Blurサイズ問題の修正

__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/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()
@@ -75,6 +76,7 @@ class AsyncBakeGenerator:
         detections_path: str,
         output_path: str,
         blur_size: int,
+        display_scale: float,
         fmt: str,
     ):
         import time
@@ -88,6 +90,7 @@ 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,
         )
@@ -121,7 +119,6 @@ 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 +138,6 @@ 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/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(
@@ -255,6 +253,7 @@ class InferenceClient:
         detections_path: str,
         output_path: str,
         blur_size: int,
+        display_scale: float,
         fmt: str,
     ) -> str:
         """
@@ -271,6 +270,7 @@ class InferenceClient:
             "detections_path": detections_path,
             "output_path": output_path,
             "blur_size": blur_size,
+            "display_scale": display_scale,
             "format": fmt,
         }
 

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 = {
@@ -86,20 +87,27 @@ class SEQUENCER_OT_bake_and_swap_blur_source(Operator):
         bake_format = scene.facemask_bake_format
         output_path = _output_path(video_strip, detections_path, bake_format)
         blur_size = int(scene.facemask_bake_blur_size)
+        display_scale = float(scene.facemask_bake_display_scale)
 
         # Reuse baked cache when parameters match and file still exists.
         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
+        try:
+            cached_display_scale_f = float(cached_display_scale)
+        except (TypeError, ValueError):
+            cached_display_scale_f = None
         if (
             cached_baked_path
             and os.path.exists(cached_baked_path)
             and cached_format == bake_format
             and cached_blur_size_int == blur_size
+            and cached_display_scale_f == display_scale
         ):
             if video_strip.get(KEY_MODE) != "baked":
                 video_strip[KEY_MODE] = "baked"
@@ -126,6 +134,7 @@ class SEQUENCER_OT_bake_and_swap_blur_source(Operator):
                 strip[KEY_MODE] = "baked"
                 strip[KEY_FORMAT] = bake_format
                 strip[KEY_BLUR_SIZE] = blur_size
+                strip[KEY_DISPLAY_SCALE] = display_scale
                 _set_strip_source(strip, result_path)
                 print(f"[FaceMask] Bake completed and source swapped: {result_path}")
             elif status == "error":
@@ -153,6 +162,7 @@ class SEQUENCER_OT_bake_and_swap_blur_source(Operator):
                 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,

operators/generate_mask.py

@@ -110,7 +110,6 @@ class SEQUENCER_OT_generate_face_mask(Operator):
         # Get parameters from scene properties
         conf_threshold = scene.facemask_conf_threshold
         iou_threshold = scene.facemask_iou_threshold
-        mask_scale = scene.facemask_mask_scale
 
         # Start generation
         generator.start(
@@ -121,7 +120,6 @@ class SEQUENCER_OT_generate_face_mask(Operator):
             fps=fps,
             conf_threshold=conf_threshold,
             iou_threshold=iou_threshold,
-            mask_scale=mask_scale,
             on_complete=on_complete,
             on_progress=on_progress,
         )

panels/vse_panel.py

@@ -74,7 +74,6 @@ 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."""
@@ -225,6 +224,7 @@ 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")
         col.prop(context.scene, "facemask_bake_format")
 
         # Source status

server/detector.py

@@ -1,28 +1,104 @@
 """
-YOLOv8 Face Detector using PyTorch with ROCm support.
+YOLOv8 Pose Head Detector using PyTorch with ROCm support.
 
-This module provides high-performance face detection using
+Detects human heads from all angles (frontal, profile, rear) by using
-YOLOv8-face model with AMD GPU (ROCm) acceleration.
+YOLOv8 pose estimation and extracting head bounding boxes from keypoints.
 """
 
 import os
 from typing import List, Tuple, Optional
-from pathlib import Path
 import numpy as np
 
 
-class YOLOFaceDetector:
+# COCO pose keypoint indices
-    """
+_HEAD_KP = [0, 1, 2, 3, 4]      # nose, left_eye, right_eye, left_ear, right_ear
-    YOLOv8 face detector with PyTorch ROCm support.
+_SHOULDER_KP = [5, 6]            # left_shoulder, right_shoulder
+_KP_CONF_THRESH = 0.3
 
-    Features:
+
-    - ROCm GPU acceleration for AMD GPUs
+def _head_bbox_from_pose(
-    - High accuracy face detection
+    kp_xy: np.ndarray,
-    - Automatic NMS for overlapping detections
+    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
+
+        # Head radius: inter-landmark span ≈ 80% of head width, so expand by ~1.25
+        # Shift center upward slightly to include scalp
+        r = max(span * 1.25, person_w * 0.20)
+        x1 = int(cx - r)
+        y1 = int(cy - r * 1.15)   # extra margin above (scalp)
+        x2 = int(cx + r)
+        y2 = int(cy + r * 0.85)   # less margin below (chin)
+        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 * 0.5, person_w * 0.20)
+        cy = cy_sh - r * 1.3   # head center is above shoulders
+        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 * 0.35, 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)
+    so that detection works regardless of the person's facing direction.
     """
 
-    # Default model path relative to this file
+    # Standard Ultralytics model — auto-downloaded on first use
-    DEFAULT_MODEL = "yolov8n-face-lindevs.pt"
+    DEFAULT_MODEL = os.path.join("models", "yolov8n-pose.pt")
 
     def __init__(
         self,
@@ -31,15 +107,6 @@ 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 +116,20 @@ class YOLOFaceDetector:
 
     @property
     def model(self):
-        """Lazy-load YOLO model."""
+        """Lazy-load YOLO pose model."""
         if self._model is None:
             from ultralytics import YOLO
             import torch
 
-            # Determine model path
+            # Use provided path or let Ultralytics auto-download the default
-            if self._model_path is None:
+            if self._model_path is not None:
-                # Assuming models are in ../models relative to server/detector.py
+                if not os.path.exists(self._model_path):
-                models_dir = Path(__file__).parent.parent / "models"
+                    raise FileNotFoundError(f"Model not found: {self._model_path}")
-                model_path = str(models_dir / self.DEFAULT_MODEL)
-            else:
                 model_path = self._model_path
+            else:
+                model_path = self.DEFAULT_MODEL
+                os.makedirs(os.path.dirname(model_path), exist_ok=True)
 
-            if not os.path.exists(model_path):
-                raise FileNotFoundError(f"Model not found: {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 +138,47 @@ 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] Pose 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 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()
+
+            # Extract keypoints for this person
+            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(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 +186,6 @@ 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 +201,6 @@ 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 +212,13 @@ 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 +230,6 @@ class YOLOFaceDetector:
         if not frames:
             return []
 
-        # Run batch inference
         try:
             results = self.model.predict(
                 frames,
@@ -175,7 +243,6 @@ 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 +254,7 @@ 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 +264,11 @@ 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 +281,19 @@ 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 +302,12 @@ class YOLOFaceDetector:
 
 
 # Singleton instance
-_detector: Optional[YOLOFaceDetector] = None
+_detector: Optional[YOLOPoseHeadDetector] = None
 
 
-def get_detector(**kwargs) -> YOLOFaceDetector:
+def get_detector(**kwargs) -> YOLOPoseHeadDetector:
-    """Get or create the global YOLO detector instance."""
+    """Get or create the global YOLO pose head detector instance."""
     global _detector
     if _detector is None:
-        _detector = YOLOFaceDetector(**kwargs)
+        _detector = YOLOPoseHeadDetector(**kwargs)
     return _detector

server/main.py

@@ -83,7 +83,6 @@ class GenerateRequest(BaseModel):
     end_frame: int
     conf_threshold: float = 0.5
     iou_threshold: float = 0.45
-    mask_scale: float = 1.5
 
 
 class BakeRequest(BaseModel):
@@ -91,6 +90,7 @@ class BakeRequest(BaseModel):
     detections_path: str
     output_path: str
     blur_size: int = 50
+    display_scale: float = 1.0
     format: str = "mp4"
 
 
@@ -305,20 +305,15 @@ 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 +351,7 @@ 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:
@@ -435,9 +430,9 @@ 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
@@ -492,11 +487,8 @@ def process_bake_task(task_id: str, req: BakeRequest):
                         process_queue.put((idx, frame))
                         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
@@ -504,42 +496,45 @@ 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))
                         continue
 
-                    roi_x1 = max(0, min_x - blur_margin)
-                    roi_y1 = max(0, min_y - blur_margin)
-                    roi_x2 = min(src_width, max_x + blur_margin)
-                    roi_y2 = min(src_height, max_y + 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
-
-                    roi_mask = np.zeros((roi_height, roi_width), dtype=np.uint8)
                     for x, y, w, h in valid_boxes:
-                        center = (x + w // 2 - roi_x1, y + h // 2 - roi_y1)
+                        # display_scale で表示サイズを決定
-                        axes = (max(1, w // 2), max(1, h // 2))
+                        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:
+                            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]
+                        roi_blurred = cv2.GaussianBlur(roi_src, (blur_size, blur_size), 0)
+
+                        # 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)
+                        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)
 
-                    roi_mask = cv2.GaussianBlur(roi_mask, (feather_kernel, feather_kernel), 0)
+                        roi_alpha = (roi_mask.astype(np.float32) / 255.0)[..., np.newaxis]
-                    roi_src = frame[roi_y1:roi_y2, roi_x1:roi_x2]
+                        roi_composed = roi_src.astype(np.float32) * (1.0 - roi_alpha) + roi_blurred.astype(np.float32) * roi_alpha
-                    roi_blurred = cv2.GaussianBlur(roi_src, (blur_size, blur_size), 0)
+                        frame[roi_y1:roi_y2, roi_x1:roi_x2] = np.clip(roi_composed, 0, 255).astype(np.uint8)
 
-                    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))
 
             except Exception as e: