ID 的绑定写法细节、以及「歌词未给定时由我补位」这一处理,均为工程收口,官方无对应明文。
  • 出处纪律:正文未使用官方外编号词法(无 object N / background N / 小写 picture n);url 未进正文;未把内部规范表述为官方规则。

四、§1.7 机械验收(21 项逐项实测)

  1. ✓ 形态判定正确:单件 ① 型,与「未要求分镜脚本」的输入一致;未硬拆链路四层(§0.2 第 3/6 条)。
  2. ✓ 六段齐全且顺序正确(doc2:250),实测段序:subject_definitions → summary → retention_analysis → detailed_description → overall_soundscape → non_diegetic_music。
  3. ✓ 字段名与官方一致无拼写错;全能参考六段无指令行,H4 不适用。
  4. ✓ 四标签使用:<Subject 1> / <Video 1> / <Audio 1> / <Audio 2>,grep 实测各标签定义与引用编号一致,无幽灵编号、无幽灵素材(3 条素材全部被认领);关系标记只用了 8 取值中的 fully_preserved / attribute_transfer / reference。
  5. ✓ references 数组顺序 = 菜单 order 1→3,三条记录均携带原 url 原文未改链;<Video 1> 独立成条、<Picture N> 无条目属合法形态(本单图片为 0)。
  6. ✓ 非锚点素材认领:<Video 1> 独立定义条目;<Audio 1>/<Audio 2> 独立定义条目;<Subject 1> 定义句以序号内联引用 <Video 1> 认领来源。
  7. ✓ 两句对白均用 <d>[English] …</d> 包裹,<d> 内只有语言标签与台词;逐字保真按 V 口径执行(见交付说明)。
  8. ✓ 未引用任何标记为不可用的素材——本单 3 条素材 note 均为空、全部可用,无 excluded。
  9. ✓ 无设定冲突;未添加未经授权的剧情结果(男主唱只跳舞与唱歌,女和声只加入副歌,无拒绝/离开/新关系类推断)。
  10. ✓ 正文为散文 + 官方字段名,无 JSON 通道要求;无残留模板占位符(grep 实测 0 处字面 Shot N / S.SS)。
  11. ✓ H13 实测:视频 1 ≤ 3,音频 2 ≤ 3,文件总数 3 ≤ 12,音频非单独输入。
  12. ✓ 正文层编号词法实测:除官方四标签与 (S1) (S2) 外 0 处其它编号词法(grep object / background / picture 均 0 处)。
  13. ✓ H1 实测:最终 Prompt 共 4902 字符(含标点与空白,len() 口径)≤ 7000。
  14. ✓ H2 实测:单段 10 s,落在 4–15 s;正文无独立「总时长 X 秒」指令行(时长只出现在 summary 概述与切镜时刻中);未提及帧率,H3 不触发。
  15. ✓ H5 实测:[Shot 1] 无时间戳;[Shot 2] At 00:05.000 毫秒 3 位、唯一一个切镜时刻、落在 0–10 s 内且递增。H5b 实测:① 型正文 0 处 [N 秒-M 秒] 窗式记号(grep 复核)。
  16. ✓ H6 实测:detailed_description 共 410 个英文词(按空白分词,对该字段正文计量),落在 350–500 区间(无台词密集/视频编辑豁免需要)。
  17. ✓ H10 实测:overall_soundscape 3 句(1–4);non_diegetic_music 1 句(1–3);两者均非 N/A,环境声具体(轮胎/通风/灯管电流声),配乐写了 BPM、乐器与进入时刻。
  18. ✓ H11 实测:[Shot 1] 5 s 承载 12 词唱词、[Shot 2] 5 s 承载 4 词唱词,无 3 s 镜塞大段词的失衡;H16 实测:<Video 1> 定义为形象/动作来源(非分镜草稿图),全片 0 处秒级对齐表述,3 条素材全部被认领。
  19. ✓ 本节自报数字均先实测后写盘:4902 字符、410 词、3 句、1 句、retention_analysis 0 处 (Sx)、正文 0 处窗式记号,全部由脚本对最终交付原文计量得出;资产层 JSON 无 schema 外字段(「素材落位说明」写在说明层),url 字段只放 url。
  20. ✓ retention_analysis 逐标签一行且沿用 subject_definitions 含义,grep 实测 0 处 (Sx);关系标记只挂在有参考来源的标签上(4 条全部有来源)。
  21. ✓ summary 任务类型标记与输入对齐:音色参考 ×2 → [reference generation + audio reference];无帧锚点输入 → 未标 keyframe completion。