Update app.py

app.py

@@ -6,7 +6,7 @@ from typing import Dict, List, Optional, Tuple, Union
 import gradio as gr
 import spaces
 import torch
-import face_recognition
+from facenet_pytorch import MTCNN, InceptionResnetV1
 import numpy as np
 from PIL import Image
 from transformers import AutoProcessor, LlavaOnevisionForConditionalGeneration
@@ -15,6 +15,9 @@ MODEL_ID = os.environ.get("MODEL_ID", "BSC-LT/salamandra-7b-vision")
 DTYPE = torch.float16 if torch.cuda.is_available() else torch.float32
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 
+mtcnn = MTCNN(image_size=160, margin=0, post_process=True, device=DEVICE)
+facenet = InceptionResnetV1(pretrained='vggface2').eval().to(DEVICE)
+
 _model = None
 _processor = None
 
@@ -93,18 +96,29 @@ def describe_batch(images: List[Image.Image], context_json: str,
     return outputs
 
 
-def face_image_embedding(image: Image.Image) -> List[float]:
+def face_image_embedding(image: Image.Image) -> List[float] | None:
     try:
-        face_model = face_recognition
-        encs = face_model.face_encodings(image)
-        if encs:
-            embeddings = [(e / np.linalg.norm(e)).astype(float).tolist() for e in encs]
-            return embeddings
-        return None
+        # detectar y extraer cara
+        face = mtcnn(image)
+
+        if face is None:
+            return None
+
+        # FaceNet espera tensor shape (1,3,160,160)
+        face = face.unsqueeze(0).to(DEVICE)
+
+        # obtener embedding
+        with torch.no_grad():
+            emb = facenet(face).cpu().numpy()[0]
+
+        # normalizar igual que tu código original
+        emb = emb / np.linalg.norm(emb)
+
+        return emb.astype(float).tolist()
+
     except Exception as e:
         print(f"Fallo embedding cara: {e}")
-    
-    return None
+        return None
 
 
 # ----------------------------- UI & Endpoints --------------------------------