姿勢推定と頭部検知を合成する手段を追加

__init__.py

@@ -40,15 +40,6 @@ def register():
         step=0.01,
     )
 
-    bpy.types.Scene.facemask_mask_scale = FloatProperty(
-        name="Mask Scale",
-        description="Scale factor for mask region (1.0 = exact face size)",
-        default=1.5,
-        min=1.0,
-        max=3.0,
-        step=0.1,
-    )
-
     bpy.types.Scene.facemask_cache_dir = StringProperty(
         name="Cache Directory",
         description="Optional cache root directory (empty = default .mask_cache)",
@@ -64,6 +55,15 @@ def register():
         max=501,
     )
 
+    bpy.types.Scene.facemask_bake_display_scale = FloatProperty(
+        name="Mask Scale",
+        description="Scale factor for the blur mask ellipse at bake time (1.0 = raw detection size)",
+        default=1.3,
+        min=0.5,
+        max=3.0,
+        step=0.1,
+    )
+
     bpy.types.Scene.facemask_bake_format = EnumProperty(
         name="Bake Format",
         description="Output format for baked blur video",
@@ -91,9 +91,9 @@ def unregister():
     # Unregister scene properties
     del bpy.types.Scene.facemask_conf_threshold
     del bpy.types.Scene.facemask_iou_threshold
-    del bpy.types.Scene.facemask_mask_scale
     del bpy.types.Scene.facemask_cache_dir
     del bpy.types.Scene.facemask_bake_blur_size
+    del bpy.types.Scene.facemask_bake_display_scale
     del bpy.types.Scene.facemask_bake_format
 
 

core/__init__.py

@@ -1,5 +1,5 @@
 """Core module exports."""
 
-from .async_bake_generator import AsyncBakeGenerator, get_bake_generator
+from .async_bake_generator import AsyncBakeGenerator as AsyncBakeGenerator, get_bake_generator as get_bake_generator
-from .async_generator import AsyncMaskGenerator, get_generator
+from .async_generator import AsyncMaskGenerator as AsyncMaskGenerator, get_generator as get_generator
-from .compositor_setup import create_mask_blur_node_tree, get_or_create_blur_node_tree
+from .compositor_setup import create_mask_blur_node_tree as create_mask_blur_node_tree, get_or_create_blur_node_tree as get_or_create_blur_node_tree

core/async_bake_generator.py

@@ -32,6 +32,7 @@ class AsyncBakeGenerator:
         detections_path: str,
         output_path: str,
         blur_size: int,
+        display_scale: float,
         fmt: str,
         on_complete: Optional[Callable] = None,
         on_progress: Optional[Callable] = None,
@@ -53,7 +54,7 @@ class AsyncBakeGenerator:
 
         self.worker_thread = threading.Thread(
             target=self._worker,
-            args=(video_path, detections_path, output_path, blur_size, fmt),
+            args=(video_path, detections_path, output_path, blur_size, display_scale, fmt),
             daemon=True,
         )
         self.worker_thread.start()
@@ -63,18 +64,102 @@ class AsyncBakeGenerator:
             first_interval=0.1,
         )
 
+    def start_images(
+        self,
+        image_dir: str,
+        filenames: list,
+        output_dir: str,
+        detections_path: str,
+        blur_size: int,
+        display_scale: float,
+        on_complete=None,
+        on_progress=None,
+    ):
+        """画像シーケンスのぼかしBakeを非同期で開始する。"""
+        global bpy
+        import bpy as _bpy
+        bpy = _bpy
+
+        if self.is_running:
+            raise RuntimeError("Blur bake already in progress")
+
+        self.is_running = True
+        self.total_frames = len(filenames)
+        self.current_frame = 0
+        self._on_complete = on_complete
+        self._on_progress = on_progress
+
+        self.worker_thread = threading.Thread(
+            target=self._worker_images,
+            args=(image_dir, filenames, output_dir, detections_path, blur_size, display_scale),
+            daemon=True,
+        )
+        self.worker_thread.start()
+        bpy.app.timers.register(self._check_progress, first_interval=0.1)
+
     def cancel(self):
         """Cancel the current bake processing."""
         self.is_running = False
         if self.worker_thread and self.worker_thread.is_alive():
             self.worker_thread.join(timeout=2.0)
 
+    def _worker_images(
+        self,
+        image_dir: str,
+        filenames: list,
+        output_dir: str,
+        detections_path: str,
+        blur_size: int,
+        display_scale: float,
+    ):
+        import time
+        from .inference_client import get_client
+
+        task_id = None
+        try:
+            client = get_client()
+            task_id = client.bake_image_blur(
+                image_dir=image_dir,
+                filenames=filenames,
+                output_dir=output_dir,
+                detections_path=detections_path,
+                blur_size=blur_size,
+                display_scale=display_scale,
+            )
+            while self.is_running:
+                status = client.get_task_status(task_id)
+                state = status.get("status")
+                total = status.get("total", 0)
+                if total > 0:
+                    self.total_frames = total
+                progress = status.get("progress", 0)
+                if progress >= 0:
+                    self.progress_queue.put(("progress", progress))
+                if state == "completed":
+                    result_path = status.get("result_path", output_dir)
+                    self.result_queue.put(("done", result_path))
+                    return
+                if state == "failed":
+                    self.result_queue.put(("error", status.get("message", "Unknown error")))
+                    return
+                if state == "cancelled":
+                    self.result_queue.put(("cancelled", None))
+                    return
+                time.sleep(0.5)
+
+            if task_id:
+                client.cancel_task(task_id)
+            self.result_queue.put(("cancelled", None))
+        except Exception as e:
+            self.result_queue.put(("error", str(e)))
+
     def _worker(
         self,
         video_path: str,
         detections_path: str,
         output_path: str,
         blur_size: int,
+        display_scale: float,
         fmt: str,
     ):
         import time
@@ -88,6 +173,7 @@ class AsyncBakeGenerator:
                 detections_path=detections_path,
                 output_path=output_path,
                 blur_size=blur_size,
+                display_scale=display_scale,
                 fmt=fmt,
             )
 

core/async_generator.py

@@ -44,7 +44,6 @@ class AsyncMaskGenerator:
         fps: float,
         conf_threshold: float = 0.5,
         iou_threshold: float = 0.45,
-        mask_scale: float = 1.5,
         on_complete: Optional[Callable] = None,
         on_progress: Optional[Callable] = None,
     ):
@@ -94,7 +93,6 @@ class AsyncMaskGenerator:
                 fps,
                 conf_threshold,
                 iou_threshold,
-                mask_scale,
             ),
             daemon=True,
         )
@@ -106,12 +104,189 @@ class AsyncMaskGenerator:
             first_interval=0.1,
         )
     
+    def start_images(
+        self,
+        image_dir: str,
+        filenames: list,
+        output_dir: str,
+        start_index: int,
+        end_index: int,
+        conf_threshold: float = 0.5,
+        iou_threshold: float = 0.45,
+        on_complete=None,
+        on_progress=None,
+    ):
+        """画像シーケンスの顔検出を非同期で開始する。"""
+        global bpy
+        import bpy as _bpy
+        bpy = _bpy
+
+        if self.is_running:
+            raise RuntimeError("Mask generation already in progress")
+
+        self.is_running = True
+        self.total_frames = end_index - start_index + 1
+        self.current_frame = 0
+        self._on_complete = on_complete
+        self._on_progress = on_progress
+
+        os.makedirs(output_dir, exist_ok=True)
+
+        self.worker_thread = threading.Thread(
+            target=self._worker_images,
+            args=(image_dir, filenames, output_dir, start_index, end_index,
+                  conf_threshold, iou_threshold),
+            daemon=True,
+        )
+        self.worker_thread.start()
+        bpy.app.timers.register(self._check_progress, first_interval=0.1)
+
+    def start_augment_pose(
+        self,
+        detections_path: str,
+        total_frames: int,
+        conf_threshold: float = 0.5,
+        iou_threshold: float = 0.45,
+        on_complete=None,
+        on_progress=None,
+    ):
+        """既存キャッシュへの pose 補完を非同期で開始する。"""
+        global bpy
+        import bpy as _bpy
+        bpy = _bpy
+
+        if self.is_running:
+            raise RuntimeError("Mask generation already in progress")
+
+        self.is_running = True
+        self.total_frames = total_frames
+        self.current_frame = 0
+        self._on_complete = on_complete
+        self._on_progress = on_progress
+
+        self.worker_thread = threading.Thread(
+            target=self._worker_augment_pose,
+            args=(detections_path, conf_threshold, iou_threshold),
+            daemon=True,
+        )
+        self.worker_thread.start()
+        bpy.app.timers.register(self._check_progress, first_interval=0.1)
+
+    def _worker_augment_pose(
+        self,
+        detections_path: str,
+        conf_threshold: float,
+        iou_threshold: float,
+    ):
+        """client.augment_pose() を呼んで task_id でポーリング。"""
+        import time
+        from .inference_client import get_client
+
+        try:
+            client = get_client()
+            task_id = client.augment_pose(
+                detections_path=detections_path,
+                conf_threshold=conf_threshold,
+                iou_threshold=iou_threshold,
+            )
+
+            while self.is_running:
+                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":
+                    progress = status.get("progress", self.total_frames)
+                    if progress >= 0:
+                        self.progress_queue.put(("progress", progress))
+                    result_path = status.get("result_path", detections_path)
+                    self.result_queue.put(("done", result_path))
+                    return
+                elif state == "failed":
+                    self.result_queue.put(("error", status.get("message", "Unknown error")))
+                    return
+                elif state == "cancelled":
+                    self.result_queue.put(("cancelled", None))
+                    return
+
+                progress = status.get("progress", 0)
+                if progress >= 0:
+                    self.progress_queue.put(("progress", progress))
+                time.sleep(0.5)
+
+            client.cancel_task(task_id)
+            self.result_queue.put(("cancelled", None))
+        except Exception as e:
+            import traceback
+            print(f"[FaceMask] Error: {e}")
+            traceback.print_exc()
+            self.result_queue.put(("error", str(e)))
+
     def cancel(self):
         """Cancel the current processing."""
         self.is_running = False
         if self.worker_thread and self.worker_thread.is_alive():
             self.worker_thread.join(timeout=2.0)
 
+    def _worker_images(
+        self,
+        image_dir: str,
+        filenames: list,
+        output_dir: str,
+        start_index: int,
+        end_index: int,
+        conf_threshold: float,
+        iou_threshold: float,
+    ):
+        import time
+        from .inference_client import get_client
+
+        try:
+            client = get_client()
+            task_id = client.generate_mask_images(
+                image_dir=image_dir,
+                filenames=filenames,
+                output_dir=output_dir,
+                start_index=start_index,
+                end_index=end_index,
+                conf_threshold=conf_threshold,
+                iou_threshold=iou_threshold,
+            )
+            while self.is_running:
+                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":
+                    progress = status.get("progress", self.total_frames)
+                    if progress >= 0:
+                        self.progress_queue.put(("progress", 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":
+                    self.result_queue.put(("error", status.get("message", "Unknown error")))
+                    return
+                elif state == "cancelled":
+                    self.result_queue.put(("cancelled", None))
+                    return
+                progress = status.get("progress", 0)
+                if progress >= 0:
+                    self.progress_queue.put(("progress", progress))
+                time.sleep(0.5)
+
+            client.cancel_task(task_id)
+            self.result_queue.put(("cancelled", None))
+        except Exception as e:
+            self.result_queue.put(("error", str(e)))
+
     def _worker(
         self,
         video_path: str,
@@ -121,7 +296,6 @@ class AsyncMaskGenerator:
         fps: float,
         conf_threshold: float,
         iou_threshold: float,
-        mask_scale: float,
     ):
         """
         Worker thread function. Delegates to inference server and polls status.
@@ -141,7 +315,6 @@ class AsyncMaskGenerator:
                 end_frame=end_frame,
                 conf_threshold=conf_threshold,
                 iou_threshold=iou_threshold,
-                mask_scale=mask_scale,
             )
             print(f"[FaceMask] Task started: {task_id}")
             

