Update generator.py

generator.py

@@ -101,7 +101,8 @@ class RetroArtConverter:
             Generate depth map using available depth detector.
             Supports: LeresDetector, ZoeDetector, or MidasDetector.
             """
-            if self.depth_detector is not None:
+            # --- FIX 1: Check for self.zoe_depth, not self.depth_detector ---
+            if self.zoe_depth is not None:
                 try:
                     if image.mode != 'RGB':
                         image = image.convert('RGB')
@@ -120,14 +121,15 @@ class RetroArtConverter:
                     image_for_depth = image.resize(size_for_depth, Image.LANCZOS)
                     
                     if target_width != orig_width or target_height != orig_height:
-                        print(f"[DEPTH] Resized for {self.depth_type.upper()}Detector: {orig_width}x{orig_height} -> {target_width}x{target_height}")
+                        # --- FIX 2: Use "ZOE" instead of undefined self.depth_type ---
+                        print(f"[DEPTH] Resized for ZOEDetector: {orig_width}x{orig_height} -> {target_width}x{target_height}")
                     
                     # Use torch.no_grad() and clear cache
                     with torch.no_grad():
-                        # --- FIX: Move model to GPU for inference and back to CPU ---
-                        self.depth_detector.to(self.device)
-                        depth_image = self.depth_detector(image_for_depth)
-                        self.depth_detector.to("cpu")
+                        # --- FIX 1: Use self.zoe_depth ---
+                        self.zoe_depth.to(self.device)
+                        depth_image = self.zoe_depth(image_for_depth)
+                        self.zoe_depth.to("cpu")
                     
                     # ADDED: Clear GPU cache after depth detection
                     if torch.cuda.is_available():
@@ -137,11 +139,13 @@ class RetroArtConverter:
                     if depth_width != orig_width or depth_height != orig_height:
                         depth_image = depth_image.resize((int(orig_width), int(orig_height)), Image.LANCZOS)
                     
-                    print(f"[DEPTH] {self.depth_type.upper()} depth map generated: {orig_width}x{orig_height}")
+                    # --- FIX 2: Use "ZOE" instead of undefined self.depth_type ---
+                    print(f"[DEPTH] ZOE depth map generated: {orig_width}x{orig_height}")
                     return depth_image
                     
                 except Exception as e:
-                    print(f"[DEPTH] {self.depth_type.upper()}Detector failed ({e}), falling back to grayscale depth")
+                    # --- FIX 2: Use "ZOE" instead of undefined self.depth_type ---
+                    print(f"[DEPTH] ZOEDetector failed ({e}), falling back to grayscale depth")
                     # ADDED: Clear cache on error
                     if torch.cuda.is_available():
                         torch.cuda.empty_cache()
@@ -237,7 +241,8 @@ class RetroArtConverter:
         
         # Generate depth map
         print("Generating depth map...")
-        depth_image, depth_array = self.get_depth_map(resized_image)
+        # --- FIX 3: get_depth_map only returns one value ---
+        depth_image = self.get_depth_map(resized_image)
         
         if depth_image is None:
             raise RuntimeError("Failed to generate depth map")
@@ -267,7 +272,8 @@ class RetroArtConverter:
                     face_bbox_original = face.bbox
                     
                     print(f"  [OK] Face detected")
-                    print(f"  - Embedding shape: {face_embeddings.shape}")
+                    # --- FIX 4: Clarify this is the numpy shape ---
+                    print(f"  - Embedding shape (numpy): {face_embeddings.shape}")
                     print(f"  - Keypoints: {face.kps.shape}")
                     print(f"  - Bbox: {face_bbox_original}")
                     
@@ -285,7 +291,7 @@ class RetroArtConverter:
                         print("  [ADAPTIVE] Low confidence - increasing identity weight")
                         identity_preservation = max(identity_preservation, ADAPTIVE_PARAMS['low_confidence']['identity_preservation'])
                         identity_control_scale = max(identity_control_scale, ADAPTIVE_PARAMS['low_confidence']['identity_control_scale'])
-                    
+                
                 else:
                     print("  No faces detected in image")
                     
@@ -370,14 +376,16 @@ class RetroArtConverter:
             if face_embeddings is not None:
                 print(f"Adding face embeddings for IP-Adapter...")
                 
-                # Just pass the embeddings - pipeline does everything!
-                pipe_kwargs["image_embeds"] = face_embeddings  # numpy array (512,)
+                # --- FIX 4: Convert numpy array to torch tensor, add batch dim, and move to device ---
+                face_embeds_tensor = torch.tensor(face_embeddings, dtype=self.dtype, device=self.device).unsqueeze(0)
+                pipe_kwargs["image_embeds"] = face_embeds_tensor
                 
                 # Control IP-Adapter strength
                 boosted_scale = identity_preservation * IDENTITY_BOOST_MULTIPLIER
                 pipe_kwargs["ip_adapter_scale"] = boosted_scale
                 
-                print(f"  - Face embeddings shape: {face_embeddings.shape}")
+                # --- FIX 4: Update log to show tensor shape ---
+                print(f"  - Face embeddings tensor shape: {face_embeds_tensor.shape}")
                 print(f"  - IP-Adapter scale: {boosted_scale:.2f}")
                 print(f"  [OK] Face embeddings configured")
             else: