feat: 中間データのmsgpack移行

core/async_bake_generator.py

@@ -29,7 +29,7 @@ class AsyncBakeGenerator:
     def start(
         self,
         video_path: str,
-        mask_path: str,
+        detections_path: str,
         output_path: str,
         blur_size: int,
         fmt: str,
@@ -53,7 +53,7 @@ class AsyncBakeGenerator:
 
         self.worker_thread = threading.Thread(
             target=self._worker,
-            args=(video_path, mask_path, output_path, blur_size, fmt),
+            args=(video_path, detections_path, output_path, blur_size, fmt),
             daemon=True,
         )
         self.worker_thread.start()
@@ -72,7 +72,7 @@ class AsyncBakeGenerator:
     def _worker(
         self,
         video_path: str,
-        mask_path: str,
+        detections_path: str,
         output_path: str,
         blur_size: int,
         fmt: str,
@@ -85,7 +85,7 @@ class AsyncBakeGenerator:
             client = get_client()
             task_id = client.bake_blur(
                 video_path=video_path,
-                mask_path=mask_path,
+                detections_path=detections_path,
                 output_path=output_path,
                 blur_size=blur_size,
                 fmt=fmt,

core/async_generator.py

@@ -9,8 +9,7 @@ Blender's UI remains responsive via bpy.app.timers.
 import os
 import threading
 import queue
-from functools import partial
+from typing import Optional, Callable
-from typing import Optional, Callable, Tuple
 from pathlib import Path
 
 # Will be imported when running inside Blender
@@ -151,8 +150,19 @@ class AsyncMaskGenerator:
                 status = client.get_task_status(task_id)
                 state = status.get("status")
 
+                total = status.get("total", 0)
+                if total > 0:
+                    self.total_frames = total
+                
                 if state == "completed":
-                    self.result_queue.put(("done", output_dir))
+                    final_progress = status.get("progress", self.total_frames)
+                    if final_progress >= 0:
+                        self.progress_queue.put(("progress", final_progress))
+                    result_path = status.get(
+                        "result_path",
+                        os.path.join(output_dir, "detections.msgpack"),
+                    )
+                    self.result_queue.put(("done", result_path))
                     return
                 
                 elif state == "failed":
@@ -167,7 +177,7 @@ class AsyncMaskGenerator:
                 
                 # Report progress
                 progress = status.get("progress", 0)
-                if progress > 0:
+                if progress >= 0:
                     self.progress_queue.put(("progress", progress))
                 
                 time.sleep(0.5)
@@ -207,6 +217,16 @@ class AsyncMaskGenerator:
                 msg_type, data = self.result_queue.get_nowait()
                 self.is_running = False
 
+                # Ensure UI receives a final progress update before completion.
+                if (
+                    msg_type == "done"
+                    and self.total_frames > 0
+                    and self.current_frame < self.total_frames
+                    and self._on_progress
+                ):
+                    self.current_frame = self.total_frames
+                    self._on_progress(self.current_frame, self.total_frames)
+                
                 if self._on_complete:
                     self._on_complete(msg_type, data)
                 

core/inference_client.py

@@ -252,7 +252,7 @@ class InferenceClient:
     def bake_blur(
         self,
         video_path: str,
-        mask_path: str,
+        detections_path: str,
         output_path: str,
         blur_size: int,
         fmt: str,
@@ -268,7 +268,7 @@ class InferenceClient:
 
         data = {
             "video_path": video_path,
-            "mask_path": mask_path,
+            "detections_path": detections_path,
             "output_path": output_path,
             "blur_size": blur_size,
             "format": fmt,

core/utils.py

@@ -78,6 +78,11 @@ def get_cache_dir_for_strip(strip_name: str) -> str:
     return os.path.join(get_cache_root(), strip_name)
 
 
+def get_detections_path_for_strip(strip_name: str) -> str:
+    """Get msgpack detection cache path for a specific strip."""
+    return os.path.join(get_cache_dir_for_strip(strip_name), "detections.msgpack")
+
+
 def get_cache_info(strip_name: Optional[str] = None) -> Tuple[str, int, int]:
     """
     Get cache directory information.

flake.nix

@@ -61,26 +61,22 @@
             # venvをアクティベート
             source "$VENV_DIR/bin/activate"
 
-            # 必要なパッケージのインストール確認とインストール
+            # PyTorch ROCm版の導入（GPU未認識時のみ）
             if ! python -c "import torch; print(torch.cuda.is_available())" 2>/dev/null | grep -q "True"; then
-              echo "[Setup] Installing Python dependencies..."
+              echo "[Setup] Installing PyTorch ROCm dependencies..."
-              # まずPyTorch ROCm版をインストール（ROCm 7.0 nightly - ROCm 7.1.1環境で動作確認済み）
               pip install --quiet --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.0
-              # 次に通常のPyPIから他のパッケージをインストール
-              pip install --quiet \
-                ultralytics \
-                opencv-python-headless \
-                numpy \
-                fastapi \
-                uvicorn \
-                pydantic
-              # opencv-pythonがインストールされていたら削除（headless版のみ使用）
-              pip uninstall -y opencv-python opencv 2>/dev/null || true
-              # opencv-python-headlessを再インストールして確実にする
-              pip install --quiet --force-reinstall opencv-python-headless
-              echo "[Setup] Dependencies installed successfully"
             fi
 
+            # プロジェクト依存（requirements.txt）を同期
+            if [ -f "$PWD/requirements.txt" ]; then
+              echo "[Setup] Syncing Python dependencies from requirements.txt..."
+              pip install --quiet -r "$PWD/requirements.txt"
+            fi
+
+            # opencv-pythonが入っていた場合はheadlessに統一
+            pip uninstall -y opencv-python opencv 2>/dev/null || true
+            pip install --quiet --upgrade opencv-python-headless
+
             # Pythonパスにカレントディレクトリを追加
             export PYTHONPATH="$PWD:$PYTHONPATH"
 

operators/apply_blur.py

@@ -12,6 +12,7 @@ from bpy.types import Operator
 
 from ..core.async_bake_generator import get_bake_generator
 from ..core.async_generator import get_generator as get_mask_generator
+from ..core.utils import get_detections_path_for_strip
 
 
 KEY_ORIGINAL = "facemask_original_filepath"
@@ -28,19 +29,9 @@ FORMAT_EXT = {
 }
 
 
-def _find_mask_strip(seq_editor, strip_name: str):
+def _output_path(video_strip, detections_path: str, fmt: str) -> str:
-    return seq_editor.strips.get(f"{strip_name}_mask")
-
-
-def _resolve_mask_path(mask_strip) -> str:
-    if mask_strip.type == "MOVIE":
-        return bpy.path.abspath(mask_strip.filepath)
-    return ""
-
-
-def _output_path(video_strip, mask_path: str, fmt: str) -> str:
     ext = FORMAT_EXT.get(fmt, "mp4")
-    out_dir = os.path.dirname(mask_path)
+    out_dir = os.path.dirname(detections_path)
     safe_name = video_strip.name.replace("/", "_").replace("\\", "_")
     return os.path.join(out_dir, f"{safe_name}_blurred.{ext}")
 
@@ -83,22 +74,17 @@ class SEQUENCER_OT_bake_and_swap_blur_source(Operator):
         scene = context.scene
         video_strip = seq_editor.active_strip
 
-        mask_strip = _find_mask_strip(seq_editor, video_strip.name)
-        if not mask_strip:
-            self.report({"ERROR"}, f"Mask strip not found: {video_strip.name}_mask")
-            return {"CANCELLED"}
-
         video_path = bpy.path.abspath(video_strip.filepath)
-        mask_path = _resolve_mask_path(mask_strip)
+        detections_path = get_detections_path_for_strip(video_strip.name)
         if not os.path.exists(video_path):
             self.report({"ERROR"}, f"Source video not found: {video_path}")
             return {"CANCELLED"}
-        if not mask_path or not os.path.exists(mask_path):
+        if not os.path.exists(detections_path):
-            self.report({"ERROR"}, f"Mask video not found: {mask_path}")
+            self.report({"ERROR"}, f"Detection cache not found: {detections_path}")
             return {"CANCELLED"}
 
         bake_format = scene.facemask_bake_format
-        output_path = _output_path(video_strip, mask_path, bake_format)
+        output_path = _output_path(video_strip, detections_path, bake_format)
         blur_size = int(scene.facemask_bake_blur_size)
 
         # Reuse baked cache when parameters match and file still exists.
@@ -164,7 +150,7 @@ class SEQUENCER_OT_bake_and_swap_blur_source(Operator):
         try:
             bake_generator.start(
                 video_path=video_path,
-                mask_path=mask_path,
+                detections_path=detections_path,
                 output_path=output_path,
                 blur_size=blur_size,
                 fmt=bake_format.lower(),

operators/generate_mask.py

@@ -55,8 +55,7 @@ class SEQUENCER_OT_generate_face_mask(Operator):
         # Check cache - if masks already exist, use them
         expected_frame_count = strip.frame_final_end - strip.frame_final_start + 1
         if self._check_cache(output_dir, expected_frame_count):
-            self.report({'INFO'}, f"Using cached masks from {output_dir}")
+            self.report({'INFO'}, f"Using cached detections from {output_dir}")
-            self._add_mask_strip(context, strip.name, output_dir)
             return {'FINISHED'}
         
         # Get frame range
@@ -71,20 +70,26 @@ class SEQUENCER_OT_generate_face_mask(Operator):
             self.report({'WARNING'}, "Mask generation already in progress")
             return {'CANCELLED'}
         
-        # Store strip name for callback
-        strip_name = strip.name
-        
         def on_complete(status, data):
             """Called when mask generation completes."""
+            wm = context.window_manager
+            wm.mask_total = max(wm.mask_total, generator.total_frames)
+            if status == "done":
+                wm.mask_progress = wm.mask_total
+            elif status in {"error", "cancelled"}:
+                wm.mask_progress = min(wm.mask_progress, wm.mask_total)
+
             if status == "done":
-                # Add mask strip to sequence editor
-                self._add_mask_strip(context, strip_name, data)
                 print(f"[FaceMask] Mask generation completed: {data}")
             elif status == "error":
                 print(f"[FaceMask] Error: {data}")
             elif status == "cancelled":
                 print("[FaceMask] Generation cancelled")
 
+            for area in context.screen.areas:
+                if area.type == 'SEQUENCE_EDITOR':
+                    area.tag_redraw()
+        
         def on_progress(current, total):
             """Called on progress updates."""
             # Update window manager properties for UI
@@ -143,120 +148,28 @@ class SEQUENCER_OT_generate_face_mask(Operator):
         if not os.path.exists(cache_dir):
             return False
 
-        # Check for MP4 video (new format)
+        detections_path = os.path.join(cache_dir, "detections.msgpack")
-        mask_video = os.path.join(cache_dir, "mask.mp4")
+        if not os.path.exists(detections_path):
-        if os.path.exists(mask_video):
-            # Prefer frame-count verification when cv2 is available, but do not
-            # hard-fail on Blender Python environments without cv2.
-            try:
-                import cv2
-
-                cap = cv2.VideoCapture(mask_video)
-                if cap.isOpened():
-                    frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
-                    cap.release()
-                    # Accept cache if at least 90% of frames exist
-                    return frame_count >= expected_frames * 0.9
-                cap.release()
             return False
+
+        # Quick sanity check: non-empty file
+        try:
+            if os.path.getsize(detections_path) <= 0:
+                return False
+        except OSError:
+            return False
+
+        # Optional frame count verification if msgpack is available
+        try:
+            import msgpack
+
+            with open(detections_path, "rb") as f:
+                payload = msgpack.unpackb(f.read(), raw=False)
+            frames = payload.get("frames", [])
+            return len(frames) >= expected_frames * 0.9
         except Exception:
-                # Fallback: treat existing MP4 cache as valid when cv2 is unavailable.
             return True
 
-        # Fallback: check for PNG sequence (backward compatibility)
-        mask_files = [f for f in os.listdir(cache_dir)
-                      if f.startswith("mask_") and f.endswith(".png")]
-
-        # Accept cache if at least 90% of frames exist
-        return len(mask_files) >= expected_frames * 0.9
-    
-    def _add_mask_strip(self, context, source_strip_name: str, mask_path: str):
-        """Add mask video as a new strip.
-
-        Args:
-            context: Blender context
-            source_strip_name: Name of the source video strip
-            mask_path: Path to mask video file or directory (for backward compatibility)
-        """
-        scene = context.scene
-        seq_editor = scene.sequence_editor
-
-        if not seq_editor:
-            return
-
-        # Find source strip (Blender 5.0 uses 'strips' instead of 'sequences')
-        source_strip = seq_editor.strips.get(source_strip_name)
-        if not source_strip:
-            return
-
-        # Check if mask_path is a video file or directory (backward compatibility)
-        if os.path.isfile(mask_path):
-            # New format: single MP4 file
-            mask_video = mask_path
-        else:
-            # Old format: directory with PNG sequence (backward compatibility)
-            mask_video = os.path.join(mask_path, "mask.mp4")
-            if not os.path.exists(mask_video):
-                # Fallback to PNG sequence
-                mask_files = sorted([
-                    f for f in os.listdir(mask_path)
-                    if f.startswith("mask_") and f.endswith(".png")
-                ])
-                if not mask_files:
-                    return
-                first_mask = os.path.join(mask_path, mask_files[0])
-                self._add_mask_strip_png_sequence(context, source_strip_name, mask_path, mask_files, first_mask)
-                return
-
-        # Find an empty channel
-        used_channels = {s.channel for s in seq_editor.strips}
-        new_channel = source_strip.channel + 1
-        while new_channel in used_channels:
-            new_channel += 1
-
-        # Add movie strip (Blender 5.0 API)
-        mask_strip = seq_editor.strips.new_movie(
-            name=f"{source_strip_name}_mask",
-            filepath=mask_video,
-            channel=new_channel,
-            frame_start=source_strip.frame_final_start,
-        )
-
-        # Set blend mode for mask
-        mask_strip.blend_type = 'ALPHA_OVER'
-        mask_strip.blend_alpha = 0.5
-
-    def _add_mask_strip_png_sequence(self, context, source_strip_name, mask_dir, mask_files, first_mask):
-        """Backward compatibility: Add PNG sequence as mask strip."""
-        scene = context.scene
-        seq_editor = scene.sequence_editor
-        source_strip = seq_editor.strips.get(source_strip_name)
-
-        if not source_strip:
-            return
-
-        # Find an empty channel
-        used_channels = {s.channel for s in seq_editor.strips}
-        new_channel = source_strip.channel + 1
-        while new_channel in used_channels:
-            new_channel += 1
-
-        # Add image sequence (Blender 5.0 API)
-        mask_strip = seq_editor.strips.new_image(
-            name=f"{source_strip_name}_mask",
-            filepath=first_mask,
-            channel=new_channel,
-            frame_start=source_strip.frame_final_start,
-        )
-
-        # Add remaining frames
-        for mask_file in mask_files[1:]:
-            mask_strip.elements.append(mask_file)
-
-        # Set blend mode for mask
-        mask_strip.blend_type = 'ALPHA_OVER'
-        mask_strip.blend_alpha = 0.5
-
 
 class SEQUENCER_OT_cancel_mask_generation(Operator):
     """Cancel ongoing mask generation."""

panels/vse_panel.py

@@ -5,12 +5,18 @@ Provides a sidebar panel in the Video Sequence Editor
 for controlling mask generation and blur application.
 """
 
+import os
 import bpy
 from bpy.types import Panel
 
 from ..core.async_bake_generator import get_bake_generator
 from ..core.async_generator import get_generator
-from ..core.utils import get_server_status, get_cache_info, format_size
+from ..core.utils import (
+    get_server_status,
+    get_cache_info,
+    format_size,
+    get_detections_path_for_strip,
+)
 
 
 class SEQUENCER_PT_face_mask(Panel):
@@ -186,19 +192,19 @@ class SEQUENCER_PT_face_mask(Panel):
         row = box.row()
         row.label(text=f"Strip: {strip.name}")
         
-        # Check for existing mask
+        detections_path = get_detections_path_for_strip(strip.name)
-        seq_editor = context.scene.sequence_editor
+        has_mask = bpy.path.abspath(detections_path) and os.path.exists(
-        mask_name = f"{strip.name}_mask"
+            bpy.path.abspath(detections_path)
-        has_mask = mask_name in seq_editor.strips
+        )
         
         if has_mask:
             row = box.row()
-            row.label(text="✓ Mask exists", icon='CHECKMARK')
+            row.label(text="✓ Detection cache exists", icon='CHECKMARK')
         
         # Generate button
         op = box.operator(
             "sequencer.generate_face_mask",
-            text="Generate Face Mask" if not has_mask else "Regenerate Mask",
+            text="Generate Detection Cache" if not has_mask else "Regenerate Cache",
             icon='FACE_MAPS',
         )
     
@@ -207,13 +213,13 @@ class SEQUENCER_PT_face_mask(Panel):
         box = layout.box()
         box.label(text="Blur Bake", icon='MATFLUID')
 
-        # Check for mask strip
+        detections_path = get_detections_path_for_strip(strip.name)
-        seq_editor = context.scene.sequence_editor
+        has_mask = bpy.path.abspath(detections_path) and os.path.exists(
-        mask_name = f"{strip.name}_mask"
+            bpy.path.abspath(detections_path)
-        has_mask = mask_name in seq_editor.strips
+        )
         
         if not has_mask:
-            box.label(text="Generate a mask first", icon='INFO')
+            box.label(text="Generate detection cache first", icon='INFO')
             return
 
         # Bake parameters

requirements.txt

@@ -0,0 +1,7 @@
+ultralytics
+opencv-python-headless
+msgpack
+numpy
+fastapi
+uvicorn
+pydantic

server/main.py

@@ -33,7 +33,6 @@ fix_library_path()
 
 import threading
 import uuid
-import queue
 import traceback
 from typing import Dict, Optional, List
 from pathlib import Path
@@ -43,6 +42,7 @@ from pydantic import BaseModel
 import uvicorn
 import cv2
 import numpy as np
+import msgpack
 
 # Add project root to path for imports if needed
 sys.path.append(str(Path(__file__).parent.parent))
@@ -86,7 +86,7 @@ class GenerateRequest(BaseModel):
 
 class BakeRequest(BaseModel):
     video_path: str
-    mask_path: str
+    detections_path: str
     output_path: str
     blur_size: int = 50
     format: str = "mp4"
@@ -122,185 +122,79 @@ def _build_video_writer(
 
     raise RuntimeError(f"Failed to create video writer for format='{fmt}'")
 
