feat(#1754): 渲染期画面扰动——让批量视频本身更不同 (#1755)

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

apps/worker/video_processing/unified_render_service.py

@@ -1308,10 +1308,15 @@ class UnifiedRenderService:
         if freeze_seconds > 0:
             filters.append(f"tpad=stop_mode=clone:stop_duration={freeze_seconds:.3f}")
 
-        # 调速 — 与 filter_complex 路径一致
+        # 视觉扰动（plan 级别，直通模式同样适用）
+        vp = self._get_visual_perturbation()
+
+        # 调速 — 与 filter_complex 路径一致（叠加视觉扰动 speed_factor）
         speed = UnifiedRenderService._clip_speed(clip)
-        if abs(speed - 1.0) >= 1e-6:
-            filters.append(f"setpts=PTS/{speed:.4f}")
+        vp_speed = vp.get("speed_factor", 1.0) if vp else 1.0
+        effective_speed = speed * vp_speed
+        if abs(effective_speed - 1.0) >= 1e-6:
+            filters.append(f"setpts=PTS/{effective_speed:.4f}")
 
         # 倒放滤镜
         reverse_config = ReverseConfig.from_dict(clip.config.get("reverse"))
@@ -1320,6 +1325,10 @@ class UnifiedRenderService:
             if reverse_filter:
                 filters.append(reverse_filter)
 
+        # 视觉扰动：hflip（在 scale 之前）
+        if vp:
+            self._apply_visual_perturbation_pre_scale(filters, vp)
+
         # scale + pad（等比缩放+留黑边）
         if role in ("overlay", "corner_voice"):
             pip_w = int(self.output_width * _PIP_SCALE)
@@ -1334,6 +1343,10 @@ class UnifiedRenderService:
             filters.append(f"scale={self.output_width}:{self.output_height}:force_original_aspect_ratio=decrease")
             filters.append(f"pad={self.output_width}:{self.output_height}:trunc((ow-iw)/2):trunc((oh-ih)/2):black")
 
+        # 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)
+
         # 调色滤镜
         color_grade = ColorGradeConfig.from_dict(clip.config.get("color_grade"))
         if color_grade.enabled and color_grade.has_effect():
@@ -1729,6 +1742,8 @@ class UnifiedRenderService:
         # Step 1: 预处理每个 clip — trim + scale + setpts
         # 为每个 clip 生成预处理后的标签 [v0], [v1], ...
         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()
         for i, clip in enumerate(all_clips):
             label = f"v{i}"
             role = _resolve_layer_role(clip.clip_type, clip.config)
@@ -1749,10 +1764,12 @@ 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
+            # 调速 — 基于 setpts 改变播放速度（叠加视觉扰动 speed_factor）
             speed = UnifiedRenderService._clip_speed(clip)
-            if abs(speed - 1.0) >= 1e-6:
-                filters.append(f"setpts=PTS/{speed:.4f}")
+            vp_speed = vp.get("speed_factor", 1.0) if vp else 1.0
+            effective_speed = speed * vp_speed
+            if abs(effective_speed - 1.0) >= 1e-6:
+                filters.append(f"setpts=PTS/{effective_speed:.4f}")
 
             # 倒放滤镜（在 trim 之后、scale 之前应用）
             reverse_config = ReverseConfig.from_dict(clip.config.get("reverse"))
@@ -1761,6 +1778,10 @@ class UnifiedRenderService:
                 if reverse_filter:
                     filters.append(reverse_filter)
 
+            # 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)
+
             # scale
             if role in ("overlay", "corner_voice"):
                 pip_w = int(self.output_width * _PIP_SCALE)
@@ -1775,6 +1796,10 @@ class UnifiedRenderService:
                 filters.append(f"scale={self.output_width}:{self.output_height}:force_original_aspect_ratio=decrease")
                 filters.append(f"pad={self.output_width}:{self.output_height}:trunc((ow-iw)/2):trunc((oh-ih)/2):black")
 
+            # 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)
+
             # 调色滤镜（每个 clip 独立的 color grade 配置）
             color_grade = ColorGradeConfig.from_dict(clip.config.get("color_grade"))
             if color_grade.enabled and color_grade.has_effect():
@@ -2190,6 +2215,35 @@ class UnifiedRenderService:
         """
         return _clip_playback_speed_pure(getattr(clip, "playback_speed", 1.0))
 
+    def _get_visual_perturbation(self) -> dict:
+        # 读取当前 plan 的视觉扰动参数（plan.config.visual_perturbation）
+        perturbation = (self.plan.config or {}).get("visual_perturbation") or {}
+        if not perturbation:
+            return {}
+        return {
+            "hflip": bool(perturbation.get("hflip", False)),
+            "zoom_ratio": max(1.0, min(1.2, float(perturbation.get("zoom_ratio", 1.0) or 1.0))),
+            "speed_factor": max(0.8, min(1.2, float(perturbation.get("speed_factor", 1.0) or 1.0))),
+            "brightness_shift": max(-30, min(30, int(perturbation.get("brightness_shift", 0) or 0))),
+        }
+
+    def _apply_visual_perturbation_pre_scale(self, filters: list[str], perturbation: dict) -> None:
+        # scale+pad 之前的扰动（hflip），就地修改 filters
+        if perturbation.get("hflip"):
+            filters.append("hflip")
+
+    def _apply_visual_perturbation_post_scale(self, filters: list[str], perturbation: dict) -> None:
+        # scale+pad 之后的扰动（zoom/brightness），就地修改 filters
+        zoom = perturbation.get("zoom_ratio", 1.0)
+        if zoom > 1.0 + 1e-6:
+            zoomed_w = int(self.output_width * zoom)
+            zoomed_h = int(self.output_height * zoom)
+            filters.append(f"scale={zoomed_w}:{zoomed_h}")
+            filters.append(f"crop={self.output_width}:{self.output_height}")
+        brightness = perturbation.get("brightness_shift", 0)
+        if brightness != 0:
+            filters.append(f"eq=brightness={brightness / 100.0:.3f}")
+
     @staticmethod
     def _clip_volume(clip: ResolvedClip) -> float:
         """获取 clip 的音量（config.volume）。缺省 1.0 原声，0.0 静音。"""