core/batch_processor.py

@@ -0,0 +1,280 @@
+"""
+Batch processor for sequential Generate+Bake across multiple VSE strips.
+
+Uses timer-based async chaining so Blender's UI stays responsive.
+"""
+
+from typing import List, Optional, Callable, Any
+
+# Lazy-imported inside Blender
+bpy = None
+
+
+class _DummyOperator:
+    """Dummy operator object for _start_bake_impl calls."""
+
+    def report(self, level, msg):
+        print(f"[FaceMask] Batch: {msg}")
+
+
+class BatchProcessor:
+    """Manages sequential Generate Detection Cache → Bake across a list of strips."""
+
+    def __init__(self):
+        self.is_running: bool = False
+        self._mode: str = "full"  # "full" or "mask_only"
+        self._strip_names: List[str] = []
+        self._current_idx: int = 0
+        self._context: Any = None
+        self._cancelled: bool = False
+        self._results: List[dict] = []
+        self._on_item_complete: Optional[Callable] = None  # (idx, total, name, status)
+        self._on_all_complete: Optional[Callable] = None   # (results)
+
+    # ------------------------------------------------------------------
+    # Public API
+    # ------------------------------------------------------------------
+
+    def start(self, context, strips, on_item_complete=None, on_all_complete=None, mode="full"):
+        """Start batch processing for the given strips.
+
+        mode:
+          "full"      - マスク生成（キャッシュなければ）→ Bake
+          "mask_only" - キャッシュを無視してマスク生成のみ（Bakeしない）
+        """
+        global bpy
+        import bpy as _bpy
+        bpy = _bpy
+
+        if self.is_running:
+            raise RuntimeError("Batch already running")
+
+        self.is_running = True
+        self._mode = mode
+        self._strip_names = [s.name for s in strips]
+        self._current_idx = 0
+        self._context = context
+        self._cancelled = False
+        self._results = []
+        self._on_item_complete = on_item_complete
+        self._on_all_complete = on_all_complete
+
+        wm = context.window_manager
+        wm.batch_current = 0
+        wm.batch_total = len(self._strip_names)
+        wm.batch_current_name = ""
+
+        bpy.app.timers.register(self._process_next, first_interval=0.0)
+
+    def cancel(self):
+        """Cancel batch. Stops currently running mask gen / bake."""
+        self._cancelled = True
+        from .async_generator import get_generator
+        from .async_bake_generator import get_bake_generator
+        gen = get_generator()
+        bake_gen = get_bake_generator()
+        if gen.is_running:
+            gen.cancel()
+        if bake_gen.is_running:
+            bake_gen.cancel()
+
+    # ------------------------------------------------------------------
+    # Internal: queue stepping
+    # ------------------------------------------------------------------
+
+    def _process_next(self):
+        """Process the next strip in the queue (called via timer)."""
+        if self._cancelled:
+            self._finish()
+            return None
+
+        if self._current_idx >= len(self._strip_names):
+            self._finish()
+            return None
+
+        strip_name = self._strip_names[self._current_idx]
+        seq_editor = self._context.scene.sequence_editor
+        strip = seq_editor.strips.get(strip_name)
+
+        if strip is None:
+            print(f"[FaceMask] Batch: strip not found, skipping: {strip_name}")
+            self._results.append({"strip": strip_name, "status": "skipped"})
+            if self._on_item_complete:
+                self._on_item_complete(self._current_idx, len(self._strip_names), strip_name, "skipped")
+            self._current_idx += 1
+            bpy.app.timers.register(self._process_next, first_interval=0.0)
+            return None
+
+        # Update wm progress labels
+        wm = self._context.window_manager
+        wm.batch_current = self._current_idx + 1
+        wm.batch_current_name = strip_name
+        for area in self._context.screen.areas:
+            if area.type == "SEQUENCE_EDITOR":
+                area.tag_redraw()
+
+        if self._mode == "mask_only":
+            # キャッシュを無視して常にマスク生成（Bakeしない）
+            self._start_mask_gen(strip)
+        else:
+            from .utils import check_detection_cache
+            if not check_detection_cache(strip.name):
+                self._start_mask_gen(strip)
+            else:
+                self._start_bake(strip)
+
+        return None  # one-shot timer
+
+    def _schedule_next(self):
+        bpy.app.timers.register(self._process_next, first_interval=0.0)
+
+    # ------------------------------------------------------------------
+    # Mask generation
+    # ------------------------------------------------------------------
+
+    def _start_mask_gen(self, strip):
+        from ..operators.generate_mask import start_mask_gen_for_strip
+
+        strip_name = strip.name
+
+        def on_complete(status, data):
+            self._on_mask_done(strip_name, status, data)
+
+        def on_progress(current, total):
+            wm = self._context.window_manager
+            wm.mask_progress = current
+            wm.mask_total = max(total, 1)
+            for area in self._context.screen.areas:
+                if area.type == "SEQUENCE_EDITOR":
+                    area.tag_redraw()
+
+        try:
+            start_mask_gen_for_strip(self._context, strip, on_complete, on_progress)
+            print(f"[FaceMask] Batch: started mask gen for {strip_name}")
+        except Exception as e:
+            print(f"[FaceMask] Batch: failed to start mask gen for {strip_name}: {e}")
+            self._on_mask_done(strip_name, "error", str(e))
+
+    def _on_mask_done(self, strip_name, status, data):
+        if self._cancelled or status == "cancelled":
+            self._results.append({"strip": strip_name, "status": "cancelled"})
+            self._finish()
+            return
+
+        if status == "error":
+            print(f"[FaceMask] Batch: mask gen failed for {strip_name}: {data}")
+            self._results.append({"strip": strip_name, "status": "error", "reason": str(data)})
+            if self._on_item_complete:
+                self._on_item_complete(self._current_idx, len(self._strip_names), strip_name, "error")
+            self._current_idx += 1
+            self._schedule_next()
+            return
+
+        # Mask gen succeeded
+        if self._mode == "mask_only":
+            # Bakeしない：結果を記録して次へ
+            self._results.append({"strip": strip_name, "status": "done"})
+            if self._on_item_complete:
+                self._on_item_complete(self._current_idx, len(self._strip_names), strip_name, "done")
+            self._current_idx += 1
+            self._schedule_next()
+            return
+
+        # full mode: proceed to bake
+        seq_editor = self._context.scene.sequence_editor
+        strip = seq_editor.strips.get(strip_name)
+        if strip is None:
+            print(f"[FaceMask] Batch: strip removed after mask gen: {strip_name}")
+            self._results.append({"strip": strip_name, "status": "skipped"})
+            if self._on_item_complete:
+                self._on_item_complete(self._current_idx, len(self._strip_names), strip_name, "skipped")
+            self._current_idx += 1
+            self._schedule_next()
+            return
+
+        self._start_bake(strip)
+
+    # ------------------------------------------------------------------
+    # Bake
+    # ------------------------------------------------------------------
+
+    def _start_bake(self, strip):
+        from .async_bake_generator import get_bake_generator
+        from ..operators.apply_blur import _start_bake_impl
+
+        strip_name = strip.name
+
+        def on_complete_extra(status, data):
+            self._on_bake_done(strip_name, status, data)
+
+        bake_gen = get_bake_generator()
+        result = _start_bake_impl(
+            _DummyOperator(),
+            self._context,
+            force=False,
+            strip=strip,
+            on_complete_extra=on_complete_extra,
+        )
+
+        if result == {"CANCELLED"}:
+            # Error starting bake
+            print(f"[FaceMask] Batch: bake failed to start for {strip_name}")
+            self._results.append({"strip": strip_name, "status": "error", "reason": "bake failed to start"})
+            if self._on_item_complete:
+                self._on_item_complete(self._current_idx, len(self._strip_names), strip_name, "error")
+            self._current_idx += 1
+            self._schedule_next()
+        elif not bake_gen.is_running:
+            # Cache hit: on_complete_extra was NOT called by _start_bake_impl
+            print(f"[FaceMask] Batch: bake cache hit for {strip_name}")
+            self._on_bake_done(strip_name, "done", None)
+
+    def _on_bake_done(self, strip_name, status, data):
+        if self._cancelled or status == "cancelled":
+            self._results.append({"strip": strip_name, "status": "cancelled"})
+            self._finish()
+            return
+
+        if status == "error":
+            print(f"[FaceMask] Batch: bake failed for {strip_name}: {data}")
+            self._results.append({"strip": strip_name, "status": "error", "reason": str(data)})
+        else:
+            self._results.append({"strip": strip_name, "status": "done"})
+            print(f"[FaceMask] Batch: completed {strip_name}")
+
+        if self._on_item_complete:
+            self._on_item_complete(self._current_idx, len(self._strip_names), strip_name, status)
+
+        self._current_idx += 1
+        self._schedule_next()
+
+    # ------------------------------------------------------------------
+    # Finish
+    # ------------------------------------------------------------------
+
+    def _finish(self):
+        self.is_running = False
+        wm = self._context.window_manager
+        wm.batch_current = 0
+        wm.batch_total = 0
+        wm.batch_current_name = ""
+
+        print(f"[FaceMask] Batch: all done. Results: {self._results}")
+
+        if self._on_all_complete:
+            self._on_all_complete(self._results)
+
+        for area in self._context.screen.areas:
+            if area.type == "SEQUENCE_EDITOR":
+                area.tag_redraw()
+
+
+# Singleton
+_batch_processor: Optional[BatchProcessor] = None
+
+
+def get_batch_processor() -> BatchProcessor:
+    global _batch_processor
+    if _batch_processor is None:
+        _batch_processor = BatchProcessor()
+    return _batch_processor

core/compositor_setup.py

@@ -8,7 +8,7 @@ only to masked regions of a video strip.
 def create_mask_blur_node_tree(
     name: str = "FaceMaskBlur",
     blur_size: int = 50,
-) -> "bpy.types.NodeTree":
+) -> "bpy.types.NodeTree":  # noqa: F821
     """
     Create a compositing node tree for mask-based blur.
     
@@ -107,10 +107,10 @@ def create_mask_blur_node_tree(
 
 
 def setup_strip_compositor_modifier(
-    strip: "bpy.types.Strip",
+    strip: "bpy.types.Strip",  # noqa: F821
-    mask_strip: "bpy.types.Strip",
+    mask_strip: "bpy.types.Strip",  # noqa: F821
-    node_tree: "bpy.types.NodeTree",
+    node_tree: "bpy.types.NodeTree",  # noqa: F821
-) -> "bpy.types.SequenceModifier":
+) -> "bpy.types.SequenceModifier":  # noqa: F821
     """
     Add a Compositor modifier to a strip using the mask-blur node tree.
     
@@ -148,7 +148,7 @@ def setup_strip_compositor_modifier(
     return modifier
 
 
-def get_or_create_blur_node_tree(blur_size: int = 50) -> "bpy.types.NodeTree":
+def get_or_create_blur_node_tree(blur_size: int = 50) -> "bpy.types.NodeTree":  # noqa: F821
     """
     Get existing or create new blur node tree with specified blur size.
     

core/inference_client.py

@@ -204,7 +204,6 @@ class InferenceClient:
         end_frame: int,
         conf_threshold: float,
         iou_threshold: float,
-        mask_scale: float,
     ) -> str:
         """
         Request mask generation.
@@ -222,7 +221,6 @@ class InferenceClient:
             "end_frame": end_frame,
             "conf_threshold": conf_threshold,
             "iou_threshold": iou_threshold,
-            "mask_scale": mask_scale,
         }
 
         req = urllib.request.Request(
@@ -239,6 +237,36 @@ class InferenceClient:
         except urllib.error.HTTPError as e:
             raise RuntimeError(f"Server error: {e.read().decode('utf-8')}")
 
+    def augment_pose(
+        self,
+        detections_path: str,
+        conf_threshold: float,
+        iou_threshold: float,
+    ) -> str:
+        """既存キャッシュに pose 推定結果を追加合成する。task_id を返す。"""
+        if not self.is_server_running():
+            self.start_server()
+
+        data = {
+            "detections_path": detections_path,
+            "conf_threshold": conf_threshold,
+            "iou_threshold": iou_threshold,
+        }
+
+        req = urllib.request.Request(
+            f"{self.SERVER_URL}/augment_pose",
+            data=json.dumps(data).encode("utf-8"),
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+
+        try:
+            with urllib.request.urlopen(req) as response:
+                result = json.loads(response.read().decode("utf-8"))
+                return result["id"]
+        except urllib.error.HTTPError as e:
+            raise RuntimeError(f"Server error: {e.read().decode('utf-8')}")
+
     def get_task_status(self, task_id: str) -> Dict[str, Any]:
         """Get status of a task."""
         try:
@@ -249,12 +277,30 @@ class InferenceClient:
         except urllib.error.HTTPError:
             return {"status": "unknown"}
 
+    def get_video_info(self, video_path: str) -> Dict[str, Any]:
+        """Get video metadata from the inference server."""
+        if not self.is_server_running():
+            self.start_server()
+
+        req = urllib.request.Request(
+            f"{self.SERVER_URL}/video_info",
+            data=json.dumps({"video_path": video_path}).encode("utf-8"),
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+        try:
+            with urllib.request.urlopen(req) as response:
+                return json.loads(response.read().decode("utf-8"))
+        except urllib.error.HTTPError as e:
+            raise RuntimeError(f"Server error: {e.read().decode('utf-8')}")
+
     def bake_blur(
         self,
         video_path: str,
         detections_path: str,
         output_path: str,
         blur_size: int,
+        display_scale: float,
         fmt: str,
     ) -> str:
         """
@@ -271,6 +317,7 @@ class InferenceClient:
             "detections_path": detections_path,
             "output_path": output_path,
             "blur_size": blur_size,
+            "display_scale": display_scale,
             "format": fmt,
         }
 
@@ -288,6 +335,76 @@ class InferenceClient:
         except urllib.error.HTTPError as e:
             raise RuntimeError(f"Server error: {e.read().decode('utf-8')}")
 
+    def generate_mask_images(
+        self,
+        image_dir: str,
+        filenames: list,
+        output_dir: str,
+        start_index: int,
+        end_index: int,
+        conf_threshold: float,
+        iou_threshold: float,
+    ) -> str:
+        """画像シーケンスの顔検出タスクを開始して task_id を返す。"""
+        if not self.is_server_running():
+            self.start_server()
+
+        data = {
+            "image_dir": image_dir,
+            "filenames": filenames,
+            "output_dir": output_dir,
+            "start_index": start_index,
+            "end_index": end_index,
+            "conf_threshold": conf_threshold,
+            "iou_threshold": iou_threshold,
+        }
+        req = urllib.request.Request(
+            f"{self.SERVER_URL}/generate_images",
+            data=json.dumps(data).encode("utf-8"),
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+        try:
+            with urllib.request.urlopen(req) as response:
+                result = json.loads(response.read().decode("utf-8"))
+                return result["id"]
+        except urllib.error.HTTPError as e:
+            raise RuntimeError(f"Server error: {e.read().decode('utf-8')}")
+
+    def bake_image_blur(
+        self,
+        image_dir: str,
+        filenames: list,
+        output_dir: str,
+        detections_path: str,
+        blur_size: int,
+        display_scale: float,
+    ) -> str:
+        """画像シーケンスのぼかしBakeタスクを開始して task_id を返す。"""
+        if not self.is_server_running():
+            self.start_server()
+
+        data = {
+            "image_dir": image_dir,
+            "filenames": filenames,
+            "output_dir": output_dir,
+            "detections_path": detections_path,
+            "blur_size": blur_size,
+            "display_scale": display_scale,
+        }
+        req = urllib.request.Request(
+            f"{self.SERVER_URL}/bake_image_blur",
+            data=json.dumps(data).encode("utf-8"),
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+        try:
+            with urllib.request.urlopen(req) as response:
+                result = json.loads(response.read().decode("utf-8"))
+                return result["id"]
+        except urllib.error.HTTPError as e:
+            raise RuntimeError(f"Server error: {e.read().decode('utf-8')}")
+
     def cancel_task(self, task_id: str):
         """Cancel a task."""
         try:

core/utils.py

@@ -83,6 +83,15 @@ def get_detections_path_for_strip(strip_name: str) -> str:
     return os.path.join(get_cache_dir_for_strip(strip_name), "detections.msgpack")
 
 
+def check_detection_cache(strip_name: str) -> bool:
+    """Detection cache ファイルが存在し有効かどうか確認する。"""
+    path = get_detections_path_for_strip(strip_name)
+    try:
+        return os.path.exists(path) and os.path.getsize(path) > 0
+    except OSError:
+        return False
+
+
 def get_cache_info(strip_name: Optional[str] = None) -> Tuple[str, int, int]:
     """
     Get cache directory information.

operators/__init__.py

@@ -3,15 +3,18 @@
 from . import generate_mask
 from . import apply_blur
 from . import clear_cache
+from . import batch_bake
 
 
 def register():
     generate_mask.register()
     apply_blur.register()
     clear_cache.register()
+    batch_bake.register()
 
 
 def unregister():
+    batch_bake.unregister()
     clear_cache.unregister()
     apply_blur.unregister()
     generate_mask.unregister()

operators/apply_blur.py

@@ -20,6 +20,7 @@ KEY_BAKED = "facemask_baked_filepath"
 KEY_MODE = "facemask_source_mode"
 KEY_FORMAT = "facemask_bake_format"
 KEY_BLUR_SIZE = "facemask_bake_blur_size"
+KEY_DISPLAY_SCALE = "facemask_bake_display_scale"
 
 
 FORMAT_EXT = {
@@ -36,6 +37,12 @@ def _output_path(video_strip, detections_path: str, fmt: str) -> str:
     return os.path.join(out_dir, f"{safe_name}_blurred.{ext}")
 
 
+def _output_dir_for_images(strip, detections_path: str) -> str:
+    out_dir = os.path.dirname(detections_path)
+    safe_name = strip.name.replace("/", "_").replace("\\", "_")
+    return os.path.join(out_dir, f"{safe_name}_blurred")
+
+
 def _reload_movie_strip(strip):
     if hasattr(strip, "reload"):
         try:
@@ -44,67 +51,80 @@ def _reload_movie_strip(strip):
             pass
 
 
-def _set_strip_source(strip, filepath: str):
+def _set_strip_source(strip, path: str):
-    strip.filepath = filepath
+    if strip.type == "IMAGE":
+        strip.directory = path
+    else:
+        strip.filepath = path
         _reload_movie_strip(strip)
 
