fix(ditto-emotion): GPU表情公式配套——全兜底置空+强度保底0.35+对齐固化 (#2292)

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

apps/web/src/pages/ai-avatar/api/aiAvatar.ts

@@ -126,4 +126,5 @@ export const finalizeRenderJob = (renderId: string, payload?: { cover_url?: stri
   apiClient.post<{ video_id: string; cover_url: string; status: string }>(
     `/ai-avatar/render/${renderId}/finalize`,
     payload ?? {},
+    { timeout: 120_000 }, // 后端兜底抽帧/写入成片库可能耗时较长
   )

packages/application/ditto_emotion_service.py

@@ -4,7 +4,8 @@
 1. 分句：先按 。！？； 句末标点切分；整段无句末标点且超长时按逗号等语义停顿二次切分
 2. 调 DoubaoClient.chat_completion 分析每句表情（emo: 0-7 八种全部开放, intensity: 0-1）
 3. 结果 LRU 缓存（文案 hash → 情绪列表）
-4. LLM 失败/超时/格式错 → 返回空列表（降级中性表情，不阻塞生成）
+4. LLM 失败/超时/格式错 → analyze 层预分句全中性兜底；build_timeline 识别全兜底特征后
+   返回空串不传 emo_timeline（避免压制 GPU 端自身表情逻辑），不阻塞生成
 5. TTS 完成后按字数比例或 sentence_timings 对齐成秒级 timeline
 """
 
@@ -31,6 +32,11 @@ EMO_SURPRISE = 6
 EMO_CONTEMPT = 7
 ALLOWED_EMOS = set(range(8))  # 8 种表情全部开放（0-7）；超出范围的非法值仍会被降级为中性
 
+# 传给 Ditto 的 emo_timeline 强度下限：GPU 端表情混合公式修复后，intensity>=0.3
+# 目标情绪才能压过中性；LLM 常给陈述句打 0.1-0.2，信号太弱、表情效果一眼假。
+# 构造 timeline entry 时统一保底到该值（不影响 analyze 层 LLM 原始输出与日志排查）
+MIN_TIMELINE_INTENSITY = 0.35
+
 # ── 分句正则 ─────────────────────────────────────────────────────
 _SENT_SPLIT_RE = re.compile(r"(?<=[。！？；!?;])\s*")
 # 二次切分：逗号/分号/顿号/冒号等语义停顿（lookbehind 保留标点在前句末尾）
@@ -82,7 +88,8 @@ class EmotionTimelineEntry:
         self.start = round(start, 2)
         self.end = round(end, 2)
         self.emo = emo
-        self.intensity = round(intensity, 2)
+        # 强度保底：低于 MIN_TIMELINE_INTENSITY 的信号在 GPU 混合公式中压不过中性
+        self.intensity = round(max(intensity, MIN_TIMELINE_INTENSITY), 2)
 
     def to_dict(self) -> dict[str, Any]:
         return {
@@ -361,6 +368,18 @@ class DittoEmotionService:
         segments = self.analyze(text)
         if not segments:
             return ""
+        # LLM 失败/异常时 _cached_analyze 会兜底返回"全中性(emo=4)+低强度(<=0.1)"的
+        # 预分句结果；把这种"明确的全中性指令"传给 GPU 会压制 Ditto 端
+        # use_script_emo/emo_global 的自身处理，比不传更糟糕。识别全兜底特征后
+        # 返回空串（payload 不带 emo_timeline），交由 GPU 端自行决定表情。
+        # 注意：LLM 真实分析即使全中性，intensity 一般 >0.1，不会被误伤。
+        if all(s.emo == EMO_NEUTRAL and s.intensity <= 0.1 for s in segments):
+            logger.info(
+                "[ditto_emotion] 分析结果全为中性低强度（疑似 LLM 兜底），"
+                "emo_timeline 置空交由 GPU 端处理: text=%.80s",
+                text,
+            )
+            return ""
         if sentence_timings:
             entries = align_timeline_by_timings(segments, sentence_timings, audio_duration)
         else:

tests/unit/test_ditto_emotion_service.py

@@ -8,8 +8,11 @@ from unittest.mock import MagicMock, patch
 import pytest
 
 from packages.application.ditto_emotion_service import (
+    EMO_ANGER,
     EMO_HAPPY,
     EMO_NEUTRAL,
+    EMO_SAD,
+    EMO_SURPRISE,
     DittoEmotionService,
     EmotionSegment,
     _parse_emotion_json,
@@ -362,3 +365,112 @@ class TestWarnIfSuspiciousAllNeutral:
         with caplog.at_level("WARNING", logger="packages.application.ditto_emotion_service"):
             _warn_if_suspicious_all_neutral("家人们！", [])
         assert not caplog.records
+
+
+# ── build_timeline 全中性兜底置空（GPU端表情公式配套修复1）─────────
+class TestBuildTimelineAllNeutralFallback:
+    """LLM 兜底特征（全 emo=4 且 intensity<=0.1）→ 不传 emo_timeline，交 GPU 端自行处理."""
+
+    def test_llm_failure_fallback_returns_empty(self):
+        """LLM 调用失败 → _cached_analyze 兜底全中性0.1 → build_timeline 置空."""
+        import packages.application.ditto_emotion_service as m
+
+        svc = _make_service(enabled=True)
+        m._cached_analyze.cache_clear()
+        try:
+            with patch.object(svc, "_call_llm", return_value=[]):
+                result = svc.build_timeline("全兜底置空检测专用独一无二文案甲。全兜底置空检测专用文案乙。", 5.0)
+                assert result == ""
+        finally:
+            m._cached_analyze.cache_clear()
+
+    def test_all_neutral_low_intensity_returns_empty(self):
+        """LLM 返回全 emo=4 且 intensity<=0.1（兜底特征）→ 置空."""
+        import packages.application.ditto_emotion_service as m
+
+        svc = _make_service(enabled=True)
+        m._cached_analyze.cache_clear()
+        fake_reply = json.dumps(
+            [
+                {"text": "低强度中性检测专用文案甲", "emo": 4, "intensity": 0.1},
+                {"text": "低强度中性检测专用文案乙", "emo": 4, "intensity": 0.05},
+            ]
+        )
+        try:
+            with patch.object(svc, "_call_llm", return_value=_parse_emotion_json(fake_reply)):
+                assert svc.build_timeline("低强度中性检测专用文案甲。低强度中性检测专用文案乙。", 5.0) == ""
+        finally:
+            m._cached_analyze.cache_clear()
+
+    def test_all_neutral_but_higher_intensity_kept(self):
+        """全中性但 intensity>0.1（LLM 真实输出特征）→ 正常生成 timeline，不误伤."""
+        import packages.application.ditto_emotion_service as m
+
+        svc = _make_service(enabled=True)
+        m._cached_analyze.cache_clear()
+        fake_reply = json.dumps(
+            [
+                {"text": "高强度中性保留检测专用文案甲", "emo": 4, "intensity": 0.3},
+                {"text": "高强度中性保留检测专用文案乙", "emo": 4, "intensity": 0.2},
+            ]
+        )
+        try:
+            with patch.object(svc, "_call_llm", return_value=_parse_emotion_json(fake_reply)):
+                result = svc.build_timeline("高强度中性保留检测专用文案甲。高强度中性保留检测专用文案乙。", 4.0)
+                data = json.loads(result)
+                assert len(data) == 2
+                assert all(item["emo"] == 4 for item in data)
+        finally:
+            m._cached_analyze.cache_clear()
+
+
+# ── timeline entry 强度保底 0.35（GPU端表情公式配套修复3）─────────
+class TestTimelineIntensityFloor:
+    def test_low_intensity_floored(self):
+        segs = [EmotionSegment("ab", EMO_HAPPY, 0.2)]
+        entries = align_timeline_by_length(segs, 2.0)
+        assert entries[0].intensity == 0.35
+
+    def test_floor_boundary_values(self):
+        segs = [
+            EmotionSegment("a", EMO_HAPPY, 0.1),
+            EmotionSegment("b", EMO_SAD, 0.34),
+            EmotionSegment("c", EMO_ANGER, 0.35),
+        ]
+        entries = align_timeline_by_length(segs, 3.0)
+        assert entries[0].intensity == 0.35
+        assert entries[1].intensity == 0.35
+        assert entries[2].intensity == 0.35
+
+    def test_high_intensity_unchanged(self):
+        segs = [EmotionSegment("ab", EMO_HAPPY, 0.6)]
+        entries = align_timeline_by_length(segs, 2.0)
+        assert entries[0].intensity == 0.6
+
+    def test_floor_applies_in_timings_path(self):
+        segs = [EmotionSegment("hello", EMO_SAD, 0.15)]
+        timings = [{"start": 0.0, "end": 2.0}]
+        entries = align_timeline_by_timings(segs, timings, 2.0)
+        assert entries[0].intensity == 0.35
+
+
+# ── timings 数量不匹配回退字数占比（GPU端表情公式配套修复2，固化行为）───
+class TestAlignTimingsMismatchCharRatio:
+    def test_mismatch_falls_back_to_char_ratio_with_emo_mapping(self):
+        """LLM 情绪分句（按逗号）与 TTS 分句（句末标点）数量不一致时，
+        必须回退到按情绪分句字数占比分配，保证 emo 与时间段一一对应."""
+        segs = [
+            EmotionSegment("a" * 6, EMO_HAPPY, 0.5),
+            EmotionSegment("b" * 2, EMO_SAD, 0.5),
+            EmotionSegment("c" * 2, EMO_SURPRISE, 0.5),
+        ]
+        timings = [{"start": 0.0, "end": 10.0}]  # len 1 != 3
+        entries = align_timeline_by_timings(segs, timings, 10.0)
+        assert len(entries) == 3
+        assert entries[0].start == 0.0
+        assert entries[0].end == 6.0
+        assert entries[1].start == 6.0
+        assert entries[1].end == 8.0
+        assert entries[2].start == 8.0
+        assert entries[2].end == 10.0
+        assert [e.emo for e in entries] == [EMO_HAPPY, EMO_SAD, EMO_SURPRISE]