TS-21.md 9.02 KB

TS-21 交付物

交付说明

  • 交付形态判定(第 0 步):① 型。理由:raw_brief 明确要求 "I need an English structured prompt",并点名 "official six-field Ref2VA structure as JSON",且要求 "Keep the body text in English"——英文结构化需求命中 ① 型识别特征。
  • 模式判定:全能参考(多素材身份/风格锁定:产品外观来源 + 出镜者形象与口播节奏来源)。raw_brief 说 "the headset look and the reviewer's identity must stay consistent",即主体身份锁定 → 全能参考,六段结构。
  • 输出 schema:按调用方指定 {"prompt": ..., "references": [...]},顶层仅这两个键,无 schema 外字段、无解释性文字混入 JSON;prompt 为单字符串(六段以空行分隔),JSON 可解析。
  • 段落顺序(doc2:250):subject_definitions → summary → retention_analysis → detailed_description → overall_soundscape → non_diegetic_music。
  • H4 适用性:instruction 第一行条款的适用域是 ① 型 I2VA/FL2VA/L2VA(SKILL §0.4 H4 适用列);全能参考六段无指令行,故本交付不写 instruction 行,直接从 subject_definitions 起。
  • 外观描述的取舍(零臆造 vs. 具体可辨识):出镜者外观的唯一来源是 reviewer_video、产品外观的唯一来源是 headset_product。raw_brief 没有给任何性别、发色、服装、颜色、材质描述,本稿不替用户编造(编造出与参考素材冲突的外观,恰恰破坏它明确要求的 "identity must stay consistent")。因此两份素材的具象特征以「从来源素材原样承接、两镜之间不改变」的方式写进 <Subject N> 定义句,并列出被承接的属性类别(发型与发长 / 服装 / 坐姿;外壳颜色 / 耳罩形状 / 头梁 / 刻印 logo)——既给了可辨识锚点,又不虚构图里无法核对的取值。
  • <Video 1> 不自动产生 <Audio 1>(H15):reviewer_video 是普通参考视频,职责为出镜者形象与口播节奏,故立为 <Video 1>,不因它含声而另设 <Audio N>;其音轨未被声明为复制/引用对象。
  • 口播内容不臆造:raw_brief 只给了动作(拿起、翻转展示耳罩与刻印 logo、戴上),没有给出任何台词原文。按零臆造纪律,本稿不替用户编造口播内容,因此目标视频内无发声事件 → 说话人 ID 不分配(从不发声的角色不获得 (Sx)),也无 <d> 对白。<Video 1> 的「口播节奏」只作为 weak_reference 的节奏/出镜方式参考进入 retention_analysis,不产生台词。
  • 素材落位:headset_product(产品外观来源)→ 在 <Subject 2> 定义句内以 <Picture 1> 序号内联引用认领,不另起独立 <Picture N> 条目(role 为「产品外观来源」,非目标帧锚点);reviewer_video → <Video 1> 独立条目。无幽灵编号。
  • 关系标记(H9 八个固定取值):<Subject 1> = fully_preserved(出镜者身份保留);<Subject 2> = fully_preserved(耳机外观保留);<Video 1> = weak_reference(只取节奏与出镜方式,不复制画面内容)。
  • non_diegetic_music = N/A:需求未要求配乐,显式写 N/A 而非留空。环境声已具体描写(1-4 句内)。
  • duration_seconds 取值:12(B 级内部规范:由请求体指定、不写进正文)。单段 12 秒,落在 4–15 秒内,无需拆分。帧率未指定,正文不写帧率(H3:不写不扣分)。
  • 切镜:三个动作节点(拿起 / 翻转展示 / 戴上)切两镜,[Shot 1] 无时间戳,后续 At MM:SS.mmm 毫秒 3 位、严格递增、落在 12 秒内;① 型内零处 [N 秒-M 秒] 窗式记号(H5b)。

最终提示词原文(JSON 交付)