 
-class SEQUENCER_OT_bake_and_swap_blur_source(Operator):
+def _start_bake_impl(operator, context, force: bool = False, strip=None, on_complete_extra=None):
-    """Bake masked blur and replace active strip source with baked video."""
+    """Bakeの共通実装。force=True でキャッシュを無視して再Bakeする。
 
-    bl_idname = "sequencer.bake_and_swap_blur_source"
+    strip: 処理対象のstrip。None の場合は active_strip を使用。
-    bl_label = "Bake & Swap Source"
+    on_complete_extra: 非同期Bake完了時に追加で呼ばれるコールバック (status, data)。
-    bl_description = "Bake masked blur to video and swap active strip source"
+                       キャッシュヒット即時完了の場合は呼ばれない。
-    bl_options = {"REGISTER", "UNDO"}
+    MOVIE / IMAGE 両対応。
-
+    """
-    @classmethod
-    def poll(cls, context):
-        if not context.scene.sequence_editor:
-            return False
-        # Prevent overlapping heavy tasks
-        if get_mask_generator().is_running:
-            return False
-        if get_bake_generator().is_running:
-            return False
-        strip = context.scene.sequence_editor.active_strip
-        return bool(strip and strip.type == "MOVIE")
-
-    def execute(self, context):
     seq_editor = context.scene.sequence_editor
     scene = context.scene
-        video_strip = seq_editor.active_strip
+    video_strip = strip if strip is not None else seq_editor.active_strip
+    is_image = video_strip.type == "IMAGE"
 
-        video_path = bpy.path.abspath(video_strip.filepath)
     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 os.path.exists(detections_path):
-            self.report({"ERROR"}, f"Detection cache not found: {detections_path}")
+        operator.report({"ERROR"}, f"Detection cache not found: {detections_path}")
         return {"CANCELLED"}
 
+    blur_size = int(scene.facemask_bake_blur_size)
+    display_scale = float(scene.facemask_bake_display_scale)
+
+    if is_image:
+        image_dir = bpy.path.abspath(video_strip.directory)
+        filenames = [elem.filename for elem in video_strip.elements]
+        if not os.path.isdir(image_dir):
+            operator.report({"ERROR"}, f"Image directory not found: {image_dir}")
+            return {"CANCELLED"}
+        output_dir = _output_dir_for_images(video_strip, detections_path)
+        original_source = image_dir
+        bake_format = None  # IMAGE strips don't use format
+    else:
+        video_path = bpy.path.abspath(video_strip.filepath)
+        if not os.path.exists(video_path):
+            operator.report({"ERROR"}, f"Source video not found: {video_path}")
+            return {"CANCELLED"}
         bake_format = scene.facemask_bake_format
         output_path = _output_path(video_strip, detections_path, bake_format)
-        blur_size = int(scene.facemask_bake_blur_size)
+        original_source = video_path
 
-        # Reuse baked cache when parameters match and file still exists.
+    if not force:
+        # パラメータが一致するキャッシュがあればswapのみ
         cached_baked_path = video_strip.get(KEY_BAKED)
-        cached_format = video_strip.get(KEY_FORMAT)
         cached_blur_size = video_strip.get(KEY_BLUR_SIZE)
+        cached_display_scale = video_strip.get(KEY_DISPLAY_SCALE)
         try:
             cached_blur_size_int = int(cached_blur_size)
         except (TypeError, ValueError):
             cached_blur_size_int = None
-        if (
+        try:
-            cached_baked_path
+            cached_display_scale_f = float(cached_display_scale)
-            and os.path.exists(cached_baked_path)
+        except (TypeError, ValueError):
-            and cached_format == bake_format
+            cached_display_scale_f = None
+
+        cache_exists = (
+            cached_baked_path and os.path.exists(cached_baked_path)
             and cached_blur_size_int == blur_size
-        ):
+            and cached_display_scale_f == display_scale
+        )
+        if not is_image:
+            cache_exists = cache_exists and video_strip.get(KEY_FORMAT) == bake_format
+
+        if cache_exists:
             if video_strip.get(KEY_MODE) != "baked":
                 video_strip[KEY_MODE] = "baked"
                 _set_strip_source(video_strip, cached_baked_path)
-            self.report({"INFO"}, "Using cached baked blur")
+            operator.report({"INFO"}, "Using cached baked blur")
             return {"FINISHED"}
 
     bake_generator = get_bake_generator()
@@ -117,17 +137,18 @@ class SEQUENCER_OT_bake_and_swap_blur_source(Operator):
             return
 
         if status == "done":
-                result_path = data or output_path
+            result = data or (output_dir if is_image else output_path)
-                original_path = strip.get(KEY_ORIGINAL)
             current_mode = strip.get(KEY_MODE, "original")
-                if not original_path or current_mode != "baked":
+            if not strip.get(KEY_ORIGINAL) or current_mode != "baked":
-                    strip[KEY_ORIGINAL] = video_path
+                strip[KEY_ORIGINAL] = original_source
-                strip[KEY_BAKED] = result_path
+            strip[KEY_BAKED] = result
             strip[KEY_MODE] = "baked"
-                strip[KEY_FORMAT] = bake_format
             strip[KEY_BLUR_SIZE] = blur_size
-                _set_strip_source(strip, result_path)
+            strip[KEY_DISPLAY_SCALE] = display_scale
-                print(f"[FaceMask] Bake completed and source swapped: {result_path}")
+            if not is_image:
+                strip[KEY_FORMAT] = bake_format
+            _set_strip_source(strip, result)
+            print(f"[FaceMask] Bake completed and source swapped: {result}")
         elif status == "error":
             print(f"[FaceMask] Bake failed: {data}")
         elif status == "cancelled":
@@ -137,6 +158,9 @@ class SEQUENCER_OT_bake_and_swap_blur_source(Operator):
             if area.type == "SEQUENCE_EDITOR":
                 area.tag_redraw()
 
+        if on_complete_extra:
+            on_complete_extra(status, data)
+
     def on_progress(current, total):
         wm.bake_progress = current
         wm.bake_total = max(total, 1)
@@ -148,20 +172,108 @@ class SEQUENCER_OT_bake_and_swap_blur_source(Operator):
     wm.bake_total = 1
 
     try:
+        if is_image:
+            bake_generator.start_images(
+                image_dir=image_dir,
+                filenames=filenames,
+                output_dir=output_dir,
+                detections_path=detections_path,
+                blur_size=blur_size,
+                display_scale=display_scale,
+                on_complete=on_complete,
+                on_progress=on_progress,
+            )
+        else:
             bake_generator.start(
                 video_path=video_path,
                 detections_path=detections_path,
                 output_path=output_path,
                 blur_size=blur_size,
+                display_scale=display_scale,
                 fmt=bake_format.lower(),
                 on_complete=on_complete,
                 on_progress=on_progress,
             )
     except Exception as e:
-            self.report({"ERROR"}, f"Failed to start bake: {e}")
+        operator.report({"ERROR"}, f"Failed to start bake: {e}")
         return {"CANCELLED"}
 
-        self.report({"INFO"}, "Started blur bake in background")
+    operator.report({"INFO"}, "Started blur bake in background")
+    return {"FINISHED"}
+
+
+class SEQUENCER_OT_bake_and_swap_blur_source(Operator):
+    """Bake masked blur (reuse cache if parameters match)."""
+
+    bl_idname = "sequencer.bake_and_swap_blur_source"
+    bl_label = "Bake"
+    bl_description = "Bake masked blur to video and swap active strip source"
+    bl_options = {"REGISTER", "UNDO"}
+
+    @classmethod
+    def poll(cls, context):
+        if not context.scene.sequence_editor:
+            return False
+        if get_mask_generator().is_running:
+            return False
+        if get_bake_generator().is_running:
+            return False
+        strip = context.scene.sequence_editor.active_strip
+        return bool(strip and strip.type in {"MOVIE", "IMAGE"})
+
+    def execute(self, context):
+        return _start_bake_impl(self, context, force=False)
+
+
+class SEQUENCER_OT_force_rebake_blur(Operator):
+    """Force re-bake, ignoring any existing cached result."""
+
+    bl_idname = "sequencer.force_rebake_blur"
+    bl_label = "Re-bake"
+    bl_description = "Discard cached bake and re-bake from scratch"
+    bl_options = {"REGISTER", "UNDO"}
+
+    @classmethod
+    def poll(cls, context):
+        if not context.scene.sequence_editor:
+            return False
+        if get_mask_generator().is_running:
+            return False
+        if get_bake_generator().is_running:
+            return False
+        strip = context.scene.sequence_editor.active_strip
+        return bool(strip and strip.type in {"MOVIE", "IMAGE"})
+
+    def execute(self, context):
+        return _start_bake_impl(self, context, force=True)
+
+
+class SEQUENCER_OT_swap_to_baked_blur(Operator):
+    """Swap active strip source to already-baked video (no re-bake)."""
+
+    bl_idname = "sequencer.swap_to_baked_blur"
+    bl_label = "Swap to Baked"
+    bl_description = "Switch active strip source to the baked video without re-baking"
+    bl_options = {"REGISTER", "UNDO"}
+
+    @classmethod
+    def poll(cls, context):
+        if not context.scene.sequence_editor:
+            return False
+        if get_bake_generator().is_running:
+            return False
+        strip = context.scene.sequence_editor.active_strip
+        if not strip or strip.type not in {"MOVIE", "IMAGE"}:
+            return False
+        baked_path = strip.get(KEY_BAKED)
+        return bool(baked_path and os.path.exists(baked_path))
+
+    def execute(self, context):
+        strip = context.scene.sequence_editor.active_strip
+        baked_path = strip.get(KEY_BAKED)
+        _set_strip_source(strip, baked_path)
+        strip[KEY_MODE] = "baked"
+        self.report({"INFO"}, "Swapped to baked source")
         return {"FINISHED"}
 
 
@@ -169,7 +281,7 @@ class SEQUENCER_OT_restore_original_source(Operator):
     """Restore active strip source filepath to original video."""
 
     bl_idname = "sequencer.restore_original_source"
-    bl_label = "Restore Original Source"
+    bl_label = "Restore Original"
     bl_description = "Restore active strip to original source filepath"
     bl_options = {"REGISTER", "UNDO"}
 
@@ -180,7 +292,9 @@ class SEQUENCER_OT_restore_original_source(Operator):
         if get_bake_generator().is_running:
             return False
         strip = context.scene.sequence_editor.active_strip
-        if not strip or strip.type != "MOVIE":
+        if not strip or strip.type not in {"MOVIE", "IMAGE"}:
+            return False
+        if strip.get(KEY_MODE, "original") == "original":
             return False
         return bool(strip.get(KEY_ORIGINAL))
 
@@ -205,7 +319,7 @@ class SEQUENCER_OT_apply_mask_blur(Operator):
 
     bl_idname = "sequencer.apply_mask_blur"
     bl_label = "Apply Mask Blur"
-    bl_description = "Compatibility alias for Bake & Swap Source"
+    bl_description = "Compatibility alias for Bake"
     bl_options = {"REGISTER", "UNDO"}
 
     @classmethod
@@ -236,6 +350,8 @@ class SEQUENCER_OT_cancel_bake_blur(Operator):
 
 classes = [
     SEQUENCER_OT_bake_and_swap_blur_source,
+    SEQUENCER_OT_force_rebake_blur,
+    SEQUENCER_OT_swap_to_baked_blur,
     SEQUENCER_OT_restore_original_source,
     SEQUENCER_OT_cancel_bake_blur,
     SEQUENCER_OT_apply_mask_blur,

operators/batch_bake.py

@@ -0,0 +1,191 @@
+"""
+Batch Bake operator: sequentially Generate Detection Cache → Bake
+for all selected MOVIE strips in the VSE.
+"""
+
+import os
+import bpy
+from bpy.props import IntProperty, StringProperty
+from bpy.types import Operator
+
+from ..core.batch_processor import get_batch_processor
+from ..core.async_generator import get_generator as get_mask_generator
+from ..core.async_bake_generator import get_bake_generator
+from .apply_blur import KEY_ORIGINAL, KEY_MODE, _set_strip_source
+
+
+class SEQUENCER_OT_batch_bake_selected(Operator):
+    """Generate detection cache and bake blur for all selected MOVIE/IMAGE strips."""
+
+    bl_idname = "sequencer.batch_bake_selected"
+    bl_label = "Batch Bake Selected"
+    bl_description = "Generate detection cache and bake blur for all selected MOVIE/IMAGE strips"
+    bl_options = {"REGISTER"}
+
+    @classmethod
+    def poll(cls, context):
+        if not context.scene.sequence_editor:
+            return False
+        if get_batch_processor().is_running:
+            return False
+        if get_mask_generator().is_running:
+            return False
+        if get_bake_generator().is_running:
+            return False
+        seq_editor = context.scene.sequence_editor
+        return any(s.select and s.type in {"MOVIE", "IMAGE"} for s in seq_editor.strips)
+
+    def execute(self, context):
+        seq_editor = context.scene.sequence_editor
+        strips = [s for s in seq_editor.strips if s.select and s.type in {"MOVIE", "IMAGE"}]
+
+        if not strips:
+            self.report({"WARNING"}, "No MOVIE or IMAGE strips selected")
+            return {"CANCELLED"}
+
+        batch = get_batch_processor()
+
+        def on_item_complete(idx, total, strip_name, status):
+            pass  # wm properties already updated by BatchProcessor
+
+        def on_all_complete(results):
+            done = sum(1 for r in results if r["status"] == "done")
+            total = len(results)
+            print(f"[FaceMask] Batch finished: {done}/{total} strips completed")
+
+        wm = context.window_manager
+        wm.batch_current = 0
+        wm.batch_total = len(strips)
+        wm.batch_current_name = ""
+
+        batch.start(context, strips, on_item_complete=on_item_complete, on_all_complete=on_all_complete)
+        self.report({"INFO"}, f"Batch bake started for {len(strips)} strips")
+        return {"FINISHED"}
+
+
+class SEQUENCER_OT_batch_regenerate_cache(Operator):
+    """Regenerate detection cache for all selected MOVIE/IMAGE strips (ignore existing cache)."""
+
+    bl_idname = "sequencer.batch_regenerate_cache"
+    bl_label = "Batch Regenerate Cache"
+    bl_description = "Regenerate detection cache for all selected MOVIE/IMAGE strips"
+    bl_options = {"REGISTER"}
+
+    @classmethod
+    def poll(cls, context):
+        if not context.scene.sequence_editor:
+            return False
+        if get_batch_processor().is_running:
+            return False
+        if get_mask_generator().is_running:
+            return False
+        if get_bake_generator().is_running:
+            return False
+        seq_editor = context.scene.sequence_editor
+        return any(s.select and s.type in {"MOVIE", "IMAGE"} for s in seq_editor.strips)
+
+    def execute(self, context):
+        seq_editor = context.scene.sequence_editor
+        strips = [s for s in seq_editor.strips if s.select and s.type in {"MOVIE", "IMAGE"}]
+
+        if not strips:
+            self.report({"WARNING"}, "No MOVIE or IMAGE strips selected")
+            return {"CANCELLED"}
+
+        batch = get_batch_processor()
+
+        def on_all_complete(results):
+            done = sum(1 for r in results if r["status"] == "done")
+            print(f"[FaceMask] Batch regenerate finished: {done}/{len(results)} strips")
+
+        batch.start(
+            context,
+            strips,
+            on_all_complete=on_all_complete,
+            mode="mask_only",
+        )
+        self.report({"INFO"}, f"Batch regenerate cache started for {len(strips)} strips")
+        return {"FINISHED"}
+
+
+class SEQUENCER_OT_batch_restore_original(Operator):
+    """Restore original source for all selected MOVIE/IMAGE strips."""
+
+    bl_idname = "sequencer.batch_restore_original"
+    bl_label = "Batch Restore Original"
+    bl_description = "Restore original source filepath for all selected MOVIE/IMAGE strips"
+    bl_options = {"REGISTER", "UNDO"}
+
+    @classmethod
+    def poll(cls, context):
+        if not context.scene.sequence_editor:
+            return False
+        if get_batch_processor().is_running:
+            return False
+        seq_editor = context.scene.sequence_editor
+        return any(s.select and s.type in {"MOVIE", "IMAGE"} for s in seq_editor.strips)
+
+    def execute(self, context):
+        seq_editor = context.scene.sequence_editor
+        strips = [s for s in seq_editor.strips if s.select and s.type in {"MOVIE", "IMAGE"}]
+
+        restored = 0
+        skipped = 0
+        for strip in strips:
+            original_path = strip.get(KEY_ORIGINAL)
+            if not original_path or not os.path.exists(original_path):
+                skipped += 1
+                continue
+            if strip.get(KEY_MODE, "original") != "original":
+                _set_strip_source(strip, original_path)
+                strip[KEY_MODE] = "original"
+            restored += 1
+
+        self.report(
+            {"INFO"},
+            f"Restored {restored} strip(s)"
+            + (f", skipped {skipped} (no original stored)" if skipped else ""),
+        )
+        return {"FINISHED"}
+
+
+class SEQUENCER_OT_cancel_batch_bake(Operator):
+    """Cancel ongoing batch bake."""
+
+    bl_idname = "sequencer.cancel_batch_bake"
+    bl_label = "Cancel Batch Bake"
+    bl_description = "Cancel the current batch bake process"
+    bl_options = {"REGISTER"}
+
+    def execute(self, context):
+        batch = get_batch_processor()
+        if batch.is_running:
+            batch.cancel()
+            self.report({"INFO"}, "Batch bake cancelled")
+        else:
+            self.report({"WARNING"}, "No batch bake in progress")
+        return {"FINISHED"}
+
+
+classes = [
+    SEQUENCER_OT_batch_bake_selected,
+    SEQUENCER_OT_batch_regenerate_cache,
+    SEQUENCER_OT_batch_restore_original,
+    SEQUENCER_OT_cancel_batch_bake,
+]
+
+
+def register():
+    for cls in classes:
+        bpy.utils.register_class(cls)
+    bpy.types.WindowManager.batch_current = IntProperty(default=0)
+    bpy.types.WindowManager.batch_total = IntProperty(default=0)
+    bpy.types.WindowManager.batch_current_name = StringProperty(default="")
+
+
+def unregister():
+    del bpy.types.WindowManager.batch_current_name
+    del bpy.types.WindowManager.batch_total
+    del bpy.types.WindowManager.batch_current
+    for cls in reversed(classes):
+        bpy.utils.unregister_class(cls)

