diff --git a/.env.example b/.env.example
index 4d9a061..a3d0872 100644
--- a/.env.example
+++ b/.env.example
@@ -21,6 +21,6 @@ AI_PRIVACY_ENABLE_IDENTITY_EXCLUSION=true
 AI_PRIVACY_IDENTITY_RECHECK_INTERVAL=12
 AI_PRIVACY_IDENTITY_NEGATIVE_RECHECK_INTERVAL=90
 AI_PRIVACY_ME_IMAGE_PATH=me.jpeg
-AI_PRIVACY_FACE_MODEL_PATH=models/yolov8n-face.pt
+AI_PRIVACY_FACE_MODEL_PATH=models/yolov12n-face.pt
 AI_PRIVACY_FACE_DETECTION_MODEL_PATH=models/face_detection_yunet_2023mar.onnx
 AI_PRIVACY_FACE_RECOGNITION_MODEL_PATH=models/face_recognition_sface_2021dec.onnx
diff --git a/README.md b/README.md
index 8da9aac..432117b 100644
--- a/README.md
+++ b/README.md
@@ -24,7 +24,7 @@ Mac 로컬 실행은 위와 같이 CPU 모드를 사용한다. 운영 환경은
 
 ## AI privacy runtime
 
-`privacy_blur`는 모든 WebRTC 스트림이 YOLOv8n-Face, YuNet, SFace 모델 세트를
+`privacy_blur`는 모든 WebRTC 스트림이 YOLOv12n-Face, YuNet, SFace 모델 세트를
 공유하고 각 스트림은 독립적인 얼굴 추적 상태만 유지한다. 공유 워커가 짧은 시간
 동안 들어온 프레임을 최대 `AI_PRIVACY_MAX_BATCH_SIZE`개까지 모아 한 번에 추론하며,
 대기열은 `AI_PRIVACY_MAX_PENDING_FRAMES`로 제한된다.
diff --git a/app/core/config.py b/app/core/config.py
index 8b36f40..119622d 100644
--- a/app/core/config.py
+++ b/app/core/config.py
@@ -27,7 +27,7 @@ class Settings(BaseSettings):
     ai_privacy_identity_recheck_interval: int = 12
     ai_privacy_identity_negative_recheck_interval: int = 90
     ai_privacy_me_image_path: str | None = None
-    ai_privacy_face_model_path: str = "models/yolov8n-face.pt"
+    ai_privacy_face_model_path: str = "models/yolov12n-face.pt"
     ai_privacy_face_detection_model_path: str = (
         "models/face_detection_yunet_2023mar.onnx"
     )
diff --git a/models/yolov12n-face.pt b/models/yolov12n-face.pt
new file mode 100644
index 0000000..10b0411
Binary files /dev/null and b/models/yolov12n-face.pt differ
diff --git a/privacy_blur/config.py b/privacy_blur/config.py
index 91bdfc8..2552dae 100644
--- a/privacy_blur/config.py
+++ b/privacy_blur/config.py
@@ -7,7 +7,7 @@
 
 @dataclass(frozen=True, slots=True)
 class PrivacyBlurConfig:
-    face_model_path: Path = Path("models/yolov8n-face.pt")
+    face_model_path: Path = Path("models/yolov12n-face.pt")
     face_imgsz: int = 640
     face_conf_threshold: float = 0.22
     face_iou_threshold: float = 0.45
diff --git a/privacy_blur/runtime.py b/privacy_blur/runtime.py
index cb3bb9b..344ae7a 100644
--- a/privacy_blur/runtime.py
+++ b/privacy_blur/runtime.py
@@ -1015,6 +1015,7 @@ def _blur_frames(
             y = torch.arange(height, device=self.device, dtype=dtype).view(height, 1)
             x = torch.arange(width, device=self.device, dtype=dtype).view(1, width)
 
+            max_face_size = 0.0
             for batch_index, frame_index in enumerate(indices):
                 for box in boxes_per_frame[frame_index]:
                     box_x, box_y, box_w, box_h = box[:4].astype(float)
@@ -1039,14 +1040,12 @@ def _blur_frames(
                     masks[batch_index, 0] = torch.maximum(
                         masks[batch_index, 0], ellipse.to(dtype)
                     )
+                    max_face_size = max(max_face_size, max(box_w, box_h))
 
-            blurred = self._separable_gaussian(
-                images, _make_odd(self.config.blur_kernel)
-            )
-            feathered = self._separable_gaussian(
-                masks,
-                _make_odd(self.config.feather * 2 + 1),
-            ).clamp(0, 1)
+            blur_kernel = self._adaptive_blur_kernel(max_face_size)
+            blurred = self._separable_gaussian(images, blur_kernel)
+            feather_kernel = _make_odd(self._adaptive_feather(max_face_size))
+            feathered = self._separable_gaussian(masks, feather_kernel).clamp(0, 1)
             protected = images * (1.0 - feathered) + blurred * feathered
             protected = protected.clamp(0, 255).round().to(torch.uint8)
             protected_numpy = protected.permute(0, 2, 3, 1).contiguous().cpu().numpy()
@@ -1054,6 +1053,20 @@ def _blur_frames(
                 outputs[frame_index] = protected_numpy[batch_index]
         return outputs
 
+    def _adaptive_blur_kernel(self, max_face_size: float) -> int:
+        base = self.config.blur_kernel
+        reference = float(self.config.small_face_threshold)
+        ratio = max_face_size / reference
+        scaled = base * max(1.0, min(ratio, 3.0))
+        capped = min(scaled, 151.0)
+        return _make_odd(int(capped))
+
+    def _adaptive_feather(self, max_face_size: float) -> int:
+        base = self.config.feather
+        reference = float(self.config.small_face_threshold)
+        ratio = max_face_size / reference
+        return int(base * max(1.0, min(ratio, 2.0)))
+
     def _separable_gaussian(
         self, tensor: torch.Tensor, kernel_size: int
     ) -> torch.Tensor:
diff --git a/tests/test_performance_runtime.py b/tests/test_performance_runtime.py
index 706ed24..8021498 100644
--- a/tests/test_performance_runtime.py
+++ b/tests/test_performance_runtime.py
@@ -17,7 +17,7 @@
 
 def make_test_config(**overrides) -> PrivacyBlurConfig:
     defaults = {
-        "face_model_path": ROOT / "models/yolov8n-face.pt",
+        "face_model_path": ROOT / "models/yolov12n-face.pt",
         "me_image_path": ROOT / "me.jpeg",
         "face_detection_model_path": ROOT / "models/face_detection_yunet_2023mar.onnx",
         "face_recognition_model_path": ROOT
diff --git a/tests/test_privacy_blur.py b/tests/test_privacy_blur.py
index c89897e..a633134 100644
--- a/tests/test_privacy_blur.py
+++ b/tests/test_privacy_blur.py
@@ -15,7 +15,7 @@ def _build_config(**overrides):
     from privacy_blur import PrivacyBlurConfig
 
     defaults = {
-        "face_model_path": ROOT / "models/yolov8n-face.pt",
+        "face_model_path": ROOT / "models/yolov12n-face.pt",
         "face_detection_model_path": ROOT / "models/face_detection_yunet_2023mar.onnx",
         "face_recognition_model_path": ROOT
         / "models/face_recognition_sface_2021dec.onnx",