+def _scale_bbox(
+    x: int,
+    y: int,
+    w: int,
+    h: int,
+    scale: float,
+    frame_width: int,
+    frame_height: int,
+) -> Optional[List[int]]:
+    """Scale bbox around center and clamp to frame boundaries."""
+    if w <= 0 or h <= 0:
+        return None
+
+    center_x = x + (w * 0.5)
+    center_y = y + (h * 0.5)
+    scaled_w = max(1, int(w * scale))
+    scaled_h = max(1, int(h * scale))
+
+    x1 = max(0, int(center_x - scaled_w * 0.5))
+    y1 = max(0, int(center_y - scaled_h * 0.5))
+    x2 = min(frame_width, x1 + scaled_w)
+    y2 = min(frame_height, y1 + scaled_h)
+    out_w = x2 - x1
+    out_h = y2 - y1
+    if out_w <= 0 or out_h <= 0:
+        return None
+    return [x1, y1, out_w, out_h]
+
+
 def process_video_task(task_id: str, req: GenerateRequest):
-    """Background task to process video with async MP4 output."""
+    """Background task to detect faces and save bbox cache as msgpack."""
-    writer = None
+    cap = None
-    write_queue = None
-    writer_thread = None
 
     try:
         tasks[task_id].status = TaskStatus.PROCESSING
         cancel_event = cancel_events.get(task_id)
 
-        # Verify video exists
         if not os.path.exists(req.video_path):
             tasks[task_id].status = TaskStatus.FAILED
             tasks[task_id].message = f"Video not found: {req.video_path}"
             return
 
-        # Initialize detector (will load model on first run)
         print(f"Loading detector for task {task_id}...")
         detector = get_detector(
             conf_threshold=req.conf_threshold,
-            iou_threshold=req.iou_threshold
+            iou_threshold=req.iou_threshold,
         )
         _ = detector.model
 
-        # Open video
         cap = cv2.VideoCapture(req.video_path)
         if not cap.isOpened():
             tasks[task_id].status = TaskStatus.FAILED
             tasks[task_id].message = "Failed to open video"
             return
 
-        # Get video properties
+        fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
-        fps = cap.get(cv2.CAP_PROP_FPS)
         width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
         height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
         total_video_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
         end_frame = min(req.end_frame, total_video_frames - 1)
         frames_to_process = end_frame - req.start_frame + 1
+        if frames_to_process <= 0:
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = "Invalid frame range"
+            return
 
         tasks[task_id].total = frames_to_process
-
-        # Ensure output directory exists
         os.makedirs(req.output_dir, exist_ok=True)
+        output_msgpack_path = os.path.join(req.output_dir, "detections.msgpack")
 
-        # Setup MP4 writer (grayscale)
-        output_video_path = os.path.join(req.output_dir, "mask.mp4")
-        fourcc = cv2.VideoWriter_fourcc(*'mp4v')
-        writer = cv2.VideoWriter(output_video_path, fourcc, fps, (width, height), isColor=False)
-
-        if not writer.isOpened():
-            tasks[task_id].status = TaskStatus.FAILED
-            tasks[task_id].message = "Failed to create video writer"
-            cap.release()
-            return
-
-        # Async writer setup
-        write_queue = queue.Queue(maxsize=30)  # Buffer up to 30 frames
-        writer_running = threading.Event()
-        writer_running.set()
-
-        def async_writer():
-            """Background thread for writing frames to video."""
-            while writer_running.is_set() or not write_queue.empty():
-                try:
-                    mask = write_queue.get(timeout=0.1)
-                    if mask is not None:
-                        writer.write(mask)
-                    write_queue.task_done()
-                except queue.Empty:
-                    continue
-
-        writer_thread = threading.Thread(target=async_writer, daemon=True)
-        writer_thread.start()
-
-        print(f"Starting processing: {req.video_path} ({frames_to_process} frames) -> {output_video_path}")
-
-        # Batch processing configuration
-        BATCH_SIZE = 5  # Optimal batch size for 4K video (72.9% improvement)
-        frame_buffer = []
-        TEMPORAL_SIDE_WEIGHT = 0.7
-        TEMPORAL_CENTER_WEIGHT = 1.0
-
-        # Temporal blending state (streaming, low-memory)
-        prev_mask = None
-        curr_mask = None
-        wrote_first_frame = False
-
-        def _scale_mask(mask: np.ndarray, weight: float) -> np.ndarray:
-            """Scale mask intensity for temporal blending."""
-            if weight == 1.0:
-                return mask
-            return cv2.convertScaleAbs(mask, alpha=weight, beta=0)
-
-        def _blend_edge(base: np.ndarray, neighbor: np.ndarray) -> np.ndarray:
-            """Blend for first/last frame (one-sided temporal context)."""
-            base_w = _scale_mask(base, TEMPORAL_CENTER_WEIGHT)
-            neighbor_w = _scale_mask(neighbor, TEMPORAL_SIDE_WEIGHT)
-            return cv2.max(base_w, neighbor_w)
-
-        def _blend_middle(prev: np.ndarray, cur: np.ndarray, nxt: np.ndarray) -> np.ndarray:
-            """Blend for middle frames (previous/current/next temporal context)."""
-            prev_w = _scale_mask(prev, TEMPORAL_SIDE_WEIGHT)
-            cur_w = _scale_mask(cur, TEMPORAL_CENTER_WEIGHT)
-            nxt_w = _scale_mask(nxt, TEMPORAL_SIDE_WEIGHT)
-            return cv2.max(cur_w, cv2.max(prev_w, nxt_w))
-
-        def push_mask_temporal(raw_mask: np.ndarray):
-            """Push mask and emit blended output in frame order."""
-            nonlocal prev_mask, curr_mask, wrote_first_frame
-
-            if prev_mask is None:
-                prev_mask = raw_mask
-                return
-
-            if curr_mask is None:
-                curr_mask = raw_mask
-                return
-
-            if not wrote_first_frame:
-                write_queue.put(_blend_edge(prev_mask, curr_mask))
-                wrote_first_frame = True
-
-            # Emit blended current frame using prev/current/next
-            write_queue.put(_blend_middle(prev_mask, curr_mask, raw_mask))
-
-            # Slide temporal window
-            prev_mask = curr_mask
-            curr_mask = raw_mask
-
-        def flush_temporal_tail():
-            """Flush remaining masks after all frames are processed."""
-            if prev_mask is None:
-                return
-
-            # Single-frame case
-            if curr_mask is None:
-                write_queue.put(_scale_mask(prev_mask, TEMPORAL_CENTER_WEIGHT))
-                return
-
-            # Two-frame case
-            if not wrote_first_frame:
-                write_queue.put(_blend_edge(prev_mask, curr_mask))
-
-            # Always emit last frame with one-sided blend
-            write_queue.put(_blend_edge(curr_mask, prev_mask))
-
-        def process_batch():
-            """Process accumulated batch of frames."""
-            if not frame_buffer:
-                return
-
-            # Batch inference at full resolution
-            batch_detections = detector.detect_batch(frame_buffer)
-
-            # Generate masks for each frame
-            for i, detections in enumerate(batch_detections):
-                frame = frame_buffer[i]
-
-                # Generate mask at original resolution
-                mask = detector.generate_mask(
-                    frame.shape,
-                    detections,
-                    mask_scale=req.mask_scale
-                )
-
-                # Temporal blend before async write
-                push_mask_temporal(mask)
-
-            # Clear buffer
-            frame_buffer.clear()
-
-        # Seek once to the starting frame. Avoid random-access seek on every frame.
         if req.start_frame > 0:
             seek_ok = cap.set(cv2.CAP_PROP_POS_FRAMES, req.start_frame)
             if not seek_ok:
-                print(
-                    f"[FaceMask] Warning: CAP_PROP_POS_FRAMES seek failed, "
-                    f"fallback to sequential skip ({req.start_frame} frames)"
-                )
                 for _ in range(req.start_frame):
                     ret, _ = cap.read()
                     if not ret:
@@ -310,49 +204,81 @@ def process_video_task(task_id: str, req: GenerateRequest):
                         )
                         return
 
-        # Process loop with batching
+        frame_buffer: List[np.ndarray] = []
+        frame_detections: List[List[List[float]]] = []
+        batch_size = 5
         current_count = 0
-        for frame_idx in range(req.start_frame, end_frame + 1):
+
+        def process_batch():
+            nonlocal current_count
+            if not frame_buffer:
+                return
+
+            batch_detections = detector.detect_batch(frame_buffer)
+            for detections in batch_detections:
+                packed_detections: List[List[float]] = []
+                for x, y, w, h, conf in detections:
+                    scaled = _scale_bbox(
+                        int(x),
+                        int(y),
+                        int(w),
+                        int(h),
+                        float(req.mask_scale),
+                        width,
+                        height,
+                    )
+                    if scaled is None:
+                        continue
+                    packed_detections.append(
+                        [scaled[0], scaled[1], scaled[2], scaled[3], float(conf)]
+                    )
+                frame_detections.append(packed_detections)
+                current_count += 1
+                tasks[task_id].progress = current_count
+
+            frame_buffer.clear()
+
+        print(
+            f"Starting detection cache generation: {req.video_path} "
+            f"({frames_to_process} frames) -> {output_msgpack_path}"
+        )
+
+        for _ in range(req.start_frame, end_frame + 1):
             if cancel_event and cancel_event.is_set():
                 tasks[task_id].status = TaskStatus.CANCELLED
                 tasks[task_id].message = "Cancelled by user"
                 break
 
-            # Read next frame sequentially (after one-time initial seek)
             ret, frame = cap.read()
+            if not ret:
+                break
 
-            if ret:
-                # Store frame for batch processing
             frame_buffer.append(frame)
-
+            if len(frame_buffer) >= batch_size:
-                # Process batch when full
-                if len(frame_buffer) >= BATCH_SIZE:
                 process_batch()
 
-            # Update progress
-            current_count += 1
-            tasks[task_id].progress = current_count
-
-        # Process remaining frames in buffer
         if frame_buffer:
             process_batch()
-        flush_temporal_tail()
-
-        # Cleanup
-        writer_running.clear()
-        write_queue.join()  # Wait for all frames to be written
-        if writer_thread:
-            writer_thread.join(timeout=5)
-
-        cap.release()
-        if writer:
-            writer.release()
 
         if tasks[task_id].status == TaskStatus.PROCESSING:
+            payload = {
+                "version": 1,
+                "video_path": req.video_path,
+                "start_frame": req.start_frame,
+                "end_frame": req.start_frame + len(frame_detections) - 1,
+                "width": width,
+                "height": height,
+                "fps": fps,
+                "mask_scale": float(req.mask_scale),
+                "frames": frame_detections,
+            }
+            with open(output_msgpack_path, "wb") as f:
+                f.write(msgpack.packb(payload, use_bin_type=True))
+
             tasks[task_id].status = TaskStatus.COMPLETED
-            tasks[task_id].result_path = output_video_path  # Return video path
+            tasks[task_id].result_path = output_msgpack_path
-            tasks[task_id].message = "Processing completed successfully"
+            tasks[task_id].message = "Detection cache completed"
-            print(f"Task {task_id} completed: {output_video_path}")
+            print(f"Task {task_id} completed: {output_msgpack_path}")
 
     except Exception as e:
         tasks[task_id].status = TaskStatus.FAILED
@@ -360,15 +286,15 @@ def process_video_task(task_id: str, req: GenerateRequest):
         print(f"Error in task {task_id}: {e}")
         traceback.print_exc()
     finally:
-        # Cleanup
+        if cap:
+            cap.release()
         if task_id in cancel_events:
             del cancel_events[task_id]
 
 
 def process_bake_task(task_id: str, req: BakeRequest):
-    """Background task to bake blur into a regular video file."""
+    """Background task to bake blur using bbox detections in msgpack."""
     src_cap = None
-    mask_cap = None
     writer = None
 
     try:
@@ -379,58 +305,55 @@ def process_bake_task(task_id: str, req: BakeRequest):
             tasks[task_id].status = TaskStatus.FAILED
             tasks[task_id].message = f"Video not found: {req.video_path}"
             return
-
+        if not os.path.exists(req.detections_path):
-        if not os.path.exists(req.mask_path):
             tasks[task_id].status = TaskStatus.FAILED
-            tasks[task_id].message = f"Mask video not found: {req.mask_path}"
+            tasks[task_id].message = f"Detections file not found: {req.detections_path}"
             return
 
         src_cap = cv2.VideoCapture(req.video_path)
-        mask_cap = cv2.VideoCapture(req.mask_path)
-
         if not src_cap.isOpened():
             tasks[task_id].status = TaskStatus.FAILED
             tasks[task_id].message = "Failed to open source video"
             return
-        if not mask_cap.isOpened():
+
+        with open(req.detections_path, "rb") as f:
+            payload = msgpack.unpackb(f.read(), raw=False)
+        frames = payload.get("frames")
+        if not isinstance(frames, list):
             tasks[task_id].status = TaskStatus.FAILED
-            tasks[task_id].message = "Failed to open mask video"
+            tasks[task_id].message = "Invalid detections format: 'frames' is missing"
             return
 
         src_fps = src_cap.get(cv2.CAP_PROP_FPS) or 30.0
         src_width = int(src_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
         src_height = int(src_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
         src_frames = int(src_cap.get(cv2.CAP_PROP_FRAME_COUNT))
-        mask_frames = int(mask_cap.get(cv2.CAP_PROP_FRAME_COUNT))
-
         if src_width <= 0 or src_height <= 0:
             tasks[task_id].status = TaskStatus.FAILED
             tasks[task_id].message = "Invalid source video dimensions"
             return
 
-        total = min(src_frames, mask_frames) if src_frames > 0 and mask_frames > 0 else 0
+        total = min(src_frames, len(frames)) if src_frames > 0 else len(frames)
         if total <= 0:
             tasks[task_id].status = TaskStatus.FAILED
-            tasks[task_id].message = "Source/mask frame count is zero"
+            tasks[task_id].message = "Source/detections frame count is zero"
             return
         tasks[task_id].total = total
 
         output_dir = os.path.dirname(req.output_path)
         if output_dir:
             os.makedirs(output_dir, exist_ok=True)
-
         writer = _build_video_writer(req.output_path, req.format, src_fps, src_width, src_height)
 
-        # Kernel size must be odd and >= 1
         blur_size = max(1, int(req.blur_size))
         if blur_size % 2 == 0:
             blur_size += 1
+        feather_radius = max(3, min(25, blur_size // 3))
+        feather_kernel = feather_radius * 2 + 1
 
-        print(f"[FaceMask] Starting blur bake: {req.video_path} + {req.mask_path} -> {req.output_path}")
-        if src_frames != mask_frames:
         print(
-                f"[FaceMask] Warning: frame count mismatch "
+            f"[FaceMask] Starting blur bake (bbox-msgpack): {req.video_path} + "
-                f"(src={src_frames}, mask={mask_frames}), processing {total} frames"
+            f"{req.detections_path} -> {req.output_path}"
         )
 
         for idx in range(total):
@@ -440,29 +363,61 @@ def process_bake_task(task_id: str, req: BakeRequest):
                 break
 
             src_ok, src_frame = src_cap.read()
-            mask_ok, mask_frame = mask_cap.read()
+            if not src_ok:
-            if not src_ok or not mask_ok:
                 break
 
-            if mask_frame.ndim == 3:
+            frame_boxes = frames[idx] if idx < len(frames) else []
-                mask_gray = cv2.cvtColor(mask_frame, cv2.COLOR_BGR2GRAY)
+            if not frame_boxes:
-            else:
+                writer.write(src_frame)
-                mask_gray = mask_frame
+                tasks[task_id].progress = idx + 1
+                continue
 
-            if mask_gray.shape[0] != src_height or mask_gray.shape[1] != src_width:
+            mask_gray = np.zeros((src_height, src_width), dtype=np.uint8)
-                mask_gray = cv2.resize(
+            for box in frame_boxes:
-                    mask_gray,
+                if not isinstance(box, list) or len(box) < 4:
-                    (src_width, src_height),
+                    continue
-                    interpolation=cv2.INTER_LINEAR,
+                x, y, w, h = int(box[0]), int(box[1]), int(box[2]), int(box[3])
+                if w <= 0 or h <= 0:
+                    continue
+                center = (x + w // 2, y + h // 2)
+                axes = (max(1, w // 2), max(1, h // 2))
+                cv2.ellipse(mask_gray, center, axes, 0, 0, 360, 255, -1)
+
+            if cv2.countNonZero(mask_gray) == 0:
+                writer.write(src_frame)
+                tasks[task_id].progress = idx + 1
+                continue
+
+            mask_gray = cv2.GaussianBlur(mask_gray, (feather_kernel, feather_kernel), 0)
+            _, mask_binary = cv2.threshold(mask_gray, 2, 255, cv2.THRESH_BINARY)
+            non_zero_coords = cv2.findNonZero(mask_binary)
+            if non_zero_coords is None:
+                writer.write(src_frame)
+                tasks[task_id].progress = idx + 1
+                continue
+
+            x, y, w, h = cv2.boundingRect(non_zero_coords)
+            blur_margin = max(1, (blur_size // 2) + feather_radius)
+            x1 = max(0, x - blur_margin)
+            y1 = max(0, y - blur_margin)
+            x2 = min(src_width, x + w + blur_margin)
+            y2 = min(src_height, y + h + blur_margin)
+
+            roi_src = src_frame[y1:y2, x1:x2]
+            roi_mask = mask_gray[y1:y2, x1:x2]
+            if roi_src.size == 0:
+                writer.write(src_frame)
+                tasks[task_id].progress = idx + 1
+                continue
+
+            roi_blurred = cv2.GaussianBlur(roi_src, (blur_size, blur_size), 0)
+            roi_alpha = (roi_mask.astype(np.float32) / 255.0)[..., np.newaxis]
+            roi_composed = (roi_src.astype(np.float32) * (1.0 - roi_alpha)) + (
+                roi_blurred.astype(np.float32) * roi_alpha
             )
-
+            output_frame = src_frame.copy()
-            blurred = cv2.GaussianBlur(src_frame, (blur_size, blur_size), 0)
+            output_frame[y1:y2, x1:x2] = np.clip(roi_composed, 0, 255).astype(np.uint8)
-            alpha = (mask_gray.astype(np.float32) / 255.0)[..., np.newaxis]
+            writer.write(output_frame)
-            composed = (src_frame.astype(np.float32) * (1.0 - alpha)) + (
-                blurred.astype(np.float32) * alpha
-            )
-            writer.write(np.clip(composed, 0, 255).astype(np.uint8))
-
             tasks[task_id].progress = idx + 1
 
         if tasks[task_id].status == TaskStatus.PROCESSING:
@@ -479,8 +434,6 @@ def process_bake_task(task_id: str, req: BakeRequest):
     finally:
         if src_cap:
             src_cap.release()
-        if mask_cap:
-            mask_cap.release()
         if writer:
             writer.release()
         if task_id in cancel_events: