Upload app.py with huggingface_hub

app.py

@@ -124,11 +124,13 @@ def _load_detector():
         _DETECTOR = "unavailable"
 
 
+_FACE_PAD = 0.10   # 10% proportional padding — matches LAGENDA benchmark MAE=3.555
+
+
 def _detect_faces(img_rgb: np.ndarray,
-                  min_face_px: int = 20,
-                  margin: int = 15) -> list[tuple[int, int, int, int]]:
+                  min_face_px: int = 20) -> list[tuple[int, int, int, int]]:
     """
-    Returns list of (x0, y0, x1, y1) with margin padding, sorted by area desc.
+    Returns list of (x0, y0, x1, y1) with 10% proportional padding, sorted by area desc.
     Falls back to empty list if YuNet is unavailable.
     """
     if _DETECTOR == "unavailable" or _DETECTOR is None:
@@ -147,11 +149,13 @@ def _detect_faces(img_rgb: np.ndarray,
     bboxes = []
     for face in faces:
         x, y, fw, fh = face[:4].astype(int)
-        # Add margin, clamp to image bounds
-        x0 = max(0, x - margin)
-        y0 = max(0, y - margin)
-        x1 = min(w, x + fw + margin)
-        y1 = min(h, y + fh + margin)
+        # 10% proportional padding (matches training/benchmark setup)
+        pw = int(fw * _FACE_PAD)
+        ph = int(fh * _FACE_PAD)
+        x0 = max(0, x - pw)
+        y0 = max(0, y - ph)
+        x1 = min(w, x + fw + pw)
+        y1 = min(h, y + fh + ph)
         if (x1 - x0) >= min_face_px and (y1 - y0) >= min_face_px:
             bboxes.append((x0, y0, x1, y1))
 
@@ -271,7 +275,7 @@ Upload a photo. **YuNet** auto-detects faces, then **FaceAge-DINOv3** predicts a
 
 | | |
 |--|--|
-| 🏆 LAGENDA 84k MAE | **3.760** (beats MiVOLO v2 measured 3.859) |
+| 🏆 LAGENDA 84k MAE | **3.555** (beats MiVOLO v2 paper 3.650, face-only) |
 | 🧠 Backbone | DINOv3-ViT-L/16 (Meta AI, 307M params) |
 | ⚡ Speed | ~100 ms / face on CPU (ONNX FP32) |
 | 🔍 Detector | YuNet (OpenCV, ~350 KB) |