Update generator.py

generator.py

@@ -153,6 +153,7 @@ class RetroArtConverter:
                         image = image.convert('RGB')
                     
                     orig_width, orig_height = image.size
+                    # **FIX 1 START: Ensure all size variables are standard Python int**
                     orig_width = int(orig_width)
                     orig_height = int(orig_height)
                     
@@ -163,17 +164,25 @@ class RetroArtConverter:
                     target_width = int(max(64, target_width))
                     target_height = int(max(64, target_height))
                     
+                    # Create an explicit tuple of standard ints
+                    size_for_depth = (int(target_width), int(target_height))
+                    
+                    # Always resize using the explicit int tuple to avoid numpy.int64 issues
+                    # This replaces the conditional resize
+                    image_for_depth = image.resize(size_for_depth, Image.LANCZOS)
+                    
                     if target_width != orig_width or target_height != orig_height:
-                        image = image.resize((int(target_width), int(target_height)), Image.LANCZOS)
                         print(f"[DEPTH] Resized for ZoeDetector: {orig_width}x{orig_height} -> {target_width}x{target_height}")
                     
                     # FIXED: Add torch.no_grad() wrapper
                     with torch.no_grad():
-                        depth_image = self.zoe_depth(image)
+                        depth_image = self.zoe_depth(image_for_depth) # Use the correctly-typed resized image
                     
                     depth_width, depth_height = depth_image.size
                     if depth_width != orig_width or depth_height != orig_height:
+                        # Resize back to the original size that get_depth_map received
                         depth_image = depth_image.resize((int(orig_width), int(orig_height)), Image.LANCZOS)
+                    # **FIX 1 END**
                     
                     print(f"[DEPTH] Zoe depth map generated: {orig_width}x{orig_height}")
                     return depth_image
@@ -192,6 +201,10 @@ class RetroArtConverter:
     def add_trigger_word(self, prompt):
         """Add trigger word to prompt if not present"""
         if TRIGGER_WORD.lower() not in prompt.lower():
+            # **FIX 3 START: Handle empty or blank prompt**
+            if not prompt or not prompt.strip():
+                return TRIGGER_WORD
+            # **FIX 3 END**
             return f"{TRIGGER_WORD}, {prompt}"
         return prompt
     
@@ -437,6 +450,11 @@ class RetroArtConverter:
         prompt = sanitize_text(prompt)
         negative_prompt = sanitize_text(negative_prompt)
         
+        # **FIX 3 START: Ensure blank negative prompts are empty strings for Compel**
+        if not negative_prompt or not negative_prompt.strip():
+            negative_prompt = ""
+        # **FIX 3 END**
+        
         # Apply parameter validation
         if consistency_mode:
             print("\n[CONSISTENCY] Validating and adjusting parameters...")
@@ -446,7 +464,7 @@ class RetroArtConverter:
                     identity_control_scale, depth_control_scale, consistency_mode
                 )
         
-        # Add trigger word
+        # Add trigger word (handles blank prompt fix)
         prompt = self.add_trigger_word(prompt)
         
         # Calculate optimal size with flexible aspect ratio support