Update generator.py

generator.py

@@ -766,14 +766,26 @@ class RetroArtConverter:
         if hasattr(self.pipe, 'text_encoder'):
             pipe_kwargs["clip_skip"] = 2
         
-        # --- FIX START: Configure ControlNet inputs dynamically ---
         control_images = []
         conditioning_scales = []
         scale_debug_str = []
         
+        # Helper function to ensure control image has correct dimensions
+        def ensure_correct_size(img, target_w, target_h, name="control"):
+            """Ensure image matches target dimensions exactly"""
+            if img is None:
+                return Image.new("RGB", (target_w, target_h), (0,0,0))
+            
+            if img.size != (target_w, target_h):
+                print(f"  [RESIZE] {name}: {img.size} -> ({target_w}, {target_h})")
+                img = img.resize((target_w, target_h), Image.LANCZOS)
+            return img
+        
         # 1. InstantID (Identity)
         if self.instantid_active:
             if has_detected_faces and face_kps_image is not None:
+                # Ensure face keypoints image has correct size
+                face_kps_image = ensure_correct_size(face_kps_image, target_width, target_height, "InstantID")
                 control_images.append(face_kps_image)
                 conditioning_scales.append(identity_control_scale)
                 scale_debug_str.append(f"Identity: {identity_control_scale:.2f}")
@@ -816,24 +828,33 @@ class RetroArtConverter:
 
         # 2. Depth
         if self.depth_active:
+            # Ensure depth image has correct size
+            depth_image = ensure_correct_size(depth_image, target_width, target_height, "Depth")
             control_images.append(depth_image)
             conditioning_scales.append(depth_control_scale)
             scale_debug_str.append(f"Depth: {depth_control_scale:.2f}")
 
         # 3. OpenPose (Expression)
         if self.openpose_active:
-            control_images.append(openpose_image) # This is already a blank map if it failed
+            # Ensure openpose image has correct size
+            openpose_image = ensure_correct_size(openpose_image, target_width, target_height, "OpenPose")
+            control_images.append(openpose_image)
             conditioning_scales.append(expression_control_scale)
             scale_debug_str.append(f"Expression: {expression_control_scale:.2f}")
 
+        # Final validation: ensure all control images have identical dimensions
         if control_images:
+            expected_size = (target_width, target_height)
+            for idx, img in enumerate(control_images):
+                if img.size != expected_size:
+                    print(f"  [WARNING] Control image {idx} size mismatch: {img.size} vs expected {expected_size}")
+                    control_images[idx] = img.resize(expected_size, Image.LANCZOS)
+            
             pipe_kwargs["control_image"] = control_images
             pipe_kwargs["controlnet_conditioning_scale"] = conditioning_scales
-            print(f"Active ControlNets: {len(control_images)}")
+            print(f"Active ControlNets: {len(control_images)} (all {target_width}x{target_height})")
         else:
             print("No active ControlNets, running standard Img2Img")
-        # --- FIX END ---
-        
         
         # Generate
         print(f"Generating with LCM: Steps={num_inference_steps}, CFG={guidance_scale}, Strength={strength}")