feat: stream subject packs by generation batch

RULES.md

@@ -11,11 +11,11 @@
 - 详见 `CLAUDE.md` 立项决策段 + `.memory/plan.md` 七步管线拆解
 - 风格：`04-Dark-Gallery-Ambient`（路径：`~/Projects/research/20260305-网页风格库/04-Dark-Gallery-Ambient.md`）
 - 第一冲刺：步骤 1-4（下载 / 拆轨 / 关键帧 / ASR+翻译）
-- 当前产品方向（2026-05-19 再确认）：信息流广告快速复刻默认进入“三字段候选生成”工作流。主界面为“左侧素材输入列 + 右侧信息流复刻工作表”。用户粘贴 TK 链接或上传视频后点击“开始分析”，系统自动下载源视频；下载完成后并行启动两条路：音频文案路提取原音频文案/字幕，并分析讲话人、语速节奏、背景音乐/环境声/音效；视频视觉路自动抽取参考帧。源视频工作区右侧主体链路是“参考帧池 → 转换层 → 主体元素”：参考帧池竖向排列；转换层只保留真人重构、卡通重构、元素重构、自主描述四个入口，每个入口最多拖入 3 张参考帧，拖入只加入参考队列，不自动生成；用户放好参考和文字后点击生成，右侧主体元素区按重构类型分组展示全新 6 视图主体。这四类都属于参考重构，不抠图、不复制原人、不复刻原画面。旧下方“相似主体 / 主体模板库”不再作为主路径。波形下方的画面胶片只是临时预览，点击只跳转原视频时间点，双击或拖进参考帧池才正式加入关键帧，已加入的胶片直接显示“已添加”。产品图上传后独立形成产品资产包，自动识别视角/结构/比例并补缺角度。分镜工作台按逐句时间轴默认只露“文案 / 场景一句话 / 人物+产品+动作”，产品素材池、批量控制、三字段、视频候选和高级区都必须可折叠；视频候选无内容时默认不占大面积，有候选时默认只显示迷你缩略条，展开后才显示 4-grid。单条默认生成 4 个视频候选，顶部支持整片批量生成候选；首尾帧、视觉规划、产品出现方式和旧 6 字段保留在“高级”抽屉与后端 quick-plan 自动展开中，不能再作为客户默认闸门。
+- 当前产品方向（2026-05-19 再确认）：信息流广告快速复刻默认进入“三字段候选生成”工作流。主界面为“左侧素材输入列 + 右侧信息流复刻工作表”。用户粘贴 TK 链接或上传视频后点击“开始分析”，系统自动下载源视频；下载完成后并行启动两条路：音频文案路提取原音频文案/字幕，并分析讲话人、语速节奏、背景音乐/环境声/音效；视频视觉路自动抽取参考帧。源视频工作区右侧主体链路是“参考帧池 → 转换层 → 主体元素”：参考帧池竖向排列；转换层只保留真人重构、卡通重构、元素重构、自主描述四个入口，每个入口最多拖入 3 张参考帧，拖入只加入参考队列，不自动生成；用户放好参考和文字后点击生成，右侧主体元素区按每次生成的套图文件夹展示全新 6 视图主体，当前套图在最上层展开，其他套图顺位进入下方可滚动列表，同一重构方向允许保留多套。这四类都属于参考重构，不抠图、不复制原人、不复刻原画面。旧下方“相似主体 / 主体模板库”不再作为主路径。波形下方的画面胶片只是临时预览，点击只跳转原视频时间点，双击或拖进参考帧池才正式加入关键帧，已加入的胶片直接显示“已添加”。产品图上传后独立形成产品资产包，自动识别视角/结构/比例并补缺角度。分镜工作台按逐句时间轴默认只露“文案 / 场景一句话 / 人物+产品+动作”，产品素材池、批量控制、三字段、视频候选和高级区都必须可折叠；视频候选无内容时默认不占大面积，有候选时默认只显示迷你缩略条，展开后才显示 4-grid。单条默认生成 4 个视频候选，顶部支持整片批量生成候选；首尾帧、视觉规划、产品出现方式和旧 6 字段保留在“高级”抽屉与后端 quick-plan 自动展开中，不能再作为客户默认闸门。
 
 ## 部署事实
 - 平台：VPS `76.13.31.179`（Ubuntu 24.04 / Docker Compose / Coolify Traefik）
-- 发布状态：已部署并验证（2026-05-19，转换层改为真人重构 / 卡通重构 / 元素重构 / 自主描述四个入口，每个入口最多 3 张参考帧；拖入只加入参考队列，点击生成后固定生成全新 6 视图，主体元素按重构类型分组展示；胶片双击/拖拽加入参考帧池 + 胶片缓存复用 + 音频解析失败可重试，右侧三栏主体管线：竖向参考帧池 + 转换层 + 主体元素，参考帧缩略图保持小尺寸 9:16 比例 + hover 左侧紧凑预览 + 转换层多参考滚动，旧主体模板区移出主路径 + 逐句时间轴移到原版视频下方并支持双行文案 + 波形同框时间对齐画面胶片 + 胶片密度按钮上移波形顶部 + 去分隔线 + 胶片上下错落 + body 顶层原位大放大 + 隐藏源视频工作区音频解析摘要卡 + 隐藏工作区顶部状态提示条 + 三字段候选生成工作流 + 折叠紧凑候选区）；`https://marketing.skg.com` 已启用应用内登录页，未登录 API 返回 401，认证后首页 200；容器内 `/health` 返回 `ok:true`
+- 发布状态：已部署并验证（2026-05-19，主体元素改为按套图文件夹分组展示，主体生成接口提交后立即返回 queued 占位并后台逐视角生成、逐张回填；转换层为真人重构 / 卡通重构 / 元素重构 / 自主描述四个入口，每个入口最多 3 张参考帧；拖入只加入参考队列，点击生成后固定生成全新 6 视图；胶片双击/拖拽加入参考帧池 + 胶片缓存复用 + 音频解析失败可重试，右侧三栏主体管线：竖向参考帧池 + 转换层 + 主体元素，参考帧缩略图保持小尺寸 9:16 比例 + hover 左侧紧凑预览 + 转换层多参考滚动，旧主体模板区移出主路径 + 逐句时间轴移到原版视频下方并支持双行文案 + 波形同框时间对齐画面胶片 + 胶片密度按钮上移波形顶部 + 去分隔线 + 胶片上下错落 + body 顶层原位大放大 + 隐藏源视频工作区音频解析摘要卡 + 隐藏工作区顶部状态提示条 + 三字段候选生成工作流 + 折叠紧凑候选区）；`https://marketing.skg.com` 已启用应用内登录页，未登录 API 返回 401，认证后首页 200；容器内 `/health` 返回 `ok:true`
 - 主站 / 前端：`https://marketing.skg.com`
 - API / 后端：`https://marketing.skg.com/api`
 - 代码仓库 / Gitea：`https://git.kang-kang.com/kangwan/20260512-skg-tk`

api/main.py

@@ -289,6 +289,7 @@ AssetSize = Literal["source", "1024", "1536", "2048"]
 AssetQuality = Literal["hd"]
 SubjectKind = Literal["object", "living"]
 SubjectView = str
+SubjectAssetStatus = Literal["queued", "in_progress", "completed", "failed"]
 SceneMode = Literal["remove_subject", "similar", "style"]
 SceneStyle = Literal["source", "premium_product", "clean_studio", "warm_lifestyle", "cinematic"]
 SceneAssetRole = Literal["scene", "first_frame", "last_frame"]
@@ -462,6 +463,13 @@ class SubjectAsset(BaseModel):
     size: AssetSize = "source"
     source_frame_indices: list[int] = Field(default_factory=list)
     ai_completed: bool = True
+    status: SubjectAssetStatus = "completed"
+    progress: int = 100
+    error: str = ""
+    pack_id: str = ""
+    pack_label: str = ""
+    pack_mode: str = ""
+    pack_created_at: float = 0.0
     created_at: float = 0.0
 
 
@@ -1371,6 +1379,26 @@ async def lifespan(_: FastAPI):
                         audio_script=audio_script,
                         message="服务重启 · 上次音频处理已中断，可重新处理",
                     )
+                subject_generation_interrupted = False
+                recovered_frames = []
+                for f in job.frames:
+                    for e in f.elements or []:
+                        recovered_assets = []
+                        for asset in e.subject_assets or []:
+                            if asset.status in {"queued", "in_progress"}:
+                                recovered_assets.append(asset.model_copy(update={
+                                    "status": "failed",
+                                    "progress": 100,
+                                    "error": "服务重启 · 上次主体生成已中断，可重新生成",
+                                    "ai_completed": False,
+                                }))
+                                subject_generation_interrupted = True
+                            else:
+                                recovered_assets.append(asset)
+                        e.subject_assets = recovered_assets
+                    recovered_frames.append(f)
+                if subject_generation_interrupted:
+                    update(job, frames=recovered_frames, message="服务重启 · 上次主体生成已中断，可重新生成")
                 JOBS[p.name] = job
             except Exception:
                 pass
@@ -4793,6 +4821,11 @@ class GenerateSubjectAssetsReq(BaseModel):
     subject_profile: SubjectProfilePreference | None = None
     prompt: str = ""
     replace_views: bool = False
+    source_subject_brief: str = ""
+    pack_id: str = ""
+    pack_label: str = ""
+    pack_mode: str = ""
+    pack_created_at: float = 0.0
 
 
 def _subject_profile_prompt_clause(profile: SubjectProfilePreference | None) -> str:
@@ -5252,8 +5285,195 @@ def cutout_element(job_id: str, idx: int, element_id: str) -> Job:
     return job
 
 
+def _subject_source_indices(req: GenerateSubjectAssetsReq, idx: int) -> list[int]:
+    source_indices = [int(x) for x in (req.source_frame_indices or [idx]) if isinstance(x, int) or str(x).isdigit()]
+    if idx not in source_indices:
+        source_indices = [idx] + source_indices
+    return list(dict.fromkeys(source_indices))[:12]
+
+
+def _normalize_subject_pack_id(value: str, idx: int, element_id: str) -> str:
+    cleaned = "".join(ch for ch in (value or "").strip() if ch.isalnum() or ch in {"_", "-"})
+    return cleaned[:96] or f"subject_pack_{idx:03d}_{element_id}_{uuid.uuid4().hex[:8]}"
+
+
+def _update_subject_asset_status(
+    job_id: str,
+    idx: int,
+    element_id: str,
+    asset_id: str,
+    *,
+    status: SubjectAssetStatus,
+    progress: int,
+    error: str = "",
+    message: str = "",
+) -> None:
+    job = JOBS.get(job_id)
+    if not job:
+        return
+    new_frames = []
+    for f in job.frames:
+        if f.index == idx:
+            for e in f.elements:
+                if e.id == element_id:
+                    updated_assets = []
+                    for asset in e.subject_assets or []:
+                        if asset.id == asset_id:
+                            updated_assets.append(asset.model_copy(update={
+                                "status": status,
+                                "progress": max(0, min(100, int(progress))),
+                                "error": error,
+                                "ai_completed": status == "completed",
+                            }))
+                        else:
+                            updated_assets.append(asset)
+                    e.subject_assets = updated_assets
+        new_frames.append(f)
+    update(job, frames=new_frames, message=message or job.message, error=error if status == "failed" else job.error)
+
+
+def _subject_assets_background_worker(
+    job_id: str,
+    idx: int,
+    element_id: str,
+    req: GenerateSubjectAssetsReq,
+    queued: list[tuple[SubjectView, str, str]],
+) -> None:
+    if req.reconstruction_mode == "similar" and not req.source_subject_brief.strip():
+        try:
+            req.source_subject_brief = _describe_source_subject(job_id, _subject_source_indices(req, idx))
+        except Exception as e:
+            print(f"[subject assets] source brief failed job={job_id} error={e}", flush=True)
+    for position, (view, view_label, placeholder_id) in enumerate(queued, start=1):
+        _update_subject_asset_status(
+            job_id,
+            idx,
+            element_id,
+            placeholder_id,
+            status="in_progress",
+            progress=10,
+            message=f"主体资产生成中 · {view_label} · {position}/{len(queued)}",
+        )
+        one_req = req.model_copy(deep=True)
+        one_req.views = [view]
+        one_req.replace_views = True
+        try:
+            _generate_subject_assets_sync(job_id, idx, element_id, one_req)
+        except HTTPException as e:
+            detail = str(e.detail)
+            _update_subject_asset_status(
+                job_id,
+                idx,
+                element_id,
+                placeholder_id,
+                status="failed",
+                progress=100,
+                error=detail,
+                message=f"主体资产生成失败 · {view_label}",
+            )
+        except Exception as e:
+            detail = str(e)
+            _update_subject_asset_status(
+                job_id,
+                idx,
+                element_id,
+                placeholder_id,
+                status="failed",
+                progress=100,
+                error=detail,
+                message=f"主体资产生成失败 · {view_label}",
+            )
+
+
 @app.post("/jobs/{job_id}/frames/{idx}/elements/{element_id}/subject-assets", response_model=Job)
 def generate_subject_assets(job_id: str, idx: int, element_id: str, req: GenerateSubjectAssetsReq) -> Job:
+    """提交主体多视角生成任务，立即返回占位卡；后台逐张生成并逐张写回。"""
+    job = JOBS.get(job_id)
+    if not job:
+        raise HTTPException(404, "job not found")
+    frame = _find_frame(job, idx)
+    el = next((e for e in frame.elements if e.id == element_id), None)
+    if not el:
+        raise HTTPException(404, "element not found")
+
+    views = _subject_view_labels(req.subject_kind, req.views)
+    source_indices = _subject_source_indices(req, idx)
+    target_views = {view for view, _label in views}
+    now = time.time()
+    explicit_pack_id = bool((req.pack_id or "").strip())
+    pack_id = _normalize_subject_pack_id(req.pack_id, idx, element_id)
+    pack_label = (req.pack_label or "").strip()[:120] or f"{el.name_zh} · 主体套图"
+    pack_mode = (req.pack_mode or "").strip()[:40] or req.subject_style
+    pack_created_at = req.pack_created_at or now
+    placeholders: list[SubjectAsset] = []
+    queued: list[tuple[SubjectView, str, str]] = []
+    for view, view_label in views:
+        asset_id = f"subject_{idx:03d}_{element_id}_{view}_{uuid.uuid4().hex[:8]}"
+        placeholders.append(SubjectAsset(
+            id=asset_id,
+            view=view,
+            label=f"{el.name_zh} · {view_label}",
+            url="",
+            width=0,
+            height=0,
+            background=req.background,
+            quality=req.quality,
+            size=req.size,
+            source_frame_indices=source_indices,
+            ai_completed=False,
+            status="queued",
+            progress=0,
+            error="",
+            pack_id=pack_id,
+            pack_label=pack_label,
+            pack_mode=pack_mode,
+            pack_created_at=pack_created_at,
+            created_at=now,
+        ))
+        queued.append((view, view_label, asset_id))
+
+    new_frames = []
+    for f in job.frames:
+        if f.index == idx:
+            for e in f.elements:
+                if e.id == element_id:
+                    e.subject_kind = req.subject_kind
+                    e.cutout_background = req.background
+                    current_assets = e.subject_assets or []
+                    if req.replace_views:
+                        for old_asset in current_assets:
+                            should_replace = old_asset.view in target_views and (
+                                old_asset.pack_id == pack_id if explicit_pack_id else True
+                            )
+                            if should_replace and old_asset.url:
+                                _delete_subject_asset_file(job_id, old_asset.id)
+                        current_assets = [
+                            asset for asset in current_assets
+                            if not (
+                                asset.view in target_views and (
+                                    asset.pack_id == pack_id if explicit_pack_id else True
+                                )
+                            )
+                        ]
+                    e.subject_assets = current_assets + placeholders
+        new_frames.append(f)
+    update(job, frames=new_frames, message=f"主体资产已提交 · {el.name_zh} · {len(placeholders)} 张逐张生成中", error="")
+
+    worker_req = req.model_copy(deep=True)
+    worker_req.views = [view for view, _label in views]
+    worker_req.pack_id = pack_id
+    worker_req.pack_label = pack_label
+    worker_req.pack_mode = pack_mode
+    worker_req.pack_created_at = pack_created_at
+    threading.Thread(
+        target=_subject_assets_background_worker,
+        args=(job_id, idx, element_id, worker_req, queued),
+        daemon=True,
+    ).start()
+    return job
+
+
+def _generate_subject_assets_sync(job_id: str, idx: int, element_id: str, req: GenerateSubjectAssetsReq) -> Job:
     """为一个主体生成多视角资产包。
     如果传入 source_frame_indices 或内置 character_id，则把多张参考图作为独立 image[] 证据提交。"""
     import time as _time
@@ -5265,10 +5485,7 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
     if not el:
         raise HTTPException(404, "element not found")
 
-    source_indices = [int(x) for x in (req.source_frame_indices or [idx]) if isinstance(x, int) or str(x).isdigit()]
+    source_indices = _subject_source_indices(req, idx)
-    if idx not in source_indices:
-        source_indices = [idx] + source_indices
-    source_indices = list(dict.fromkeys(source_indices))[:12]
 
     similar_mode = req.reconstruction_mode == "similar"
     character_reference_paths: list[Path] = []
@@ -5311,7 +5528,11 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
     tmp_focus: Path | None = None
     model_src: Path | list[Path] | None = None
     frame_reference_paths = [p for p in (_source_frame_path(job_id, i) for i in source_indices) if p.exists()]
-    source_subject_brief = _describe_source_subject(job_id, source_indices) if similar_mode else ""
+    source_subject_brief = (
+        _ensure_english(req.source_subject_brief.strip())
+        if similar_mode and req.source_subject_brief.strip()
+        else (_describe_source_subject(job_id, source_indices) if similar_mode else "")
+    )
     source_subject_clause = (
         f"Source video role brief from selected keyframes: {source_subject_brief}. "
         "Use this brief to preserve role category, creator-ad energy, camera readability, and broad styling, while creating a new non-identical subject. "
@@ -5484,6 +5705,13 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
                 quality=req.quality,
                 size=req.size,
                 source_frame_indices=source_indices,
+                status="completed",
+                progress=100,
+                error="",
+                pack_id=req.pack_id,
+                pack_label=req.pack_label,
+                pack_mode=req.pack_mode,
+                pack_created_at=req.pack_created_at or _time.time(),
                 created_at=_time.time(),
             ))
     finally:
@@ -5509,10 +5737,21 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
                     current_assets = e.subject_assets or []
                     if req.replace_views:
                         replaced_views = {asset.view for asset in generated}
+                        replace_pack_id = (req.pack_id or "").strip()
                         for old_asset in current_assets:
-                            if old_asset.view in replaced_views:
+                            should_replace = old_asset.view in replaced_views and (
+                                old_asset.pack_id == replace_pack_id if replace_pack_id else True
+                            )
+                            if should_replace:
                                 _delete_subject_asset_file(job_id, old_asset.id)
-                        current_assets = [asset for asset in current_assets if asset.view not in replaced_views]
+                        current_assets = [
+                            asset for asset in current_assets
+                            if not (
+                                asset.view in replaced_views and (
+                                    asset.pack_id == replace_pack_id if replace_pack_id else True
+                                )
+                            )
+                        ]
                     final_assets = current_assets + generated
                     e.subject_assets = final_assets
                     if req.subject_kind == "living":

web/app/page.tsx

@@ -892,7 +892,12 @@ export default function Home() {
       .filter((item) => {
         const runningVideo = !!item.generated_videos?.some((v) => v.status === "queued" || v.status === "in_progress")
         const runningAudio = item.audio_script?.status === "rewriting"
-        return runningVideo || runningAudio || !TERMINAL.includes(item.status)
+        const runningSubject = item.frames.some((frame) =>
+          frame.elements?.some((element) =>
+            element.subject_assets?.some((asset) => asset.status === "queued" || asset.status === "in_progress"),
+          ),
+        )
+        return runningVideo || runningAudio || runningSubject || !TERMINAL.includes(item.status)
       })
       .map((item) => item.id)
 
@@ -913,7 +918,14 @@ export default function Home() {
   }, [
     job?.id,
     job?.status,
-    jobs.map((item) => `${item.id}:${item.status}:${item.progress}:${item.audio_script?.status ?? ""}:${item.audio_script?.voice_url ?? ""}:${item.generated_videos?.map((v) => `${v.id}:${v.status}:${v.progress}`).join(",")}`).join("|"),
+    jobs.map((item) => {
+      const subjectState = item.frames.flatMap((frame) =>
+        frame.elements?.flatMap((element) =>
+          element.subject_assets?.map((asset) => `${asset.id}:${asset.status ?? "completed"}:${asset.progress ?? 100}:${asset.url ?? ""}`) ?? [],
+        ) ?? [],
+      ).join(",")
+      return `${item.id}:${item.status}:${item.progress}:${item.audio_script?.status ?? ""}:${item.audio_script?.voice_url ?? ""}:${item.generated_videos?.map((v) => `${v.id}:${v.status}:${v.progress}`).join(",")}:${subjectState}`
+    }).join("|"),
   ])
 
   const [pinnedNodes, setPinnedNodes] = useState<Set<string>>(() => new Set(loadNodePins()))

web/components/ad-recreation-board.tsx

@@ -214,6 +214,20 @@ type ResolvedSubjectProfile = {
   promptSummary: string
   payload: SubjectProfilePreference
 }
+type SubjectAssetPack = {
+  key: string
+  id: string
+  label: string
+  mode: SubjectReconstructionMode
+  frame: KeyFrame
+  element: KeyElement
+  createdAt: number
+  assets: SubjectAsset[]
+  total: number
+  completed: number
+  failed: number
+  running: boolean
+}
 type StoryboardVisualMode = NonNullable<StoryboardScene["visual_mode"]>
 type RowPlanPatch = Partial<Pick<AudioStoryboardRow, "visualMode" | "needsProduct" | "needsSubject" | "subjectDescription" | "subjectDescriptionZh" | "skgCopy" | "skgCopyZh" | "sceneOneLine" | "sceneOneLineZh" | "actionOneLine" | "actionOneLineZh" | "visualPlan" | "visualPlanZh" | "firstFramePlan" | "firstFramePlanZh" | "lastFramePlan" | "lastFramePlanZh" | "productIntegration" | "productIntegrationZh" | "productPlacement" | "productPlacementZh">>
 type WorkflowStepId = "input" | "source" | "audio" | "visual" | "subject" | "product" | "script" | "scene" | "video"
@@ -1119,9 +1133,46 @@ function buildSimilarSubjectPrompt(
 }
 
 function subjectAssetUrl(job: Job, asset: SubjectAsset) {
+  if (!asset.url && asset.status && asset.status !== "completed") return ""
   return apiAssetUrl(asset.url) || resolveImageRefUrl(job.id, { kind: "asset", frame_idx: 0, element_id: asset.id })
 }
 
+function subjectAssetStatus(asset: SubjectAsset) {
+  return asset.status ?? (asset.url ? "completed" : "completed")
+}
+
+function subjectAssetIsRunning(asset: SubjectAsset) {
+  const status = subjectAssetStatus(asset)
+  return status === "queued" || status === "in_progress"
+}
+
+function subjectAssetStatusLabel(asset: SubjectAsset) {
+  const status = subjectAssetStatus(asset)
+  if (status === "queued") return "排队中"
+  if (status === "in_progress") return `生成中 ${asset.progress ?? 0}%`
+  if (status === "failed") return "失败"
+  return asset.width && asset.height ? `${asset.width}x${asset.height}` : undefined
+}
+
+function subjectAssetPackKey(frame: KeyFrame, element: KeyElement, asset: SubjectAsset) {
+  return `${frame.index}:${element.id}:${asset.pack_id || `legacy-${element.id}`}`
+}
+
+function subjectAssetPackSortAssets(assets: SubjectAsset[]) {
+  return [...assets].sort((a, b) => {
+    const ai = SUBJECT_VIEW_ORDER.indexOf(a.view)
+    const bi = SUBJECT_VIEW_ORDER.indexOf(b.view)
+    if (ai !== bi) return (ai === -1 ? 99 : ai) - (bi === -1 ? 99 : bi)
+    return (a.created_at || 0) - (b.created_at || 0)
+  })
+}
+
+function subjectAssetPackSummary(pack: SubjectAssetPack) {
+  if (pack.running) return `${pack.completed}/${pack.total} 生成中`
+  if (pack.failed) return `${pack.completed}/${pack.total} · 失败 ${pack.failed}`
+  return `${pack.completed || pack.total} 张`
+}
+
 function characterPreviewImage(character?: { primary_image?: string; images?: Array<{ id: string; view?: string; filename: string; label?: string }> } | null) {
   if (!character?.images?.length) return null
   return character.images.find((image) => image.id === character.primary_image)
@@ -3156,6 +3207,7 @@ function SourceSubjectPipeline({
   const [cartoonStyleOpen, setCartoonStyleOpen] = useState(false)
   const [subjectBusyFor, setSubjectBusyFor] = useState<{ jobId: string; jobLabel: string; mode: SubjectReconstructionMode; viewCount: number; sourceCount: number; profileLabel: string } | null>(null)
   const [subjectAssetBusy, setSubjectAssetBusy] = useState<string | null>(null)
+  const [expandedSubjectPackKey, setExpandedSubjectPackKey] = useState<string | null>(null)
   const [lastSubjectProfile, setLastSubjectProfile] = useState<ResolvedSubjectProfile | null>(null)
   const subjectBusy = !!subjectBusyFor
   const selectedSubjectViews = RECONSTRUCTION_SUBJECT_VIEW_VALUES
@@ -3182,25 +3234,65 @@ function SourceSubjectPipeline({
     }
     return items
   }, [frames])
-  const visibleActorAssets = useMemo(() => {
+  const subjectAssetPacks = useMemo<SubjectAssetPack[]>(() => {
-    const items: Array<{ frame: KeyFrame; element: KeyElement; mode: SubjectReconstructionMode; asset: SubjectAsset }> = []
+    const packs = new Map<string, SubjectAssetPack>()
     for (const source of actorSources) {
-      const latestByView = new Map<string, SubjectAsset>()
       for (const asset of source.element.subject_assets ?? []) {
+        const key = subjectAssetPackKey(source.frame, source.element, asset)
+        const rawMode = asset.pack_mode as SubjectReconstructionMode | undefined
+        const packMode = rawMode && RECONSTRUCTION_MODES.some((item) => item.value === rawMode) ? rawMode : source.mode
+        const createdAt = asset.pack_created_at || asset.created_at || 0
+        const existing = packs.get(key)
+        if (existing) {
+          existing.assets.push(asset)
+          existing.createdAt = Math.min(existing.createdAt || createdAt, createdAt)
+        } else {
+          packs.set(key, {
+            key,
+            id: asset.pack_id || key,
+            label: asset.pack_label || `${reconstructionModeConfig(packMode).label}套图`,
+            mode: packMode,
+            frame: source.frame,
+            element: source.element,
+            createdAt,
+            assets: [asset],
+            total: 0,
+            completed: 0,
+            failed: 0,
+            running: false,
+          })
+        }
+      }
+    }
+    return [...packs.values()].map((pack) => {
+      const latestByView = new Map<string, SubjectAsset>()
+      for (const asset of pack.assets) {
         const current = latestByView.get(asset.view)
         if (!current || (asset.created_at || 0) >= (current.created_at || 0)) latestByView.set(asset.view, asset)
       }
-      for (const asset of latestByView.values()) items.push({ ...source, asset })
+      const assets = subjectAssetPackSortAssets([...latestByView.values()])
-    }
+      const completed = assets.filter((asset) => subjectAssetStatus(asset) === "completed").length
-    return items.sort((a, b) => {
+      const failed = assets.filter((asset) => subjectAssetStatus(asset) === "failed").length
+      const running = assets.some(subjectAssetIsRunning)
+      return { ...pack, assets, total: assets.length, completed, failed, running }
+    }).sort((a, b) => {
       const mi = RECONSTRUCTION_MODES.findIndex((item) => item.value === a.mode)
       const mj = RECONSTRUCTION_MODES.findIndex((item) => item.value === b.mode)
       if (mi !== mj) return mi - mj
-      const ai = SUBJECT_VIEW_ORDER.indexOf(a.asset.view)
+      return (b.createdAt || 0) - (a.createdAt || 0)
-      const bi = SUBJECT_VIEW_ORDER.indexOf(b.asset.view)
-      return (ai === -1 ? 99 : ai) - (bi === -1 ? 99 : bi)
     })
   }, [actorSources])
+  const activeSubjectPack = useMemo(
+    () => subjectAssetPacks.find((pack) => pack.key === expandedSubjectPackKey) ?? subjectAssetPacks[0] ?? null,
+    [expandedSubjectPackKey, subjectAssetPacks],
+  )
+  const runningActorModes = useMemo(() => {
+    const next = new Set<SubjectReconstructionMode>()
+    for (const pack of subjectAssetPacks) {
+      if (pack.running) next.add(pack.mode)
+    }
+    return next
+  }, [subjectAssetPacks])
 
   useEffect(() => {
     setConversionFrameIndicesByMode({ ...EMPTY_RECONSTRUCTION_FRAME_MAP })
@@ -3210,8 +3302,15 @@ function SourceSubjectPipeline({
     setSubjectAssetBusy(null)
     setActiveDropMode(null)
     setCartoonStyleOpen(false)
+    setExpandedSubjectPackKey(null)
   }, [job.id])
 
+  useEffect(() => {
+    if (expandedSubjectPackKey && !subjectAssetPacks.some((pack) => pack.key === expandedSubjectPackKey)) {
+      setExpandedSubjectPackKey(null)
+    }
+  }, [expandedSubjectPackKey, subjectAssetPacks])
+
   useEffect(() => {
     setConversionFrameIndicesByMode((current) => {
       const next = {} as Record<SubjectReconstructionMode, number[]>
@@ -3233,6 +3332,10 @@ function SourceSubjectPipeline({
       toast.warning("主体套图正在生成中，完成后再重生。")
       return
     }
+    if (runningActorModes.has(mode)) {
+      toast.warning(`${reconstructionModeConfig(mode).label}还有主体图正在逐张生成。`)
+      return
+    }
     const sourceFrames = sourceIndices
       .map((index) => frames.find((frame) => frame.index === index))
       .filter((frame): frame is KeyFrame => !!frame)
@@ -3288,10 +3391,18 @@ function SourceSubjectPipeline({
         views: selectedSubjectViews,
         subject_profile: requestProfile?.payload ?? null,
         prompt: buildSimilarSubjectPrompt(subjectStyle, userDirection, null, requestProfile),
-        replace_views: true,
+        replace_views: false,
+        pack_label: `${reconstructionModeConfig(mode).label} ${new Date().toLocaleTimeString("zh-CN", { hour: "2-digit", minute: "2-digit", hour12: false })}`,
+        pack_mode: mode,
       })
       onJobUpdate(updated)
-      toast.success(`${reconstructionModeConfig(mode).label}已生成：${selectedSubjectViews.length} 张`)
+      const updatedFrame = updated.frames.find((frame) => frame.index === baseFrame.index)
+      const updatedElement = updatedFrame?.elements?.find((item) => item.id === element.id)
+      const newestAsset = [...(updatedElement?.subject_assets ?? [])].sort((a, b) => (b.pack_created_at || b.created_at || 0) - (a.pack_created_at || a.created_at || 0))[0]
+      if (updatedFrame && updatedElement && newestAsset) {
+        setExpandedSubjectPackKey(subjectAssetPackKey(updatedFrame, updatedElement, newestAsset))
+      }
+      toast.success(`${reconstructionModeConfig(mode).label}已提交：${selectedSubjectViews.length} 张会逐张出来`)
     } catch (e) {
       try {
         onJobUpdate(await getJob(requestJobId))
@@ -3359,9 +3470,13 @@ function SourceSubjectPipeline({
           requestProfile,
         ),
         replace_views: true,
+        pack_id: asset.pack_id ?? "",
+        pack_label: asset.pack_label ?? "",
+        pack_mode: asset.pack_mode ?? mode,
+        pack_created_at: asset.pack_created_at ?? asset.created_at ?? 0,
       })
       onJobUpdate(updated)
-      toast.success("已重新生成这张主体元素")
+      toast.success("已提交重生，这张主体元素会生成完成后替换")
     } catch (e) {
       toast.error("主体元素重生失败：" + (e instanceof Error ? e.message : String(e)))
     } finally {
@@ -3497,6 +3612,7 @@ function SourceSubjectPipeline({
               const canGenerate = mode === "custom"
                 ? Boolean(reconstructionDirections.custom.trim() || modeFrames.length)
                 : modeFrames.length > 0
+              const modeRunning = runningActorModes.has(mode)
               return (
                 <div
                   key={mode}
@@ -3606,11 +3722,11 @@ function SourceSubjectPipeline({
                   <button
                     type="button"
                     onClick={() => void generateSubjectPack(mode)}
-                    disabled={subjectBusy || !canGenerate}
+                    disabled={subjectBusy || modeRunning || !canGenerate}
                     className="skg-primary-action mt-2 inline-flex h-8 w-full items-center justify-center gap-1 px-3 text-[10.5px] font-semibold transition disabled:cursor-not-allowed disabled:opacity-40"
                   >
-                    {subjectBusyFor?.mode === mode ? <Loader2 className="h-3.5 w-3.5 animate-spin" /> : <Sparkles className="h-3.5 w-3.5" />}
+                    {subjectBusyFor?.mode === mode || modeRunning ? <Loader2 className="h-3.5 w-3.5 animate-spin" /> : <Sparkles className="h-3.5 w-3.5" />}
-                    {subjectBusyFor?.mode === mode ? `生成中 · ${subjectBusyFor.sourceCount || "描述"} 参考` : `生成${modeConfig.label} 6视图`}
+                    {subjectBusyFor?.mode === mode || modeRunning ? "逐张生成中" : `生成${modeConfig.label} 6视图`}
                   </button>
                 </div>
               )
@@ -3628,7 +3744,7 @@ function SourceSubjectPipeline({
         <div className="mb-2 flex items-center justify-between gap-2">
           <SectionTitle icon={<Sparkles className="h-4 w-4" />} title="主体元素" />
           <span className="rounded-md border border-white/10 bg-black/35 px-2 py-1 text-[10px] text-white/42">
-            {visibleActorAssets.length ? `${visibleActorAssets.length} 张` : "待生成"}
+            {subjectAssetPacks.length ? `${subjectAssetPacks.length} 套` : "待生成"}
           </span>
         </div>
         <div className="min-h-[410px] rounded-md border border-white/10 bg-black/32 p-2 2xl:min-h-[500px]">
@@ -3638,52 +3754,90 @@ function SourceSubjectPipeline({
               <span className="mt-1 block text-cyan-50/58">主体设定：{subjectBusyFor.profileLabel}</span>
             </div>
           ) : null}
-          {visibleActorAssets.length ? (
+          {subjectAssetPacks.length ? (
-            <div className="space-y-3">
+            <div className="space-y-2">
-              {RECONSTRUCTION_MODES.map((modeConfig) => {
+              {activeSubjectPack ? (
-                const items = visibleActorAssets.filter((item) => item.mode === modeConfig.value)
+                <div className="rounded-md border border-[#d6b36a]/28 bg-[#d6b36a]/[0.07] p-2">
-                if (!items.length) return null
+                  <div className="mb-2 flex items-center justify-between gap-2">
-                return (
+                    <div className="min-w-0">
-                  <div key={modeConfig.value} className="space-y-1.5">
+                      <div className="truncate text-[11px] font-semibold text-white">{activeSubjectPack.label}</div>
-                    <div className="flex items-center justify-between gap-2 text-[10px] text-white/44">
+                      <div className="mt-0.5 text-[9.5px] text-white/42">
-                      <span>{modeConfig.label}</span>
+                        {reconstructionModeConfig(activeSubjectPack.mode).label} · {subjectAssetPackSummary(activeSubjectPack)}
-                      <span>{items.length} 张</span>
+                      </div>
-                    </div>
-                    <div className="grid grid-cols-[repeat(auto-fill,minmax(76px,1fr))] gap-2 2xl:grid-cols-[repeat(auto-fill,minmax(88px,1fr))]">
-                      {items.map((item) => {
-                        const { asset } = item
-                        const busyMode = subjectAssetBusy?.endsWith(asset.id) ? subjectAssetBusy.split(":")[0] : ""
-                        return (
-                          <MediaAssetTile
-                            key={asset.id}
-                            src={subjectAssetUrl(job, asset)}
-                            href={subjectAssetUrl(job, asset)}
-                            alt={asset.label || asset.view}
-                            label={asset.label || subjectViewLabel(asset.view)}
-                            meta={asset.width && asset.height ? `${asset.width}x${asset.height}` : undefined}
-                            className="aspect-[9/16] bg-white"
-                            objectFit="contain"
-                            title={asset.label || subjectViewLabel(asset.view)}
-                            actions={[{
-                              key: "regen",
-                              label: "重新生成这一张",
-                              icon: <RefreshCw className="h-3 w-3" />,
-                              tone: "cyan",
-                              busy: busyMode === "regen",
-                              disabled: !!subjectAssetBusy || subjectBusy,
-                              onClick: () => void regenerateSubjectAsset(item),
-                            }]}
-                            onDelete={() => void deleteActorAsset(item)}
-                            deleting={busyMode === "delete"}
-                            deleteDisabled={!!subjectAssetBusy || subjectBusy}
-                            deleteLabel="删除这一张"
-                          />
-                        )
-                      })}
                     </div>
+                    <span className="shrink-0 rounded-full border border-white/10 bg-black/35 px-2 py-0.5 text-[9px] text-white/46">
+                      {activeSubjectPack.assets.length} 张
+                    </span>
                   </div>
-                )
+                  <div className="grid max-h-[300px] grid-cols-[repeat(auto-fill,minmax(76px,1fr))] gap-2 overflow-y-auto pr-0.5 2xl:grid-cols-[repeat(auto-fill,minmax(88px,1fr))]">
-              })}
+                    {activeSubjectPack.assets.map((asset) => {
+                      const busyMode = subjectAssetBusy?.endsWith(asset.id) ? subjectAssetBusy.split(":")[0] : ""
+                      const status = subjectAssetStatus(asset)
+                      const running = subjectAssetIsRunning(asset)
+                      const failed = status === "failed"
+                      const mediaUrl = subjectAssetUrl(job, asset)
+                      const item = { frame: activeSubjectPack.frame, element: activeSubjectPack.element, mode: activeSubjectPack.mode, asset }
+                      return (
+                        <MediaAssetTile
+                          key={asset.id}
+                          src={mediaUrl || undefined}
+                          href={mediaUrl || undefined}
+                          alt={asset.label || asset.view}
+                          label={asset.label || subjectViewLabel(asset.view)}
+                          meta={subjectAssetStatusLabel(asset)}
+                          className={`aspect-[9/16] ${running ? "bg-black/40" : failed ? "bg-rose-950/30" : "bg-white"}`}
+                          objectFit="contain"
+                          busy={running}
+                          emptyText={failed ? "失败" : running ? "生成中" : undefined}
+                          title={asset.label || subjectViewLabel(asset.view)}
+                          actions={[{
+                            key: "regen",
+                            label: "重新生成这一张",
+                            icon: <RefreshCw className="h-3 w-3" />,
+                            tone: "cyan",
+                            busy: busyMode === "regen",
+                            disabled: !!subjectAssetBusy || subjectBusy || running,
+                            onClick: () => void regenerateSubjectAsset(item),
+                          }]}
+                          onDelete={() => void deleteActorAsset(item)}
+                          deleting={busyMode === "delete"}
+                          deleteDisabled={!!subjectAssetBusy || subjectBusy || running}
+                          deleteLabel="删除这一张"
+                        />
+                      )
+                    })}
+                  </div>
+                </div>
+              ) : null}
+              <div className="max-h-32 overflow-auto rounded-md border border-white/10 bg-black/24 p-1.5">
+                <div className="grid grid-cols-[repeat(auto-fill,minmax(112px,1fr))] gap-1.5">
+                  {subjectAssetPacks.map((pack, index) => {
+                    const active = activeSubjectPack?.key === pack.key
+                    return (
+                      <button
+                        key={pack.key}
+                        type="button"
+                        onClick={() => setExpandedSubjectPackKey(pack.key)}
+                        className={`min-w-0 rounded-md border px-2 py-1.5 text-left transition ${
+                          active
+                            ? "border-[#d6b36a]/70 bg-[#d6b36a]/14 text-white"
+                            : "border-white/10 bg-black/28 text-white/58 hover:border-white/22 hover:text-white"
+                        }`}
+                        title={pack.label}
+                      >
+                        <div className="flex items-center gap-1.5">
+                          <Package className="h-3.5 w-3.5 shrink-0" />
+                          <span className="min-w-0 truncate text-[10px] font-semibold">{pack.label || `套图 ${index + 1}`}</span>
+                        </div>
+                        <div className="mt-1 flex items-center justify-between gap-2 text-[9px] text-white/38">
+                          <span>{reconstructionModeConfig(pack.mode).label}</span>
+                          <span className="font-mono">{subjectAssetPackSummary(pack)}</span>
+                        </div>
+                      </button>
+                    )
+                  })}
+                </div>
+              </div>
             </div>
           ) : (
             <div className="flex h-40 items-center justify-center rounded border border-dashed border-white/12 px-3 text-center text-[10.5px] leading-snug text-white/34">
@@ -4022,9 +4176,11 @@ function SourceReferenceBuildPanel({
         subject_profile: requestProfile.payload,
         prompt: buildSimilarSubjectPrompt(subjectStyle, subjectDirection, selectedTemplatePrompt, requestProfile),
         replace_views: true,
+        pack_label: `主体视图 ${new Date().toLocaleTimeString("zh-CN", { hour: "2-digit", minute: "2-digit", hour12: false })}`,
+        pack_mode: "realistic",
       })
       onJobUpdate(updated)
-      toast.success(`相似主体 ${selectedSubjectViews.length} 张高清白底图已生成`)
+      toast.success(`相似主体已提交：${selectedSubjectViews.length} 张会逐张出来`)
     } catch (e) {
       try {
         onJobUpdate(await getJob(requestJobId))
@@ -4060,9 +4216,13 @@ function SourceReferenceBuildPanel({
         subject_profile: requestProfile.payload,
         prompt: buildSimilarSubjectPrompt(subjectStyle, subjectDirection, selectedTemplatePrompt, requestProfile),
         replace_views: true,
+        pack_id: asset.pack_id ?? "",
+        pack_label: asset.pack_label ?? "",
+        pack_mode: asset.pack_mode ?? "realistic",
+        pack_created_at: asset.pack_created_at ?? asset.created_at ?? 0,
       })
       onJobUpdate(updated)
-      toast.success("已重新生成这张主体视图")
+      toast.success("已提交重生，这张主体视图会生成完成后替换")
     } catch (e) {
       toast.error("单张主体视图重生失败：" + (e instanceof Error ? e.message : String(e)))
     } finally {
@@ -4335,16 +4495,22 @@ function SourceReferenceBuildPanel({
             <div className="mb-2 grid grid-cols-[repeat(auto-fill,minmax(96px,1fr))] gap-2">
               {visibleActorAssets.map((asset) => {
                 const busyMode = subjectAssetBusy?.endsWith(asset.id) ? subjectAssetBusy.split(":")[0] : ""
+                const status = subjectAssetStatus(asset)
+                const running = subjectAssetIsRunning(asset)
+                const failed = status === "failed"
+                const mediaUrl = subjectAssetUrl(job, asset)
                 return (
                   <MediaAssetTile
                     key={asset.id}
-                    src={subjectAssetUrl(job, asset)}
+                    src={mediaUrl || undefined}
-                    href={subjectAssetUrl(job, asset)}
+                    href={mediaUrl || undefined}
                     alt={asset.label || asset.view}
                     label={asset.label || asset.view || "主体视图预览"}
-                    meta={asset.width && asset.height ? `${asset.width}x${asset.height}` : undefined}
+                    meta={subjectAssetStatusLabel(asset)}
-                    className="aspect-[9/16] w-20 bg-white 2xl:w-24"
+                    className={`aspect-[9/16] w-20 2xl:w-24 ${running ? "bg-black/40" : failed ? "bg-rose-950/30" : "bg-white"}`}
                     objectFit="contain"
+                    busy={running}
+                    emptyText={failed ? "失败" : running ? "生成中" : undefined}
                     title={asset.label || asset.view}
                     actions={[{
                       key: "regen",
@@ -4352,12 +4518,12 @@ function SourceReferenceBuildPanel({
                       icon: <RefreshCw className="h-3 w-3" />,
                       tone: "cyan",
                       busy: busyMode === "regen",
-                      disabled: !!subjectAssetBusy,
+                      disabled: !!subjectAssetBusy || running,
                       onClick: () => void regenerateSubjectAsset(asset),
                     }]}
                     onDelete={() => void deleteActorAsset(asset)}
                     deleting={busyMode === "delete"}
-                    deleteDisabled={!!subjectAssetBusy}
+                    deleteDisabled={!!subjectAssetBusy || running}
                     deleteLabel="删除这一张"
                   />
                 )

web/lib/api.ts

@@ -701,6 +701,7 @@ export type AssetBackground = "white" | "black"
 export type AssetSize = "source" | "1024" | "1536" | "2048"
 export type SubjectKind = "object" | "living"
 export type SubjectView = string
+export type SubjectAssetStatus = "queued" | "in_progress" | "completed" | "failed"
 export type SceneMode = "remove_subject" | "similar" | "style"
 export type SceneStyle = "source" | "premium_product" | "clean_studio" | "warm_lifestyle" | "cinematic"
 export type SceneAssetRole = "scene" | "first_frame" | "last_frame"
@@ -754,6 +755,13 @@ export interface SubjectAsset {
   size: AssetSize
   source_frame_indices?: number[]
   ai_completed?: boolean
+  status?: SubjectAssetStatus
+  progress?: number
+  error?: string
+  pack_id?: string
+  pack_label?: string
+  pack_mode?: string
+  pack_created_at?: number
   created_at: number
 }
 
@@ -1500,6 +1508,10 @@ export async function generateSubjectAssets(
     subject_profile?: SubjectProfilePreference | null
     prompt?: string
     replace_views?: boolean
+    pack_id?: string
+    pack_label?: string
+    pack_mode?: string
+    pack_created_at?: number
   } = {},
 ): Promise<Job> {
   const res = await fetch(`${API_BASE}/jobs/${jobId}/frames/${frameIdx}/elements/${elementId}/subject-assets`, {
@@ -1519,6 +1531,10 @@ export async function generateSubjectAssets(
       subject_profile: body.subject_profile ?? null,
       prompt: body.prompt ?? "",
       replace_views: body.replace_views ?? false,
+      pack_id: body.pack_id ?? "",
+      pack_label: body.pack_label ?? "",
+      pack_mode: body.pack_mode ?? "",
+      pack_created_at: body.pack_created_at ?? 0,
     }),
   })
   if (!res.ok) {