Merge pull request 'fix(viral-video) v1.6.1: P0 渲染卡死根因修复 + P1 VLM 快速降级' (#2148) from fix/2148-p0-render-hang-p1-vlm-fast-fail into develop

P0 根因修复（v9 渲染卡死 20+ min）：
1. Seedance 2.5 content 数组协议错误：之前把参考音频/视频放在顶层 reference_audios/reference_videos（非官方字段），
   且首帧图不带 role 导致服务端识别为 first_frame 模式（该模式强制 ratio=adaptive），传入 ratio=9:16 导致任务
   静默排队/失败轮询到 900s 超时。
2. 修复：所有参考（图/音/视）放进 content 数组并带正确 role（reference_image/reference_audio/reference_video）；
   首张图根据是否有额外参考自动选择 first_frame vs reference_image 模式。
3. ratio 智能选择：纯首帧无参考→adaptive；omni_reference→按用户 ratio（默认 9:16）；
   若创建任务因 ratio 返回 HTTP 400，自动回退 adaptive 重试一次。
4. 轮询 total_timeout 从 900s 收紧到 480s（8min），下载超时 120s；
   celery 任务加 soft_time_limit（render 660s/analyze 240s/generate_copy 180s/pipeline 900s）防挂死。
5. 轮询每 5 次打一次 info 日志，失败时完整打印 error.code/message。
6. _hb_stop 提前声明避免 finally UnboundLocalError。

P1 VLM lite 快速降级：
- lite timeout 从 45s 收紧到 15s，pro 降级固定 25s，总耗时 ≤40s（之前 45+60≈100s）。
- 日志：创建任务时打印 mode（first_frame/omni_ref）、ratio、参考数量。

.env.example

@@ -212,9 +212,14 @@ COSYVOICE_CLONE_MODEL=voice-enrollment
 
 DOUBAO_API_KEY=your-doubao-api-key
 DOUBAO_MODEL=doubao-seed-1-6-250615
+DOUBAO_FAST_MODEL=doubao-1-5-pro-32k-250115
 DOUBAO_BASE_URL=https://ark.cn-beijing.volces.com/api/v3
 DOUBAO_TIMEOUT=30
 DOUBAO_MAX_RETRIES=2
+# 视觉模型：pro 精度高，lite 速度快（viral-video 商品识别默认用 lite 提速）
+DOUBAO_VISION_MODEL=doubao-1-5-vision-pro-250328
+DOUBAO_VISION_LITE_MODEL=doubao-1-5-vision-lite-250315
+DOUBAO_VISION_USE_LITE=true
 
 # ==================== 积分/会员系统 (#1895) ====================
 # 积分系统总开关：默认 false（暂停积分系统）。

alembic/versions/090_viral_video_phase_msg.py

@@ -0,0 +1,35 @@
+"""viral video add phase_message column (#2134)
+
+Revision ID: 090_viral_video_phase_msg
+Revises: 089_viral_video_cols
+Create Date: 2026-10-02
+
+#2134 阶段细粒度提示：viral_video 表新增 phase_message 列（中文阶段提示文案）。
+current_stage 列已在之前版本存在，本迁移只补 phase_message。
+幂等 ADD COLUMN IF NOT EXISTS。
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "090_viral_video_phase_msg"
+down_revision = "089_viral_video_cols"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    # SQLite/PostgreSQL 兼容的幂等添加列
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    cols = {c["name"] for c in inspector.get_columns("viral_video_jobs")}
+    if "phase_message" not in cols:
+        op.add_column(
+            "viral_video_jobs",
+            sa.Column("phase_message", sa.String(length=500), nullable=False, server_default=""),
+        )
+
+
+def downgrade() -> None:
+    op.drop_column("viral_video_jobs", "phase_message")

alembic/versions/091_viral_video_stage.py

@@ -0,0 +1,49 @@
+"""viral video add current_stage column (#2137 follow-up)
+
+Revision ID: 091_viral_video_stage
+Revises: 090_viral_video_phase_msg
+Create Date: 2026-10-02
+
+#2137 follow-up fix: 090 migration missed current_stage column on viral_video_jobs,
+causing UndefinedColumn errors and 500s on all authenticated viral-video endpoints.
+Idempotently add current_stage and double-check phase_message.
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "091_viral_video_stage"
+down_revision = "090_viral_video_phase_msg"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    cols = {c["name"] for c in inspector.get_columns("viral_video_jobs")}
+    if "current_stage" not in cols:
+        op.add_column(
+            "viral_video_jobs",
+            sa.Column(
+                "current_stage",
+                sa.String(length=200),
+                nullable=False,
+                server_default="",
+            ),
+        )
+    if "phase_message" not in cols:
+        op.add_column(
+            "viral_video_jobs",
+            sa.Column(
+                "phase_message",
+                sa.String(length=500),
+                nullable=False,
+                server_default="",
+            ),
+        )
+
+
+def downgrade() -> None:
+    op.drop_column("viral_video_jobs", "current_stage")

alembic/versions/092_viral_video_heartbeat.py

@@ -0,0 +1,42 @@
+"""viral_video_jobs 增加 heartbeat_at 列（worker 心跳，用于僵尸任务超时回收）
+
+Revision ID: 092_viral_video_heartbeat
+Revises: 091_viral_video_stage
+Create Date: 2026-10-02
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "092_viral_video_heartbeat"
+down_revision = "091_viral_video_stage"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    cols = {c["name"] for c in inspector.get_columns("viral_video_jobs")}
+    if "heartbeat_at" not in cols:
+        op.add_column("viral_video_jobs", sa.Column("heartbeat_at", sa.DateTime(), nullable=True))
+        op.execute(
+            "UPDATE viral_video_jobs SET heartbeat_at = updated_at " "WHERE status = 'running' AND heartbeat_at IS NULL"
+        )
+        try:
+            op.create_index("ix_viral_video_jobs_heartbeat_at", "viral_video_jobs", ["heartbeat_at"])
+        except Exception:
+            pass
+
+
+def downgrade() -> None:
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    cols = {c["name"] for c in inspector.get_columns("viral_video_jobs")}
+    if "heartbeat_at" in cols:
+        try:
+            op.drop_index("ix_viral_video_jobs_heartbeat_at", table_name="viral_video_jobs")
+        except Exception:
+            pass
+        op.drop_column("viral_video_jobs", "heartbeat_at")

apps/api/app/api/routes/viral_video.py

@@ -128,6 +128,8 @@ def _to_response(job) -> ViralVideoJobResponse:
         style_guide=job.style_guide,
         style_template_id=job.style_template_id,
         status=job.status,
+        current_stage=getattr(job, "current_stage", "") or "",
+        phase_message=getattr(job, "phase_message", "") or "",
         image_analysis=getattr(job, "image_analysis", None),
         storyboard=getattr(job, "storyboard", None),
         generated_copy_text=getattr(job, "generated_copy_text", "") or "",
@@ -398,28 +400,42 @@ def retry_viral_video_job(
     authenticated_user: AuthenticatedUser = Depends(get_current_user),
     session: Session = Depends(get_db_session),
 ) -> ViralVideoJobResponse:
-    """重试失败的爆款视频任务。"""
+    """重试失败的爆款视频任务（也支持对僵尸/超时 running 任务强制重置后重试）。"""
+    from datetime import datetime, timezone
+
     repo = _get_job_repo(session)
     job = repo.get(job_id)
     if job is None:
         raise HTTPException(status_code=404, detail="任务不存在")
     if job.user_id != authenticated_user.user.id:
         raise HTTPException(status_code=403, detail="无权操作此任务")
-    if job.status != ViralVideoStatus.FAILED:
-        raise HTTPException(status_code=409, detail="只有失败的任务可以重试")
+
+    # 判定是否为僵尸 running 任务：running 超过 10 分钟且心跳停止超过 2 分钟
+    now = datetime.now(timezone.utc)
+    is_stale_running = False
+    if job.status == ViralVideoStatus.RUNNING and job.started_at is not None:
+        hb = getattr(job, "heartbeat_at", None) or job.updated_at
+        if (now - job.started_at).total_seconds() > 10 * 60 and hb is not None and (now - hb).total_seconds() > 2 * 60:
+            is_stale_running = True
+
+    if job.status != ViralVideoStatus.FAILED and not is_stale_running:
+        raise HTTPException(status_code=409, detail="只有失败或超时的任务可以重试")
 
     # 重置状态
     job.retry_count += 1
     job.status = ViralVideoStatus.PENDING
-    job.error_msg = ""
+    job.error_msg = "" if not is_stale_running else "任务执行超时，已重置重试"
     job.started_at = None
     job.completed_at = None
+    job.current_stage = ""
+    job.phase_message = ""
+    job.heartbeat_at = None
     repo.update(job)
 
     # 重新入队
     try:
         celery_app.send_task("worker.run_viral_video_pipeline", args=[job.id])
-        logger.info("[爆款视频] 重试入队: job_id=%s retry_count=%d", job.id, job.retry_count)
+        logger.info("[爆款视频] 重试入队: job_id=%s retry_count=%d stale=%s", job.id, job.retry_count, is_stale_running)
     except Exception as e:
         logger.error("[爆款视频] 重试入队失败: %s", e, exc_info=True)
         job.mark_failed(f"重试入队失败: {e}")

apps/api/app/schemas/viral_video.py

@@ -201,6 +201,10 @@ class ViralVideoJobResponse(BaseModel):
     style_guide: dict | None = None
     style_template_id: str = ""
     status: str
+    current_stage: str = (
+        ""  # 细粒度阶段 snake_case（analyzing_images/parsing_intent/generating_script/reviewing/tts_synthesizing/rendering_video/uploading）
+    )
+    phase_message: str = ""  # 中文阶段提示文案（前端轮询/SSE 直接展示）
     image_analysis: dict | None = None
     # v1.6 编导脚本（推荐前端使用）
     copy_result: dict | None = None