{
  "prompt": "subject_definitions:\n<Subject 1> is the reviewer presenting the product on camera, carried over from <Video 1>: the same person as in <Video 1>, with the same hairstyle and the same hair length, the same clothing, and the same seated on-camera manner, held unchanged across both shots so that the reviewer's identity stays consistent from the first frame to the last.\n<Subject 2> is the headset the reviewer presents, shown in <Picture 1>: the same shell colour, the same earcup shape and cushioning, the same headband, and the same engraved logo as in <Picture 1>, held unchanged across both shots; the engraved logo sits on the outer face of the earcup, and it is this logo that the reviewer turns toward the camera.\n<Video 1> is the reference video providing the reviewer's on-camera presence, framing and pacing: the target video borrows only the reviewer's presentation rhythm and the way a product is handled for the camera, and does not copy any image content from it.\n\nsummary:\n[reference generation] A twelve-second product presentation video. A reviewer picks up the headset from the table in front of them, turns it over in their hands to show the earcups and the engraved logo to the camera, and then puts it on. The headset's look and the reviewer's identity stay consistent across both shots, each carried over unchanged from its own reference.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - the reviewer's identity as established in <Video 1>, including hairstyle, hair length, clothing and seated on-camera manner, is retained unchanged in both shots.\n<Subject 2> (appears in [Shot 1], [Shot 2]): fully_preserved - the headset's shell colour, earcup shape and cushioning, headband and engraved logo as established in <Picture 1> are retained unchanged in both shots.\n<Video 1>: weak_reference - used only for the reviewer's presentation pacing and the way a product is handled for the camera; no framing or image content from the reference video is reproduced.\n\ndetailed_description:\nLive-action, cinematic, clean commercial product footage: soft, controlled studio light over a plain table surface, with the headset kept as the brightest plane in every frame and the background held in gentle, even falloff. [Shot 1] The camera holds a Static Shot at tabletop height, framing the reviewer from the waist up behind the plain table, their forearms resting on the surface and the headset lying face-up in front of them with the headband running toward the camera. The reviewer reaches forward with one hand, slides their fingers under the headband and lifts the headset off the table in one smooth motion, bringing it up into the centre of the frame; their other hand comes up to meet it and takes hold of one earcup, steadying it so the shell does not swing. The reviewer's identity, hairstyle, hair length, clothing and seated manner stay exactly as carried over from <Video 1>, and the headset's shell colour, earcup shape, headband and engraved logo stay exactly as carried over from <Picture 1>. [Shot 2] At 00:04.500, the camera cuts to a slow Push In to a medium close-up as the reviewer turns the headset over in both hands: they rotate it so the outer face of the earcup comes around toward the lens and hold it there, the engraved logo catching a thin highlight along its recessed edge, then roll their wrists to bring the padded headband up past the top of the frame and back down, showing the cushioning on the underside and the second earcup before letting the headset settle back into both hands. They lift the headset to their head, seat one earcup over one ear with the first hand and the second earcup with the other, and press the headband down gently until it rests evenly on their hair; their hands stay at the earcups for a beat, then slide away and drop back to the table. The shot ends with the headset on their head, the engraved logo facing outward, both hands flat on the table surface, and the studio light still holding the headset as the brightest plane in the frame.\noverall_soundscape:\nThe room is quiet, with a low, even air-handling hum behind the table. Small, close sounds carry clearly: the headset's shell ticking once as it is lifted off the table surface, the soft compressing sound of the cushions seating over the ears, and the headband settling into hair with a muted brush.\nnon_diegetic_music:\nN/A",
  "references": [
    {
      "order": 1,
      "id": "headset_product",
      "type": "image",
      "role": "产品外观来源",
      "url": "<OSS>/h3_v4_ts21/headset.png",
      "note": ""
    },
    {
      "order": 2,
      "id": "reviewer_video",
      "type": "video",
      "role": "出镜者形象与口播节奏来源",
      "url": "<OSS>/h3_v4_ts21/reviewer.mp4",
      "note": ""
    }
  ]
}

资产层清单(与 JSON references 同一数组,此处单列以示数据完整性)

order id type 正文层落位 url(原样)
1 headset_product image <Subject 2> 定义句内联引用(<Picture 1> 序号),不另起独立条目 <OSS>/h3_v4_ts21/headset.png
2 reviewer_video video <Video 1> 独立条目(retention_analysis 标 weak_reference) <OSS>/h3_v4_ts21/reviewer.mp4