fix(viral-video): 6 E2E bug fixes (fusion_level/concat/ingest/duplicated URL/TTS/Seedance ratio) (#2113)

Co-authored-by: xiaoxia <dev@xiaoxiajianji.com>
Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>

apps/api/app/api/routes/upload.py

@@ -191,6 +191,23 @@ def _find_duplicate_asset(
     return None
 
 
+
+def _get_existing_asset_url(existing: Any, storage_service: Any) -> str:
+    """安全获取已存在素材的公网 URL，兼容 domain Asset（无 file_url 字段）和 ORM model。"""
+    # Domain Asset 只有 storage_key 字段；ORM model 有 file_url 但存的也是 storage_key
+    key = ""
+    for attr in ("storage_key", "file_url"):
+        v = getattr(existing, attr, None)
+        if v:
+            key = v
+            break
+    if not key:
+        return ""
+    try:
+        return storage_service.get_url(key) or ""
+    except Exception:
+        return ""
+
 def _create_pending_asset(
     asset_repository,
     project_id,
@@ -390,7 +407,7 @@ async def prepare_direct_upload(
                 duplicated=True,
                 skip_transfer=True,
                 asset_id=existing.id,
-                url=existing.file_url or storage_service.get_url(existing.storage_key) or "",
+                url=_get_existing_asset_url(existing, storage_service),
             )
 
     file_id = uuid4().hex[:8]

apps/api/app/schemas/viral_video.py

@@ -8,7 +8,7 @@ from pydantic import BaseModel, Field, field_validator
 
 # ── 枚举常量 ─────────────────────────────────────────────────────────────
 
-VALID_FUSION_LEVELS = ("ai_full", "ai_polish", "user_primary")
+VALID_FUSION_LEVELS = ("ai_full", "full_ai", "ai_polish", "user_primary")
 VALID_STYLE_STRENGTHS = ("light", "medium", "strict")
 VALID_STAGES = (
     "image_analysis",
@@ -50,6 +50,9 @@ class CreateViralVideoRequest(BaseModel):
     @field_validator("fusion_level")
     @classmethod
     def _validate_fusion_level(cls, v: str) -> str:
+        # 兼容前端历史写法 full_ai（等价 ai_full）
+        if v == "full_ai":
+            return "ai_full"
         if v not in VALID_FUSION_LEVELS:
             raise ValueError(f"fusion_level 必须是 {VALID_FUSION_LEVELS} 之一")
         return v

apps/web/src/api/viral-video/types.ts

@@ -1,8 +1,8 @@
-export type FusionLevel = "full_ai" | "polish" | "as_is"
+export type FusionLevel = "ai_full" | "ai_polish" | "user_primary"
 export const FUSION_LEVELS: { value: FusionLevel; label: string; desc: string }[] = [
-  { value: "full_ai", label: "AI 全写", desc: "给我方向，全由AI创作" },
-  { value: "polish", label: "AI润色", desc: "我写草稿，AI帮我润色" },
-  { value: "as_is", label: "按我写的来", desc: "几乎不改我的文案" },
+  { value: "ai_full", label: "AI 全写", desc: "给我方向，全由AI创作" },
+  { value: "ai_polish", label: "AI润色", desc: "我写草稿，AI帮我润色" },
+  { value: "user_primary", label: "按我写的来", desc: "几乎不改我的文案" },
 ]
 
 export type StyleStrength = "light" | "medium" | "strict"

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

@@ -121,7 +121,7 @@ const emptyTask = (id: string, title: string): TabTask => ({
   title,
   images: [],
   refVideo: null,
-  fusionLevel: "full_ai",
+  fusionLevel: "ai_full",
   styleStrength: "medium",
   voiceId: "",
   userCopy: "",
@@ -425,7 +425,7 @@ const ViralVideoPage: React.FC = () => {
         images: imgUrls,
         reference_video_url: task.refVideo?.ossUrl || undefined,
         style_strength: task.refVideo?.ossUrl ? task.styleStrength : undefined,
-        user_copy_text: task.fusionLevel !== "full_ai" ? task.userCopy : undefined,
+        user_copy_text: task.fusionLevel !== "ai_full" ? task.userCopy : undefined,
         fusion_level: task.fusionLevel,
         // 注意：STEP1「开始分析」阶段还没有真正开始生成视频，后端会跑到 WAIT_USER_CONFIRM 暂停。
         // 配音和视频参数在 STEP2 确认文案时由 confirm-intent 继续使用；但后端接口目前只接收一次 payload，
@@ -816,14 +816,14 @@ const ViralVideoPage: React.FC = () => {
             </div>
             <div className="vv-form-row">
               <label className="vv-label">
-                {task.fusionLevel === "full_ai"
+                {task.fusionLevel === "ai_full"
                   ? "你想拍什么？（给 AI 方向即可，可选）"
                   : "写下你的文案/草稿"}
               </label>
               <textarea
                 className="vv-textarea"
                 placeholder={
-                  task.fusionLevel === "full_ai"
+                  task.fusionLevel === "ai_full"
                     ? "例：夏季上新的连衣裙，突出清爽、显白、百搭"
                     : "例：姐妹们！这件连衣裙夏天穿太清爽了…"
                 }

apps/worker/video_processing/concat_engine.py

@@ -553,7 +553,20 @@ def concat_video_files(
     if work_dir is None:
         work_dir = output_path.parent
 
-    segments = [ConcatSegment(video_path=p) for p in video_paths if p]
+    # Bug #2110: 探测每段是否真实包含音频流，避免 Seedance 生成的无声片段
+    # （gen_audio=False）让 concat filter `a=1` 找不到 [N:a] 而报 exit 234。
+    from video_processing.ffmpeg_utils import probe_has_audio as _probe_has_audio
+
+    segments: list[ConcatSegment] = []
+    for p in video_paths:
+        if not p:
+            continue
+        try:
+            has_audio = _probe_has_audio(p)
+        except Exception:
+            has_audio = True  # 探测失败保守认为有音频
+        segments.append(ConcatSegment(video_path=p, has_audio=has_audio))
+
     config = ConcatConfig(segments=segments, force_reencode=force_reencode)
 
     engine = ConcatEngine(work_dir)

apps/worker/worker_app/tasks/viral_video.py

@@ -385,15 +385,28 @@ def _step_tts(job: ViralVideoJob, copy_text: str):
         from apps.worker.services.tts_service_factory import get_tts_service
 
         tts_service = get_tts_service()
-        # 兼容老接口：部分 provider 只接收 text 参数
+        # Bug #2110: persona_id 透传给 voice_id（空则用 CosyVoice 默认 longxiaochun_v3），
+        # 统一输出 mp3 给后续 ffmpeg 混音（之前默认 wav 导致部分 provider/后处理不兼容）。
+        voice_id = (job.persona_id or "").strip()
         try:
-            result = tts_service.synthesize(text=copy_text, voice_id=job.persona_id or "default")
+            result = tts_service.synthesize(
+                text=copy_text,
+                voice_id=voice_id or "longxiaochun_v3",
+                format="mp3",
+            )
         except TypeError:
-            result = tts_service.synthesize(text=copy_text)
+            # 老 provider 只支持 text 参数
+            try:
+                result = tts_service.synthesize(text=copy_text, voice_id=voice_id or "longxiaochun_v3")
+            except TypeError:
+                result = tts_service.synthesize(text=copy_text)
         if result is None:
             return None
         p = _Path(result) if not isinstance(result, _Path) else result
         if p.exists():
+            logger.info(
+                "[爆款视频] TTS 合成完成: voice=%s path=%s size=%d", voice_id or "longxiaochun_v3", p, p.stat().st_size
+            )
             return p
         logger.warning("[爆款视频] TTS 返回路径不存在: %s", p)
         return None

packages/adapters/sqlalchemy_impl/asset_repository.py

@@ -502,8 +502,19 @@ class SQLAlchemyAssetRepository:
         return [self._to_domain(m) for m in models]
 
     def find_by_storage_key(self, storage_key: str) -> Asset | None:
-        """按 storage_key（对应 DB 中的 file_url）查找素材。"""
-        model = self.session.query(AssetModel).filter(AssetModel.file_url == storage_key).first()
+        """按 storage_key 查找素材。
+
+        Bug #2110: 历史数据 file_url 列可能是旧路径（assets/...），新代码统一写入
+        storage_key 列。双列 OR 查询，避免占位 asset 因路径错配导致 ingest 兜底新建
+        第二条 READY 记录，原占位卡 PROCESSING → 前端缩略图出现后消失。
+        """
+        if not storage_key:
+            return None
+        model = (
+            self.session.query(AssetModel)
+            .filter((AssetModel.storage_key == storage_key) | (AssetModel.file_url == storage_key))
+            .first()
+        )
         if model is None:
             return None
         return self._to_domain(model)

packages/shared/ai_client.py

@@ -248,7 +248,7 @@ class DoubaoClient:
         *,
         image_url: str | None = None,
         duration: int = 5,
-        ratio: str = "9:16",
+        ratio: str | None = "9:16",
         resolution: str = "720p",
         generate_audio: bool = False,
         watermark: bool = False,
@@ -287,11 +287,13 @@ class DoubaoClient:
             "model": video_model,
             "content": content,
             "generate_audio": generate_audio,
-            "ratio": ratio,
             "duration": int(duration),
             "resolution": resolution,
             "watermark": watermark,
         }
+        # Bug #2110: ratio=None 时不传（首帧图生视频跟随原图比例，传 ratio 会 400 InvalidParameter）
+        if ratio:
+            create_payload["ratio"] = ratio
 
         headers = {
             "Authorization": f"Bearer {self.api_key}",
@@ -299,12 +301,13 @@ 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",
+            "Seedance 创建任务请求: url=%s model=%s duration=%ds ratio=%s gen_audio=%s image_url=%s",
             create_url,
             video_model,
             duration,
-            ratio,
+            ratio or "(follow-image)",
             generate_audio,
+            bool(image_url),
         )
 
         # 1) 创建任务（带重试）
@@ -313,7 +316,13 @@ class DoubaoClient:
         for attempt in range(self.max_retries + 1):
             try:
                 resp = httpx.post(create_url, headers=headers, json=create_payload, timeout=self.timeout)
-                if resp.status_code >= 400:
+                # 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。
+                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",
@@ -357,7 +366,11 @@ class DoubaoClient:
         while time.time() < deadline:
             try:
                 resp = httpx.get(poll_url, headers=headers, timeout=self.timeout)
-                resp.raise_for_status()
+                try:
+                    if int(getattr(resp, "status_code", 200)) >= 400:
+                        resp.raise_for_status()
+                except (TypeError, ValueError):
+                    pass
                 data = resp.json()
                 status = data.get("status", "")
                 last_status = status

packages/shared/ai_service.py

@@ -543,29 +543,35 @@ def call_video_generation(
     *,
     image_url: str | None = None,
     duration: int = 5,
-    ratio: str = "9:16",
+    ratio: str | None = "9:16",
     resolution: str = "720p",
     output_dir: str | None = None,
 ) -> str | None:
     """调用 Seedance 2.5 生成视频段，返回本地 MP4 路径；失败返回 None。
 
     封装 ai_client.video_generation：提交异步任务→轮询→下载到本地。
+    Bug #2110: 首帧参考图模式下不传 ratio（API 要求跟随首帧图比例，传 ratio=9:16
+    会返回 400 InvalidParameter）。
     """
     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
+    # 首帧模式：不强制 ratio，让模型跟随首帧图比例
+    effective_ratio = None if image_url else ratio
     try:
-        return client.video_generation(
+        kwargs: dict = dict(
             prompt=prompt,
             image_url=image_url,
             duration=duration,
-            ratio=ratio,
             resolution=resolution,
             generate_audio=False,  # 我们自己混 TTS
             watermark=False,
             output_dir=output_dir,
         )
+        if effective_ratio:
+            kwargs["ratio"] = effective_ratio
+        return client.video_generation(**kwargs)
     except Exception as e:
         logger.error("[ai_service] call_video_generation 异常: %s", e, exc_info=True)
         return None

tests/unit/test_ai_client_video.py

@@ -46,6 +46,8 @@ class TestVideoGenerationHappyPath:
         fake_task_resp = MagicMock()
         fake_task_resp.json.return_value = {"id": "task-001"}
         fake_task_resp.raise_for_status = MagicMock()
+        fake_task_resp.status_code = 200
+        fake_task_resp.text = ""
 
         fake_poll_resp = MagicMock()
         fake_poll_resp.json.return_value = {
@@ -53,6 +55,8 @@ class TestVideoGenerationHappyPath:
             "content": {"video_url": "https://cdn.example.com/v.mp4"},
         }
         fake_poll_resp.raise_for_status = MagicMock()
+        fake_poll_resp.status_code = 200
+        fake_poll_resp.text = ""
 
         class FakeStreamResponse:
             def __init__(self):

tests/unit/test_prepare_dedup_1714.py

@@ -144,6 +144,9 @@ def _storage():
         "expires_at": "2026-01-01T00:00:00Z",
         "fields": {"key": "uploads/abc/test.mp4"},
     }
+    # Bug #2110: duplicated 命中时 _get_existing_asset_url 调用 get_url 返回公网 URL 字符串，
+    # Mock 默认返回 MagicMock，会让 DirectUploadPrepareResponse.url: str 校验失败。
+    s.get_url.return_value = ""
     return s
 
 

tests/unit/test_viral_video_p0.py

@@ -120,7 +120,7 @@ class TestTTSPath:
         """get_tts_service 抛 ImportError 时 _step_tts 返回 None。"""
         from apps.worker.worker_app.tasks import viral_video as vv
 
-        with patch("services.tts_service_factory.get_tts_service", side_effect=ImportError("no tts")):
+        with patch("apps.worker.services.tts_service_factory.get_tts_service", side_effect=ImportError("no tts")):
             assert vv._step_tts(mock_job, "文案") is None
 
     def test_tts_returns_none_when_path_not_exists(self, mock_job, tmp_path):
@@ -128,7 +128,7 @@ class TestTTSPath:
 
         fake_service = MagicMock()
         fake_service.synthesize.return_value = str(tmp_path / "not_exist.mp3")
-        with patch("services.tts_service_factory.get_tts_service", return_value=fake_service):
+        with patch("apps.worker.services.tts_service_factory.get_tts_service", return_value=fake_service):
             assert vv._step_tts(mock_job, "文案") is None
 
     def test_tts_returns_path_when_exists(self, mock_job, tmp_path):
@@ -138,8 +138,11 @@ class TestTTSPath:
         audio.write_bytes(b"ID3fake")
         fake_service = MagicMock()
         fake_service.synthesize.return_value = audio
-        with patch("services.tts_service_factory.get_tts_service", return_value=fake_service):
+        with patch("apps.worker.services.tts_service_factory.get_tts_service", return_value=fake_service):
             result = vv._step_tts(mock_job, "文案")
+            # Bug #2110: 校验传入了 voice_id+format=mp3
+            call_kwargs = fake_service.synthesize.call_args.kwargs
+            assert call_kwargs.get("format") == "mp3"
             assert isinstance(result, Path)
             assert result.exists()