apps/web/src/pages/viral-video/ViralVideoPage.tsx

@@ -9,8 +9,6 @@ import {
   CaretRightOutlined,
   VideoCameraOutlined,
   SoundOutlined,
-  BulbOutlined,
-  ThunderboltOutlined,
   ReloadOutlined,
   WarningFilled,
   SearchOutlined,
@@ -23,6 +21,7 @@ import {
   LoadingOutlined,
   CheckCircleFilled,
   EditOutlined,
+  HistoryOutlined,
 } from "@ant-design/icons"
 import { Select, Input, message } from "antd"
 import { uploadAssetDirect, getAssetLibraries, getAssetsByKind, type AssetItem } from "@/api/assets"
@@ -72,7 +71,7 @@ type UploadedImage = {
   localId: string
   url: string
   name: string
-  size: number
+  size: number | undefined
   uploading: boolean
   pct: number
   ossUrl?: string
@@ -122,6 +121,8 @@ type TabTask = {
   // STEP3 设置
   videoRatio: string
   videoModel: string
+  quality: string
+  upscale: string
   // UI 状态
   uiStep: UIStep
   imageAnalysis: ImageAnalysisResult | null
@@ -150,14 +151,61 @@ const TARGET_CUSTOMERS = [
   "都市蓝领",
   "都市银发",
 ]
-const PERSONAS = ["亲切邻家姐姐", "专业顾问", "搞笑段子手", "真实测评博主", "闺蜜种草", "行业专家"]
+const PERSONAS = [
+  "通用个人IP",
+  "老板型IP",
+  "专家型IP",
+  "顾问型IP",
+  "创始人IP",
+  "创业者IP",
+  "从业者经验派",
+  "避坑顾问型",
+  "知识科普型",
+  "测评种草型",
+]
 const STRUCTURES = [
-  "痛点开场→卖点→促单",
-  "反差对比→证据→促单",
-  "故事讲述→共鸣→促单",
-  "专家背书→数据→促单",
+  "反差破局+亮明观点+还原现状",
+  "痛点场景植入+给出通关秘籍",
+  "反常识开场+扎心现状+制造焦虑",
+  "核心难题+化解焦虑+描绘愿景",
+  "痛点前置+前后对比+给出期许",
+  "悬念拉满+揭露危害+明确目的",
+  "情绪共鸣+勾起好奇+拔高价值",
+  "颠覆认知+现状拆解+赋予价值",
+  "结果先行+消除顾虑+保留悬念",
+  "真实困境+认知反转+传授秘籍",
+  "反差事实+案例佐证+戳破真相",
+  "预埋悬念+标杆案例+引出核心",
+  "危机渲染+放大痛点+输出干货",
+  "逆袭开场+成功诀窍+低价引流",
+  "反常识观点+分析危害+描述现状",
+  "正反对比+解析原因+简易方法",
+  "真实场景+内心刻画+共情引导+展望未来",
+  "连续否定+代入感受+打破认知+精准戳痛",
+  "揭秘行业+类比施压+放大焦虑+精准锁客",
+]
+const PURPOSES = [
+  "获客引流",
+  "账号涨粉",
+  "活动通知",
+  "口播带货",
+  "促销转化",
+  "场景种草",
+  "痛点解决",
+  "功能演示",
+  "门店发现",
+  "到店实录",
+  "招牌体验",
+  "同城团购",
+  "品牌主张",
+  "品牌故事",
+  "生活方式",
+  "创意概念",
+  "对话短剧",
+  "反转短剧",
+  "悬念短剧",
+  "情绪短片",
 ]
-const PURPOSES = ["品牌种草", "新品上市", "促销转化", "活动引流"]
 const DURATIONS = [15, 20, 30, 45, 60]
 const RATIOS = [
   { v: "9:16", label: "9:16 竖屏（抖音/视频号）" },
@@ -168,6 +216,18 @@ const MODELS = [
   { v: "seedance-2.5", label: "Seedance 2.5（推荐）" },
   { v: "seedance-2.0", label: "Seedance 2.0" },
 ]
+const QUALITY_OPTIONS = [
+  { v: "480p", label: "480p（快速）" },
+  { v: "720p", label: "720p（清晰）" },
+  { v: "1080p", label: "1080p（高清）" },
+]
+const UPSCALE_OPTIONS = [
+  { v: "default", label: "默认" },
+  { v: "sd", label: "普清" },
+  { v: "hd", label: "高清" },
+  { v: "uhd", label: "超清" },
+  { v: "omnipotent", label: "全能" },
+]
 const VOICE_TIPS =
   "支持 MP3/WAV/M4A/AAC/OGG 格式，最大 10MB，时长 ≤30 秒。建议清晰人声、无背景音乐、环境安静；录音请保持距麦克风 15-20cm，音量适中。"
 
@@ -306,7 +366,8 @@ const MOCK_STORYBOARD: Storyboard = {
     "还在为餐桌选不到好桌子发愁？这张北美黑胡桃木餐桌，一家人坐下来吃饭刚刚好。全实木、无贴皮，纹理好看又耐刮。点小黄车，给家里添一张好桌子。",
 }
 
-const fmtSize = (bytes: number) => {
+const fmtSize = (bytes: number | undefined) => {
+  if (!bytes || bytes <= 0) return "--"
   if (bytes < 1024) return `${bytes}B`
   if (bytes < 1024 * 1024) return `${(bytes / 1024).toFixed(1)}KB`
   return `${(bytes / 1024 / 1024).toFixed(1)}MB`
@@ -327,11 +388,13 @@ const emptyTask = (id: string, title: string): TabTask => ({
   targetCustomer: "",
   language: "中文(普通话)",
   viralStructure: STRUCTURES[0],
-  marketingPurpose: PURPOSES[0],
+  marketingPurpose: "",
   duration: 15,
   persona: "",
   videoRatio: "9:16",
   videoModel: "seedance-2.5",
+  quality: "480p",
+  upscale: "default",
   uiStep: "step1_upload",
   imageAnalysis: null,
   storyboard: null,
@@ -343,6 +406,32 @@ const emptyTask = (id: string, title: string): TabTask => ({
 })
 
 /* ─────────── 主页面 ─────────── */
+
+function getCopyStageText(stage: string | undefined): string {
+  switch (stage) {
+    case "image_analysis":
+    case "video_analysis":
+    case "analyzing_images":
+      return "🔍 正在分析商品特征..."
+    case "intent_parsing":
+    case "generating_outline":
+    case "planning":
+      return "📋 正在构思脚本框架..."
+    case "script_generation":
+    case "generating_storyboard":
+    case "writing_shots":
+      return "🎬 正在生成分镜脚本..."
+    case "review":
+    case "generating_copy":
+    case "writing_copy":
+      return "✍️ 正在撰写口播文案..."
+    case "finalizing":
+      return "✨ 正在优化细节..."
+    default:
+      return "🎬 AI 正在创作视频分镜脚本..."
+  }
+}
+
 const ViralVideoPage: React.FC = () => {
   const [tasks, setTasks] = useState<TabTask[]>(() => [emptyTask("t1", "生成 1")])
   const [activeId, setActiveId] = useState<string>("t1")
@@ -493,8 +582,6 @@ const ViralVideoPage: React.FC = () => {
   )
 
   /* ── 图片 ── */
-  const canEditAssets = task.uiStep === "step1_upload" || task.uiStep === "failed"
-
   const appendImagesFromFiles = useCallback(
     async (files: FileList | File[]) => {
       const remaining = 9 - task.images.length
@@ -530,7 +617,15 @@ const ViralVideoPage: React.FC = () => {
           setTask((t) => ({
             ...t,
             images: t.images.map((im) =>
-              im.localId === item.localId ? { ...im, uploading: false, pct: 100, ossUrl: url } : im,
+              im.localId === item.localId
+                ? {
+                    ...im,
+                    uploading: false,
+                    pct: 100,
+                    ossUrl: url,
+                    size: item.file!.size || im.size,
+                  }
+                : im,
             ),
           }))
         } catch {
@@ -551,7 +646,7 @@ const ViralVideoPage: React.FC = () => {
         localId: `lib_${a.id}_${Math.random().toString(36).slice(2, 6)}`,
         url: a.thumbnail_url || a.file_url || "",
         name: a.name || "未命名",
-        size: a.file_size || 0,
+        size: a.file_size || undefined,
         uploading: false,
         pct: 100,
         ossUrl: a.file_url || undefined,
@@ -788,7 +883,9 @@ const ViralVideoPage: React.FC = () => {
         duration: task.duration,
         persona_id: task.persona || undefined,
         user_copy_text:
-          task.fusionLevel === "user_primary" && task.userCopy ? task.userCopy : undefined,
+          (task.fusionLevel === "user_primary" || task.fusionLevel === "ai_polish") && task.userCopy
+            ? task.userCopy
+            : undefined,
         fusion_level: task.fusionLevel,
         style_strength: task.refVideo?.ossUrl ? task.styleStrength : undefined,
         voice_id: task.refAudio?.source === "preset" ? task.refAudio.id : undefined,
@@ -895,10 +992,6 @@ const ViralVideoPage: React.FC = () => {
   const outputUrl = task.job?.output_url || task.job?.result_video_url
   const errMsg = task.videoError || task.job?.error_message || task.job?.error_msg
 
-  const canEditStep2 =
-    task.uiStep !== "step2_generating" &&
-    task.uiStep !== "step3_generating" &&
-    task.uiStep !== "step3_done"
   const step2Enabled =
     task.uiStep === "step1_done" ||
     task.uiStep === "step2_generating" ||
@@ -943,24 +1036,61 @@ const ViralVideoPage: React.FC = () => {
         </div>
         {products.map((p, i) => (
           <div key={i} className="vv-recog-item">
-            {p.name && (
+            <div className="vv-recog-line">
+              <span className="vv-recog-k">图片{i + 1}：</span>
+              <span>
+                {p.name || "未识别"}
+                {p.spec && <span className="vv-recog-meta">（{p.spec}）</span>}
+                {p.brand && <span className="vv-recog-meta"> · {p.brand}</span>}
+                {p.category && <span className="vv-recog-meta"> · {p.category}</span>}
+              </span>
+            </div>
+            {featureText(p.key_features ?? p.features) && (
               <div className="vv-recog-line">
-                <span className="vv-recog-k">商品：</span>
-                <span>{p.name}</span>
-                {p.spec && <span style={{ color: "#6b7280" }}>（{p.spec}）</span>}
-                {p.brand && <span style={{ color: "#6b7280" }}> · {p.brand}</span>}
+                <span className="vv-recog-k">核心特征：</span>
+                <span className="vv-recog-v">{featureText(p.key_features ?? p.features)}</span>
               </div>
             )}
-            {featureText(p.features) && (
+            {p.colors && p.colors.length > 0 && (
               <div className="vv-recog-line">
-                <span className="vv-recog-k">视觉特征：</span>
-                <span>{featureText(p.features)}</span>
+                <span className="vv-recog-k">主色调：</span>
+                <span className="vv-recog-v">{p.colors.join(" / ")}</span>
               </div>
             )}
-            {p.label_text && (
+            {p.material_or_texture && (
               <div className="vv-recog-line">
-                <span className="vv-recog-k">品牌/标签文字：</span>
-                <span>{p.label_text}</span>
+                <span className="vv-recog-k">材质/纹理：</span>
+                <span className="vv-recog-v">{p.material_or_texture}</span>
+              </div>
+            )}
+            {p.visual_style && (
+              <div className="vv-recog-line">
+                <span className="vv-recog-k">视觉风格：</span>
+                <span className="vv-recog-v">{p.visual_style}</span>
+              </div>
+            )}
+            {p.scene && (
+              <div className="vv-recog-line">
+                <span className="vv-recog-k">场景：</span>
+                <span className="vv-recog-v">{p.scene}</span>
+              </div>
+            )}
+            {p.target_audience_hint && (
+              <div className="vv-recog-line">
+                <span className="vv-recog-k">目标人群：</span>
+                <span className="vv-recog-v">{p.target_audience_hint}</span>
+              </div>
+            )}
+            {(p.text_on_image || p.label_text) && (
+              <div className="vv-recog-line">
+                <span className="vv-recog-k">包装文字：</span>
+                <span className="vv-recog-v">{p.text_on_image || p.label_text}</span>
+              </div>
+            )}
+            {p.selling_points && (
+              <div className="vv-recog-line">
+                <span className="vv-recog-k">卖点：</span>
+                <span className="vv-recog-v">{p.selling_points}</span>
               </div>
             )}
           </div>
@@ -1009,8 +1139,11 @@ const ViralVideoPage: React.FC = () => {
     if (task.uiStep === "step2_generating") {
       return (
         <div className="vv-copy-box vv-copy-loading">
-          <LoadingOutlined style={{ color: "#7c3aed", fontSize: 18, marginRight: 8 }} />
-          <span>AI 正在创作视频分镜脚本…</span>
+          <span className="vv-spinner" />
+          <span className="vv-copy-loading-text">
+            {task.job?.progress_message || getCopyStageText(task.job?.progress_stage)}
+          </span>
+          <span className="vv-copy-loading-hint">预计 10-30 秒，请稍候</span>
         </div>
       )
     }
@@ -1049,160 +1182,160 @@ const ViralVideoPage: React.FC = () => {
     ) {
       return null
     }
-    const locked = task.uiStep === "step3_generating" || task.uiStep === "step3_done"
+    const locked = false
     const isMockPreview = sb === MOCK_STORYBOARD
 
     return (
       <div className="vv-copy-box vv-storyboard">
-        <div className="vv-copy-head">
-          <EditOutlined style={{ color: "#7c3aed" }} />
-          <span className="vv-copy-title">AI 分镜脚本</span>
-          {isMockPreview && (
-            <span className="vv-copy-tag vv-copy-tag-preview">
-              示例预览 · 等后端输出新结构后自动替换
-            </span>
-          )}
-        </div>
-
-        {/* 1. 总览卡片 */}
-        <div className="vv-sb-overview">
-          <div className="vv-sb-ov-field">
-            <label>主题</label>
+        <div className="vv-sb-doc">
+          {/* 视频总览 */}
+          <h4 className="vv-sb-h">视频总览</h4>
+          <div className="vv-sb-kv">
+            <span className="vv-sb-k">整体主题：</span>
             <input
-              className="vv-input vv-sb-input"
+              className="vv-sb-inline vv-sb-inline-input"
               value={sb.overview.theme}
               disabled={locked}
               onChange={(e) => updateOverview({ theme: e.target.value })}
             />
           </div>
-          <div className="vv-sb-ov-field">
-            <label>总时长</label>
+          <div className="vv-sb-kv">
+            <span className="vv-sb-k">总时长：</span>
             <input
-              className="vv-input vv-sb-input"
+              className="vv-sb-inline vv-sb-inline-input"
               value={sb.overview.total_duration}
               disabled={locked}
               onChange={(e) => updateOverview({ total_duration: e.target.value })}
             />
           </div>
-          <div className="vv-sb-ov-field">
-            <label>画幅</label>
+          <div className="vv-sb-kv">
+            <span className="vv-sb-k">画幅：</span>
             <Select
-              className="vv-select vv-select-purple vv-sb-input"
-              style={{ width: "100%" }}
+              className="vv-select vv-sb-inline-select"
+              style={{ width: 200 }}
               value={sb.overview.aspect_ratio}
               disabled={locked}
               onChange={(v) => updateOverview({ aspect_ratio: v })}
-              options={RATIOS.map((r) => ({ value: r.v, label: r.label }))}
+              options={RATIOS.map((r) => ({ value: r.v, label: r.v }))}
+              variant="borderless"
             />
           </div>
-        </div>
 
-        {/* 2. 场景与光线 */}
-        <div className="vv-sb-section">
-          <div className="vv-sb-section-title">
-            <BulbOutlined style={{ color: "#7c3aed" }} /> 场景与光线
-          </div>
+          {/* 场景与光线 */}
+          <h4 className="vv-sb-h">场景与光线</h4>
           <textarea
-            className="vv-textarea"
+            className="vv-textarea vv-sb-doc-ta"
             value={sb.scene_and_lighting}
             disabled={locked}
             onChange={(e) => updateStoryboard({ scene_and_lighting: e.target.value })}
             placeholder="描述整体场景氛围、光线方向与色温…"
           />
-        </div>
 
-        {/* 3. 逐镜头卡片 */}
-        <div className="vv-sb-section">
-          <div className="vv-sb-section-title">
-            <VideoCameraOutlined style={{ color: "#7c3aed" }} /> 逐镜头拆解（{sb.shots.length} 镜）
-          </div>
-          <div className="vv-sb-shots">
+          {/* 逐镜头 */}
+          <h4 className="vv-sb-h">逐秒镜头拆解</h4>
+          <div className="vv-sb-doc-shots">
             {sb.shots.map((sh, idx) => {
               const refImg =
                 typeof sh.reference_image_index === "number"
                   ? task.images[sh.reference_image_index]
                   : undefined
               return (
-                <div key={idx} className="vv-sb-shot">
-                  <div className="vv-sb-shot-head">
-                    <span className="vv-sb-shot-num">镜头 {idx + 1}</span>
+                <div key={idx} className="vv-sb-doc-shot">
+                  <div className="vv-sb-doc-shot-head">
                     <input
-                      className="vv-sb-time"
+                      className="vv-sb-time-doc"
                       value={sh.time_range}
                       disabled={locked}
                       onChange={(e) => updateShot(idx, { time_range: e.target.value })}
+                      placeholder="0-3秒"
                     />
                   </div>
-                  <div className="vv-sb-shot-grid">
-                    <div className="vv-sb-cell">
-                      <label>景别 / 运镜</label>
-                      <textarea
-                        className="vv-textarea vv-sb-cell-ta"
-                        value={sh.shot_type_angle_movement}
-                        disabled={locked}
-                        onChange={(e) =>
-                          updateShot(idx, { shot_type_angle_movement: e.target.value })
-                        }
-                      />
-                    </div>
-                    <div className="vv-sb-cell vv-sb-cell-wide">
-                      <label>场景与对白</label>
-                      <textarea
-                        className="vv-textarea vv-sb-cell-ta"
-                        value={sh.scene_and_dialogue}
-                        disabled={locked}
-                        onChange={(e) => updateShot(idx, { scene_and_dialogue: e.target.value })}
-                      />
-                    </div>
-                    <div className="vv-sb-cell">
-                      <label>动作细节</label>
-                      <textarea
-                        className="vv-textarea vv-sb-cell-ta"
-                        value={sh.action_details}
-                        disabled={locked}
-                        onChange={(e) => updateShot(idx, { action_details: e.target.value })}
-                      />
-                    </div>
-                    <div className="vv-sb-cell">
-                      <label>音效 / BGM</label>
-                      <textarea
-                        className="vv-textarea vv-sb-cell-ta"
-                        value={sh.audio_bgm}
-                        disabled={locked}
-                        onChange={(e) => updateShot(idx, { audio_bgm: e.target.value })}
-                      />
-                    </div>
-                    <div className="vv-sb-cell">
-                      <label>转场</label>
-                      <textarea
-                        className="vv-textarea vv-sb-cell-ta"
-                        value={sh.transition}
-                        disabled={locked}
-                        onChange={(e) => updateShot(idx, { transition: e.target.value })}
-                      />
-                    </div>
-                    <div className="vv-sb-cell vv-sb-ref">
-                      <label>参考图</label>
-                      <div className="vv-sb-ref-box">
-                        {refImg?.url ? (
-                          <img src={refImg.url} alt={`ref-${idx}`} className="vv-sb-ref-img" />
-                        ) : (
-                          <span className="vv-sb-ref-empty">未指定</span>
+                  <div className="vv-sb-kv">
+                    <span className="vv-sb-k">景别/角度与运镜：</span>
+                    <textarea
+                      className="vv-textarea vv-sb-doc-ta vv-sb-doc-ta-sm"
+                      value={sh.shot_type_angle_movement}
+                      disabled={locked}
+                      onChange={(e) =>
+                        updateShot(idx, { shot_type_angle_movement: e.target.value })
+                      }
+                    />
+                  </div>
+                  <div className="vv-sb-kv">
+                    <span className="vv-sb-k">场景与对白：</span>
+                    <textarea
+                      className="vv-textarea vv-sb-doc-ta"
+                      value={sh.scene_and_dialogue}
+                      disabled={locked}
+                      onChange={(e) => updateShot(idx, { scene_and_dialogue: e.target.value })}
+                    />
+                  </div>
+                  <div className="vv-sb-kv">
+                    <span className="vv-sb-k">动作与真人细节：</span>
+                    <textarea
+                      className="vv-textarea vv-sb-doc-ta vv-sb-doc-ta-sm"
+                      value={sh.action_details}
+                      disabled={locked}
+                      onChange={(e) => updateShot(idx, { action_details: e.target.value })}
+                    />
+                  </div>
+                  <div className="vv-sb-kv">
+                    <span className="vv-sb-k">音效/BGM：</span>
+                    <textarea
+                      className="vv-textarea vv-sb-doc-ta vv-sb-doc-ta-sm"
+                      value={sh.audio_bgm}
+                      disabled={locked}
+                      onChange={(e) => updateShot(idx, { audio_bgm: e.target.value })}
+                    />
+                  </div>
+                  <div className="vv-sb-kv">
+                    <span className="vv-sb-k">转场：</span>
+                    <textarea
+                      className="vv-textarea vv-sb-doc-ta vv-sb-doc-ta-sm"
+                      value={sh.transition}
+                      disabled={locked}
+                      onChange={(e) => updateShot(idx, { transition: e.target.value })}
+                    />
+                  </div>
+                  <div className="vv-sb-kv vv-sb-kv-ref">
+                    <span className="vv-sb-k">参考图片：</span>
+                    {refImg?.url ? (
+                      <span className="vv-sb-ref-badge">
+                        <img src={refImg.url} alt={`ref-${idx}`} className="vv-sb-ref-thumb" />
+                        {!locked && (
+                          <button
+                            className="vv-sb-ref-x"
+                            onClick={() => updateShot(idx, { reference_image_index: undefined })}
+                            aria-label="移除参考图"
+                          >
+                            ×
+                          </button>
                         )}
-                      </div>
-                    </div>
+                      </span>
+                    ) : !locked && task.images.length > 0 ? (
+                      <span className="vv-sb-ref-pick">
+                        {task.images.map((im, i) => (
+                          <button
+                            key={im.localId}
+                            className="vv-sb-ref-opt"
+                            onClick={() => updateShot(idx, { reference_image_index: i })}
+                            title={`@图片${i + 1}`}
+                          >
+                            <img src={im.url} alt={`img${i}`} />
+                          </button>
+                        ))}
+                      </span>
+                    ) : (
+                      <span className="vv-sb-muted">未指定</span>
+                    )}
                   </div>
                 </div>
               )
             })}
           </div>
-        </div>
 
-        {/* 4. 硬性约束 */}
-        <div className="vv-sb-section">
-          <div className="vv-sb-section-title">
-            <ThunderboltOutlined style={{ color: "#7c3aed" }} /> 硬性约束
-          </div>
+          {/* 硬性约束 */}
+          <h4 className="vv-sb-h">硬性约束</h4>
           <div className="vv-sb-taglist">
             {sb.hard_constraints.map((c, i) => (
               <span key={i} className="vv-sb-tag vv-sb-tag-hard">
@@ -1229,13 +1362,9 @@ const ViralVideoPage: React.FC = () => {
               </button>
             )}
           </div>
-        </div>
 
-        {/* 5. 负面提示词 */}
-        <div className="vv-sb-section">
-          <div className="vv-sb-section-title">
-            <CloseOutlined style={{ color: "#ef4444" }} /> 负面提示词
-          </div>
+          {/* 负面提示词 */}
+          <h4 className="vv-sb-h">负面提示词</h4>
           <div className="vv-sb-taglist">
             {sb.negative_prompts.map((c, i) => (
               <span key={i} className="vv-sb-tag vv-sb-tag-neg">
@@ -1262,20 +1391,18 @@ const ViralVideoPage: React.FC = () => {
               </button>
             )}
           </div>
-        </div>
 
-        {/* 6. 完整口播稿（折叠） */}
-        <div className="vv-sb-section">
-          <button className="vv-sb-collapse" onClick={() => setVoOpen(!voOpen)}>
-            <span>
+          {/* 完整口播稿（可折叠） */}
+          <h4 className="vv-sb-h">
+            <button className="vv-sb-doc-collapse" onClick={() => setVoOpen(!voOpen)} type="button">
               <SoundOutlined style={{ color: "#7c3aed", marginRight: 6 }} />
               完整口播稿（TTS 合成使用）
-            </span>
-            <span className={`vv-sb-caret ${voOpen ? "open" : ""}`}>▾</span>
-          </button>
+              <span className={`vv-sb-caret ${voOpen ? "open" : ""}`}>▾</span>
+            </button>
+          </h4>
           {voOpen && (
             <textarea
-              className="vv-textarea vv-sb-vo"
+              className="vv-textarea vv-sb-doc-ta vv-sb-vo"
               value={sb.voiceover_script}
               disabled={locked}
               onChange={(e) => updateStoryboard({ voiceover_script: e.target.value })}
@@ -1284,16 +1411,18 @@ const ViralVideoPage: React.FC = () => {
           )}
         </div>
 
-        {/* 7. 重新生成按钮（ghost） */}
-        {task.uiStep === "step2_copy_ready" && (
-          <button
-            className="vv-btn vv-btn-ghost vv-btn-block"
-            onClick={handleGenerateCopy}
-            style={{ marginTop: 12 }}
-          >
-            <ReloadOutlined /> 重新生成文案
-          </button>
-        )}
+        <div className="vv-sb-actions">
+          {isMockPreview && (
+            <span className="vv-copy-tag vv-copy-tag-preview">
+              示例预览 · 后端新结构就绪后自动替换
+            </span>
+          )}
+          {task.uiStep === "step2_copy_ready" && (
+            <button className="vv-btn vv-btn-ghost" onClick={handleGenerateCopy}>
+              <ReloadOutlined /> 重新生成文案
+            </button>
+          )}
+        </div>
       </div>
     )
   }
@@ -1341,17 +1470,12 @@ const ViralVideoPage: React.FC = () => {
             <div className="vv-section-body" style={{ paddingTop: 4 }}>
               {/* 图片按钮 */}
               <div className="vv-action-row">
-                <button
-                  className="vv-action-btn"
-                  onClick={() => imgInputRef.current?.click()}
-                  disabled={!canEditAssets}
-                >
+                <button className="vv-action-btn" onClick={() => imgInputRef.current?.click()}>
                   <UploadOutlined /> 本地上传
                 </button>
                 <button
                   className="vv-action-btn"
                   onClick={() => setAssetPicker({ open: true, kind: "image", multiple: true })}
-                  disabled={!canEditAssets}
                 >
                   <FolderOpenOutlined /> 从素材库选择图片
                 </button>
@@ -1373,7 +1497,7 @@ const ViralVideoPage: React.FC = () => {
                   已选 <strong style={{ color: "#111827" }}>{task.images.length}</strong>/9
                 </span>
                 {task.images.length > 0 && (
-                  <button className="vv-link-btn" onClick={clearImages} disabled={!canEditAssets}>
+                  <button className="vv-link-btn" onClick={clearImages}>
                     清空图片
                   </button>
                 )}
@@ -1406,7 +1530,7 @@ const ViralVideoPage: React.FC = () => {
                         <button
                           className="vv-icon-btn"
                           onClick={() => moveImg(i, -1)}
-                          disabled={i === 0 || im.uploading || !canEditAssets}
+                          disabled={i === 0 || im.uploading}
                           title="上移"
                         >
                           <ArrowUpOutlined />
@@ -1414,7 +1538,7 @@ const ViralVideoPage: React.FC = () => {
                         <button
                           className="vv-icon-btn"
                           onClick={() => moveImg(i, 1)}
-                          disabled={i === task.images.length - 1 || im.uploading || !canEditAssets}
+                          disabled={i === task.images.length - 1 || im.uploading}
                           title="下移"
                         >
                           <ArrowDownOutlined />
@@ -1422,7 +1546,7 @@ const ViralVideoPage: React.FC = () => {
                         <button
                           className="vv-icon-btn vv-icon-btn-danger"
                           onClick={() => removeImg(im.localId)}
-                          disabled={im.uploading || !canEditAssets}
+                          disabled={!!im.uploading}
                           title="删除"
                         >
                           <CloseOutlined />
@@ -1441,11 +1565,7 @@ const ViralVideoPage: React.FC = () => {
                   <span className="vv-subblock-hint">最大 200MB · 最长 120 秒</span>
                 </div>
                 <div className="vv-action-row">
-                  <button
-                    className="vv-action-btn"
-                    onClick={() => videoInputRef.current?.click()}
-                    disabled={!canEditAssets}
-                  >
+                  <button className="vv-action-btn" onClick={() => videoInputRef.current?.click()}>
                     <UploadOutlined /> 本地上传视频
                   </button>
                   <button
@@ -1453,7 +1573,6 @@ const ViralVideoPage: React.FC = () => {
                     onClick={() =>
                       setTask({ voicePanel: task.voicePanel === "douyin" ? null : "douyin" })
                     }
-                    disabled={!canEditAssets}
                   >
                     <LinkOutlined /> 抖音链接识别
                   </button>
@@ -1504,11 +1623,7 @@ const ViralVideoPage: React.FC = () => {
                         </div>
                       )}
                     </div>
-                    <button
-                      className="vv-link-btn"
-                      onClick={removeRefVideo}
-                      disabled={!canEditAssets}
-                    >
+                    <button className="vv-link-btn" onClick={removeRefVideo}>
                       删除重传
                     </button>
                     {task.refVideo.ossUrl && (
@@ -1522,7 +1637,6 @@ const ViralVideoPage: React.FC = () => {
                               key={s.value}
                               className={`vv-pill ${task.styleStrength === s.value ? "active" : ""}`}
                               onClick={() => setTask({ styleStrength: s.value })}
-                              disabled={!canEditAssets}
                             >
                               {s.label}
                             </button>
@@ -1545,7 +1659,6 @@ const ViralVideoPage: React.FC = () => {
                   <button
                     className={`vv-audio-cell ${task.refAudio?.source === "preset" ? "selected" : ""}`}
                     onClick={() => setVoicePickerOpen(true)}
-                    disabled={!canEditAssets}
                   >
                     <FileTextOutlined className="vv-audio-ic" />
                     <span>选择内置音色</span>
@@ -1553,7 +1666,6 @@ const ViralVideoPage: React.FC = () => {
                   <button
                     className={`vv-audio-cell ${task.voicePanel === "upload" ? "active" : ""} ${task.refAudio?.source === "upload" ? "selected" : ""}`}
                     onClick={() => voiceInputRef.current?.click()}
-                    disabled={!canEditAssets}
                   >
                     <UploadOutlined className="vv-audio-ic" />
                     <span>本地上传</span>
@@ -1561,7 +1673,6 @@ const ViralVideoPage: React.FC = () => {
                   <button
                     className={`vv-audio-cell ${task.voicePanel === "library" ? "active" : ""} ${task.refAudio?.source === "library" ? "selected" : ""}`}
                     onClick={() => setAssetPicker({ open: true, kind: "voice", multiple: false })}
-                    disabled={!canEditAssets}
                   >
                     <FolderOpenOutlined className="vv-audio-ic" />
                     <span>从素材库选择</span>
@@ -1569,7 +1680,6 @@ const ViralVideoPage: React.FC = () => {
                   <button
                     className={`vv-audio-cell ${task.voicePanel === "record" ? "active" : ""} ${task.refAudio?.source === "clone" ? "selected" : ""}`}
                     onClick={() => setCloneModalOpen(true)}
-                    disabled={!canEditAssets}
                   >
                     <AudioMutedOutlined className="vv-audio-ic" />
                     <span>直接录音</span>
@@ -1589,11 +1699,7 @@ const ViralVideoPage: React.FC = () => {
                   <div className="vv-selected-audio">
                     <SoundOutlined style={{ color: "#7c3aed", marginRight: 6 }} />
                     <span style={{ flex: 1 }}>已选：{task.refAudio.name}</span>
-                    <button
-                      className="vv-link-btn"
-                      onClick={() => setTask({ refAudio: null })}
-                      disabled={!canEditAssets}
-                    >
+                    <button className="vv-link-btn" onClick={() => setTask({ refAudio: null })}>
                       移除
                     </button>
                   </div>
@@ -1689,7 +1795,7 @@ const ViralVideoPage: React.FC = () => {
         </div>
 
         {/* ═══════════ 中列 STEP 2 ═══════════ */}
-        <div className={`vv-col ${step2Enabled ? "" : "vv-col-disabled"}`}>
+        <div className="vv-col">
           <div className="vv-section">
             <div className="vv-step-subhead">
               <span className={`vv-step-badge ${step2Enabled ? "" : "vv-step-badge-dim"}`}>2</span>
@@ -1705,22 +1811,27 @@ const ViralVideoPage: React.FC = () => {
                     key={f.value}
                     className={`vv-fusion-btn ${task.fusionLevel === f.value ? "active" : ""}`}
                     onClick={() => setTask({ fusionLevel: f.value })}
-                    disabled={!canEditStep2 || !step2Enabled}
                   >
                     <strong>{f.label}</strong>
                     {f.desc}
                   </button>
                 ))}
               </div>
-              {task.fusionLevel === "user_primary" && (
+              {(task.fusionLevel === "ai_polish" || task.fusionLevel === "user_primary") && (
                 <div className="vv-form-row">
-                  <label className="vv-label">写下你的文案/草稿</label>
+                  <label className="vv-label">
+                    {task.fusionLevel === "ai_polish" ? "写下你的文案草稿" : "写下你的文案/草稿"}
+                  </label>
                   <textarea
+                    key={task.fusionLevel}
                     className="vv-textarea"
-                    placeholder="请输入你想要的原始文案，AI 将严格保留你的表达"
+                    placeholder={
+                      task.fusionLevel === "ai_polish"
+                        ? "请输入你的文案草稿，AI将在此基础上润色优化..."
+                        : "请输入你想要的原始文案，AI将严格保留你的表达..."
+                    }
                     value={task.userCopy}
                     onChange={(e) => setTask({ userCopy: e.target.value })}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
               )}
@@ -1736,77 +1847,71 @@ const ViralVideoPage: React.FC = () => {
                     placeholder="如：家具店、宠物店、装修公司"
                     value={task.industry}
                     onChange={(e) => setTask({ industry: e.target.value })}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">目标客户</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
                     placeholder="请选择目标客户"
                     value={task.targetCustomer || undefined}
                     onChange={(v) => setTask({ targetCustomer: v })}
                     options={TARGET_CUSTOMERS.map((i) => ({ value: i, label: i }))}
                     allowClear
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">台词语言</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
                     value={task.language}
                     onChange={(v) => setTask({ language: v })}
                     options={LANGUAGES.map((i) => ({ value: i, label: i }))}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">我的人设</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
-                    placeholder="请选择人设"
+                    placeholder="请选择人设类型"
+                    optionFilterProp="children"
                     value={task.persona || undefined}
                     onChange={(v) => setTask({ persona: v })}
                     options={PERSONAS.map((i) => ({ value: i, label: i }))}
-                    allowClear
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">爆款结构</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
                     value={task.viralStructure}
                     onChange={(v) => setTask({ viralStructure: v })}
                     options={STRUCTURES.map((i) => ({ value: i, label: i }))}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">营销目的</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
-                    value={task.marketingPurpose}
+                    placeholder="请选择营销目的"
+                    value={task.marketingPurpose || undefined}
                     onChange={(v) => setTask({ marketingPurpose: v })}
                     options={PURPOSES.map((i) => ({ value: i, label: i }))}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row" style={{ gridColumn: "1 / -1" }}>
                   <label className="vv-label">文案视频时长</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
                     value={task.duration}
                     onChange={(v) => setTask({ duration: v })}
                     options={DURATIONS.map((n) => ({ value: n, label: `${n}秒` }))}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
               </div>
@@ -1860,43 +1965,68 @@ const ViralVideoPage: React.FC = () => {
         </div>
 
         {/* ═══════════ 右列 STEP 3 ═══════════ */}
-        <div className={`vv-col ${step3Enabled ? "" : "vv-col-disabled"}`}>
+        <div className="vv-col">
           <div className="vv-section">
-            <div className="vv-step-subhead">
-              <span className={`vv-step-badge ${step3Enabled ? "" : "vv-step-badge-dim"}`}>3</span>
-              <span className="vv-step-subtitle">STEP 3 视频设置并生成视频</span>
+            <div className="vv-step-subhead vv-step3-head">
+              <div className="vv-step3-head-left">
+                <span className={`vv-step-badge ${step3Enabled ? "" : "vv-step-badge-dim"}`}>
+                  3
+                </span>
+                <span className="vv-step-subtitle">STEP 3 生成视频</span>
+              </div>
+              {step3Enabled && (
+                <button
+                  className="vv-btn vv-btn-ghost vv-btn-xs vv-history-btn"
+                  onClick={() => message.info("历史作品功能开发中")}
+                >
+                  <HistoryOutlined /> 历史作品
+                </button>
+              )}
               {!step3Enabled && <span className="vv-step-lock">请先完成 STEP 2</span>}
             </div>
+            <div className="vv-step3-divider" />
             <div className="vv-section-body">
-              <div className="vv-form-row">
-                <label className="vv-label">视频比例</label>
-                <Select
-                  className="vv-select"
-                  style={{ width: "100%" }}
-                  value={task.videoRatio}
-                  onChange={(v) => setTask({ videoRatio: v })}
-                  options={RATIOS.map((r) => ({ value: r.v, label: r.label }))}
-                  disabled={
-                    !step3Enabled ||
-                    task.uiStep === "step3_generating" ||
-                    task.uiStep === "step3_done"
-                  }
-                />
-              </div>
-              <div className="vv-form-row">
-                <label className="vv-label">生成模型</label>
-                <Select
-                  className="vv-select"
-                  style={{ width: "100%" }}
-                  value={task.videoModel}
-                  onChange={(v) => setTask({ videoModel: v })}
-                  options={MODELS.map((m) => ({ value: m.v, label: m.label }))}
-                  disabled={
-                    !step3Enabled ||
-                    task.uiStep === "step3_generating" ||
-                    task.uiStep === "step3_done"
-                  }
-                />
+              <div className="vv-step3-grid">
+                <div className="vv-form-row">
+                  <label className="vv-label">视频比例</label>
+                  <Select
+                    className="vv-select vv-select-step3"
+                    style={{ width: "100%" }}
+                    value={task.videoRatio}
+                    onChange={(v) => setTask({ videoRatio: v })}
+                    options={RATIOS.map((r) => ({ value: r.v, label: r.v }))}
+                  />
+                </div>
+                <div className="vv-form-row">
+                  <label className="vv-label">视频模型</label>
+                  <Select
+                    className="vv-select vv-select-step3"
+                    style={{ width: "100%" }}
+                    value={task.videoModel}
+                    onChange={(v) => setTask({ videoModel: v })}
+                    options={MODELS.map((m) => ({ value: m.v, label: m.label }))}
+                  />
+                </div>
+                <div className="vv-form-row">
+                  <label className="vv-label">清晰度</label>
+                  <Select
+                    className="vv-select vv-select-step3"
+                    style={{ width: "100%" }}
+                    value={task.quality}
+                    onChange={(v) => setTask({ quality: v })}
+                    options={QUALITY_OPTIONS.map((m) => ({ value: m.v, label: m.label }))}
+                  />
+                </div>
+                <div className="vv-form-row">
+                  <label className="vv-label">成片超分</label>
+                  <Select
+                    className="vv-select vv-select-step3"
+                    style={{ width: "100%" }}
+                    value={task.upscale}
+                    onChange={(v) => setTask({ upscale: v })}
+                    options={UPSCALE_OPTIONS.map((m) => ({ value: m.v, label: m.label }))}
+                  />
+                </div>
               </div>
             </div>
           </div>

packages/adapters/sqlalchemy_impl/models.py

@@ -950,6 +950,9 @@ class ViralVideoJobModel(Base):
     video_model = Column(String(100), nullable=False, default="")
     # 结果与状态
     status = Column(String(30), nullable=False, default="pending", index=True)
+    current_stage = Column(String(200), nullable=False, default="")  # 细粒度阶段 snake_case
+    phase_message = Column(String(500), nullable=False, default="")  # 阶段中文提示文案
+    heartbeat_at = Column(DateTime, nullable=True, index=True)  # worker 心跳，用于僵尸任务超时回收
     intent_result = Column(JSON, nullable=True)
     image_analysis = Column(JSON, nullable=True)
     storyboard = Column(JSON, nullable=True)

packages/adapters/sqlalchemy_impl/viral_video_repository.py

@@ -37,6 +37,9 @@ def _to_domain(model: ViralVideoJobModel) -> ViralVideoJob:
         video_ratio=getattr(model, "video_ratio", "9:16") or "9:16",
         video_model=getattr(model, "video_model", "") or "",
         status=ViralVideoStatus(model.status) if model.status else ViralVideoStatus.PENDING,
+        current_stage=getattr(model, "current_stage", "") or "",
+        phase_message=getattr(model, "phase_message", "") or "",
+        heartbeat_at=getattr(model, "heartbeat_at", None),
         intent_result=dict(model.intent_result) if model.intent_result else None,
         image_analysis=dict(model.image_analysis) if getattr(model, "image_analysis", None) else None,
         storyboard=list(model.storyboard) if getattr(model, "storyboard", None) else None,
@@ -83,6 +86,9 @@ class SQLAlchemyViralVideoJobRepository:
             video_ratio=job.video_ratio,
             video_model=job.video_model,
             status=job.status,
+            current_stage=job.current_stage or "",
+            phase_message=job.phase_message or "",
+            heartbeat_at=job.heartbeat_at,
             intent_result=job.intent_result,
             image_analysis=job.image_analysis,
             storyboard=job.storyboard,
@@ -106,6 +112,9 @@ class SQLAlchemyViralVideoJobRepository:
         if model is None:
             raise ValueError(f"ViralVideoJob {job.id} not found")
         model.status = job.status
+        model.current_stage = job.current_stage or ""
+        model.phase_message = job.phase_message or ""
+        model.heartbeat_at = job.heartbeat_at
         model.intent_result = job.intent_result
         model.image_analysis = job.image_analysis
         model.storyboard = job.storyboard

packages/config/base.py

@@ -90,11 +90,14 @@ class SharedSettings(BaseSettings):
 
     # ── 豆包大模型（火山引擎方舟） ────────────────────────────────────────
     doubao_api_key: str = ""
-    doubao_model: str = "doubao-seed-1-6-250615"
+    doubao_model: str = "doubao-seed-1-6-250615"  # 推理模型（通用兜底）
+    doubao_fast_model: str = "doubao-1-5-pro-32k-250115"  # 快速结构化输出模型（编导脚本/意图解析/审核）
     doubao_base_url: str = "https://ark.cn-beijing.volces.com/api/v3"
     doubao_timeout: int = 30
     doubao_max_retries: int = 2
-    doubao_vision_model: str = "doubao-1-5-vision-pro-250915"
+    doubao_vision_model: str = "doubao-1-5-vision-pro-250328"  # 高精度视觉（备用）
+    doubao_vision_lite_model: str = "doubao-1-5-vision-lite-250315"  # 快速视觉（商品识别默认，速度优先）
+    doubao_vision_use_lite: bool = True  # viral-video 图片分析默认用 lite 提速
     doubao_embedding_model: str = "doubao-embedding-large-text-240915"
     doubao_video_model: str = "doubao-seedance-2-5-260628"
     doubao_video_timeout: int = 600  # 视频生成轮询总超时（秒）

packages/domain/viral_video.py

@@ -117,6 +117,9 @@ class ViralVideoJob:
     # 状态
     id: str = field(default_factory=lambda: uuid4().hex)
     status: ViralVideoStatus = ViralVideoStatus.PENDING
+    current_stage: str = ""  # 细粒度阶段（ViralVideoStage.value，snake_case）
+    phase_message: str = ""  # 阶段中文提示文案，前端轮询直接展示
+    heartbeat_at: datetime | None = None  # worker 心跳时间，用于超时僵尸任务检测
     result_video_url: str = ""
     credits_cost: int = 0
     error_msg: str = ""
@@ -138,9 +141,19 @@ class ViralVideoJob:
         ):
             raise ValueError(f"Cannot transition from {self.status} to running")
         self.status = ViralVideoStatus.RUNNING
+        now = datetime.now(timezone.utc)
         if self.started_at is None:
-            self.started_at = datetime.now(timezone.utc)
-        self.updated_at = datetime.now(timezone.utc)
+            self.started_at = now
+        self.heartbeat_at = now
+        self.updated_at = now
+
+    def touch_heartbeat(self) -> None:
+        """更新心跳时间（worker 在长任务中周期性调用，用于超时检测）。"""
+        now = datetime.now(timezone.utc)
+        if self.started_at is None:
+            self.started_at = now
+        self.heartbeat_at = now
+        self.updated_at = now
 
     def mark_image_analyzed(self) -> None:
         if self.status not in (ViralVideoStatus.PENDING, ViralVideoStatus.RUNNING):

packages/shared/ai_client.py

@@ -40,6 +40,8 @@ class DoubaoClient:
         self.timeout: int = settings.doubao_timeout
         self.max_retries: int = settings.doubao_max_retries
         self.vision_model: str = settings.doubao_vision_model
+        self.vision_lite_model: str = settings.doubao_vision_lite_model
+        self.fast_model: str = settings.doubao_fast_model
 
     def embed_text(self, text: str, timeout: int | None = None) -> list[float] | None:
         """调用豆包文本 Embedding API，返回浮点向量；失败返回 None。"""
@@ -92,6 +94,7 @@ class DoubaoClient:
         messages: list[dict[str, str]],
         temperature: float = 0.7,
         max_tokens: int = 1024,
+        model: str | None = None,
     ) -> Optional[str]:
         """调用 Chat Completion 接口.
 
@@ -112,7 +115,7 @@ class DoubaoClient:
             "Content-Type": "application/json",
         }
         payload: dict[str, Any] = {
-            "model": self.model,
+            "model": model or self.model,
             "messages": messages,
             "temperature": temperature,
             "max_tokens": max_tokens,
@@ -154,11 +157,12 @@ class DoubaoClient:
         max_tokens: int = 2048,
         temperature: float = 0.3,
         timeout: int | None = None,
+        model: str | None = None,
     ) -> Optional[str]:
         """调用豆包视觉理解 API（OpenAI 兼容多模态格式）.
 
         将 images 附加到最后一条 user message 的 content 中，
-        使用 vision_model（默认 doubao-1-5-vision-pro-250915）。
+        使用 vision_model（默认 doubao-1-5-vision-pro-250328）。
 
         Args:
             messages: 对话消息列表。最后一条 user message 会被注入图片内容。
@@ -204,7 +208,7 @@ class DoubaoClient:
             "Content-Type": "application/json",
         }
         payload: dict[str, Any] = {
-            "model": self.vision_model,
+            "model": model or self.vision_model,
             "messages": vision_messages,
             "temperature": temperature,
             "max_tokens": max_tokens,
@@ -258,26 +262,13 @@ class DoubaoClient:
         reference_audios: list[str] | None = None,
         reference_videos: list[str] | None = None,
     ) -> str | None:
-        """调用 Seedance 2.5 文生/图生视频（异步任务→轮询→下载），返回本地 MP4 路径；失败返回 None。
+        """调用 Seedance 2.5 生视频（异步任务→轮询→下载），返回本地 MP4 路径；失败返回 None。
 
-        v1.6: 支持多参考图（产品素材）+ 参考音频（TTS口型驱动）+ 参考视频，单次生成最长 30 秒。
-
-        Args:
-            prompt: 文本提示词（含完整编导脚本：总览+场景光线+逐镜头时间轴+硬约束+负面词）
-            image_url: 首帧参考图 URL（可选，提供则走图生视频首帧模式，ratio 跟随首帧）
-            duration: 视频时长 4~30 秒
-            ratio: 宽高比 16:9/9:16/1:1/4:3/3:4/21:9/adaptive；image_url 存在时自动忽略
-            resolution: 480p/720p/1080p
-            generate_audio: 是否让模型原生合成音效/BGM（v1.6 默认 True，配合 reference_audios 做口型驱动）
-            watermark: 是否加水印
-            output_dir: 下载目录，默认 /tmp
-            model: 指定模型 ID；空则用 settings.doubao_video_model
-            reference_images: 多参考图 URL 列表（产品素材，最多30张；注意 image_url 为首帧单独传）
-            reference_audios: 参考音频 URL 列表（TTS口播，驱动口型，最多10段）
-            reference_videos: 参考视频 URL 列表（风格参考）
-
-        Returns:
-            本地 MP4 文件路径，失败返回 None。
+        【v1.6.1 修复】严格按官方 content 数组协议构造请求：
+        - 所有参考（图/音/视）必须放进 content 数组并带 role 字段，不能放顶层 reference_audios/reference_videos（非官方字段，会被忽略或导致异常）。
+        - 首帧图（first_frame 模式）Seedance 2.5 强制 ratio=adaptive；走 omni_reference（参考生视频）模式时才能指定 9:16/1:1 等具体比例。
+          判定：传了参考音频/视频或 ≥1 张多参考图时，走 omni_reference（首张图 role=reference_image）；纯首帧无参考时走 first_frame（ratio 强制 adaptive）。
+        - 创建任务若因 ratio 报错（HTTP 400），自动回退到 ratio=adaptive 重试一次。
         """
         if not self.is_available:
             return None
@@ -286,18 +277,48 @@ class DoubaoClient:
 
         settings = get_shared_settings()
         poll_interval = getattr(settings, "doubao_video_poll_interval", 10) or 10
-        total_timeout = getattr(settings, "doubao_video_timeout", 900) or 900
+        # 收紧总超时：轮询 8min + 下载 2min = 最长 ~10min，防止出现 20min 卡死
+        total_timeout = getattr(settings, "doubao_video_timeout", 480) or 480
         default_video_model = getattr(settings, "doubao_video_model", None) or "doubao-seedance-2-5-260628"
         video_model = model or default_video_model
 
+        ref_audios = [u for u in (reference_audios or [])[:10] if u and isinstance(u, str)]
+        ref_videos = [u for u in (reference_videos or [])[:3] if u and isinstance(u, str)]
+        ref_imgs = [u for u in (reference_images or [])[:9] if u and isinstance(u, str)]
+
+        # 判断任务模式：有参考音/视/多图 → omni_reference（支持指定 ratio）；纯首帧 → first_frame（ratio=adaptive）
+        has_extra_refs = bool(ref_audios or ref_videos or ref_imgs)
+        is_first_frame_mode = bool(image_url) and not has_extra_refs
+        # 最终 ratio：first_frame 模式强制 adaptive，否则按用户传值（默认 9:16）
+        final_ratio = "adaptive" if is_first_frame_mode else (ratio or "9:16")
+
+        # 构造 content 数组：text + 图 + 音 + 视
         content: list[dict[str, Any]] = [{"type": "text", "text": prompt.strip()}]
         if image_url:
-            content.append({"type": "image_url", "image_url": {"url": image_url}})
-        # v1.6: 多参考图（产品素材）
-        if reference_images:
-            for url in reference_images[:30]:
-                if url and isinstance(url, str):
-                    content.append({"type": "image_url", "image_url": {"url": url}})
+            if has_extra_refs:
+                # omni_reference：首张图作为 reference_image，允许指定 ratio
+                content.append(
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": image_url},
+                        "role": "reference_image",
+                    }
+                )
+            else:
+                # 纯首帧：不带 role，服务端识别为 first_frame（或显式 role=first_frame）
+                content.append(
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": image_url},
+                        "role": "first_frame",
+                    }
+                )
+        for u in ref_imgs:
+            content.append({"type": "image_url", "image_url": {"url": u}, "role": "reference_image"})
+        for u in ref_audios:
+            content.append({"type": "audio_url", "audio_url": {"url": u}, "role": "reference_audio"})
+        for u in ref_videos:
+            content.append({"type": "video_url", "video_url": {"url": u}, "role": "reference_video"})
 
         create_payload: dict[str, Any] = {
             "model": video_model,
@@ -306,18 +327,8 @@ class DoubaoClient:
             "duration": int(duration),
             "resolution": resolution,
             "watermark": bool(watermark),
+            "ratio": final_ratio,
         }
-        # v1.6: 参考音频（TTS 驱动口型）
-        if reference_audios:
-            create_payload["reference_audios"] = [
-                {"url": u, "role": "audio_url"} for u in reference_audios[:10] if u and isinstance(u, str)
-            ]
-        # v1.6: 参考视频（风格参考）
-        if reference_videos:
-            create_payload["reference_videos"] = [{"url": u} for u in reference_videos[:5] if u and isinstance(u, str)]
-        # Bug #2110 / v1.6: ratio=None 时不传（首帧图生视频跟随原图比例）
-        if ratio and not image_url:
-            create_payload["ratio"] = ratio
 
         headers = {
             "Authorization": f"Bearer {self.api_key}",
@@ -325,78 +336,94 @@ class DoubaoClient:
         }
         create_url = f"{self.base_url}/contents/generations/tasks"
         logger.info(
-            "Seedance 创建任务请求: url=%s model=%s duration=%ds ratio=%s gen_audio=%s image_url=%s ref_imgs=%d ref_audios=%d ref_videos=%d",
-            create_url,
+            "Seedance 创建任务: model=%s dur=%ds ratio=%s mode=%s gen_audio=%s img=%d aud=%d vid=%d",
             video_model,
             duration,
-            ratio or "(follow-image)",
+            final_ratio,
+            "first_frame" if is_first_frame_mode else "omni_ref",
             generate_audio,
-            bool(image_url),
-            len(reference_images or []),
-            len(reference_audios or []),
-            len(reference_videos or []),
+            (1 if image_url else 0) + len(ref_imgs),
+            len(ref_audios),
+            len(ref_videos),
         )
 
-        # 1) 创建任务（带重试）
-        task_id: str | None = None
-        last_error: Exception | None = None
-        for attempt in range(self.max_retries + 1):
-            try:
-                resp = httpx.post(create_url, headers=headers, json=create_payload, timeout=self.timeout)
-                # 测试环境下 MagicMock().status_code 是 MagicMock，与 int 比较会抛 TypeError；
-                # 用显式 int() 转换+类型判断，避免误判。
+        def _do_create(payload: dict) -> tuple[str | None, Exception | None, int, str]:
+            """返回 (task_id, last_err, status_code, body_text)。"""
+            last_err: Exception | None = None
+            for attempt in range(self.max_retries + 1):
                 try:
-                    _status = int(resp.status_code)
-                except (TypeError, ValueError):
-                    _status = 200
-                if _status >= 400:
-                    # 把响应体完整打出来（通常含 error.code/message，能直接定位：模型未开通/Key 无权限/模型 ID 错误）
-                    logger.error(
-                        "Seedance 创建任务 HTTP %d: body=%s",
-                        resp.status_code,
-                        (resp.text or "")[:1000],
-                    )
-                resp.raise_for_status()
-                data = resp.json()
-                task_id = data.get("id")
-                if task_id:
-                    break
-                last_error = RuntimeError(f"create task returned no id: {str(data)[:200]}")
-            except Exception as e:
-                last_error = e
-                if attempt < self.max_retries:
-                    wait = 0.5 * (2**attempt)
-                    logger.warning(
-                        "Seedance 创建任务失败，%.1fs 后重试 (%d/%d): %s", wait, attempt + 1, self.max_retries + 1, e
-                    )
-                    time.sleep(wait)
+                    resp = httpx.post(create_url, headers=headers, json=payload, timeout=self.timeout)
+                    sc = int(getattr(resp, "status_code", 0) or 0)
+                    body = (getattr(resp, "text", "") or "")[:1500]
+                    if sc >= 400:
+                        logger.error("Seedance 创建任务 HTTP %d: body=%s", sc, body)
+                        try:
+                            resp.raise_for_status()
+                        except Exception as ee:
+                            last_err = ee
+                            if attempt < self.max_retries:
+                                time.sleep(0.5 * (2**attempt))
+                                continue
+                            return None, last_err, sc, body
+                    data = resp.json()
+                    tid = data.get("id")
+                    if tid:
+                        return tid, None, sc, body
+                    last_err = RuntimeError(f"create ok but no id: {str(data)[:300]}")
+                except Exception as e:
+                    last_err = e
+                    if attempt < self.max_retries:
+                        wait = 0.5 * (2**attempt)
+                        logger.warning(
+                            "Seedance 创建任务失败，%.1fs 后重试 (%d/%d): %s",
+                            wait,
+                            attempt + 1,
+                            self.max_retries + 1,
+                            e,
+                        )
+                        time.sleep(wait)
+            return None, last_err, 0, ""
+
+        # 第一次尝试
+        task_id, last_err, sc, body = _do_create(create_payload)
+
+        # ratio 兜底：HTTP 400 且 body 提到 ratio / adaptive → 回退 adaptive 再试一次
+        if (
+            not task_id
+            and sc == 400
+            and final_ratio != "adaptive"
+            and (
+                "ratio" in (body or "").lower()
+                or "aspect" in (body or "").lower()
+                or "adaptive" in (body or "").lower()
+            )
+        ):
+            logger.warning("Seedance 创建因 ratio 失败，回退 ratio=adaptive 重试")
+            create_payload["ratio"] = "adaptive"
+            task_id, last_err, sc2, body2 = _do_create(create_payload)
+
         if not task_id:
             logger.error(
-                "Seedance 创建任务最终失败: model=%s base_url=%s err=%s  【排查建议】"
-                "1) 确认方舟控制台已开通 Doubao-Seedance-2.5 模型；"
-                "2) DOUBAO_API_KEY 对应的账号有该模型调用权限；"
-                "3) DOUBAO_BASE_URL 必须为 https://ark.cn-beijing.volces.com/api/v3；"
-                "4) 若控制台用「推理接入点」(endpoint)，请把 DOUBAO_VIDEO_MODEL 改为 ep-xxx 接入点 ID。",
+                "Seedance 创建任务最终失败: model=%s base_url=%s err=%s body=%s  【排查】"
+                "1) 方舟控制台已开通 doubao-seedance-2-5-260628；2) API Key 有该模型权限；"
+                "3) DOUBAO_BASE_URL=https://ark.cn-beijing.volces.com/api/v3；4) 参考素材 URL 公网可访问。",
                 video_model,
                 self.base_url,
-                last_error,
+                last_err,
+                (body or "")[:500],
             )
             return None
 
-        logger.info(
-            "Seedance 任务已创建: task_id=%s model=%s duration=%ds gen_audio=%s",
-            task_id,
-            video_model,
-            duration,
-            generate_audio,
-        )
+        logger.info("Seedance 任务已创建: task_id=%s ratio=%s", task_id, create_payload["ratio"])
 
         # 2) 轮询状态
         poll_url = f"{create_url}/{task_id}"
         deadline = time.time() + total_timeout
         video_url: str | None = None
         last_status: str = "queued"
+        poll_count = 0
         while time.time() < deadline:
+            poll_count += 1
             try:
                 resp = httpx.get(poll_url, headers=headers, timeout=self.timeout)
                 try:
@@ -411,48 +438,71 @@ class DoubaoClient:
                     content_obj = data.get("content") or {}
                     video_url = content_obj.get("video_url")
                     if video_url:
+                        logger.info("Seedance 任务成功: task_id=%s polls=%d", task_id, poll_count)
                         break
-                    last_error = RuntimeError(f"task succeeded but no video_url: {str(data)[:300]}")
+                    last_err = RuntimeError(f"task succeeded but no video_url: {str(data)[:300]}")
+                    logger.error("Seedance succeeded 但无 video_url: %s", last_err)
                     break
                 if status == "failed":
                     err = data.get("error") or {}
-                    last_error = RuntimeError(f"task failed: {err.get('code','')} {err.get('message','')}")
+                    last_err = RuntimeError(f"task failed: code={err.get('code','')} msg={err.get('message','')}")
+                    logger.error("Seedance 任务失败 task_id=%s: %s", task_id, last_err)
                     break
                 if status in ("expired", "cancelled"):
-                    last_error = RuntimeError(f"task {status}")
+                    last_err = RuntimeError(f"task {status}")
+                    logger.error("Seedance 任务 %s: task_id=%s", status, task_id)
                     break
-                # queued / running: 继续轮询
+                # 每 5 次轮询打一次 info 日志，便于观察进度
+                if poll_count % 5 == 0:
+                    logger.info("Seedance 轮询中: task_id=%s status=%s polls=%d", task_id, status, poll_count)
             except httpx.HTTPStatusError as e:
-                last_error = e
-                logger.warning(
-                    "Seedance 轮询 HTTP %d: body=%s",
-                    e.response.status_code,
-                    (e.response.text or "")[:500],
-                )
+                last_err = e
+                logger.warning("Seedance 轮询 HTTP %d: %s", e.response.status_code, (e.response.text or "")[:300])
             except Exception as e:
-                last_error = e
+                last_err = e
                 logger.debug("Seedance 轮询异常: %s", e)
             time.sleep(poll_interval)
 
         if not video_url:
-            logger.error("Seedance 任务未成功: task_id=%s status=%s err=%s", task_id, last_status, last_error)
+            logger.error(
+                "Seedance 任务未成功: task_id=%s last_status=%s polls=%d err=%s (总等待 %.0fs)",
+                task_id,
+                last_status,
+                poll_count,
+                last_err,
+                total_timeout,
+            )
             return None
 
-        # 3) 下载到本地
+        # 3) 下载到本地（下载超时收紧到 120s）
         try:
             out_dir = output_dir or "/tmp"
             os.makedirs(out_dir, exist_ok=True)
             local_path = f"{out_dir}/seedance_{task_id}_{uuid.uuid4().hex[:8]}.mp4"
-            with httpx.stream("GET", video_url, timeout=300) as r:
+            download_timeout = 120.0
+            logger.info(
+                "Seedance 开始下载: task_id=%s url=%s timeout=%.0fs", task_id, video_url[:120], download_timeout
+            )
+            with httpx.stream("GET", video_url, timeout=download_timeout) as r:
                 r.raise_for_status()
+                downloaded = 0
                 with open(local_path, "wb") as f:
                     for chunk in r.iter_bytes(chunk_size=1024 * 256):
                         if chunk:
                             f.write(chunk)
-            logger.info("Seedance 视频下载完成: %s (%d bytes)", local_path, os.path.getsize(local_path))
+                            downloaded += len(chunk)
+            size = os.path.getsize(local_path)
+            logger.info("Seedance 视频下载完成: %s size=%d bytes", local_path, size)
+            if size == 0:
+                logger.error("Seedance 下载文件大小为 0")
+                try:
+                    os.remove(local_path)
+                except Exception:
+                    pass
+                return None
             return local_path
         except Exception as e:
-            logger.error("Seedance 视频下载失败: %s", e)
+            logger.error("Seedance 视频下载失败: %s", e, exc_info=True)
             return None
 
 

packages/shared/ai_service.py

@@ -496,16 +496,32 @@ def run_generate_cover(
 # ── 通用 LLM / Vision 调用（#2039 ViralVideoOrchestrator 使用，复用现有豆包客户端）──
 
 
-def call_llm(prompt: str, temperature: float = 0.7) -> object:
-    """调用豆包大模型（文本对话），返回解析后的 JSON（dict/list）或原文字符串；失败返回 None。"""
+def call_llm(
+    prompt: str,
+    temperature: float = 0.7,
+    max_tokens: int = 2048,
+    model: str | None = None,
+    system_prompt: str | None = None,
+) -> object:
+    """调用豆包大模型（文本对话），返回解析后的 JSON（dict/list）或原文字符串；失败返回 None。
+
+    Args:
+        prompt: 用户侧提示。
+        temperature: 采样温度。
+        max_tokens: 输出上限（结构化任务默认 2048，长文案可按需加大）。
+        model: 覆盖默认模型（如 fast_model 提速用），None 走配置默认推理模型。
+        system_prompt: 覆盖默认 system prompt。
+    """
     client = get_doubao_client()
     if not client.is_available:
         return None
+    if system_prompt is None:
+        system_prompt = "你是专业的短视频内容策划助手。需要结构化输出时请严格使用 JSON。"
     messages = [
-        {"role": "system", "content": "你是专业的短视频内容策划助手。需要结构化输出时请严格使用 JSON。"},
+        {"role": "system", "content": system_prompt},
         {"role": "user", "content": prompt},
     ]
-    raw = client.chat_completion(messages, temperature=temperature, max_tokens=4096)
+    raw = client.chat_completion(messages, temperature=temperature, max_tokens=max_tokens, model=model)
     if raw is None:
         return None
     try:
@@ -514,14 +530,26 @@ def call_llm(prompt: str, temperature: float = 0.7) -> object:
         return raw
 
 
-def call_vision(image_url: str, prompt: str) -> object:
+def call_vision(
+    image_url: str,
+    prompt: str,
+    *,
+    model: str | None = None,
+    max_tokens: int = 1024,
+    temperature: float = 0.2,
+    timeout: int = 45,
+    system_prompt: str | None = None,
+) -> object:
     """调用豆包视觉大模型分析图片，返回解析后的 JSON 或原文字符串；失败返回 None。
 
-    Bug #2114 (VLM 牛头不对马嘴根因修复):
-    之前误走 client.chat_completion(用文本模型 doubao-seed-1.6)，多模态 content list 被当成
-    纯文本发给文本模型 → 模型要么看不到图、要么抛 400，静默被 except 吞掉 → 返回 None →
-    _step_image_analysis fallback 到 {"name":"未识别"} → 后续文案/分镜完全没图的信息。
-    现改走 vision_completion，走视觉模型 doubao-1-5-vision-pro-250915。
+    Args:
+        image_url: 可公网访问的图片 URL（直接传给豆包视觉模型，无需本地下载）。
+        prompt: 用户侧文本提示。
+        model: 覆盖默认视觉模型（如 vision_lite_model 提速用），None 走配置默认。
+        max_tokens: 输出上限，商品识别用 800~1200 足够，避免长输出拖慢首 token。
+        temperature: 温度。
+        timeout: 单次请求超时（秒）。
+        system_prompt: 覆盖默认 system prompt（viral-video 商品分析会传专门的详细 prompt）。
     """
     client = get_doubao_client()
     if not client.is_available:
@@ -531,28 +559,33 @@ def call_vision(image_url: str, prompt: str) -> object:
         logger.warning("[call_vision] 空 image_url，跳过视觉分析")
         return None
 
-    system_prompt = (
-        "你是资深电商视觉分析师。请严格基于用户提供的图片观察回答，"
-        "图片里没有的信息不要凭空想象或编造；看不清或无法判断时明确说"
-        "「图片中无法判断」，不要猜测。输出必须是严格 JSON，不要附加 Markdown 或解释文字。"
-    )
+    if system_prompt is None:
+        system_prompt = (
+            "你是资深电商视觉分析师。请严格基于用户提供的图片观察回答，"
+            "图片里没有的信息不要凭空想象或编造；看不清或无法判断时明确说"
+            "「无法判断」，不要猜测。输出必须是严格 JSON，不要附加 Markdown 或解释文字。"
+        )
     messages = [
         {"role": "system", "content": system_prompt},
         {"role": "user", "content": prompt},
     ]
 
+    used_model = model or getattr(client, "vision_model", "?")
     logger.info(
-        "[call_vision] 调用豆包视觉模型 vision_model=%s image_url=%s prompt_len=%d",
-        getattr(client, "vision_model", "?"),
+        "[call_vision] 调用豆包视觉模型 model=%s image_url=%s prompt_len=%d max_tokens=%d timeout=%d",
+        used_model,
         image_url[:120],
         len(prompt),
+        max_tokens,
+        timeout,
     )
     raw = client.vision_completion(
         messages=messages,
         images=[image_url],
-        temperature=0.2,
-        max_tokens=2048,
-        timeout=60,
+        temperature=temperature,
+        max_tokens=max_tokens,
+        timeout=timeout,
+        model=model,
     )
     if raw is None:
         logger.warning("[call_vision] 视觉模型返回 None (image_url=%s)", image_url[:80])
@@ -585,20 +618,20 @@ def call_video_generation(
     reference_audios: list[str] | None = None,
     reference_videos: list[str] | None = None,
 ) -> str | None:
-    """调用 Seedance 2.5 生成视频（v1.6 单次出片版），返回本地 MP4 路径；失败返回 None。
+    """调用 Seedance 2.5 生成视频（v1.6.1 单次出片版），返回本地 MP4 路径；失败返回 None。
 
-    v1.6:
-    - 默认 generate_audio=True，模型原生合成环境音效/BGM；
-    - reference_audios 传 TTS 音频 URL 数组做口型驱动；
-    - reference_images 传产品素材 URL 数组做视觉参考；
-    - 单次最长 30 秒，不分段不拼接；
-    - image_url 存在时为「首帧图生视频」模式，自动不传 ratio（Bug #2110）。
+    v1.6.1 关键约束（避免 20min 卡死）：
+    - 参考音频/视频/多图全部放进 content 数组并带 role=reference_audio/reference_video/reference_image；
+    - 纯首帧无参考时（first_frame 模式），Seedance 2.5 强制 ratio=adaptive；
+      传了参考音/视/多图时走 omni_reference 模式，ratio 可指定为 9:16（客户端内部自动判断）。
+    - ratio 默认 9:16（竖屏），客户端会根据是否有参考自动在 first_frame/adaptive 与 omni/9:16 间切换；
+      若创建任务因 ratio 报错（HTTP 400），客户端会自动回退到 adaptive 再试一次。
     """
     client = get_doubao_client()
     if not client.is_available:
         logger.warning("[ai_service] 豆包客户端未配置，跳过视频生成")
         return None
-    effective_ratio = None if image_url else ratio
+    effective_ratio = ratio or "9:16"
     try:
         kwargs: dict = dict(
             prompt=prompt,

scripts/render_env.sh

@@ -57,7 +57,7 @@ if [ "$TARGET_ENV" = "staging" ]; then
 fi
 
 # 共用 secrets 直接导出（如果存在）
-SHARED_SECRETS="OSS_ACCESS_KEY_ID OSS_ACCESS_KEY_SECRET COSYVOICE_API_KEY DASHSCOPE_API_KEY MEDIAKIT_API_KEY DOUBAO_API_KEY DOUBAO_MODEL DOUBAO_BASE_URL DOUBAO_VISION_MODEL WECHAT_APP_ID WECHAT_APP_SECRET TIKHUB_API_KEY APIZERO_API_KEY GPU_WORKER_TOKEN"
+SHARED_SECRETS="OSS_ACCESS_KEY_ID OSS_ACCESS_KEY_SECRET COSYVOICE_API_KEY DASHSCOPE_API_KEY MEDIAKIT_API_KEY DOUBAO_API_KEY DOUBAO_MODEL DOUBAO_FAST_MODEL DOUBAO_BASE_URL DOUBAO_VISION_MODEL DOUBAO_VISION_LITE_MODEL DOUBAO_VISION_USE_LITE WECHAT_APP_ID WECHAT_APP_SECRET TIKHUB_API_KEY APIZERO_API_KEY GPU_WORKER_TOKEN"
 for var in $SHARED_SECRETS; do
   value="${!var:-}"
   # 已经在环境中了，无需额外操作

tests/unit/test_2035_coverage.py

@@ -14,10 +14,13 @@ from packages.shared.ai_client import DoubaoClient
 class _FakeSettings:
     doubao_api_key = "test-key"
     doubao_model = "test-model"
+    doubao_fast_model = "test-fast-model"
     doubao_base_url = "https://ark.cn-beijing.volces.com/api/v3"
     doubao_timeout = 10
     doubao_max_retries = 0
     doubao_vision_model = "test-vision"
+    doubao_vision_lite_model = "test-vision-lite"
+    doubao_vision_use_lite = False
     doubao_embedding_model = "test-embedding"
 
 

tests/unit/test_health_routes.py

@@ -44,7 +44,7 @@ class TestCheckDatabase:
         assert result["type"] == "postgresql"
         assert result["message"] == "Database connection successful"
         mock_psycopg.connect.assert_called_once_with(
-            "postgresql+psycopg://test:test@localhost/test", connect_timeout=3
+            "postgresql://test:test@localhost/test", connect_timeout=3
         )
         mock_cur.execute.assert_called_once_with("SELECT 1")
         mock_conn.close.assert_called_once()
@@ -96,7 +96,7 @@ class TestCheckMigrations:
         assert result["status"] == "healthy"
         assert result["message"] == "Database migrations applied"
         mock_psycopg.connect.assert_called_once_with(
-            "postgresql+psycopg://test:test@localhost/test", connect_timeout=3
+            "postgresql://test:test@localhost/test", connect_timeout=3
         )
         mock_conn.close.assert_called_once()
 

tests/unit/test_viral_video_p0.py

@@ -253,8 +253,8 @@ class TestCallVideoGenerationV16:
             )
             assert result == str(out)
             kwargs = mock_client.video_generation.call_args.kwargs
-            # 首帧模式不传 ratio（Bug #2110）
-            assert "ratio" not in kwargs
+            # 图生视频也必须传 ratio（避免首帧方图导致默认输出 1:1）
+            assert kwargs.get("ratio") == "9:16", f"ratio 应透传，got {kwargs.get('ratio')!r}"
             assert kwargs["image_url"] == "https://img/x.jpg"
             assert kwargs["reference_audios"] == ["https://oss/tts.mp3"]
             assert kwargs["reference_images"] == ["https://img/r1.jpg"]

tests/unit/test_viral_video_routes.py

@@ -61,6 +61,8 @@ def _make_job(job_id: str = "job-1", user_id: str = "u1", status: str = "pending
         "video_ratio": "9:16",
         "video_model": "",
         "credits_cost": 0,
+        "current_stage": "",
+        "phase_message": "",
         "updated_at": None,
         "is_terminal": False,
         "effective_copy_text": "",