style: auto-format with black + isort [skip ci-format-check]

- _check_and_fix_tts_duration 新增 model 参数，按官方 reference_audio 硬限细分：
  * dashscope (Wan 3.0): 14.5s（硬限15s）
  * model含'2.5' (Seedance 2.5): 29s（硬限30s）
  * 其他 (Seedance 2.0/Fast/Mini): 14.5s（硬限15s）
- 调用处透传 _vv_model
- 日志增加 provider/model 字段便于排查

apps/web/src/pages/viral-video/components/StoryboardCard.tsx

@@ -79,7 +79,7 @@ export const StoryboardCard: React.FC<StoryboardCardProps> = ({
         <div className="vv-copy-box vv-copy-loading">
           <VVSpinner />
           <span className="vv-copy-loading-text">{stageMsg}</span>
-          <span className="vv-copy-loading-hint">预计 10-30 秒，请稍候</span>
+          <span className="vv-copy-loading-hint">预计 5-15 秒，请稍候</span>
         </div>
       )
     }

apps/worker/worker_app/tasks/viral_video.py

@@ -1174,6 +1174,78 @@ def _parse_script(raw: str, job: ViralVideoJob) -> dict | None:
     return _parse_script_markdown(raw, job, dur, ratio, default_theme)
 
 
+# #2308: 从旧schema混写的<visual>字段中拆分出 摄影参数 / 画面描述 / 动作细节
+_CAMERA_KEYWORDS = re.compile(
+    r"(?:景别|景|远景|全景|中景|近景|特写|大特写|俯拍|仰拍|平视|侧拍|"
+    r"机位|高度|机位高度|鸟瞰|斜拍|跟拍|固定镜头|手持|稳定器|"
+    r"推镜|拉镜|摇镜|移镜|平移|横移|升降|环绕|旋转|变焦|"
+    r"缓慢推|快速推|缓推|急推|慢推|匀速|"
+    r"[\d]+mm|f/[\d.]+|光圈|焦段|焦距|景深|浅景深)"
+)
+_ACTION_VERBS = re.compile(
+    r"(?:表情|面带|露出|微笑|皱眉|闭眼|睁眼|眼神|疲惫|放松|紧张|看向|望着|注视|凝视|"
+    r"伸手|抬手|挥手|举起|放下|拿起|握住|递|接|手扶|轻抚|抚摸|触碰|轻触|触摸|翻看|"
+    r"转身|转头|回头|点头|摇头|低头|抬头|歪头|扭头|"
+    r"走|跑|站|坐|躺|靠|倚|弯腰|侧身|前倾|后仰|走进|走出|推门|开门|关门|进入|离开|"
+    r"说话|开口|张嘴|嘴角上扬|深呼吸|吐气|流泪|"
+    r"动作|动作细节|操作|演示|接过|递上|展示|做出|比划)"
+)
+
+
+def _split_visual_field(visual: str) -> tuple[str, str, str]:
+    """把旧schema混写的<visual>文本拆成 (摄影参数, 画面描述, 动作细节)。
+
+    策略：按逗号/顿号/分号切分小片段，含摄影关键词→camera，含动作动词→action，
+    其余（主体/场景/光线/氛围）→visual。切不动时按第一个逗号前归camera、
+    最后一句含动词归action、中间归visual 兜底。
+    """
+    text = (visual or "").strip()
+    if not text:
+        return "", "", ""
+    # 去掉首尾的"镜头："/"画面："等前缀
+    text = re.sub(r"^(?:镜头|画面|视觉)\s*[:：]\s*", "", text)
+    cam_parts: list[str] = []
+    act_parts: list[str] = []
+    vis_parts: list[str] = []
+    # 按中英文逗号、顿号、分号切片
+    segs = [s.strip() for s in re.split(r"[，,、;；]\s*", text) if s.strip()]
+    if len(segs) <= 1:
+        return text, text, ""  # 切不开，整体当画面+摄影
+    for seg in segs:
+        if _CAMERA_KEYWORDS.search(seg) and not _ACTION_VERBS.search(seg):
+            cam_parts.append(seg)
+        elif _ACTION_VERBS.search(seg) and not _CAMERA_KEYWORDS.search(seg):
+            act_parts.append(seg)
+        else:
+            vis_parts.append(seg)
+    # 兜底：如果 camera 为空，取第一段
+    if not cam_parts and segs:
+        cam_parts.append(segs[0])
+        vis_parts = [s for s in vis_parts if s != segs[0]]
+    # 兜底1：最后一段含动作动词
+    if not act_parts and segs and _ACTION_VERBS.search(segs[-1]):
+        act_parts.append(segs[-1])
+        vis_parts = [s for s in vis_parts if s != segs[-1]]
+    # 兜底2：从后往前扫，找到第一个含动作动词且不含摄影关键词的段
+    if not act_parts:
+        for seg in reversed(segs):
+            if _ACTION_VERBS.search(seg) and seg not in cam_parts:
+                act_parts.append(seg)
+                vis_parts = [s for s in vis_parts if s != seg]
+                break
+    # 兜底3：包含人物主语（妈妈/店员/女生/男生/她/他/人物/模特）+动词的段视为动作
+    if not act_parts:
+        _person_re = re.compile(
+            r"(?:妈妈|店员|女生|男生|女士|男士|女孩|男孩|人物|模特|主播|她|他|主角|孩子|宝宝|老人|顾客|客人|顾客)"
+        )
+        for seg in reversed(segs):
+            if _person_re.search(seg) and seg not in cam_parts and not _CAMERA_KEYWORDS.search(seg):
+                act_parts.append(seg)
+                vis_parts = [s for s in vis_parts if s != seg]
+                break
