feat(#1970): 智能降重 PR2 - dedup_enabled 开关 + 6 维片段级微变换 (#1975)

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

apps/worker/video_processing/unified_render_service.py

@@ -171,6 +171,87 @@ class UnifiedRenderService:
         self._speed_engine = SpeedEngine()
         self._asr_timeline_cache: Any = None  # ASR 字幕结果缓存，避免重复调用
         self._asr_timeline_cached = False
+        # #1970 PR2：片段级微变换计划缓存（懒构建，dedup_enabled=False 时为 None）
+        self._micro_plan_cache: Any = None
+        self._micro_plan_loaded = False
+
+    # ── #1970 PR2 智能降重：片段级微变换 ───────────────────────────────────
+    def _dedup_enabled(self) -> bool:
+        """读取 plan.config.dedup_enabled，缺省视为 True（向后兼容）。"""
+        cfg = self.plan.config or {}
+        return bool(cfg.get("dedup_enabled", True))
+
+    def _get_micro_transform_plan(self, clip_count: int) -> Any:
+        """按 task_id+视频序号构建可复现的片段级微变换计划。
+
+        种子 hash(generation_task_id + video_index)%10000，同一任务重渲结果一致。
+        dedup_enabled=False 时返回 None，调用方不注入任何微变换。
+        P1 字幕检测：无可靠的片段文字轨道信息，hflip 一律关闭（宁可不翻转）。
+        """
+        if self._micro_plan_loaded:
+            return self._micro_plan_cache
+        self._micro_plan_loaded = True
+        if not self._dedup_enabled() or clip_count <= 0:
+            self._micro_plan_cache = None
+            return None
+        try:
+            from video_processing.micro_transform_pure import build_micro_transform_plan
+
+            cfg = self.plan.config or {}
+            task_id = str(cfg.get("generation_task_id", "") or "")
+            video_index = int(cfg.get("video_index", 0) or 0)
+            self._micro_plan_cache = build_micro_transform_plan(
+                task_id,
+                video_index,
+                clip_count,
+                clip_has_text=None,  # P1 保守策略：全部按有文字处理，不翻转
+                enable_bgm_offset=bool(cfg.get("bgm")),
+            )
+        except Exception as e:
+            logger.warning("[unified-render] 微变换计划构建失败，本次不注入: %s", e)
+            self._micro_plan_cache = None
+        return self._micro_plan_cache
+
+    @staticmethod
+    def _apply_micro_transform_video(filters: list[str], mt: Any) -> None:
+        """把片段视频微变换就地追加到 filter 链（post-scale 阶段调用）。
+
+        顺序：hflip 在 pre-scale 阶段由 _apply_micro_hflip 处理，这里只加
+        eq 亮度/对比度/饱和度。速度 setpts 与既有 clip speed 相乘（见调用点），
+        避免出现两条 setpts 互相覆盖。
+        """
+        if mt is None:
+            return
+        if abs(mt.brightness) > 1e-4 or abs(mt.contrast - 1.0) > 1e-4 or abs(mt.saturation - 1.0) > 1e-4:
+            filters.append(
+                f"eq=brightness={mt.brightness:+.4f}:" f"contrast={mt.contrast:.4f}:saturation={mt.saturation:.4f}"
+            )
+
+    @staticmethod
+    def _apply_micro_hflip(filters: list[str], mt: Any) -> None:
+        """片段级水平翻转（pre-scale 阶段）。P1 有文字/无法判定时 mt.hflip=False。"""
+        if mt is not None and mt.hflip and not mt.has_text:
+            filters.append("hflip")
+
+    @staticmethod
+    def _micro_speed_factor(mt: Any) -> float:
+        """片段微变换速度因子（0.97~1.03），无计划返回 1.0。"""
+        if mt is None:
+            return 1.0
+        return float(getattr(mt, "speed", 1.0) or 1.0)
+
+    def _get_micro_bgm_offset(self) -> float:
+        """#1970 PR2：读取本视频 BGM 起始偏移（秒），无 BGM/禁用时为 0。"""
+        if not self.plan.config:
+            return 0.0
+        try:
+            count = len([c for c in (self.plan.clips or []) if getattr(c, "clip_type", "main") != "audio"])
+            plan = self._get_micro_transform_plan(count)
+            if plan:
+                return round(float(plan.bgm_start_offset or 0.0), 3)
+        except Exception:
+            logger.debug("微变换 BGM 偏移读取失败，按 0 处理: plan_id=%s", getattr(self.plan, "id", "?"))
+        return 0.0
 
     def render(self) -> RenderResult:
         """执行渲染，返回 RenderResult.
@@ -316,6 +397,9 @@ class UnifiedRenderService:
                     ctx = RenderContext(work_dir=self.work_dir, plan_id=self.plan.id)
                     from video_processing.bgm_mixer import BGMConfig, mix_bgm_with_main
 
+                    _bgm_off = self._get_micro_bgm_offset()
+                    if _bgm_off and not (bgm_config or {}).get("audio_offset"):
+                        bgm_config = {**bgm_config, "audio_offset": _bgm_off}
                     bgm_cfg = BGMConfig.from_config_dict(self.bgm_path, bgm_config)
                     # 从直通输出中提取音频
                     main_audio_path = self.work_dir / f"pass_through_audio_{self.plan.id}.aac"
@@ -365,6 +449,7 @@ class UnifiedRenderService:
                 bgm_path=self.bgm_path,
                 bgm_config=bgm_config,
                 audio_tracks_config=audio_tracks_config,
+                bgm_start_offset=self._get_micro_bgm_offset(),
             )
             t_audio_end = time.time()
             audio_mix_ms = int((t_audio_end - t_audio_start) * 1000)
@@ -1112,6 +1197,28 @@ class UnifiedRenderService:
         if ass_path is not None:
             return False, "有字幕叠加"
 
+        # #1970 PR2：片段级微变换（变速/hflip/亮度/对比度/饱和度）需要重编码
+        try:
+            _video_sources = [c for c in (self.clips or []) if getattr(c, "clip_type", "main") != "audio"]
+            _ordinal = -1
+            for _i, _c in enumerate(_video_sources):
+                if getattr(_c, "id", None) == getattr(clip, "clip_id", None):
+                    _ordinal = _i
+                    break
+            _mt_plan = self._get_micro_transform_plan(len(_video_sources))
+            if _mt_plan and 0 <= _ordinal < len(_mt_plan.clips):
+                _mt = _mt_plan.clips[_ordinal]
+                if (
+                    abs(UnifiedRenderService._micro_speed_factor(_mt) - 1.0) >= 1e-6
+                    or (_mt.hflip and not _mt.has_text)
+                    or abs(_mt.brightness) > 1e-4
+                    or abs(_mt.contrast - 1.0) > 1e-4
+                    or abs(_mt.saturation - 1.0) > 1e-4
+                ):
+                    return False, "启用了片段级微变换"
+        except Exception:
+            logger.debug("stream copy 微变换门控检查异常，按可 copy 处理", exc_info=True)
+
         # 有调速 → 需要重编码 → 不能 copy
         speed = UnifiedRenderService._clip_speed(clip)
         if abs(speed - 1.0) >= 1e-6:
@@ -1318,11 +1425,16 @@ class UnifiedRenderService:
 
         # 视觉扰动（plan 级别，直通模式同样适用）
         vp = self._get_visual_perturbation()
+        # #1970 PR2：单片段直通；计划按源视频片段数构建，序号取 config._micro_index
+        _src_video_count = len([c for c in (self.clips or []) if getattr(c, "clip_type", "main") != "audio"])
+        mt_plan = self._get_micro_transform_plan(max(1, _src_video_count))
+        _mi = int(clip.config.get("_micro_index", 0)) if isinstance(clip.config, dict) else 0
+        mt = mt_plan.clips[_mi] if mt_plan and 0 <= _mi < len(mt_plan.clips) else None
 
-        # 调速 — 与 filter_complex 路径一致（叠加视觉扰动 speed_factor）
+        # 调速 — 与 filter_complex 路径一致（叠加视觉扰动 speed_factor 与 #1970 微变换速度）
         speed = UnifiedRenderService._clip_speed(clip)
         vp_speed = vp.get("speed_factor", 1.0) if vp else 1.0
-        effective_speed = speed * vp_speed
+        effective_speed = speed * vp_speed  # 微变换速度已烘焙进 playback_speed
         if abs(effective_speed - 1.0) >= 1e-6:
             filters.append(f"setpts=PTS/{effective_speed:.4f}")
 
@@ -1336,6 +1448,8 @@ class UnifiedRenderService:
         # 视觉扰动：hflip（在 scale 之前）
         if vp:
             self._apply_visual_perturbation_pre_scale(filters, vp)
+        # #1970 PR2：片段级 hflip（P1 保守：有文字/无法判定时不翻转）
+        UnifiedRenderService._apply_micro_hflip(filters, mt)
 
         # scale + pad（等比缩放+留黑边）
         if role in ("overlay", "corner_voice"):
@@ -1354,6 +1468,8 @@ class UnifiedRenderService:
         # 视觉扰动：zoom + brightness（在 scale+pad 之后、调色之前）
         if vp:
             self._apply_visual_perturbation_post_scale(filters, vp)
+        # #1970 PR2：片段级亮度/对比度/饱和度微调
+        UnifiedRenderService._apply_micro_transform_video(filters, mt)
 
         # 调色滤镜
         color_grade = ColorGradeConfig.from_dict(clip.config.get("color_grade"))
@@ -1450,7 +1566,8 @@ class UnifiedRenderService:
             # 音频调速（在降噪之后、音量之前，与 render_audio.py concat 路径保持一致）
             # SpeedEngine.build_audio_filter 内部已实现多级 atempo 串联，
             # 自动处理超出 [0.5, 2.0] 范围的速度（如 0.25x → atempo=0.5,atempo=0.5）。
-            speed = UnifiedRenderService._clip_speed(clip)
+            # #1970 PR2：叠加片段微变换速度因子，保持音画同步。
+            speed = UnifiedRenderService._clip_speed(clip)  # 微变换速度已烘焙进 playback_speed
             if abs(speed - 1.0) >= 1e-6:
                 try:
                     from video_processing.speed_engine import SpeedConfig, SpeedEngine
@@ -1522,11 +1639,20 @@ class UnifiedRenderService:
         支持多段裁剪：一个 clip 配置了 trim_segments 时会展开为多个 ResolvedClip。
         """
         resolved: list[ResolvedClip] = []
+        # #1970 PR2：预建片段级微变换计划，按源视频片段序号取速度因子，
+        # 烘焙进 playback_speed，保证视频 setpts 与音频 atempo 一致。
+        video_source_clips = [c for c in self.clips if getattr(c, "clip_type", "main") != "audio"]
+        mt_plan = self._get_micro_transform_plan(len(video_source_clips))
+        _video_ordinal = {id(c): i for i, c in enumerate(video_source_clips)}
+
         for clip in self.clips:
             asset_id = clip.asset_id
             if not asset_id:
                 logger.warning("片段无素材: clip_id=%s", clip.id)
                 continue
+            _mt_idx = _video_ordinal.get(id(clip), -1)
+            _mt = mt_plan.clips[_mt_idx] if mt_plan and 0 <= _mt_idx < len(mt_plan.clips) else None
+            _micro_speed = UnifiedRenderService._micro_speed_factor(_mt)
 
             local_path = self.asset_path_map.get(asset_id)
             if local_path is None or not local_path.exists():
@@ -1555,7 +1681,7 @@ class UnifiedRenderService:
                     seg_duration = seg.trim.duration
 
                     # 多段裁剪：如果段的时长超过素材实际时长，减速补偿
-                    seg_speed = configured_speed
+                    seg_speed = configured_speed * _micro_speed
                     if actual_duration > 0 and seg_duration > actual_duration + 0.05:
                         seg_speed = max(0.25, round(configured_speed * actual_duration / seg_duration, 4))
                         logger.info(
@@ -1578,7 +1704,7 @@ class UnifiedRenderService:
                         transition_effect=clip.transition_effect or "cut",
                         transition_duration=getattr(clip, "transition_duration", 0.0) or 0.0,
                         playback_speed=seg_speed,
-                        config={**clip_config, "_segment_id": seg.segment_id},
+                        config={**clip_config, "_segment_id": seg.segment_id, "_micro_index": _mt_idx},
                         actual_duration=actual_duration,
                         trim_config=seg.trim,
                     )
@@ -1633,12 +1759,13 @@ class UnifiedRenderService:
                     avail_in_asset,
                     freeze_seconds,
                 )
-            final_speed = configured_speed
+            final_speed = configured_speed * _micro_speed
 
             # freeze 标记写入 config，供视频 tpad / 音频 apad 读取
             resolved_config = dict(clip_config)
             if freeze_seconds > 0:
                 resolved_config["_freeze_seconds"] = freeze_seconds
+            resolved_config["_micro_index"] = _mt_idx
 
             rc = ResolvedClip(
                 clip_id=clip.id,
@@ -1754,9 +1881,15 @@ class UnifiedRenderService:
         preprocessed_labels: list[str] = []
         # 视觉扰动（plan 级别，所有 clip 共享同一套扰动参数）
         vp = self._get_visual_perturbation()
+        # #1970 PR2：片段级微变换（每片段独立参数，dedup_enabled=False 时为 None）
+        # 计划按源视频片段数构建，trim 多段展开时各段通过 config._micro_index 找参数
+        _src_video_count = len([c for c in (self.clips or []) if getattr(c, "clip_type", "main") != "audio"])
+        mt_plan = self._get_micro_transform_plan(_src_video_count)
         for i, clip in enumerate(all_clips):
             label = f"v{i}"
             role = _resolve_layer_role(clip.clip_type, clip.config)
+            _mi = int(clip.config.get("_micro_index", i)) if isinstance(clip.config, dict) else i
+            mt = mt_plan.clips[_mi] if mt_plan and 0 <= _mi < len(mt_plan.clips) else None
 
             filters: list[str] = []
 
@@ -1774,10 +1907,10 @@ class UnifiedRenderService:
                     filters.append(f"trim=duration={trim_dur:.3f}")
                 filters.append("setpts=PTS-STARTPTS")
 
-            # 调速 — 基于 setpts 改变播放速度（叠加视觉扰动 speed_factor）
+            # 调速 — 基于 setpts 改变播放速度（叠加视觉扰动 speed_factor 与 #1970 微变换速度）
             speed = UnifiedRenderService._clip_speed(clip)
             vp_speed = vp.get("speed_factor", 1.0) if vp else 1.0
-            effective_speed = speed * vp_speed
+            effective_speed = speed * vp_speed  # 微变换速度已烘焙进 playback_speed
             if abs(effective_speed - 1.0) >= 1e-6:
                 filters.append(f"setpts=PTS/{effective_speed:.4f}")
 
@@ -1791,6 +1924,8 @@ class UnifiedRenderService:
             # 视觉扰动：hflip（在 scale 之前，翻转原始画面）
             if vp:
                 self._apply_visual_perturbation_pre_scale(filters, vp)
+            # #1970 PR2：片段级 hflip（P1 保守：有文字/无法判定时不翻转）
+            UnifiedRenderService._apply_micro_hflip(filters, mt)
 
             # scale
             if role in ("overlay", "corner_voice"):
@@ -1809,6 +1944,8 @@ class UnifiedRenderService:
             # 视觉扰动：zoom + brightness（在 scale+pad 之后、调色之前）
             if vp:
                 self._apply_visual_perturbation_post_scale(filters, vp)
+            # #1970 PR2：片段级亮度/对比度/饱和度微调
+            UnifiedRenderService._apply_micro_transform_video(filters, mt)
 
             # 调色滤镜（每个 clip 独立的 color grade 配置）
             color_grade = ColorGradeConfig.from_dict(clip.config.get("color_grade"))