operators/generate_mask.py

@@ -7,11 +7,99 @@ from video strips in the Video Sequence Editor.
 
 import os
 import bpy
-from bpy.props import IntProperty
+from bpy.props import IntProperty, BoolProperty
 from bpy.types import Operator
 
 from ..core.async_generator import get_generator
-from ..core.utils import get_cache_dir_for_strip
+from ..core.inference_client import get_client
+from ..core.utils import get_cache_dir_for_strip, check_detection_cache
+
+
+def compute_strip_frame_range(strip, scene, client) -> tuple:
+    """(start_frame, end_frame, source_fps) を返す。失敗時は例外を送出。"""
+    video_path = bpy.path.abspath(strip.filepath)
+    video_info = client.get_video_info(video_path)
+    total_video_frames = int(video_info.get("frame_count", 0))
+    source_fps = float(video_info.get("fps", 0.0))
+    if total_video_frames <= 0:
+        raise ValueError(f"Could not read frame count from video: {video_path}")
+    if source_fps <= 0:
+        source_fps = scene.render.fps / scene.render.fps_base
+    project_fps = scene.render.fps / scene.render.fps_base
+    fps_ratio = source_fps / project_fps
+    start_frame = int(round(strip.frame_offset_start * fps_ratio))
+    end_frame = start_frame + int(round(strip.frame_final_duration * fps_ratio)) - 1
+    start_frame = max(0, min(start_frame, total_video_frames - 1))
+    end_frame = max(start_frame, min(end_frame, total_video_frames - 1))
+    return start_frame, end_frame, source_fps
+
+
+def get_image_strip_files(strip) -> tuple:
+    """IMAGE strip の (abs_image_dir, filenames_list) を返す。"""
+    image_dir = bpy.path.abspath(strip.directory)
+    filenames = [elem.filename for elem in strip.elements]
+    return image_dir, filenames
+
+
+def compute_image_strip_range(strip) -> tuple:
+    """IMAGE strip のアクティブ範囲 (start_index, end_index) を返す。"""
+    total_elements = len(strip.elements)
+    start_idx = max(0, int(strip.frame_offset_start))
+    end_idx = start_idx + int(strip.frame_final_duration) - 1
+    start_idx = min(start_idx, total_elements - 1)
+    end_idx = max(start_idx, min(end_idx, total_elements - 1))
+    return start_idx, end_idx
+
+
+def start_mask_gen_for_strip(context, strip, on_complete, on_progress):
+    """Strip のマスク生成を開始する共通処理（MOVIE / IMAGE 両対応）。
+
+    generator.is_running 等のエラー時は例外を送出する。
+    wm.mask_progress / mask_total を初期化してから generator.start*() を呼ぶ。
+    """
+    scene = context.scene
+    wm = context.window_manager
+    generator = get_generator()
+
+    if generator.is_running:
+        raise RuntimeError("Mask generation already in progress")
+
+    output_dir = get_cache_dir_for_strip(strip.name)
+    os.makedirs(output_dir, exist_ok=True)
+    wm.mask_progress = 0
+
+    if strip.type == "IMAGE":
+        image_dir, filenames = get_image_strip_files(strip)
+        if not filenames:
+            raise ValueError("Image strip has no elements")
+        start_idx, end_idx = compute_image_strip_range(strip)
+        wm.mask_total = end_idx - start_idx + 1
+        generator.start_images(
+            image_dir=image_dir,
+            filenames=filenames,
+            output_dir=output_dir,
+            start_index=start_idx,
+            end_index=end_idx,
+            conf_threshold=scene.facemask_conf_threshold,
+            iou_threshold=scene.facemask_iou_threshold,
+            on_complete=on_complete,
+            on_progress=on_progress,
+        )
+    else:
+        client = get_client()
+        start_frame, end_frame, source_fps = compute_strip_frame_range(strip, scene, client)
+        wm.mask_total = end_frame - start_frame + 1
+        generator.start(
+            video_path=bpy.path.abspath(strip.filepath),
+            output_dir=output_dir,
+            start_frame=start_frame,
+            end_frame=end_frame,
+            fps=source_fps,
+            conf_threshold=scene.facemask_conf_threshold,
+            iou_threshold=scene.facemask_iou_threshold,
+            on_complete=on_complete,
+            on_progress=on_progress,
+        )
 
 
 class SEQUENCER_OT_generate_face_mask(Operator):
@@ -22,56 +110,42 @@ class SEQUENCER_OT_generate_face_mask(Operator):
     bl_description = "Detect faces and generate mask image sequence"
     bl_options = {'REGISTER', 'UNDO'}
 
+    force: BoolProperty(
+        name="Force Regenerate",
+        description="既存のキャッシュを無視して再生成する",
+        default=False,
+    )
+
     @classmethod
     def poll(cls, context):
-        """Check if operator can run."""
         if not context.scene.sequence_editor:
             return False
-        
         strip = context.scene.sequence_editor.active_strip
         if not strip:
             return False
-        
         return strip.type in {'MOVIE', 'IMAGE'}
 
     def execute(self, context):
         strip = context.scene.sequence_editor.active_strip
-        scene = context.scene
 
-        # Get video path
+        # ファイル存在確認
         if strip.type == 'MOVIE':
             video_path = bpy.path.abspath(strip.filepath)
         else:
-            # Image sequence - get directory
             video_path = bpy.path.abspath(strip.directory)
 
         if not os.path.exists(video_path):
             self.report({'ERROR'}, f"Video file not found: {video_path}")
             return {'CANCELLED'}
 
-        # Determine output directory
+        # キャッシュ確認（force=True の場合はスキップ）
-        output_dir = self._get_cache_dir(context, strip)
+        if not self.force and check_detection_cache(strip.name):
-        
+            self.report({'INFO'}, f"Using cached detections for {strip.name}")
-        # 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 detections from {output_dir}")
             return {'FINISHED'}
 
-        # Get frame range
-        start_frame = strip.frame_final_start
-        end_frame = strip.frame_final_end
-        fps = scene.render.fps / scene.render.fps_base
-        
-        # Start async generation
         generator = get_generator()
 
-        if generator.is_running:
-            self.report({'WARNING'}, "Mask generation already in progress")
-            return {'CANCELLED'}
-        
         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":
@@ -91,85 +165,25 @@ class SEQUENCER_OT_generate_face_mask(Operator):
                     area.tag_redraw()
 
         def on_progress(current, total):
-            """Called on progress updates."""
-            # Update window manager properties for UI
             wm = context.window_manager
             wm.mask_progress = current
             wm.mask_total = total
-            
-            # Force UI redraw
             for area in context.screen.areas:
                 if area.type == 'SEQUENCE_EDITOR':
                     area.tag_redraw()
 
-        # Initialize progress
+        try:
-        wm = context.window_manager
+            start_mask_gen_for_strip(context, strip, on_complete, on_progress)
-        wm.mask_progress = 0
+        except RuntimeError as e:
-        wm.mask_total = end_frame - start_frame + 1
+            self.report({'WARNING'}, str(e))
-        
+            return {'CANCELLED'}
-        # Get parameters from scene properties
+        except Exception as e:
-        conf_threshold = scene.facemask_conf_threshold
+            self.report({'ERROR'}, f"Failed to start mask generation: {e}")
-        iou_threshold = scene.facemask_iou_threshold
+            return {'CANCELLED'}
-        mask_scale = scene.facemask_mask_scale
-
-        # Start generation
-        generator.start(
-            video_path=video_path,
-            output_dir=output_dir,
-            start_frame=0,  # Frame indices in video
-            end_frame=end_frame - start_frame,
-            fps=fps,
-            conf_threshold=conf_threshold,
-            iou_threshold=iou_threshold,
-            mask_scale=mask_scale,
-            on_complete=on_complete,
-            on_progress=on_progress,
-        )
 
         self.report({'INFO'}, f"Started mask generation for {strip.name}")
         return {'FINISHED'}
 
-    def _get_cache_dir(self, context, strip) -> str:
-        """Get or create cache directory for mask images."""
-        cache_dir = get_cache_dir_for_strip(strip.name)
-        os.makedirs(cache_dir, exist_ok=True)
-        return cache_dir
-    
-    def _check_cache(self, cache_dir: str, expected_frames: int) -> bool:
-        """Check if cached masks exist and are complete.
-
-        Args:
-            cache_dir: Path to cache directory
-            expected_frames: Number of frames expected
-
-        Returns:
-            True if cache exists and is valid
-        """
-        if not os.path.exists(cache_dir):
-            return False
-
-        detections_path = os.path.join(cache_dir, "detections.msgpack")
-        if not os.path.exists(detections_path):
-            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:
-            return True
-
 
 class SEQUENCER_OT_cancel_mask_generation(Operator):
     """Cancel ongoing mask generation."""
@@ -191,10 +205,88 @@ class SEQUENCER_OT_cancel_mask_generation(Operator):
         return {'FINISHED'}
 
 
+class SEQUENCER_OT_augment_pose_mask(Operator):
+    """Add pose-based head detections to existing detection cache."""
+
+    bl_idname = "sequencer.augment_pose_mask"
+    bl_label = "Augment with Pose"
+    bl_description = "Run pose estimation and merge results into existing detection cache"
+    bl_options = {'REGISTER'}
+
+    @classmethod
+    def poll(cls, context):
+        if not context.scene.sequence_editor:
+            return False
+        strip = context.scene.sequence_editor.active_strip
+        if not strip or strip.type != 'MOVIE':
+            return False
+        return check_detection_cache(strip.name)
+
+    def execute(self, context):
+        strip = context.scene.sequence_editor.active_strip
+        output_dir = get_cache_dir_for_strip(strip.name)
+        detections_path = os.path.join(output_dir, "detections.msgpack")
+
+        if not os.path.exists(detections_path):
+            self.report({'ERROR'}, f"Detection cache not found: {detections_path}")
+            return {'CANCELLED'}
+
+        generator = get_generator()
+        scene = context.scene
+        wm = context.window_manager
+        wm.mask_progress = 0
+        wm.mask_total = 0  # サーバー側から実際の値に更新される
+
+        def on_complete(status, data):
+            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":
+                print(f"[FaceMask] Pose augmentation completed: {data}")
+            elif status == "error":
+                print(f"[FaceMask] Error: {data}")
+            elif status == "cancelled":
+                print("[FaceMask] Pose augmentation cancelled")
+
+            for area in context.screen.areas:
+                if area.type == 'SEQUENCE_EDITOR':
+                    area.tag_redraw()
+
+        def on_progress(current, total_f):
+            wm.mask_progress = current
+            wm.mask_total = total_f
+            for area in context.screen.areas:
+                if area.type == 'SEQUENCE_EDITOR':
+                    area.tag_redraw()
+
+        try:
+            generator.start_augment_pose(
+                detections_path=detections_path,
+                total_frames=0,
+                conf_threshold=scene.facemask_conf_threshold,
+                iou_threshold=scene.facemask_iou_threshold,
+                on_complete=on_complete,
+                on_progress=on_progress,
+            )
+        except RuntimeError as e:
+            self.report({'WARNING'}, str(e))
+            return {'CANCELLED'}
+        except Exception as e:
+            self.report({'ERROR'}, f"Failed to start pose augmentation: {e}")
+            return {'CANCELLED'}
+
+        self.report({'INFO'}, f"Started pose augmentation for {strip.name}")
+        return {'FINISHED'}
+
+
 # Registration
 classes = [
     SEQUENCER_OT_generate_face_mask,
     SEQUENCER_OT_cancel_mask_generation,
+    SEQUENCER_OT_augment_pose_mask,
 ]
 
 
@@ -202,13 +294,11 @@ def register():
     for cls in classes:
         bpy.utils.register_class(cls)
 
-    # Add progress properties to window manager
     bpy.types.WindowManager.mask_progress = IntProperty(default=0)
     bpy.types.WindowManager.mask_total = IntProperty(default=0)
 
 
 def unregister():
-    # Remove properties
     del bpy.types.WindowManager.mask_progress
     del bpy.types.WindowManager.mask_total
 

panels/vse_panel.py

@@ -11,11 +11,12 @@ from bpy.types import Panel
 
 from ..core.async_bake_generator import get_bake_generator
 from ..core.async_generator import get_generator
+from ..core.batch_processor import get_batch_processor
 from ..core.utils import (
     get_server_status,
     get_cache_info,
     format_size,
-    get_detections_path_for_strip,
+    check_detection_cache,
 )
 
