Pin Stage2 default duration to 10s and add default dance-video prompt template

Duration: fixed 10s (within H3's 4-15s hard constraint), frames=240 @ 24fps. Prompt: follows the internal H3 prompt-writing methodology already sitting in the repo (docs/h3skill.md, from a parallel session) — the "all-purpose reference mode" six-section structure. Character appearance details are auto-captioned from the Stage1 composed image (new composed_images. appearanceDescriptor field); dance motion phrasing is admin-authored once per reference video (new dance_refs.motionDescriptor field), since the format requires concrete descriptors and forbids "as shown in the picture." Flagged as needing a pass through h3skill.md's own 22-item mechanical acceptance checklist before production use, not claimed as fully compliant. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SdtXTsDRz5QWMjvt1HBXLz

Pin Stage2 default duration to 10s and add default dance-video prompt template
Duration: fixed 10s (within H3's 4-15s hard constraint), frames=240 @ 24fps. Prompt: follows the internal H3 prompt-writing methodology already sitting in the repo (docs/h3skill.md, from a parallel session) — the "all-purpose reference mode" six-section structure. Character appearance details are auto-captioned from the Stage1 composed image (new composed_images. appearanceDescriptor field); dance motion phrasing is admin-authored once per reference video (new dance_refs.motionDescriptor field), since the format requires concrete descriptors and forbids "as shown in the picture." Flagged as needing a pass through h3skill.md's own 22-item mechanical acceptance checklist before production use, not claimed as fully compliant. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SdtXTsDRz5QWMjvt1HBXLz
yibo
1 parent e8e21050
Showing 2 changed files with 111 additions and 4 deletions
docs/todo/dressup-dance-room-mvcp.md
docs/todo/dressup-dance-room.md
@@ -82,6 +82,7 @@
 | `videoUrl` | string | |
 | `styleTags` | array | |
 | `durationSec` | number | |
+| `motionDescriptor` | string | **ç®¡ç†å‘˜æ‰‹å†™çš„åŠ¨ä½œæè¿°**ï¼ˆå¦‚"æ‹‰ä¸é£Žæ ¼æ‘†èƒ¯+ä¸¾è‡‚æ—‹è½¬"ï¼‰ï¼Œæ‹¼è¿› Stage2 prompt æ¨¡æ¿çš„ `detailed_description`ï¼ˆå·²ç¡®è®¤ï¼Œéœ€æ±‚æ–‡æ¡£ 4.2/6.3 èŠ‚ï¼‰ï¼Œä¸Šä¼ æ—¶å¿…å¡« |
 | `copyrightSource` | enum | `test_placeholder` \| `self_made` \| `licensed`ï¼ˆç¬¬ä¸€æœŸå…¨éƒ¨ `test_placeholder`ï¼Œ6.3/8 èŠ‚ï¼‰ |
 | `isActive` | boolean | |
 | `createdAt` | number (ts) | |
@@ -95,6 +96,7 @@
 | `garmentId` | string | å¤–é”® |
 | `imageUrl` | string | åˆæˆç»“æžœ |
 | `model` | string | å®žé™…è°ƒç”¨çš„ fal.ai æ¨¡åž‹æ ‡è¯†ï¼ˆ4.1/8 èŠ‚ï¼Œå¾…é€‰åž‹ï¼‰ |
+| `appearanceDescriptor` | string | **Stage1 å®ŒæˆåŽè‡ªåŠ¨è°ƒè§†è§‰æ¨¡åž‹æè¿°è¿™å¼ æ¢è£…å›¾**ï¼Œç”Ÿæˆ 2-3 æ¡å…·ä½“ç‰¹å¾ï¼ˆå‘è‰²/å‘åž‹ã€æœè£…é¢œè‰²æ¬¾å¼ç‰ï¼‰ï¼Œæ‹¼è¿› Stage2 prompt æ¨¡æ¿çš„ `subject_definitions`/`detailed_description`ï¼ˆå·²ç¡®è®¤ï¼Œéœ€æ±‚æ–‡æ¡£ 4.2 èŠ‚ï¼‰ã€‚ç”Ÿæˆå¤±è´¥ä¸é˜»å¡žâ€”â€”ç•™ç©ºåˆ™ Stage2 ç›´æŽ¥æ‹’ç»æäº¤å¹¶é‡è¯•æ‰“æ ‡ï¼Œè€Œä¸æ˜¯æ‹¿"å¦‚å›¾æ‰€ç¤º"è¿™ç§ä¸åˆè§„æè¿°ç¡¬å‡‘ï¼ˆH3 è§„èŒƒæ˜Žç¡®ä¸å…è®¸ï¼‰ |
 | `status` | enum | `generating` \| `done` \| `failed` |
 | `createdAt` | number (ts) | |
@@ -157,6 +159,78 @@ erDiagram
   garments ||--o{ moderation_log : "å®¡æ ¸è®°å½•"
 ```
+### M.11 Stage2 é»˜è®¤ Prompt æ¨¡æ¿ï¼ˆå·²ç¡®è®¤ï¼Œéœ€æ±‚æ–‡æ¡£ 4.2 èŠ‚ï¼‰
+
+ä¾æ®ä»“åº“é‡Œçš„ `docs/h3skill.md`ï¼ˆå†…éƒ¨ H3 æç¤ºè¯å†™ä½œæ–¹æ³•è®ºï¼‰"å…¨èƒ½å‚è€ƒ
+æ¨¡å¼å…æ®µå¼"ç»“æž„å†™æˆã€‚**`{{...}}` æ˜¯å ä½ç¬¦**ï¼Œç”± M.6/M.5 çš„
+`appearanceDescriptor`/`motionDescriptor` å—æ®µåœ¨æ‹¼è£…æ—¶å¡«å…¥ï¼š
+
+- `<Picture 1>` = Stage1 æ¢è£…åˆæˆå›¾ï¼ˆ`composed_images.imageUrl`ï¼‰
+- `<Video 1>` = å‚è€ƒè·³èˆžè§†é¢‘ï¼ˆ`dance_refs.videoUrl`ï¼‰ï¼Œåªä½œåŠ¨ä½œ/è¿é•œ/
+  èŠ‚å¥å‚è€ƒï¼Œä¸æ˜¯è§†è§‰èº«ä»½æ¥æº
+- `<Audio 1>` = å¤ç”¨ `<Video 1>` çš„é…ä¹è½¨ï¼ˆH3 è§„èŒƒé‡Œå‚è€ƒè§†é¢‘ä¸å› ä¸º
+  å«å£°å°±è‡ªåŠ¨äº§ç”Ÿ `<Audio N>`ï¼Œéœ€è¦æ˜¾å¼å®šä¹‰ï¼Œè§ `docs/h3skill.md` H15ï¼‰
+
+```
+subject_definitions:
+<Subject 1> is the young woman whose identity is defined by <Picture 1>.
+<Picture 1> is the character reference frame: preserve her
+{{appearanceDescriptor}} exactly as shown, changing only her pose and
+motion to match the choreography described below.
+<Video 1> is the reference choreography clip that defines the camera
+rhythm and dance structure the target video should follow; it is not a
+visual identity source.
+
+summary:
+[reference generation] The target video shows <Subject 1>, wearing the
+outfit from <Picture 1>, performing {{motionDescriptor}} following the
+pacing and camera movement established in <Video 1>.
+
+retention_analysis:
+<Subject 1> (appears in [Shot 1]): fully_preserved - her face, hairstyle,
+and outfit from <Picture 1> are retained throughout.
+<Video 1>: weak_reference - only the choreography rhythm and camera
+movement pattern are referenced, no visual content is retained.
+<Audio 1>: fully_copy - the background music track from <Video 1> is
+reused verbatim.
+
+detailed_description:
+A vibrant, mobile-first vertical dance video with warm, glossy studio
+lighting. [Shot 1] <Subject 1>, established from <Picture 1> with
+{{appearanceDescriptor}}, stands centered in a medium-full shot against a
+softly lit backdrop. She performs {{motionDescriptor}}, her weight
+shifting smoothly on the beat, arms tracing clean arcs in rhythm with the
+music. The camera holds mostly static with a slow, subtle Push In across
+the shot, keeping her fully in frame throughout. Her expression is
+confident and playful, eyes occasionally meeting the camera. Fabric of
+her outfit moves naturally with each step and turn. The routine follows
+the pacing and structure established in <Video 1> â€” building energy
+through the middle of the clip before settling. The ending pose returns
+close to the opening stance so the clip can loop cleanly for repeated
+live playback.
+
+overall_soundscape:
+Soft ambient room tone under the music, with faint fabric rustle as
+<Subject 1> moves and rhythmic footfall on the studio floor.
+
+non_diegetic_music:
+An upbeat track carries the routine; its tempo, instrumentation, and
+dynamics are reused from <Audio 1>, driving the pacing of her movement
+throughout.
+```
+
+**å¡«å…¥ç¤ºä¾‹**ï¼ˆä»…ç¤ºæ„ï¼ŒéžçœŸå®žç´ æï¼‰ï¼š`{{appearanceDescriptor}}` =
+"shoulder-length wavy black hair, a fitted red qipao with gold trim, and
+a slender, graceful build"ï¼›`{{motionDescriptor}}` = "a lively
+Latin-inspired routine with hip sways and raised-arm turns"ã€‚
+
+**å·²çŸ¥æœªè¿‡æœºæ¢°éªŒæ”¶çš„åœ°æ–¹**ï¼ˆå¯¹åº”éœ€æ±‚æ–‡æ¡£ 8 èŠ‚ç¬¬ 9 æ¡ï¼Œæ£å¼æŽ¥å…¥å‰éœ€è¦
+æŒ‰ `docs/h3skill.md` Â§1.7 çš„ 22 é¡¹æ¸…å•é€æ¡æ ¸å®žï¼Œè¿™é‡Œä¸å‡è£…å·²ç»è¿‡å®¡ï¼‰ï¼š
+`detailed_description` çš„è‹±æ–‡è¯æ•°éœ€è¦å¡åœ¨ 350â€“500 è¯åŒºé—´ï¼ˆH6ï¼‰â€”â€”ä¸Šé¢
+è¿™ç‰ˆæ˜¯æ‰‹å†™çš„ç¤ºä¾‹é•¿åº¦ï¼Œæ²¡æœ‰åšé€è¯è®¡æ•°ï¼›`retention_analysis` ç”¨çš„å››ä¸ª
+å–å€¼ï¼ˆ`fully_preserved`/`weak_reference`/`fully_copy`ï¼‰éƒ½åœ¨ H9 ç™½åå•
+å†…ï¼Œä½†æ•´ä½“å…æ®µçš„äº”å¤„æ ‡ç¾åŒä¹‰ï¼ˆH14ï¼‰ç‰ç»†èŠ‚å»ºè®®æ£å¼æŽ¥å…¥å‰å†æ ¸ä¸€éã€‚
+
 ---
 ## V â€” Viewï¼šå‰åŽç«¯ç•Œé¢è®¾è®¡
@@ -245,7 +319,8 @@ POST /api/admin/credit/adjust               # äººå·¥è°ƒæ•´ç§¯åˆ†ï¼ˆå®¢æœåœºæ™¯ï
 å†…éƒ¨ï¼ˆç”Ÿæˆæµæ°´çº¿ç¼–æŽ’ï¼Œä¸å¯¹å‰ç«¯æš´éœ²ï¼‰ï¼š
 generateComposedImage(characterId, garmentId)      # Stage1ï¼ŒæŸ¥ç¼“å˜(M.6å”¯ä¸€ç´¢å¼•)â†’æœªå‘½ä¸è°ƒfal.ai
-generateDanceVideo(composedImageId, refDanceVideoId, backend)  # Stage2ï¼ŒbackendæŒ‰4.3èŠ‚è·¯ç”±
+captionComposedImage(composedImageId)              # Stage1å®ŒæˆåŽè‡ªåŠ¨æ‰“æ ‡ï¼Œå†™ appearanceDescriptorï¼ˆM.11ï¼‰
+generateDanceVideo(composedImageId, refDanceVideoId, backend)  # Stage2ï¼ŒæŒ‰ M.11 æ¨¡æ¿æ‹¼ promptï¼ŒbackendæŒ‰4.3èŠ‚è·¯ç”±
 ```
 ---
@@ -304,6 +379,9 @@ sequenceDiagram
   else æœªå‘½ä¸
     S->>Fal: Stage1 æ¢è£…åˆæˆ
     Fal-->>S: composedImageUrl
+    S->>Fal: captionComposedImage è‡ªåŠ¨æ‰“æ ‡ï¼ˆM.11ï¼Œç”Ÿæˆ appearanceDescriptorï¼‰
+    Fal-->>S: 2-3 æ¡å¤–è§‚ç‰¹å¾æ–‡å—
+    S->>S: æŒ‰ M.11 æ¨¡æ¿æ‹¼è£… Stage2 prompt
     S->>H3: Stage2ï¼ˆå†·å¯åŠ¨ç®—åŠ›ï¼Œ4.3èŠ‚ï¼‰
     H3-->>S: è§†é¢‘å®Œæˆ
     S->>S: å†™å…¥ generated_videos + composed_images ç¼“å˜
@@ -88,12 +88,32 @@
 后确定），复用 `server.js` 已有的 `/fal/*` 反代通道——H3 Console 本身
 不产出图片，Stage1 无论如何都得走别的模型。
-### 4.2 Stage 2：换装参考图 + 参考跳舞视频 → 生成跳舞视频
+### 4.2 Stage 2：换装参考图 + 参考跳舞视频 → 生成跳舞视频（已确认：默认时长 10 秒）
 换装图（`reference_image_urls`）+ 参考跳舞视频（`reference_video_urls`，
 提供舞蹈动作/运镜/配乐节奏，系统随机/轮换挑选，观众不选风格）→ 生成
-10 秒左右的跳舞视频片段。产物字段是 v1 `library.json` clip 的超集，加
-`characterId`/`garmentId`/`composedImageId`/`refDanceVideoId` 溯源字段。
+**固定 10 秒**的跳舞视频片段（H3 输出时长硬约束是 4–15 秒，10 秒是我们
+在这个区间内选定的固定默认值，不随请求变化；对应 `frames = 240`，按
+`H3-CONSOLE.md` 的 frames/24fps=秒 换算公式）。产物字段是 v1
+`library.json` clip 的超集，加 `characterId`/`garmentId`/
+`composedImageId`/`refDanceVideoId` 溯源字段。
+
+**Stage2 的 prompt 遵循内部 H3 提示词写作方法论**（仓库里的
+`docs/h3skill.md`——这是另一并行工作产出的官方 H3 提示词写作规范，
+内容独立于本项目其他文档但直接适用，复用而不是另起一套），走它定义的
+"全能参考模式六段式"结构（`subject_definitions` → `summary` →
+`retention_analysis` → `detailed_description` → `overall_soundscape` →
+`non_diegetic_music`）。具体默认模板见
+`docs/todo/dressup-dance-room-mvcp.md` 新增章节，两处需要自动/人工填空
+才能拼出合规 prompt（该规范要求人物特征必须具体，不能写"如图所示"）：
+
+- **人物外观描述**（发色/发型/服装颜色款式等 2-3 条具体特征）：Stage1
+  换装图生成完成后，**自动调一次视觉模型描述该图**，生成这几条特征
+  文字，随 `composed_images` 记录一起落库，不用每次生成视频都重新
+  描述（已确认）。
+- **舞蹈动作描述**（"拉丁风格摆胯+举臂旋转"这类具体动作文字）：参考
+  跳舞视频总量有限（后台预置的一个小集合），**由管理员在上传参考视频
+  时手写一段短描述**，一次性录入、长期复用（已确认，见 6.3 节）。
 ### 4.3 双算力策略（已确认）
@@ -201,6 +221,9 @@ Stage1（换装图合成）在两条路径下大概率都是 fal.ai（H3 Console
 ### 6.3 参考跳舞视频/音乐管理
 - CRUD：上传参考跳舞视频（含配乐），命名、风格标签、时长、上下架开关。
+- **动作描述字段（已确认，4.2 节）**：上传时管理员手写一段短描述具体
+  动作（如"拉丁风格摆胯+举臂旋转"），供 Stage2 prompt 模板拼装用，
+  一次录入、后续所有用到这条参考视频的生成都复用，不必每次重写。
 - 版权来源字段：当前**先用测试素材占位**（正式版权素材待确认，见第 8
   节），后台需要能记录/后续批量替换。
@@ -288,6 +311,12 @@ Stage1（换装图合成）在两条路径下大概率都是 fal.ai（H3 Console
 7. **"我的作品"页面的第三方实名认证需求**：第一期只做"我已满18岁"
    勾选式确认（5.1 节），是否需要更严格的第三方实名认证，看实际运营
    风险再评估，不预设一定要做。
+8. **人物外观自动打标的视觉模型选型**：Stage1 完成后自动描述换装图
+   （4.2 节），具体调哪个视觉模型/接口待测试几个候选。
+9. **Stage2 默认 prompt 模板的机械验收**：`docs/h3skill.md` 附带一套
+   22 项输出前机械验收清单（含 `detailed_description` 350–500 英文词
+   区间等硬约束），MVCP 文档里给的模板是按该规范手写的示例，正式接入
+   前建议照那份清单逐条过一遍再定稿。
 ## 9. 分期建议（不是承诺的排期，是依赖顺序参考）