+    return "，".join(cam_parts), "，".join(vis_parts), "，".join(act_parts)
+
+
 def _parse_script_xml(
     xml_body: str,
     raw_full: str,
@@ -1199,7 +1271,22 @@ def _parse_script_xml(
         a = c.get("attrs", {})
         body = c.get("text", "") or ""
         voice = xp.text_of(body, "voiceover") or xp.text_of(body, "voice_text")
-        visual = xp.text_of(body, "visual") or xp.text_of(body, "shot_type_angle_movement")
+        # #2308: 优先读新 <shot>（纯6项摄影参数），fallback 到 <visual>（旧schema混写）
+        shot_meta = xp.text_of(body, "shot") or ""
+        visual_full = xp.text_of(body, "visual") or xp.text_of(body, "shot_type_angle_movement") or ""
+        action_raw = xp.text_of(body, "action_details") or ""
+        if shot_meta:
+            # 新schema：<shot>是摄影参数，<visual>是纯画面描述，<action_details>是动作细节
+            cam_text = shot_meta.strip()
+            visual_text = visual_full.strip()
+            action_text = (action_raw or "").strip()
+            if action_text in ("-", "—", "无", "无。"):
+                action_text = ""
+        else:
+            # 旧schema兼容：<visual> 混写了"景别/运镜+画面+动作"，用正则拆分
+            cam_text, visual_text, action_text = _split_visual_field(visual_full)
+            if not action_text and action_raw:
+                action_text = action_raw.strip()
         ref_raw = xp.text_of(body, "reference_image_index")
         ref_idx = xp.attr_int(ref_raw, -1) if ref_raw not in (None, "") else -1
         if not isinstance(ref_idx, int) or ref_idx < 0:
@@ -1209,9 +1296,13 @@ def _parse_script_xml(
             voice_parts.append(f"[speaker:{speaker}]{voice}")
         shot = {
             "time_range": a.get("time_range") or xp.text_of(body, "time_range") or f"{i * 3}-{(i + 1) * 3}秒",
-            "shot_type_angle_movement": visual or "中景平视，固定镜头",
-            "scene_and_dialogue": (voice or "") if not visual else f"{visual}｜{voice or ''}",
-            "action_details": xp.text_of(body, "action_details") or "",
+            # shot_type_angle_movement：纯6项摄影参数，不塞画面动作
+            "shot_type_angle_movement": cam_text or "中景平视，固定镜头",
+            # scene_and_dialogue：纯口播对白（不再拼接visual，避免重复）
+            "scene_and_dialogue": voice or "",
+            # 画面描述（拆出来的纯画面）放到 shot 里供视频prompt使用
+            "visual": visual_text,
+            "action_details": action_text,
             "audio_bgm": xp.text_of(body, "audio_bgm") or "轻快BGM",
             "transition": xp.text_of(body, "transition") or ("硬切" if i < len(clips) - 1 else "结束"),
             "reference_image_index": ref_idx,
@@ -1245,11 +1336,13 @@ _MD_TIME_SPLIT_RE = re.compile(
 )
 _MD_TIME_RE = re.compile(r"(\d+)\s*[-~到至]\s*(\d+)\s*秒?")
 _MD_FIELD_RE = {
-    "visual": re.compile(r"(?:画面|视觉|镜头画面|shot_type_angle_movement)\s*[:：]\s*([^\n]+)"),
-    "camera": re.compile(r"(?:景别|运镜|镜头运动)\s*[:：]\s*([^\n]+)"),
+    "visual": re.compile(
+        r"(?:画面(?:内容|描述)?|视觉|镜头画面|场景(?:描述)?|shot_type_angle_movement)\s*[:：]\s*([^\n]+)"
+    ),
+    "camera": re.compile(r"(?:景别(?:/角度)?(?:与运镜)?|运镜|镜头运动|镜头|shot|camera)\s*[:：]\s*([^\n]+)"),
     "voiceover": re.compile(r"""(?:口播|对白|台词|配音|旁白|voiceover)\s*[:：]\s*""" r"""(?:"([^"]+)"|'([^']+)'|(\S[^
 ]*?)(?=\n\s*(?:镜头|#|$)|\s*$))"""),
-    "action": re.compile(r"(?:动作|动作细节|action)\s*[:：]\s*([^\n]+)"),
+    "action": re.compile(r"(?:动作(?:与真人细节)?|动作细节|人物动作|action)\s*[:：]\s*([^\n]+)"),
     "audio_bgm": re.compile(r"(?:音效|BGM|背景音乐|audio)\s*[:：]\s*([^\n]+)"),
     "transition": re.compile(r"(?:转场|过渡)\s*[:：]\s*([^\n]+)"),
     "reference": re.compile(r"参考@图片(\d+)|参考图[:：]?\s*(\d+)"),
@@ -1372,14 +1465,15 @@ def _parse_script_markdown(
         if voice:
             voice_parts.append(voice)
 