 
@@ -35,9 +36,15 @@ class SEQUENCER_PT_face_mask(Panel):
         seq_editor = context.scene.sequence_editor
         # Note: Blender 5.0 uses 'strips' instead of 'sequences'
 
+        batch = get_batch_processor()
         generator = get_generator()
         bake_generator = get_bake_generator()
 
+        # Batch progress (highest priority)
+        if batch.is_running:
+            self._draw_batch_progress(layout, wm, batch, generator, bake_generator)
+            return
+
         # Show progress if generating masks
         if generator.is_running:
             self._draw_progress(layout, wm, generator)
@@ -65,6 +72,7 @@ class SEQUENCER_PT_face_mask(Panel):
         self._draw_parameters(layout, scene)
         self._draw_server_status(layout)
         self._draw_cache_info(layout, context, seq_editor)
+        self._draw_batch_controls(layout, context, seq_editor)
     
     def _draw_parameters(self, layout, scene):
         """Draw detection parameters."""
@@ -74,7 +82,6 @@ class SEQUENCER_PT_face_mask(Panel):
         col = box.column(align=True)
         col.prop(scene, "facemask_conf_threshold")
         col.prop(scene, "facemask_iou_threshold")
-        col.prop(scene, "facemask_mask_scale")
 
     def _draw_server_status(self, layout):
         """Draw server status and GPU info."""
@@ -183,6 +190,76 @@ class SEQUENCER_PT_face_mask(Panel):
             icon='CANCEL',
         )
 
+    def _draw_batch_progress(self, layout, wm, batch, generator, bake_generator):
+        """Draw batch bake progress."""
+        box = layout.box()
+        if batch._mode == "mask_only":
+            box.label(text="Batch Generating Cache...", icon='RENDER_ANIMATION')
+        else:
+            box.label(text="Batch Baking...", icon='RENDER_ANIMATION')
+
+        # Overall progress
+        total = max(wm.batch_total, 1)
+        # Show n-1/total while current strip is in progress, n/total when moving to next
+        done_count = max(wm.batch_current - 1, 0)
+        overall_factor = done_count / total
+        box.progress(
+            factor=overall_factor,
+            text=f"{wm.batch_current} / {wm.batch_total}",
+        )
+
+        if wm.batch_current_name:
+            box.label(text=f"Strip: {wm.batch_current_name}")
+
+        # Inner progress (mask gen or bake)
+        if generator.is_running:
+            inner = wm.mask_progress / max(wm.mask_total, 1)
+            box.progress(
+                factor=inner,
+                text=f"Detecting: {wm.mask_progress} / {wm.mask_total}",
+            )
+        elif bake_generator.is_running:
+            inner = wm.bake_progress / max(wm.bake_total, 1)
+            box.progress(
+                factor=inner,
+                text=f"Baking: {wm.bake_progress} / {wm.bake_total}",
+            )
+
+        box.operator(
+            "sequencer.cancel_batch_bake",
+            text="Cancel Batch",
+            icon='CANCEL',
+        )
+
+    def _draw_batch_controls(self, layout, context, seq_editor):
+        """Draw batch bake button when multiple MOVIE/IMAGE strips are selected."""
+        if not seq_editor:
+            return
+        selected_movies = [s for s in seq_editor.strips if s.select and s.type in {"MOVIE", "IMAGE"}]
+        if not selected_movies:
+            return
+        count = len(selected_movies)
+        image_count = sum(1 for s in selected_movies if s.type == "IMAGE")
+        video_count = sum(1 for s in selected_movies if s.type == "MOVIE")
+        label = f"Batch ({count} selected, image: {image_count}, video: {video_count})"
+        box = layout.box()
+        box.label(text=label, icon='RENDER_ANIMATION')
+        box.operator(
+            "sequencer.batch_bake_selected",
+            text="Batch Bake Selected",
+            icon='RENDER_ANIMATION',
+        )
+        box.operator(
+            "sequencer.batch_regenerate_cache",
+            text="Batch Regenerate Cache",
+            icon='FILE_REFRESH',
+        )
+        box.operator(
+            "sequencer.batch_restore_original",
+            text="Batch Restore Original",
+            icon='LOOP_BACK',
+        )
+
     def _draw_generation_controls(self, layout, context, strip):
         """Draw mask generation controls."""
         box = layout.box()
@@ -192,31 +269,39 @@ class SEQUENCER_PT_face_mask(Panel):
         row = box.row()
         row.label(text=f"Strip: {strip.name}")
         
-        detections_path = get_detections_path_for_strip(strip.name)
+        has_mask = check_detection_cache(strip.name)
-        has_mask = bpy.path.abspath(detections_path) and os.path.exists(
-            bpy.path.abspath(detections_path)
-        )
 
         if has_mask:
             row = box.row()
             row.label(text="✓ Detection cache exists", icon='CHECKMARK')
         
-        # Generate button
+        # Generate / Regenerate button
+        if not has_mask:
             box.operator(
                 "sequencer.generate_face_mask",
-            text="Generate Detection Cache" if not has_mask else "Regenerate Cache",
+                text="Generate Detection Cache",
                 icon='FACE_MAPS',
             )
+        else:
+            op = box.operator(
+                "sequencer.generate_face_mask",
+                text="Regenerate Cache",
+                icon='FILE_REFRESH',
+            )
+            op.force = True
+            if strip.type == 'MOVIE':
+                box.operator(
+                    "sequencer.augment_pose_mask",
+                    text="Augment with Pose",
+                    icon='MOD_ARMATURE',
+                )
     
     def _draw_blur_controls(self, layout, context, strip):
         """Draw blur application controls."""
         box = layout.box()
         box.label(text="Blur Bake", icon='MATFLUID')
 
-        detections_path = get_detections_path_for_strip(strip.name)
+        has_mask = check_detection_cache(strip.name)
-        has_mask = bpy.path.abspath(detections_path) and os.path.exists(
-            bpy.path.abspath(detections_path)
-        )
 
         if not has_mask:
             box.label(text="Generate detection cache first", icon='INFO')
@@ -225,26 +310,43 @@ class SEQUENCER_PT_face_mask(Panel):
         # Bake parameters
         col = box.column(align=True)
         col.prop(context.scene, "facemask_bake_blur_size")
+        col.prop(context.scene, "facemask_bake_display_scale")
+        if strip.type == "MOVIE":
             col.prop(context.scene, "facemask_bake_format")
 
-        # Source status
+        box.separator()
-        source_mode = strip.get("facemask_source_mode", "original")
-        if source_mode == "baked":
-            box.label(text="Source: Baked", icon='CHECKMARK')
-        else:
-            box.label(text="Source: Original", icon='FILE_MOVIE')
 
-        # Bake and restore buttons
+        baked_path = strip.get("facemask_baked_filepath", "")
+        has_baked = bool(baked_path and os.path.exists(bpy.path.abspath(baked_path)))
+        source_mode = strip.get("facemask_source_mode", "original")
+
+        if not has_baked:
+            # 初回: Bakeのみ
             box.operator(
                 "sequencer.bake_and_swap_blur_source",
-            text="Bake & Swap Source",
+                text="Bake",
                 icon='RENDER_STILL',
             )
-        box.operator(
+        else:
+            # Bake済み: ソース切り替え + Re-bake
+            row = box.row(align=True)
+            if source_mode == "baked":
+                row.operator(
                     "sequencer.restore_original_source",
-            text="Restore Original Source",
+                    text="Restore Original",
                     icon='LOOP_BACK',
                 )
+            else:
+                row.operator(
+                    "sequencer.swap_to_baked_blur",
+                    text="Swap to Baked",
+                    icon='PLAY',
+                )
+            row.operator(
+                "sequencer.force_rebake_blur",
+                text="Re-bake",
+                icon='FILE_REFRESH',
+            )
 
 
 # Registration

server/detector.py

@@ -1,28 +1,36 @@
 """
-YOLOv8 Face Detector using PyTorch with ROCm support.
+YOLOv8 Head Detector using CrowdHuman-trained model with PyTorch ROCm support.
 
-This module provides high-performance face detection using
+Directly detects human heads (frontal, profile, rear) using the Owen718
-YOLOv8-face model with AMD GPU (ROCm) acceleration.
+CrowdHuman YOLOv8 model, which was trained on dense crowd scenes.
 """
 
 import os
 from typing import List, Tuple, Optional
-from pathlib import Path
 import numpy as np
 
 
-class YOLOFaceDetector:
+def _download_model(dest_path: str):
-    """
+    """モデルが存在しない場合に手動ダウンロード手順を表示して例外を送出する。"""
-    YOLOv8 face detector with PyTorch ROCm support.
+    gdrive_id = "1qlBmiEU4GBV13fxPhLZqjhjBbREvs8-m"
+    raise RuntimeError(
+        f"モデルファイルが見つかりません: {dest_path}\n"
+        "以下の手順でダウンロードしてください:\n"
+        f"  1. https://drive.google.com/file/d/{gdrive_id} を開く\n"
+        f"  2. ダウンロードしたファイルを {dest_path} に配置する"
+    )
 
-    Features:
+
-    - ROCm GPU acceleration for AMD GPUs
+class YOLOHeadDetector:
-    - High accuracy face detection
+    """
-    - Automatic NMS for overlapping detections
+    Head detector using CrowdHuman-trained YOLOv8 model with PyTorch ROCm support.
+
+    Directly detects heads (class 0: head) without pose estimation,
+    enabling robust detection of rear-facing, side-facing, and partially
+    visible people in dense crowd scenes.
     """
 
-    # Default model path relative to this file
+    DEFAULT_MODEL = os.path.join("models", "crowdhuman_yolov8_head.pt")
-    DEFAULT_MODEL = "yolov8n-face-lindevs.pt"
 
     def __init__(
         self,
@@ -31,15 +39,6 @@ class YOLOFaceDetector:
         iou_threshold: float = 0.45,
         input_size: Tuple[int, int] = (640, 640),
     ):
-        """
-        Initialize the YOLO face detector.
-
-        Args:
-            model_path: Path to PyTorch model file. If None, uses default model.
-            conf_threshold: Confidence threshold for detections
-            iou_threshold: IoU threshold for NMS
-            input_size: Model input size (width, height)
-        """
         self.conf_threshold = conf_threshold
         self.iou_threshold = iou_threshold
         self.input_size = input_size
@@ -49,23 +48,20 @@ class YOLOFaceDetector:
 
     @property
     def model(self):
-        """Lazy-load YOLO model."""
+        """Lazy-load YOLO head detection model."""
         if self._model is None:
             from ultralytics import YOLO
             import torch
 
-            # Determine model path
+            if self._model_path is not None:
-            if self._model_path is None:
+                if not os.path.exists(self._model_path):
-                # Assuming models are in ../models relative to server/detector.py
+                    raise FileNotFoundError(f"Model not found: {self._model_path}")
-                models_dir = Path(__file__).parent.parent / "models"
-                model_path = str(models_dir / self.DEFAULT_MODEL)
-            else:
                 model_path = self._model_path
-
+            else:
+                model_path = self.DEFAULT_MODEL
                 if not os.path.exists(model_path):
-                raise FileNotFoundError(f"Model not found: {model_path}")
+                    _download_model(model_path)
 
-            # Detect device (ROCm GPU or CPU)
             if torch.cuda.is_available():
                 self._device = 'cuda'
                 device_name = torch.cuda.get_device_name(0)
@@ -74,25 +70,32 @@ class YOLOFaceDetector:
                 self._device = 'cpu'
                 print("[FaceMask] Using CPU for inference (ROCm GPU not available)")
 
-            # Load model (let Ultralytics handle device management)
             try:
                 self._model = YOLO(model_path)
-                # Don't call .to() - let predict() handle device assignment
+                print(f"[FaceMask] Head detection model loaded: {model_path}")
-                print(f"[FaceMask] Model loaded, will use device: {self._device}")
+                print(f"[FaceMask] Device: {self._device}")
             except Exception as e:
                 print(f"[FaceMask] Error loading model: {e}")
                 import traceback
                 traceback.print_exc()
                 raise
 
-            print(f"[FaceMask] YOLO model loaded: {model_path}")
-            print(f"[FaceMask] Device: {self._device}")
-
         return self._model
 
+    def _results_to_detections(self, result) -> List[Tuple[int, int, int, int, float]]:
+        """Convert a single YOLO result to (x, y, w, h, conf) tuples."""
+        if result.boxes is None:
+            return []
+        detections = []
+        for box in result.boxes:
+            conf = float(box.conf[0].cpu().numpy())
+            x1, y1, x2, y2 = box.xyxy[0].cpu().numpy()
+            detections.append((int(x1), int(y1), int(x2 - x1), int(y2 - y1), conf))
+        return detections
+
     def detect(self, frame: np.ndarray) -> List[Tuple[int, int, int, int, float]]:
         """
-        Detect faces in a frame.
+        Detect heads in a frame.
 
         Args:
             frame: BGR image as numpy array (H, W, C)
@@ -100,7 +103,6 @@ class YOLOFaceDetector:
         Returns:
             List of detections as (x, y, width, height, confidence)
         """
-        # Run inference
         import torch
         print(f"[FaceMask] Inference device: {self._device}, CUDA available: {torch.cuda.is_available()}")
         try:
@@ -116,7 +118,6 @@ class YOLOFaceDetector:
             print(f"[FaceMask] ERROR during inference: {e}")
             import traceback
             traceback.print_exc()
-            # Fallback to CPU
             print("[FaceMask] Falling back to CPU inference...")
             self._device = 'cpu'
             results = self.model.predict(
@@ -128,28 +129,13 @@ class YOLOFaceDetector:
                 device='cpu',
             )
 
-        # Extract detections
+        if results:
-        detections = []
+            return self._results_to_detections(results[0])
-        if len(results) > 0 and results[0].boxes is not None:
+        return []
-            boxes = results[0].boxes
-            for box in boxes:
-                # Get coordinates in xyxy format
-                x1, y1, x2, y2 = box.xyxy[0].cpu().numpy()
-                conf = float(box.conf[0].cpu().numpy())
-
-                # Convert to x, y, width, height
-                x = int(x1)
-                y = int(y1)
-                w = int(x2 - x1)
-                h = int(y2 - y1)
-
-                detections.append((x, y, w, h, conf))
-
-        return detections
 
     def detect_batch(self, frames: List[np.ndarray]) -> List[List[Tuple[int, int, int, int, float]]]:
         """
-        Detect faces in multiple frames at once (batch processing).
+        Detect heads in multiple frames at once (batch processing).
 
         Args:
             frames: List of BGR images as numpy arrays (H, W, C)
@@ -161,7 +147,6 @@ class YOLOFaceDetector:
         if not frames:
             return []
 
-        # Run batch inference
         try:
             results = self.model.predict(
                 frames,
@@ -175,7 +160,6 @@ class YOLOFaceDetector:
             print(f"[FaceMask] ERROR during batch inference: {e}")
             import traceback
             traceback.print_exc()
-            # Fallback to CPU
             print("[FaceMask] Falling back to CPU inference...")
             self._device = 'cpu'
             results = self.model.predict(
@@ -187,28 +171,7 @@ class YOLOFaceDetector:
                 device='cpu',
             )
 
-        # Extract detections for each frame
+        return [self._results_to_detections(r) for r in results]
-        all_detections = []
-        for result in results:
-            detections = []
-            if result.boxes is not None:
-                boxes = result.boxes
-                for box in boxes:
-                    # Get coordinates in xyxy format
-                    x1, y1, x2, y2 = box.xyxy[0].cpu().numpy()
-                    conf = float(box.conf[0].cpu().numpy())
-
-                    # Convert to x, y, width, height
-                    x = int(x1)
-                    y = int(y1)
-                    w = int(x2 - x1)
-                    h = int(y2 - y1)
-
-                    detections.append((x, y, w, h, conf))
-
-            all_detections.append(detections)
-
-        return all_detections
 
     def generate_mask(
         self,
@@ -218,11 +181,11 @@ class YOLOFaceDetector:
         feather_radius: int = 20,
     ) -> np.ndarray:
         """
-        Generate a mask image from face detections.
+        Generate a mask image from head detections.
 
         Args:
             frame_shape: Shape of the original frame (height, width, channels)
-            detections: List of face detections (x, y, w, h, conf)
+            detections: List of head detections (x, y, w, h, conf)
             mask_scale: Scale factor for mask region
             feather_radius: Radius for edge feathering
 
@@ -235,25 +198,19 @@ class YOLOFaceDetector:
         mask = np.zeros((height, width), dtype=np.uint8)
 
         for (x, y, w, h, conf) in detections:
-            # Scale the bounding box
             center_x = x + w // 2
             center_y = y + h // 2
-
             scaled_w = int(w * mask_scale)
             scaled_h = int(h * mask_scale)
 
-            # Draw ellipse for natural face shape
             cv2.ellipse(
                 mask,
                 (center_x, center_y),
                 (scaled_w // 2, scaled_h // 2),
-                0,  # angle
+                0, 0, 360,
-                0, 360,  # arc
+                255, -1,
-                255,  # color (white)
-                -1,  # filled
             )
 
-        # Apply Gaussian blur for feathering
         if feather_radius > 0 and len(detections) > 0:
             kernel_size = feather_radius * 2 + 1
             mask = cv2.GaussianBlur(mask, (kernel_size, kernel_size), 0)
@@ -262,12 +219,214 @@ class YOLOFaceDetector:
 
 
 # Singleton instance
-_detector: Optional[YOLOFaceDetector] = None
+_detector: Optional[YOLOHeadDetector] = None
 
 
-def get_detector(**kwargs) -> YOLOFaceDetector:
+def get_detector(**kwargs) -> YOLOHeadDetector:
-    """Get or create the global YOLO detector instance."""
+    """Get or create the global YOLO head detector instance."""
     global _detector
     if _detector is None:
-        _detector = YOLOFaceDetector(**kwargs)
+        _detector = YOLOHeadDetector(**kwargs)
     return _detector
+
+
+# ---------------------------------------------------------------------------
+# Pose-based head detector (YOLOv8 pose estimation)
+# ---------------------------------------------------------------------------
+
+# COCO pose keypoint indices
+_HEAD_KP = [0, 1, 2, 3, 4]      # nose, left_eye, right_eye, left_ear, right_ear
+_SHOULDER_KP = [5, 6]            # left_shoulder, right_shoulder
+_KP_CONF_THRESH = 0.3
+
+
+def _head_bbox_from_pose(
+    kp_xy: np.ndarray,
+    kp_conf: np.ndarray,
+    person_x1: float,
+    person_y1: float,
+    person_x2: float,
+    person_y2: float,
+) -> Tuple[int, int, int, int]:
+    """
+    Estimate head bounding box (x, y, w, h) from COCO pose keypoints.
+
+    Strategy:
+    1. Use head keypoints (0-4: nose, eyes, ears) if visible.
+    2. Fall back to shoulder keypoints (5-6) to infer head position.
+    3. Last resort: use top of the person bounding box.
+    """
+    person_w = max(person_x2 - person_x1, 1.0)
+
+    # --- Step 1: head keypoints ---
+    visible_head = [
+        (float(kp_xy[i][0]), float(kp_xy[i][1]))
+        for i in _HEAD_KP
+        if float(kp_conf[i]) > _KP_CONF_THRESH
+    ]
+    if visible_head:
+        xs = [p[0] for p in visible_head]
+        ys = [p[1] for p in visible_head]
+        kp_x1, kp_y1 = min(xs), min(ys)
+        kp_x2, kp_y2 = max(xs), max(ys)
+        span = max(kp_x2 - kp_x1, kp_y2 - kp_y1, 1.0)
+        cx = (kp_x1 + kp_x2) / 2.0
+        cy = (kp_y1 + kp_y2) / 2.0
+        r = max(span * 0.5, person_w * 0.10)
+        x1 = int(cx - r)
+        y1 = int(cy - r)
+        x2 = int(cx + r)
+        y2 = int(cy + r)
+        return x1, y1, x2 - x1, y2 - y1
+
+    # --- Step 2: shoulder keypoints ---
+    visible_shoulder = [
+        (float(kp_xy[i][0]), float(kp_xy[i][1]))
+        for i in _SHOULDER_KP
+        if float(kp_conf[i]) > _KP_CONF_THRESH
+    ]
+    if visible_shoulder:
+        cx = sum(p[0] for p in visible_shoulder) / len(visible_shoulder)
+        cy_sh = sum(p[1] for p in visible_shoulder) / len(visible_shoulder)
+        if len(visible_shoulder) == 2:
+            sh_width = abs(visible_shoulder[1][0] - visible_shoulder[0][0])
+        else:
+            sh_width = person_w * 0.5
+        r = max(sh_width * 0.3, person_w * 0.12)
+        cy = cy_sh - r * 1.3
+        x1 = int(cx - r)
+        y1 = int(cy - r)
+        x2 = int(cx + r)
+        y2 = int(cy + r)
+        return x1, y1, x2 - x1, y2 - y1
+
+    # --- Step 3: person bbox top ---
+    r = max(person_w * 0.15, 20.0)
+    cx = (person_x1 + person_x2) / 2.0
+    x1 = int(cx - r)
+    y1 = int(person_y1)
+    x2 = int(cx + r)
+    y2 = int(person_y1 + r * 2.0)
+    return x1, y1, x2 - x1, y2 - y1
+
+
+class YOLOPoseHeadDetector:
+    """
+    Head detector using YOLOv8 pose estimation with PyTorch ROCm support.
+
+    Extracts head bounding boxes from COCO pose keypoints (nose, eyes, ears).
+    yolov8l-pose.pt is auto-downloaded by Ultralytics on first use.
+    """
+
+    DEFAULT_MODEL = os.path.join("models", "yolov8l-pose.pt")
+
+    def __init__(
+        self,
+        model_path: Optional[str] = None,
+        conf_threshold: float = 0.25,
+        iou_threshold: float = 0.45,
+        input_size: Tuple[int, int] = (640, 640),
+    ):
+        self.conf_threshold = conf_threshold
+        self.iou_threshold = iou_threshold
+        self.input_size = input_size
+        self._model = None
+        self._model_path = model_path
+        self._device = None
+
+    @property
+    def model(self):
+        """Lazy-load YOLO pose model."""
+        if self._model is None:
+            from ultralytics import YOLO
+            import torch
+
+            model_path = self._model_path if self._model_path is not None else self.DEFAULT_MODEL
+
+            if torch.cuda.is_available():
+                self._device = 'cuda'
+                device_name = torch.cuda.get_device_name(0)
+                print(f"[FaceMask] Using ROCm GPU for pose inference: {device_name}")
+            else:
+                self._device = 'cpu'
+                print("[FaceMask] Using CPU for pose inference (ROCm GPU not available)")
+
+            try:
+                self._model = YOLO(model_path)
+                print(f"[FaceMask] Pose model loaded: {model_path}")
+                print(f"[FaceMask] Device: {self._device}")
+            except Exception as e:
+                print(f"[FaceMask] Error loading pose model: {e}")
+                import traceback
+                traceback.print_exc()
+                raise
+
+        return self._model
+
+    def _results_to_detections(self, result) -> List[Tuple[int, int, int, int, float]]:
+        """Convert a single YOLO pose result to (x, y, w, h, conf) tuples."""
+        detections = []
+        if result.boxes is None or result.keypoints is None:
+            return detections
+
+        boxes = result.boxes
+        keypoints = result.keypoints
+
+        for i, box in enumerate(boxes):
+            conf = float(box.conf[0].cpu().numpy())
+            x1, y1, x2, y2 = box.xyxy[0].cpu().numpy()
+
+            kp_data = keypoints.data[i].cpu().numpy()  # shape (17, 3): x, y, conf
+            kp_xy = kp_data[:, :2]
+            kp_conf = kp_data[:, 2]
+
+            hx, hy, hw, hh = _head_bbox_from_pose(
+                kp_xy, kp_conf,
+                float(x1), float(y1), float(x2), float(y2),
+            )
+            detections.append((hx, hy, hw, hh, conf))
+
+        return detections
+
+    def detect_batch(self, frames: List[np.ndarray]) -> List[List[Tuple[int, int, int, int, float]]]:
+        """Detect heads in multiple frames at once (batch processing)."""
+        if not frames:
+            return []
+
+        try:
+            results = self.model.predict(
+                frames,
+                conf=self.conf_threshold,
+                iou=self.iou_threshold,
+                imgsz=self.input_size[0],
+                verbose=False,
+                device=self._device,
+            )
+        except Exception as e:
+            print(f"[FaceMask] ERROR during pose batch inference: {e}")
+            import traceback
+            traceback.print_exc()
+            print("[FaceMask] Falling back to CPU inference...")
+            self._device = 'cpu'
+            results = self.model.predict(
+                frames,
+                conf=self.conf_threshold,
+                iou=self.iou_threshold,
+                imgsz=self.input_size[0],
+                verbose=False,
+                device='cpu',
+            )
+
+        return [self._results_to_detections(r) for r in results]
+
+
+# Pose detector singleton
+_pose_detector: Optional[YOLOPoseHeadDetector] = None
+
+
+def get_pose_detector(**kwargs) -> YOLOPoseHeadDetector:
+    """Get or create the global YOLO pose head detector instance."""
+    global _pose_detector
+    if _pose_detector is None:
+        _pose_detector = YOLOPoseHeadDetector(**kwargs)
+    return _pose_detector

server/main.py

@@ -31,28 +31,62 @@ def fix_library_path():
 # Fix library path BEFORE any other imports
 fix_library_path()
 
-import queue
+import queue  # noqa: E402
-import threading
+import threading  # noqa: E402
-import uuid
+import uuid  # noqa: E402
-import traceback
+import traceback  # noqa: E402
-import subprocess
+import subprocess  # noqa: E402
-from typing import Dict, Optional, List
+from typing import Dict, Optional, List  # noqa: E402
-from pathlib import Path
+from pathlib import Path  # noqa: E402
 
-from fastapi import FastAPI, HTTPException, BackgroundTasks
+from fastapi import FastAPI, HTTPException, BackgroundTasks  # noqa: E402
-from pydantic import BaseModel
+from pydantic import BaseModel  # noqa: E402
-import uvicorn
+import uvicorn  # noqa: E402
-import cv2
+import cv2  # noqa: E402
-import numpy as np
+import numpy as np  # noqa: E402
-import msgpack
+import msgpack  # noqa: E402
 
 # Add project root to path for imports if needed
 sys.path.append(str(Path(__file__).parent.parent))
 
-from server.detector import get_detector
+from server.detector import get_detector, get_pose_detector  # noqa: E402
 
 app = FastAPI(title="Face Mask Inference Server")
 
+
+def _get_r_frame_rate(video_path: str) -> tuple:
+    """ffprobe でコンテナ宣言の r_frame_rate を取得する。
+
+    Returns:
+        (fps_float, fps_str): fps_str は "120/1" のような分数文字列。
+        取得失敗時は (0.0, "")。
+    """
+    try:
+        result = subprocess.run(
+            [
+                "ffprobe", "-v", "error",
+                "-select_streams", "v:0",
+                "-show_entries", "stream=r_frame_rate",
+                "-of", "default=noprint_wrappers=1:nokey=1",
+                video_path,
+            ],
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+        if result.returncode == 0:
+            rate_str = result.stdout.strip()
+            if "/" in rate_str:
+                num, den = rate_str.split("/")
+                fps_float = float(num) / float(den)
+            else:
+                fps_float = float(rate_str)
+                rate_str = str(fps_float)
+            return fps_float, rate_str
+    except Exception:
+        pass
+    return 0.0, ""
+
 # GPU status cache
 _gpu_status_cache = None
 
@@ -83,7 +117,10 @@ class GenerateRequest(BaseModel):
     end_frame: int
     conf_threshold: float = 0.5
     iou_threshold: float = 0.45
-    mask_scale: float = 1.5
+
+
+class VideoInfoRequest(BaseModel):
+    video_path: str
 
 
 class BakeRequest(BaseModel):
@@ -91,9 +128,35 @@ class BakeRequest(BaseModel):
     detections_path: str
     output_path: str
     blur_size: int = 50
+    display_scale: float = 1.0
     format: str = "mp4"
 
 
+class GenerateImagesRequest(BaseModel):
+    image_dir: str
+    filenames: List[str]
+    output_dir: str
+    start_index: int = 0
+    end_index: int = -1
+    conf_threshold: float = 0.5
+    iou_threshold: float = 0.45
+
+
+class AugmentPoseRequest(BaseModel):
+    detections_path: str
+    conf_threshold: float = 0.5
+    iou_threshold: float = 0.45
+
+
+class BakeImagesRequest(BaseModel):
+    image_dir: str
+    filenames: List[str]
+    output_dir: str
+    detections_path: str
+    blur_size: int = 50
+    display_scale: float = 1.0
+
+
 class _FFmpegPipeWriter:
     """Write BGR frames to ffmpeg stdin."""
 
@@ -138,8 +201,33 @@ def _build_ffmpeg_vaapi_writer(
     fps: float,
     width: int,
     height: int,
+    out_fps_str: str = "",
 ) -> _FFmpegPipeWriter:
-    """Create ffmpeg h264_vaapi writer with QP=24 (balanced quality/speed)."""
+    """Create ffmpeg h264_vaapi writer with QP=24 (balanced quality/speed).
+
+    fps: ソース動画の avg_frame_rate（rawパイプの入力レート）
+    out_fps_str: 出力コンテナに宣言する r_frame_rate（"120/1" 等）。
+                 ソースと異なる場合は fps フィルタでフレームを補完する。
+    """
+    # ソースの avg_fps と出力の r_fps が有意に異なる場合のみ fps フィルタを挿入
+    needs_fps_filter = bool(out_fps_str)
+    if needs_fps_filter:
+        try:
+            if "/" in out_fps_str:
+                num, den = out_fps_str.split("/")
+                out_fps_float = float(num) / float(den)
+            else:
+                out_fps_float = float(out_fps_str)
+            needs_fps_filter = abs(out_fps_float - fps) > 0.01
+        except ValueError:
+            needs_fps_filter = False
+
+    if needs_fps_filter:
+        vf = f"format=nv12,fps={out_fps_str},hwupload"
+        print(f"[FaceMask] fps filter: {fps:.3f} -> {out_fps_str}")
+    else:
+        vf = "format=nv12,hwupload"
+
     cmd = [
         "ffmpeg",
         "-hide_banner",
@@ -160,7 +248,7 @@ def _build_ffmpeg_vaapi_writer(
         "-",
         "-an",
         "-vf",
-        "format=nv12,hwupload",
+        vf,
         "-c:v",
         "h264_vaapi",
         "-qp",
@@ -176,13 +264,14 @@ def _build_video_writer(
     fps: float,
     width: int,
     height: int,
+    out_fps_str: str = "",
 ) -> object:
     """Create writer with VAAPI preference and OpenCV fallback."""
     format_key = fmt.lower()
 
     if format_key in {"mp4", "mov"}:
         try:
-            writer = _build_ffmpeg_vaapi_writer(output_path, fps, width, height)
+            writer = _build_ffmpeg_vaapi_writer(output_path, fps, width, height, out_fps_str)
             print("[FaceMask] Using output encoder: ffmpeg h264_vaapi (-qp 24)")
             return writer
         except Exception as e:
@@ -238,6 +327,390 @@ def _scale_bbox(
     return [x1, y1, out_w, out_h]
 
 
+def _apply_face_blur_inplace(
+    frame: np.ndarray,
+    frame_boxes: list,
+    src_width: int,
+    src_height: int,
+    blur_size: int,
+    display_scale: float,
+    blur_margin: int,
+) -> None:
+    """検出済み顔領域にガウスぼかしを適用する（in-place）。"""
+    if not frame_boxes:
+        return
+
+    for box in frame_boxes:
+        if not isinstance(box, list) or len(box) < 4:
+            continue
+        x, y, w, h = int(box[0]), int(box[1]), int(box[2]), int(box[3])
+        if w <= 0 or h <= 0:
+            continue
+
+        cx = x + w / 2
+        cy = y + h / 2
+        dw = max(1, int(w * display_scale))
+        dh = max(1, int(h * display_scale))
+        dx = int(cx - dw / 2)
+        dy = int(cy - dh / 2)
+
+        roi_x1 = max(0, dx - blur_margin)
+        roi_y1 = max(0, dy - blur_margin)
+        roi_x2 = min(src_width, dx + dw + blur_margin)
+        roi_y2 = min(src_height, dy + dh + blur_margin)
+        roi_width = roi_x2 - roi_x1
+        roi_height = roi_y2 - roi_y1
+        if roi_width <= 0 or roi_height <= 0:
+            continue
+
+        roi_src = frame[roi_y1:roi_y2, roi_x1:roi_x2]
+        small_w = max(1, roi_width // 2)
+        small_h = max(1, roi_height // 2)
+        roi_small = cv2.resize(roi_src, (small_w, small_h), interpolation=cv2.INTER_LINEAR)
+        small_blur_size = max(3, (blur_size // 2) | 1)
+        roi_small_blurred = cv2.GaussianBlur(roi_small, (small_blur_size, small_blur_size), 0)
+        roi_blurred = cv2.resize(roi_small_blurred, (roi_width, roi_height), interpolation=cv2.INTER_LINEAR)
+
+        roi_mask = np.zeros((roi_height, roi_width), dtype=np.uint8)
+        center = (int(cx) - roi_x1, int(cy) - roi_y1)
+        axes = (max(1, dw // 2), max(1, dh // 2))
+        cv2.ellipse(roi_mask, center, axes, 0, 0, 360, 255, -1)
+
+        result = roi_src.copy()
+        cv2.copyTo(roi_blurred, roi_mask, result)
+        frame[roi_y1:roi_y2, roi_x1:roi_x2] = result
+
+
+def process_images_task(task_id: str, req: GenerateImagesRequest):
+    """画像シーケンスから顔を検出して msgpack キャッシュを保存する。"""
+    try:
+        tasks[task_id].status = TaskStatus.PROCESSING
+        cancel_event = cancel_events.get(task_id)
+
+        if not os.path.exists(req.image_dir):
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = f"Image directory not found: {req.image_dir}"
+            return
+        if not req.filenames:
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = "No filenames provided"
+            return
+
+        detector = get_detector(
+            conf_threshold=req.conf_threshold,
+            iou_threshold=req.iou_threshold,
+        )
+        _ = detector.model
+
+        total_files = len(req.filenames)
+        start_idx = max(0, req.start_index)
+        end_idx = req.end_index if req.end_index >= 0 else total_files - 1
+        end_idx = min(end_idx, total_files - 1)
+
+        if start_idx > end_idx:
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = "Invalid index range"
+            return
+
+        indices = list(range(start_idx, end_idx + 1))
+        tasks[task_id].total = len(indices)
+        os.makedirs(req.output_dir, exist_ok=True)
+        output_msgpack_path = os.path.join(req.output_dir, "detections.msgpack")
+
+        # 画像サイズを最初のファイルから取得
+        first_path = os.path.join(req.image_dir, req.filenames[start_idx])
+        first_img = cv2.imread(first_path)
+        if first_img is None:
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = f"Cannot read image: {first_path}"
+            return
+        height, width = first_img.shape[:2]
+
+        frame_buffer: List[np.ndarray] = []
+        frame_detections: List[List[List[float]]] = []
+        batch_size = 5
+        current_count = 0
+
+        def process_batch():
+            nonlocal current_count
+            if not frame_buffer:
+                return
+            batch_det = detector.detect_batch(frame_buffer)
+            for detections in batch_det:
+                packed: List[List[float]] = []
+                for x, y, w, h, conf in detections:
+                    bx, by, bw, bh = int(x), int(y), int(w), int(h)
+                    bx = max(0, bx)
+                    by = max(0, by)
+                    bw = min(width - bx, bw)
+                    bh = min(height - by, bh)
+                    if bw <= 0 or bh <= 0:
+                        continue
+                    packed.append([bx, by, bw, bh, float(conf)])
+                frame_detections.append(packed)
+                current_count += 1
+                tasks[task_id].progress = current_count
+            frame_buffer.clear()
+
+        print(
+            f"[FaceMask] Starting image detection: {req.image_dir} "
+            f"({len(indices)} images) -> {output_msgpack_path}"
+        )
+
+        for file_idx in indices:
+            if cancel_event and cancel_event.is_set():
+                tasks[task_id].status = TaskStatus.CANCELLED
+                tasks[task_id].message = "Cancelled by user"
+                break
+
+            img_path = os.path.join(req.image_dir, req.filenames[file_idx])
+            frame = cv2.imread(img_path)
+            if frame is None:
+                frame_detections.append([])
+                current_count += 1
+                tasks[task_id].progress = current_count
+                continue
+
+            frame_buffer.append(frame)
+            if len(frame_buffer) >= batch_size:
+                process_batch()
+
+        if frame_buffer:
+            process_batch()
+
+        if tasks[task_id].status == TaskStatus.PROCESSING:
+            payload = {
+                "version": 1,
+                "image_dir": req.image_dir,
+                "filenames": req.filenames,
+                "start_frame": start_idx,
+                "end_frame": start_idx + len(frame_detections) - 1,
+                "width": width,
+                "height": height,
+                "fps": 0.0,
+                "mask_scale": 1.0,
+                "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_msgpack_path
+            tasks[task_id].message = "Image detection cache completed"
+            print(f"[FaceMask] Image detection done: {output_msgpack_path}")
+
+    except Exception as e:
+        tasks[task_id].status = TaskStatus.FAILED
+        tasks[task_id].message = str(e)
+        traceback.print_exc()
+    finally:
+        if task_id in cancel_events:
+            del cancel_events[task_id]
+
+
+def process_bake_images_task(task_id: str, req: BakeImagesRequest):
+    """画像シーケンスに顔ぼかしを適用して新ディレクトリへ書き出す。"""
+    try:
+        tasks[task_id].status = TaskStatus.PROCESSING
+        cancel_event = cancel_events.get(task_id)
+
+        if not os.path.exists(req.image_dir):
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = f"Image directory not found: {req.image_dir}"
+            return
+        if not os.path.exists(req.detections_path):
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = f"Detections file not found: {req.detections_path}"
+            return
+
+        with open(req.detections_path, "rb") as f:
+            payload = msgpack.unpackb(f.read(), raw=False)
+        frames_detections = payload.get("frames")
+        if not isinstance(frames_detections, list):
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = "Invalid detections format: 'frames' is missing"
+            return
+
+        det_start_frame = int(payload.get("start_frame", 0))
+
+        blur_size = max(1, int(req.blur_size))
+        if blur_size % 2 == 0:
+            blur_size += 1
+        display_scale = max(0.1, float(req.display_scale))
+        blur_margin = blur_size // 2
+
+        os.makedirs(req.output_dir, exist_ok=True)
+        total = len(req.filenames)
+        tasks[task_id].total = total
+
+        print(
+            f"[FaceMask] Starting image bake: {req.image_dir} "
+            f"({total} images) -> {req.output_dir}"
+        )
+
+        for i, filename in enumerate(req.filenames):
+            if cancel_event and cancel_event.is_set():
+                tasks[task_id].status = TaskStatus.CANCELLED
+                tasks[task_id].message = "Cancelled by user"
+                return
+
+            src_path = os.path.join(req.image_dir, filename)
+            frame = cv2.imread(src_path)
+            if frame is None:
+                tasks[task_id].progress = i + 1
+                continue
+
+            h, w = frame.shape[:2]
+            det_idx = i - det_start_frame
+            frame_boxes = (
+                frames_detections[det_idx]
+                if 0 <= det_idx < len(frames_detections)
+                else []
+            )
+            _apply_face_blur_inplace(frame, frame_boxes, w, h, blur_size, display_scale, blur_margin)
+
+            out_path = os.path.join(req.output_dir, filename)
+            cv2.imwrite(out_path, frame)
+            tasks[task_id].progress = i + 1
+
+        if tasks[task_id].status == TaskStatus.PROCESSING:
+            tasks[task_id].status = TaskStatus.COMPLETED
+            tasks[task_id].result_path = req.output_dir
+            tasks[task_id].message = "Image blur bake completed"
+            print(f"[FaceMask] Image bake completed: {req.output_dir}")
+
+    except Exception as e:
+        tasks[task_id].status = TaskStatus.FAILED
+        tasks[task_id].message = str(e)
+        traceback.print_exc()
+    finally:
+        if task_id in cancel_events:
+            del cancel_events[task_id]
+
+
+def augment_pose_task(task_id: str, req: AugmentPoseRequest):
+    """Background task: run pose estimation and merge results into existing cache."""
+    cap = None
+    try:
+        tasks[task_id].status = TaskStatus.PROCESSING
+        cancel_event = cancel_events.get(task_id)
+
+        if not os.path.exists(req.detections_path):
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = f"Detections file not found: {req.detections_path}"
+            return
+
+        with open(req.detections_path, "rb") as f:
+            payload = msgpack.unpackb(f.read(), raw=False)
+
+        existing_frames: List[List[List[float]]] = payload.get("frames", [])
+        video_path = payload.get("video_path")
+        start_frame = int(payload.get("start_frame", 0))
+        total = len(existing_frames)
+
+        if not video_path:
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = "Cache does not contain video_path (image caches not supported)"
+            return
+        if not os.path.exists(video_path):
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = f"Video not found: {video_path}"
+            return
+        if total == 0:
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = "Cache has no frames"
+            return
+
+        tasks[task_id].total = total
+
+        detector = get_pose_detector(
+            conf_threshold=req.conf_threshold,
+            iou_threshold=req.iou_threshold,
+        )
+        _ = detector.model
+
+        cap = cv2.VideoCapture(video_path)
+        if not cap.isOpened():
+            tasks[task_id].status = TaskStatus.FAILED
+            tasks[task_id].message = "Failed to open video"
+            return
+
+        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+        if start_frame > 0:
+            seek_ok = cap.set(cv2.CAP_PROP_POS_FRAMES, start_frame)
+            if not seek_ok:
+                for _ in range(start_frame):
+                    ret, _ = cap.read()
+                    if not ret:
+                        tasks[task_id].status = TaskStatus.FAILED
+                        tasks[task_id].message = f"Failed to seek to start frame: {start_frame}"
+                        return
+
+        frame_buffer: List[np.ndarray] = []
+        buffer_indices: List[int] = []  # existing_frames インデックス対応
+        current_count = 0
+        batch_size = 5
+
+        def process_pose_batch():
+            nonlocal current_count
+            if not frame_buffer:
+                return
+            batch_detections = detector.detect_batch(frame_buffer)
+            for idx, detections in zip(buffer_indices, batch_detections):
+                for x, y, w, h, conf in detections:
+                    bx, by, bw, bh = int(x), int(y), int(w), int(h)
+                    bx = max(0, bx)
+                    by = max(0, by)
+                    bw = min(width - bx, bw)
+                    bh = min(height - by, bh)
+                    if bw > 0 and bh > 0:
+                        existing_frames[idx].append([bx, by, bw, bh, float(conf)])
+                current_count += 1
+                tasks[task_id].progress = current_count
+            frame_buffer.clear()
+            buffer_indices.clear()
+
+        for i in range(total):
+            if cancel_event and cancel_event.is_set():
+                tasks[task_id].status = TaskStatus.CANCELLED
+                tasks[task_id].message = "Cancelled by user"
+                break
+
+            ret, frame = cap.read()
+            if not ret:
+                break
+
+            frame_buffer.append(frame)
+            buffer_indices.append(i)
+            if len(frame_buffer) >= batch_size:
+                process_pose_batch()
+
+        if frame_buffer:
+            process_pose_batch()
+
+        if tasks[task_id].status == TaskStatus.PROCESSING:
+            payload["frames"] = existing_frames
+            with open(req.detections_path, "wb") as f:
+                f.write(msgpack.packb(payload, use_bin_type=True))
+
+            tasks[task_id].status = TaskStatus.COMPLETED
+            tasks[task_id].result_path = req.detections_path
+            tasks[task_id].message = "Pose augmentation completed"
+            print(f"[FaceMask] Pose augmentation completed: {req.detections_path}")
+
+    except Exception as e:
+        tasks[task_id].status = TaskStatus.FAILED
+        tasks[task_id].message = str(e)
+        traceback.print_exc()
+    finally:
+        if cap:
+            cap.release()
+        if task_id in cancel_events:
+            del cancel_events[task_id]
+
+
 def process_video_task(task_id: str, req: GenerateRequest):
     """Background task to detect faces and save bbox cache as msgpack."""
     cap = None
@@ -305,20 +778,15 @@ def process_video_task(task_id: str, req: GenerateRequest):
             for detections in batch_detections:
                 packed_detections: List[List[float]] = []
                 for x, y, w, h, conf in detections:
-                    scaled = _scale_bbox(
+                    # bboxをそのまま保存（表示スケールはBake時に適用）
-                        int(x),
+                    bx, by, bw, bh = int(x), int(y), int(w), int(h)
-                        int(y),
+                    bx = max(0, bx)
-                        int(w),
+                    by = max(0, by)
-                        int(h),
+                    bw = min(width - bx, bw)
-                        float(req.mask_scale),
+                    bh = min(height - by, bh)
-                        width,
+                    if bw <= 0 or bh <= 0:
-                        height,
-                    )
-                    if scaled is None:
                         continue
-                    packed_detections.append(
+                    packed_detections.append([bx, by, bw, bh, float(conf)])
-                        [scaled[0], scaled[1], scaled[2], scaled[3], float(conf)]
-                    )
                 frame_detections.append(packed_detections)
                 current_count += 1
                 tasks[task_id].progress = current_count
@@ -356,7 +824,7 @@ def process_video_task(task_id: str, req: GenerateRequest):
                 "width": width,
                 "height": height,
                 "fps": fps,
-                "mask_scale": float(req.mask_scale),
+                "mask_scale": 1.0,
                 "frames": frame_detections,
             }
             with open(output_msgpack_path, "wb") as f:
@@ -402,6 +870,9 @@ def process_bake_task(task_id: str, req: BakeRequest):
             tasks[task_id].message = "Invalid detections format: 'frames' is missing"
             return
 
+        # 検出キャッシュの開始フレーム（ソース動画のフレームインデックス）
+        det_start_frame = int(payload.get("start_frame", 0))
+
         # Get video info
         temp_cap = cv2.VideoCapture(req.video_path)
         if not temp_cap.isOpened():
@@ -415,12 +886,22 @@ def process_bake_task(task_id: str, req: BakeRequest):
         src_frames = int(temp_cap.get(cv2.CAP_PROP_FRAME_COUNT))
         temp_cap.release()
 
+        # ffprobe で r_frame_rate を取得し、出力コンテナの宣言 FPS をソースに合わせる。
+        # 例: 120fps タイムベースで記録された 60fps 動画は r_frame_rate=120/1 だが
+        # cv2 は avg_frame_rate=60fps を返すため、Bake 後に Blender がFPSを別値で認識してしまう。
+        r_fps_float, r_fps_str = _get_r_frame_rate(req.video_path)
+        if r_fps_float > 0:
+            print(f"[FaceMask] r_frame_rate={r_fps_str}, avg_fps={src_fps:.3f}")
+        else:
+            r_fps_str = ""
+
         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, len(frames_detections)) if src_frames > 0 else len(frames_detections)
+        # ソース動画の全フレームを出力（スワップ後もトリム設定が正しく機能するよう）
+        total = src_frames if src_frames > 0 else (det_start_frame + len(frames_detections))
         if total <= 0:
             tasks[task_id].status = TaskStatus.FAILED
             tasks[task_id].message = "Source/detections frame count is zero"
@@ -435,9 +916,9 @@ def process_bake_task(task_id: str, req: BakeRequest):
         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))
+        display_scale = max(0.1, float(req.display_scale))
-        feather_kernel = feather_radius * 2 + 1
+        # blur_margin は境界問題回避のための計算用余白のみ（表示には使わない）
-        blur_margin = max(1, (blur_size // 2) + feather_radius)
+        blur_margin = blur_size // 2
 
         # Queues
         queue_size = 8
@@ -451,52 +932,74 @@ def process_bake_task(task_id: str, req: BakeRequest):
 
         def _reader_worker():
             """Read frames from video."""
+            import time as _time
             cap = cv2.VideoCapture(req.video_path)
             if not cap.isOpened():
                 error_holder["error"] = "Failed to open video in reader"
                 return
 
+            t_read_total = 0.0
+            frame_count = 0
             try:
                 for idx in range(total):
                     if cancel_event and cancel_event.is_set():
                         break
 
+                    t0 = _time.perf_counter()
                     ok, frame = cap.read()
+                    t_read_total += _time.perf_counter() - t0
+
                     if not ok:
                         break
 
                     read_queue.put((idx, frame))
+                    frame_count += 1
             except Exception as e:
                 error_holder["error"] = f"Reader error: {e}"
             finally:
                 cap.release()
                 read_queue.put(None)  # Sentinel
+                if frame_count > 0:
+                    print(
+                        f"[Perf/Reader] FINAL frame={frame_count}"
+                        f"  read_avg={t_read_total/frame_count*1000:.1f}ms"
+                        f"  throughput≈{frame_count/max(t_read_total,1e-9):.1f}fps"
+                    )
 
         def _processor_worker():
             """Process frames with ROI blur."""
+            import time as _time
+            t_wait_total = 0.0
+            t_blur_total = 0.0
+            t_blend_total = 0.0
+            frame_count = 0
+            REPORT_INTERVAL = 50
+
             try:
                 while True:
                     if cancel_event and cancel_event.is_set():
                         process_queue.put(None)
                         break
 
+                    t0 = _time.perf_counter()
                     item = read_queue.get()
+                    t_wait_total += _time.perf_counter() - t0
+
                     if item is None:
                         process_queue.put(None)
                         break
 
                     idx, frame = item
-                    frame_boxes = frames_detections[idx] if idx < len(frames_detections) else []
+                    det_idx = idx - det_start_frame
+                    frame_boxes = frames_detections[det_idx] if 0 <= det_idx < len(frames_detections) else []
 
                     if not frame_boxes:
                         process_queue.put((idx, frame))
+                        frame_count += 1
                         continue
 
-                    # ROI processing (same as original)
+                    # 各人物ごとに個別ROIで処理（全員まとめると離れた人物間が巨大ROIになるため）
-                    min_x, min_y = src_width, src_height
-                    max_x, max_y = 0, 0
                     valid_boxes = []
-
                     for box in frame_boxes:
                         if not isinstance(box, list) or len(box) < 4:
                             continue
@@ -504,64 +1007,112 @@ def process_bake_task(task_id: str, req: BakeRequest):
                         if w <= 0 or h <= 0:
                             continue
                         valid_boxes.append((x, y, w, h))
-                        min_x = min(min_x, x)
-                        min_y = min(min_y, y)
-                        max_x = max(max_x, x + w)
-                        max_y = max(max_y, y + h)
 
                     if not valid_boxes:
                         process_queue.put((idx, frame))
+                        frame_count += 1
                         continue
 
-                    roi_x1 = max(0, min_x - blur_margin)
+                    for x, y, w, h in valid_boxes:
-                    roi_y1 = max(0, min_y - blur_margin)
+                        # display_scale で表示サイズを決定
-                    roi_x2 = min(src_width, max_x + blur_margin)
+                        cx = x + w / 2
-                    roi_y2 = min(src_height, max_y + blur_margin)
+                        cy = y + h / 2
+                        dw = max(1, int(w * display_scale))
+                        dh = max(1, int(h * display_scale))
+                        dx = int(cx - dw / 2)
+                        dy = int(cy - dh / 2)
+
+                        # ROIは表示サイズ + blur_margin（計算用余白、境界問題回避のみ）
+                        roi_x1 = max(0, dx - blur_margin)
+                        roi_y1 = max(0, dy - blur_margin)
+                        roi_x2 = min(src_width, dx + dw + blur_margin)
+                        roi_y2 = min(src_height, dy + dh + blur_margin)
                         roi_width = roi_x2 - roi_x1
                         roi_height = roi_y2 - roi_y1
 
                         if roi_width <= 0 or roi_height <= 0:
-                        process_queue.put((idx, frame))
                             continue
 
+                        # ブラーはROI全体で計算（余白があるので端の精度が保証される）
+                        roi_src = frame[roi_y1:roi_y2, roi_x1:roi_x2]
+
+                        # ダウンサンプル→blur→アップサンプル（同等のぼかしを1/4の計算量で実現）
+                        t1 = _time.perf_counter()
+                        small_w = max(1, roi_width // 2)
+                        small_h = max(1, roi_height // 2)
+                        roi_small = cv2.resize(roi_src, (small_w, small_h), interpolation=cv2.INTER_LINEAR)
+                        small_blur_size = max(3, (blur_size // 2) | 1)
+                        roi_small_blurred = cv2.GaussianBlur(roi_small, (small_blur_size, small_blur_size), 0)
+                        roi_blurred = cv2.resize(roi_small_blurred, (roi_width, roi_height), interpolation=cv2.INTER_LINEAR)
+                        t_blur_total += _time.perf_counter() - t1
+
+                        # 合成マスクはdisplay_scaleサイズの楕円のみ（featheringなし）
                         roi_mask = np.zeros((roi_height, roi_width), dtype=np.uint8)
-                    for x, y, w, h in valid_boxes:
+                        center = (int(cx) - roi_x1, int(cy) - roi_y1)
-                        center = (x + w // 2 - roi_x1, y + h // 2 - roi_y1)
+                        axes = (max(1, dw // 2), max(1, dh // 2))
-                        axes = (max(1, w // 2), max(1, h // 2))
                         cv2.ellipse(roi_mask, center, axes, 0, 0, 360, 255, -1)
 
-                    roi_mask = cv2.GaussianBlur(roi_mask, (feather_kernel, feather_kernel), 0)
+                        # バイナリマスクなのでcopyToで高速合成（float32変換不要）
-                    roi_src = frame[roi_y1:roi_y2, roi_x1:roi_x2]
+                        t2 = _time.perf_counter()
-                    roi_blurred = cv2.GaussianBlur(roi_src, (blur_size, blur_size), 0)
+                        result = roi_src.copy()
+                        cv2.copyTo(roi_blurred, roi_mask, result)
+                        frame[roi_y1:roi_y2, roi_x1:roi_x2] = result
+                        t_blend_total += _time.perf_counter() - t2
 
-                    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
-                    )
-
-                    frame[roi_y1:roi_y2, roi_x1:roi_x2] = np.clip(roi_composed, 0, 255).astype(np.uint8)
                     process_queue.put((idx, frame))
+                    frame_count += 1
+
+                    if frame_count % REPORT_INTERVAL == 0:
+                        n = max(frame_count, 1)
+                        fps_proc = frame_count / max(t_wait_total + t_blur_total + t_blend_total, 1e-9)
+                        print(
+                            f"[Perf/Processor] frame={frame_count}"
+                            f"  wait={t_wait_total/n*1000:.1f}ms"
+                            f"  blur={t_blur_total/n*1000:.1f}ms"
+                            f"  blend={t_blend_total/n*1000:.1f}ms"
+                            f"  ROI={roi_width}x{roi_height}"
+                            f"  throughput≈{fps_proc:.1f}fps"
+                        )
 
             except Exception as e:
                 error_holder["error"] = f"Processor error: {e}"
                 process_queue.put(None)
+            finally:
+                if frame_count > 0:
+                    n = max(frame_count, 1)
+                    print(
+                        f"[Perf/Processor] FINAL frame={frame_count}"
+                        f"  wait_avg={t_wait_total/n*1000:.1f}ms"
+                        f"  blur_avg={t_blur_total/n*1000:.1f}ms"
+                        f"  blend_avg={t_blend_total/n*1000:.1f}ms"
+                    )
 
         def _writer_worker():
             """Write frames to output."""
+            import time as _time
+            t_wait_total = 0.0
+            t_write_total = 0.0
+            frame_count = 0
             writer = None
             try:
-                writer = _build_video_writer(req.output_path, req.format, src_fps, src_width, src_height)
+                writer = _build_video_writer(req.output_path, req.format, src_fps, src_width, src_height, r_fps_str)
 
                 while True:
                     if cancel_event and cancel_event.is_set():
                         break
 
+                    t0 = _time.perf_counter()
                     item = process_queue.get()
+                    t_wait_total += _time.perf_counter() - t0
+
                     if item is None:
                         break
 
                     idx, frame = item
+                    t1 = _time.perf_counter()
                     writer.write(frame)
+                    t_write_total += _time.perf_counter() - t1
+                    frame_count += 1
 
                     with progress_lock:
                         current_progress[0] = idx + 1
@@ -575,6 +1126,13 @@ def process_bake_task(task_id: str, req: BakeRequest):
                         writer.release()
                     except Exception as e:
                         print(f"[FaceMask] Writer release error: {e}")
+                if frame_count > 0:
+                    n = max(frame_count, 1)
+                    print(
+                        f"[Perf/Writer] FINAL frame={frame_count}"
+                        f"  wait_avg={t_wait_total/n*1000:.1f}ms"
+                        f"  write_avg={t_write_total/n*1000:.1f}ms"
+                    )
 
         print(
             f"[FaceMask] Starting blur bake: {req.video_path} + "
@@ -768,6 +1326,39 @@ def get_status():
         "rocm_version": gpu_info["rocm_version"]
     }
 
+
+@app.post("/video_info")
+def get_video_info(req: VideoInfoRequest):
+    if not os.path.exists(req.video_path):
+        raise HTTPException(status_code=404, detail=f"Video not found: {req.video_path}")
+
+    cap = cv2.VideoCapture(req.video_path)
+    if not cap.isOpened():
+        raise HTTPException(status_code=400, detail="Failed to open video")
+
+    try:
+        avg_fps = float(cap.get(cv2.CAP_PROP_FPS) or 0.0)
+        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH) or 0)
+        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT) or 0)
+        frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT) or 0)
+    finally:
+        cap.release()
+
+    # Blender は r_frame_rate でタイムライン配置を計算するため、
+    # cv2 の avg_frame_rate ではなく r_frame_rate を fps として返す。
+    # 例: 120fps タイムベース記録の 60fps 動画で r_frame_rate=120 を返すことで
+    # compute_strip_frame_range の fps_ratio が Blender の解釈と一致する。
+    r_fps_float, _ = _get_r_frame_rate(req.video_path)
+    fps = r_fps_float if r_fps_float > 0 else avg_fps
+
+    return {
+        "video_path": req.video_path,
+        "fps": fps,
+        "width": width,
+        "height": height,
+        "frame_count": frame_count,
+    }
+
 @app.post("/generate", response_model=Task)
 def generate_mask_endpoint(req: GenerateRequest, background_tasks: BackgroundTasks):
     task_id = str(uuid.uuid4())
@@ -789,6 +1380,36 @@ def bake_blur_endpoint(req: BakeRequest, background_tasks: BackgroundTasks):
     background_tasks.add_task(process_bake_task, task_id, req)
     return task
 
+@app.post("/generate_images", response_model=Task)
+def generate_images_endpoint(req: GenerateImagesRequest, background_tasks: BackgroundTasks):
+    task_id = str(uuid.uuid4())
+    task = Task(id=task_id, status=TaskStatus.PENDING)
+    tasks[task_id] = task
+    cancel_events[task_id] = threading.Event()
+    background_tasks.add_task(process_images_task, task_id, req)
+    return task
+
+
+@app.post("/augment_pose", response_model=Task)
+def augment_pose_endpoint(req: AugmentPoseRequest, background_tasks: BackgroundTasks):
+    task_id = str(uuid.uuid4())
+    task = Task(id=task_id, status=TaskStatus.PENDING)
+    tasks[task_id] = task
+    cancel_events[task_id] = threading.Event()
+    background_tasks.add_task(augment_pose_task, task_id, req)
+    return task
+
+
+@app.post("/bake_image_blur", response_model=Task)
+def bake_image_blur_endpoint(req: BakeImagesRequest, background_tasks: BackgroundTasks):
+    task_id = str(uuid.uuid4())
+    task = Task(id=task_id, status=TaskStatus.PENDING)
+    tasks[task_id] = task
+    cancel_events[task_id] = threading.Event()
+    background_tasks.add_task(process_bake_images_task, task_id, req)
+    return task
+
+
 @app.get("/tasks/{task_id}", response_model=Task)
 def get_task(task_id: str):
     if task_id not in tasks: