TS-37.md
12.2 KB
TS-37 交付件
一、最终提示词原文(① 型 · 全能参考模式六段 · 12 秒)
subject_definitions: <Subject 1> is the young woman defined by <Picture 2>, in her early twenties, with the shoulder-length dark brown hair tucked behind her right ear, the cream ribbed knit sweater with rolled cuffs, and the thin gold chain necklace established in <Picture 2>.
<Picture 1> defines the coffee-shop interior the entire video takes place in: a small window-side table with a white ceramic mug on a light-wood surface, a deep-blue painted wall panel behind the table, and pale daylight coming through the window at camera left.
<Audio 1> is the voice-timbre reference for <Subject 1> (S1), providing her clear, lightly husky mid-register speaking timbre; the spoken content inside <Audio 1> is not carried into the target video.
summary: [reference generation + audio reference] A young woman sits alone at a window-side table in a coffee shop for twelve seconds, says one short line toward the phone in her hands, and the light on her face and on the tabletop shifts from warm amber to cold blue-grey across the two shots.
retention_analysis: <Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - her shoulder-length dark brown hair, cream ribbed knit sweater, and thin gold chain necklace are retained across both shots.
<Picture 1> (defines the setting of [Shot 1] and [Shot 2]): fully_preserved - the window-side table, the white ceramic mug, and the deep-blue wall panel are retained.
<Picture 2> (defines <Subject 1>): attribute_transfer - only the woman's appearance is carried into the target video; the pose, background, and framing of <Picture 2> are not reproduced.
<Audio 1> (heard as <Subject 1>'s voice): reference - only her speaking timbre is taken; the spoken content inside <Audio 1> is not reused.
detailed_description: Live-action, cinematic, with soft practical lighting and a shallow depth of field that keeps the woman sharp while the counter behind her falls into a gentle blur.
[Shot 1] The camera holds a medium shot at her eye level as she sits at the window-side table from <Picture 1>, both hands cradling her phone, warm amber pendant light falling from above onto her cheekbone, onto the cream ribbed knit sweater of <Subject 1>, and onto the rim of the white ceramic mug. The phone's screen throws a small pale rectangle of light up under her chin, and the mug's handle points toward the window at camera left. Her thumb moves once across the screen, then she lifts her head and reads the display again. <Subject 1> (S1) says: <d>[English] See you tomorrow.</d> Her lips move in time with the line while the warm light keeps her skin golden and the deep-blue wall panel behind her stays two stops darker, its paint texture swallowed by the shadow. Daylight through the window lays a soft cool edge along her left shoulder, the window glass holds a faint reflection of the pendant lamp, and the steam above the mug drifts left through the warm beam, thinning as it rises.
[Shot 2] At 00:05.500, the camera cuts to a slightly wider framing at the same eye level, and the neon sign reading "営業中" above the service counter enters the top of the frame, its red and white tubes humming faintly against the glass. The pendant light above the table has cooled to a flat blue-grey, and that colder light now lays a hard edge along her jaw, along the mug's rim, and across the light-wood surface, while the window behind her has turned to early night and carries a small doubled reflection of the sign. <Subject 1> lowers the phone a few centimetres onto the tabletop, her eyes staying on its screen, and the cold light spreads slowly from the table edge toward her rolled cuffs, dimming the amber in the knit and leaving the porcelain of the mug the brightest object on the table as the frame holds on her and on the glowing sign.
overall_soundscape: Low coffee-shop room tone runs under both shots, with the hiss and knock of an espresso machine behind the counter and a few distant cups set down on saucers. Near the window, the steam above the mug is silent and the room reads quieter than the counter. As the light cools, the machine's knocks thin out and leave only its low hum under the room tone.
non_diegetic_music: N/A
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": "shop_scene",
"type": "image",
"role": "场景参考(咖啡店内景)",
"url": "<OSS>/h3_v5_ts37/shop.png"
},
{
"id": "woman_ref",
"type": "image",
"role": "人物参考(锁定外观)",
"url": "<OSS>/h3_v5_ts37/woman.png"
},
{
"id": "voice_ref",
"type": "audio",
"role": "音色参考(只参考音色,不复制台词)",
"url": "<OSS>/h3_v5_ts37/voice.wav"
}
]
}
素材落位说明(说明层,不进 JSON):shop_scene → <Picture 1> 独立定义条目(场景/环境锚点);woman_ref → 人物立绘仅定义外观,不另起 <Picture> 独立条目,在 <Subject 1> 定义句内以 <Picture 2> 序号内联引用认领;voice_ref → <Audio 1>(音色参考,对应 <Subject 1> (S1))。三条素材 note 均为空、全部可用、全部被认领,无 excluded。
三、交付说明
- 形态判定(§0 第 0 步):需求明确要「英文结构化 Prompt(official 六段)」→ 单件 ① 型,未拆链路四层。正文英文(doc2:242 硬判据:扩写的六个段落正文统一使用英文);
<d>内台词与画面实际可见文字保留原语言——台词是英文原句、霓虹招牌是日文原文,均逐字保留。 - 模式判定:场景图 + 人物外观图 + 音色参考,多素材身份/风格锁定 → 全能参考,六段顺序按 doc2:250。
- summary 任务类型标记:
[reference generation + audio reference]。+ audio reference来自音色参考素材(§1.7 第 21 项);本单唯一帧相关输入是场景图但 raw_brief 未声明它是首帧锚点,故不标keyframe completion,也不写 I2VA 指令行。 duration_seconds取值:12(raw_brief 给定,落在 4-15s 内)。- 霓虹招牌「営業中」:按画面实际可见文字处理(doc2:154),英文双引号包围、日文原文与标点逐字保留、不翻译、不简繁/大小写改写;这类可见文字同样受口径 V 保真约束。招牌在画面中的位置(服务台上方)raw_brief 未指定,这是我落的拍摄位置假设,只占用画面位置、不改文字内容;如你方素材中招牌在别处,告诉我位置即可整句替换。
- 音色参考不带入原台词(doc2:577):
<Audio 1>定义句与retention_analysis均显式声明只取音色、其口播内容不进目标视频;绑定关系写在<Audio 1>定义句内并复用<Subject 1>的全局说话人 ID(<Subject 1> (S1)),未在音频定义处独立重新编号(H15)。 - 台词口径:「See you tomorrow.」由用户在本次输入中直接给出 → 口径 V:
<d>[English] See you tomorrow.</d>,原词、原语言、原标点(句末句号)逐字保留,未补词、未增删标点、未改写。 - 说话人 ID:
(S1)唯一发声者;(Sx)只出现在detailed_description,retention_analysis零(Sx)。 - 灯光从暖转冷的具体落法(§6 三问法):暖段光源 = 桌上方吊灯的琥珀光(从上来,打在颧骨/毛衣杯沿,使肤色偏金、背景蓝墙暗两档);冷段光源 = 同一吊灯转成扁平蓝灰 + 窗外入夜(打在下颌/杯沿/木桌面,形成硬边),切口同时带进霓虹招牌这一新信息,切口成立(§4:切镜带来新信息)。
- 运镜:
[Shot 1]与[Shot 2]各自只有一个主导机位状态(medium shot / slightly wider framing),无矛盾运镜;未使用环绕写法。 non_diegetic_music显式写N/A:raw_brief 未要求配乐,按 §6 留空会被理解为自由发挥,故显式写 N/A;环境声非 N/A(未要求全程静音)。- C 级构造声明(非官方条款):人物立绘不另起
<Picture>条目、场景图独立成条这一落位分工来自官方示例的默认形态 + 本 SKILL §2 收口;「招牌位置假设」与「无配乐 → N/A」为本单工程决定。 - H13 素材总闸:图 2(≤9)+ 音频 1(≤3,且有图/视频同单)= 3 个文件(≤12),组合合法。
- 出处纪律:正文无官方外编号词法;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)
- ✓ 形态判定正确:英文结构化需求 → 单件 ① 型,未硬拆链路四层。
- ✓ 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。
- ✓ 六个字段名拼写与官方一致、无自造字段;全能参考六段无指令行,H4 不适用。
- ✓ 标签实测:
<Subject 1>/<Picture 1>/<Picture 2>/<Audio 1>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。 - ✓
references数组顺序 = 菜单 order 1→3,三条记录均携带原 url 原文未改链;woman_ref不做<Picture>独立条目属合法形态(正文<Picture>定义条目数 1 ≠ 资产层数 3,不强制一一对应)。 - ✓ 非锚点图认领:
<Picture 2>(人物外观图)由<Subject 1>定义句以序号内联引用认领;<Picture 1>有独立定义条目。 - ✓ 唯一句台词用
<d>[English] …</d>包裹,<d>内只有语言标签与台词;口径 V 逐字保真。 - ✓ 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。
- ✓ 无设定冲突(暖光→冷光是需求给定,未与场景图事实冲突);未添加未经授权的剧情结果——她只放下手机、目光留在屏幕上,画面停在开放式状态,未写离开/回复/情绪结论。
- ✓ 正文为官方字段结构散文,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)。 - ✓ H13 实测:图 2 ≤ 9,音频 1 ≤ 3 且非单独输入,文件总数 3 ≤ 12。
- ✓ 正文层编号词法实测:除官方四标签与
(S1)外 0 处其它编号词法(grepobject/background/picture均 0 处)。 - ✓ H1 实测:最终 Prompt 共 4178 字符(含标点与空白,
len()口径)≤ 7000。 - ✓ H2 实测:单段 12 s,落在 4–15 s;正文无独立「总时长 X 秒」指令行;未提及帧率,H3 不触发。
- ✓ H5 实测:
[Shot 1]无时间戳;[Shot 2] At 00:05.500毫秒 3 位、唯一切镜时刻、递增且落在 0–12 s 内。H5b 实测:① 型正文 0 处[N 秒-M 秒]窗式记号(grep 复核)。 - ✓ H6 实测:
detailed_description共 362 个英文词(按空白分词,对该字段正文计量),落在 350–500 区间,且与 raw_brief「控制在 400 词左右」的意图一致。 - ✓ H10 实测:
overall_soundscape3 句(1–4);non_diegetic_music显式N/A(无配乐必须显式写的条件成立)。 - ✓ H11 实测:唯一台词 3 词落在 0–5.5 s 的
[Shot 1]内,无失衡;H16 实测:<Picture 1>/<Picture 2>均非分镜草稿图,全片 0 处秒级对齐表述,3 条素材全部被认领。 - ✓ 本节自报数字均先实测后写盘:4178 字符、362 词、3 句、
retention_analysis0 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,url字段只放 url。 - ✓
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)。 - ✓ summary 任务类型标记与输入对齐:音色参考 →
[reference generation + audio reference];无帧锚点输入 → 未标keyframe completion。