-        # scene_and_dialogue 保持原有拼接格式
-        visual_desc = visual or "中景平视，固定镜头"
-        scene_and_dialogue = (voice or "") if not visual_desc else f"{visual_desc}｜{voice or ''}"
+        # #2308: scene_and_dialogue 只放口播对白，画面/景别已在独立字段
+        _ = visual or camera  # 兼容保留
+        scene_and_dialogue = voice or ""
 
         shot = {
             "time_range": time_range,
             "shot_type_angle_movement": shot_type,
             "scene_and_dialogue": scene_and_dialogue,
+            "visual": visual or "",
             "action_details": action,
             "audio_bgm": audio or "轻快BGM",
             "transition": trans or ("硬切" if i < len(shot_blocks) - 1 else "结束"),
@@ -1829,11 +1923,14 @@ def _upload_tts_to_oss(job: ViralVideoJob, tts_path) -> str | None:
         return None
 
 
-def _check_and_fix_tts_duration(tts_path, target_duration: int):
-    """Bug3 fix: TTS 合成后校验音频时长，超限则加速/截断。
+def _check_and_fix_tts_duration(tts_path, target_duration: int, provider: str = "doubao", model: str = ""):
+    """#2307: TTS 合成后校验音频时长，超限则加速/截断兜底。
 
-    - 音频时长 > max(target+2, 30) → 加速到目标时长
-    - 音频时长 > 30s（Seedance 硬限制）→ 必须加速/截断到 30s 以内
+    按模型reference_audio硬限动态调整（留0.5s余量）：
+      - Wan 3.0 (dashscope): reference_audio 单段 ≤15s → 卡14.5s
+      - Seedance 2.5 (doubao, model含"2.5"): reference_audio ≤30s → 卡29s
+      - Seedance 2.0/Fast/Mini (doubao): reference_audio ≤15s → 卡14.5s
+    soft_limit=target+2（atempo 1.1-1.2x 用户听不出差异），不收紧max_chars。
     返回处理后的路径（可能覆盖原文件），失败返回原路径不阻塞。
     """
     if tts_path is None:
@@ -1863,15 +1960,33 @@ def _check_and_fix_tts_duration(tts_path, target_duration: int):
             return tts_path
 
         audio_dur = float(result.stdout.strip())
-        hard_limit = 30.0  # Seedance 硬限制
+        # 按provider+model取reference_audio硬限（留0.5s余量）
+        _prov = (provider or "doubao").lower()
+        _model_lc = (model or "").lower()
+        if _prov == "dashscope":
+            hard_limit = 14.5  # Wan 3.0 reference_audio硬限15s
+        elif "2.5" in _model_lc:
+            hard_limit = 29.0  # Seedance 2.5 reference_audio硬限30s
+        else:
+            hard_limit = 14.5  # Seedance 2.0/Fast/Mini reference_audio硬限15s
         soft_limit = float(target_duration) + 2.0
         effective_limit = min(soft_limit, hard_limit)
+        if float(target_duration) > hard_limit + 1.0:
+            logger.warning(
+                "[爆款视频] 目标时长%ds超过模型%s/%s音频硬限%.1fs，TTS将压缩到硬限",
+                target_duration,
+                _prov,
+                _model_lc or "default",
+                hard_limit,
+            )
 
         logger.info(
-            "[爆款视频] TTS 音频时长检查: audio=%.1fs target=%ds limit=%.1fs",
+            "[爆款视频] TTS 音频时长检查: audio=%.1fs target=%ds limit=%.1fs provider=%s model=%s",
             audio_dur,
             target_duration,
             effective_limit,
+            _prov,
+            _model_lc or "default",
         )
 
         if audio_dur <= effective_limit:
@@ -1879,9 +1994,9 @@ def _check_and_fix_tts_duration(tts_path, target_duration: int):
 
         # 需要处理：优先用 ffmpeg 加速（atempo）保持内容完整
         if audio_dur > hard_limit:
-            target_sec = 29.0  # 必须压缩到 30s 以内
+            target_sec = hard_limit
         else:
-            target_sec = float(target_duration)
+            target_sec = min(float(target_duration), hard_limit)
 
         speed_factor = audio_dur / target_sec
         if speed_factor > 2.0:
@@ -2067,8 +2182,14 @@ def _truncate_voiceover(text: str, max_chars: int) -> str:
 
 
 def _shot_visual(s: dict) -> str:
-    """合并镜头画面与动作描述。"""
-    visual = str(s.get("scene_and_dialogue", "") or s.get("visual", "")).strip()
+    """合并镜头画面与动作描述（#2308: scene_and_dialogue现在只放对白，画面应读visual字段）。"""
+    visual = str(s.get("visual", "") or "").strip()
+    # 兜底：旧数据/兜底脚本可能没填visual，则从scene_and_dialogue取
+    if not visual:
+        sd = str(s.get("scene_and_dialogue", "") or "").strip()
+        # scene_and_dialogue现在是纯对白（用｜分隔是旧格式），只有不含对白痕迹时才当画面
+        if "｜" not in sd and not any(c in sd for c in "，。？！") or len(sd) < 20:
+            visual = sd
     act = str(s.get("action_details", "") or "").strip()
     if act and act not in visual:
         visual = f"{visual}，{act}" if visual else act
@@ -3055,7 +3176,7 @@ def _run_render_pipeline(job_id: str, session, repo, job) -> dict:
         # Bug3 fix: TTS 音频时长校验与修正
         if tts_path is not None:
             _dur_limit = max(5, min(30, int(getattr(job, "duration", 15) or 15)))
-            tts_path = _check_and_fix_tts_duration(tts_path, _dur_limit)
+            tts_path = _check_and_fix_tts_duration(tts_path, _dur_limit, provider=_vv_provider, model=_vv_model or "")
         tts_url = _upload_tts_to_oss(job, tts_path)
         _emit_progress(job_id, ViralVideoStage.TTS, 78.0, "配音完成", {"has_tts": tts_url is not None})
 

packages/application/viral_video/reviewer.py

@@ -123,7 +123,7 @@ class Reviewer:
             messages,
             temperature=0.2,
             max_tokens=1024,
-            timeout=25,
+            timeout=15,  # #2307: 25s过长，超时降级放行
         )
         if not raw:
             return None