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/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

@@ -99,25 +99,104 @@ def _save_job(repo, job, session):
 # ── 流水线各步骤 ────────────────────────────────────────────────────────
 
 
+_IMAGE_ANALYSIS_PROMPT = """请仔细观察这张图片，只基于图片中真实可见的内容进行分析，不要凭空想象。
+
+必须输出严格的 JSON（不要 Markdown 代码块，不要额外解释），字段如下：
+{
+  "category": "产品大类，如护肤品/彩妆/食品/数码/服饰/家居等，若无法识别填『无法判断』",
+  "name": "产品名称（从包装/品牌/logo/文字推断；没有品牌时描述外观如『粉色包装面霜』）",
+  "brand": "品牌名（看 logo/包装文字；看不清填『未知』）",
+  "colors": ["主体颜色"],
+  "material_or_texture": "材质/质地描述（如玻璃瓶装/塑料软管/哑光质感/金属外壳等；无法判断填『无法判断』）",
+  "key_features": [
+    "3-5 条**图片中确实能看到**的外观特征/卖点描述（如『按压式泵头』『瓶身有金色装饰线』等），不要编图片里没有的功效"
+  ],
+  "visual_style": "视觉风格（如简约高端/粉嫩少女/国潮/科技感/生活方式实拍等）",
+  "scene": "图片中的使用/展示场景（如白底棚拍/浴室场景/户外街拍/桌面静物等；纯白底填『白底产品图』）",
+  "target_audience_hint": "从视觉推断的目标人群（如年轻女性/男性商务/亲子家庭等；不确定填『通用』）",
+  "text_on_image": "图片上出现的可读文字（品牌名/Slogan/产品名等，没有则填『无』）"
+}
+
+严格要求：
+1. 任何字段无法确认时填『无法判断』或『未知』，不要猜。
+2. key_features 只能描述图片里肉眼可见的物理外观，不要写『补水保湿』『抗衰老』这类功效词（除非包装上明确印了）。
+3. 如果图片完全不是产品图（比如风景/人像/截图），category 填『非产品图』，name 填实际看到的内容。
+"""
+
+
 def _step_image_analysis(job: ViralVideoJob) -> dict:
-    """步骤 1: 图片 VLM 分析 — 识别产品特征、场景、卖点。"""
+    """步骤 1: 图片 VLM 分析 — 识别产品特征、场景、卖点。
+
+    Bug #2114 修复：
+    1) call_vision 现已走视觉模型 doubao-1-5-vision-pro（之前误走文本模型导致完全没看图）；
+    2) Prompt 强化为结构化 JSON schema，禁止编造，强制图片可见才写；
+    3) 单张失败不影响其他图片，最终至少返回一张占位结果避免后续 NoneType；
+    4) 日志打印每张图的 URL 和模型原始返回，方便排查。
+    """
     try:
         from packages.shared.ai_service import call_vision
     except ImportError:
         logger.warning("[爆款视频] ai_service.call_vision 不可用，使用占位结果")
-        return {"products": [{"name": "产品", "features": ["特征1", "特征2"], "scene": "通用场景"}]}
+        return {
+            "products": [
+                {
+                    "name": "产品",
+                    "features": ["特征1", "特征2"],
+                    "scene": "通用场景",
+                    "_source": "fallback_import_error",
+                }
+            ]
+        }
+
+    if not job.images:
+        logger.warning("[爆款视频] 任务无 images，跳过图片分析")
+        return {"products": []}
 
     results = []
-    for img_url in job.images:
+    for idx, img_url in enumerate(job.images):
+        logger.info("[爆款视频] 图片分析 #%d img=%s", idx, img_url[:160])
         try:
-            result = call_vision(
-                image_url=img_url,
-                prompt="请分析这张产品图片，识别：1)产品名称和类别 2)主要特征和卖点 3)适用场景 4)视觉风格。以JSON格式返回。",
-            )
-            results.append(result)
+            result = call_vision(image_url=img_url, prompt=_IMAGE_ANALYSIS_PROMPT)
+            if result is None:
+                logger.warning("[爆款视频] 图片 #%d call_vision 返回 None（模型超时/Key未配置）", idx)
+                results.append(
+                    {
+                        "name": "未识别",
+                        "category": "无法判断",
+                        "key_features": [],
+                        "scene": "通用",
+                        "_source": "vision_none",
+                    }
+                )
+            elif isinstance(result, str):
+                # JSON 解析失败返回的原文，包装一下防止后续 .get 报错
+                logger.warning("[爆款视频] 图片 #%d VLM 返回非 JSON 文本，包装为 features: %s", idx, result[:200])
+                results.append(
+                    {
+                        "name": "未识别",
+                        "category": "无法判断",
+                        "key_features": [],
+                        "scene": "通用",
+                        "_raw": result[:500],
+                        "_source": "vision_text",
+                    }
+                )
+            else:
+                # dict 正常
+                result.setdefault("_source", "vision")
+                results.append(result)
         except Exception as e:
-            logger.warning("[爆款视频] 图片分析失败 img=%s: %s", img_url, e)
-            results.append({"name": "未识别", "features": [], "scene": "通用"})
+            logger.warning("[爆款视频] 图片分析失败 img=%s err=%s", img_url[:120], e, exc_info=True)
+            results.append(
+                {
+                    "name": "未识别",
+                    "category": "无法判断",
+                    "key_features": [],
+                    "scene": "通用",
+                    "_source": "vision_exception",
+                    "_error": str(e)[:200],
+                }
+            )
 
     return {"products": results}
 
@@ -157,7 +236,18 @@ def _step_intent_parsing(job: ViralVideoJob, image_analysis: dict) -> dict:
 
     products_summary = ""
     for p in image_analysis.get("products", []):
-        products_summary += f"- {p.get('name', '产品')}: {', '.join(p.get('features', []))}\n"
+        feats = p.get("key_features") or p.get("features") or []
+        extras = []
+        if p.get("brand") and p.get("brand") not in ("未知", "无法判断"):
+            extras.append(f"品牌={p['brand']}")
+        if p.get("category") and p.get("category") not in ("无法判断", "非产品图"):
+            extras.append(f"品类={p['category']}")
+        if p.get("colors"):
+            extras.append(f"颜色={','.join(p['colors'])}")
+        if p.get("scene") and p.get("scene") not in ("通用",):
+            extras.append(f"场景={p['scene']}")
+        feat_str = ", ".join([str(x) for x in feats + extras])
+        products_summary += f"- {p.get('name', '产品')}: {feat_str}\n"
 
     prompt = f"""你是一个营销文案策略师。请分析以下信息，理解用户的营销意图：
 
@@ -192,7 +282,14 @@ def _step_copy_fusion(job: ViralVideoJob, intent: dict, image_analysis: dict) ->
 
     products_desc = ""
     for p in image_analysis.get("products", []):
-        products_desc += f"{p.get('name', '产品')}（{'，'.join(p.get('features', []))}）\n"
+        feats = p.get("key_features") or p.get("features") or []
+        extras = []
+        if p.get("brand") and p.get("brand") not in ("未知", "无法判断"):
+            extras.append(f"品牌={p['brand']}")
+        if p.get("visual_style"):
+            extras.append(f"风格={p['visual_style']}")
+        feat_str = "，".join([str(x) for x in feats + extras])
+        products_desc += f"{p.get('name', '产品')}（{feat_str}）\n"
 
     if job.fusion_level == "ai_full":
         prompt = f"""请为以下产品撰写一段爆款短视频文案（{job.duration}秒）：
@@ -251,8 +348,10 @@ def _step_storyboard(job: ViralVideoJob, copy_text: str, image_analysis: dict) -
     products = image_analysis.get("products", []) if image_analysis else []
     if products:
         p0 = products[0] if isinstance(products[0], dict) else {}
-        feats = p0.get("features", []) if isinstance(p0, dict) else []
-        products_hint = f"\n首帧参考产品特征：{p0.get('name','')} - {', '.join(feats[:3])}"
+        feats = (p0.get("key_features") or p0.get("features") or []) if isinstance(p0, dict) else []
+        brand = p0.get("brand") if isinstance(p0, dict) else ""
+        brand_hint = f"（品牌={brand}）" if brand and brand not in ("未知", "无法判断") else ""
+        products_hint = f"\n首帧参考产品特征：{p0.get('name','')}{brand_hint} - {', '.join(feats[:3])}"
 
     seg_seconds = 5
     n_segments = max(2, min(6, max(1, job.duration // seg_seconds)))
@@ -385,15 +484,28 @@ 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

@@ -515,26 +515,59 @@ def call_llm(prompt: str, temperature: float = 0.7) -> object:
 
 
 def call_vision(image_url: str, prompt: str) -> 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。"""
+    """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。
+    """
     client = get_doubao_client()
     if not client.is_available:
+        logger.warning("[call_vision] 豆包客户端未配置 (DOUBAO_API_KEY 缺失)")
         return None
+    if not image_url:
+        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。"
+    )
     messages = [
-        {"role": "system", "content": "你是专业的视觉分析师。需要结构化输出时请严格使用 JSON。"},
-        {
-            "role": "user",
-            "content": [
-                {"type": "text", "text": prompt},
-                {"type": "image_url", "image_url": {"url": image_url}},
-            ],
-        },
+        {"role": "system", "content": system_prompt},
+        {"role": "user", "content": prompt},
     ]
-    raw = client.chat_completion(messages, temperature=0.3, max_tokens=2048)
+
+    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", "?"),
+        image_url[:120],
+        len(prompt),
+    )
+    raw = client.vision_completion(
+        messages=messages,
+        images=[image_url],
+        temperature=0.2,
+        max_tokens=2048,
+        timeout=60,
+    )
     if raw is None:
+        logger.warning("[call_vision] 视觉模型返回 None (image_url=%s)", image_url[:80])
         return None
+    logger.info("[call_vision] 视觉模型原始返回 (前400字): %s", raw[:400])
+    # 剥离 ```json ... ``` 包裹
+    stripped = raw.strip()
+    if stripped.startswith("```"):
+        stripped = stripped.strip("`")
+        if stripped.startswith("json"):
+            stripped = stripped[4:].lstrip()
     try:
-        return json.loads(raw)
-    except (json.JSONDecodeError, TypeError):
+        return json.loads(stripped)
+    except (json.JSONDecodeError, TypeError) as e:
+        logger.warning("[call_vision] JSON 解析失败(%s)，返回原始文本: %s", e, raw[:200])
         return raw
 
 
@@ -543,29 +576,35 @@ 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()