Update generator.py

generator.py

@@ -153,15 +153,11 @@ class RetroArtConverter:
                 if image.mode != 'RGB':
                     image = image.convert('RGB')
                 
-                # CRITICAL FIX: Convert to numpy and back to ensure clean PIL Image
-                # This removes any numpy int64 contamination in image.size
-                image_array = np.array(image)
-                clean_image = Image.fromarray(image_array)
+                # ZoeDetector handles everything internally - just pass the image
+                depth_image = self.zoe_depth(image)
                 
-                # Now ZoeDetector receives clean PIL Image with Python int dimensions
-                depth_image = self.zoe_depth(clean_image)
-                
-                print(f"[DEPTH] Zoe depth map generated: {clean_image.size[0]}x{clean_image.size[1]}")
+                # Use .width and .height properties (always Python ints, not numpy types)
+                print(f"[DEPTH] Zoe depth map generated: {image.width}x{image.height}")
                 return depth_image
                 
             except Exception as e:
@@ -599,52 +595,23 @@ class RetroArtConverter:
             try:
                 print("Encoding prompts with Compel...")
                 
-                # Encode prompts separately (Compel handles dual text encoders internally)
-                conditioning = self.compel(prompt)
-                negative_conditioning = self.compel(negative_prompt)
-                
-                # Extract embeddings
-                prompt_embeds = conditioning[0]
-                pooled_prompt_embeds = conditioning[1]
-                negative_prompt_embeds = negative_conditioning[0]
-                negative_pooled_prompt_embeds = negative_conditioning[1]
-                
-                # Force to 77 tokens (SDXL standard) to ensure compatibility
-                target_length = 77
-                
-                # Truncate or pad positive embeddings
-                if prompt_embeds.shape[1] > target_length:
-                    prompt_embeds = prompt_embeds[:, :target_length, :]
-                elif prompt_embeds.shape[1] < target_length:
-                    padding = torch.zeros(
-                        prompt_embeds.shape[0],
-                        target_length - prompt_embeds.shape[1],
-                        prompt_embeds.shape[2],
-                        dtype=prompt_embeds.dtype,
-                        device=prompt_embeds.device
-                    )
-                    prompt_embeds = torch.cat([prompt_embeds, padding], dim=1)
-                
-                # Truncate or pad negative embeddings
-                if negative_prompt_embeds.shape[1] > target_length:
-                    negative_prompt_embeds = negative_prompt_embeds[:, :target_length, :]
-                elif negative_prompt_embeds.shape[1] < target_length:
-                    padding = torch.zeros(
-                        negative_prompt_embeds.shape[0],
-                        target_length - negative_prompt_embeds.shape[1],
-                        negative_prompt_embeds.shape[2],
-                        dtype=negative_prompt_embeds.dtype,
-                        device=negative_prompt_embeds.device
-                    )
-                    negative_prompt_embeds = torch.cat([negative_prompt_embeds, padding], dim=1)
+                # Encode prompts (Compel returns tuple: (embeddings, pooled_embeddings))
+                conditioning, pooled = self.compel(prompt)
+                negative_conditioning, negative_pooled = self.compel(negative_prompt)
+                
+                # OFFICIAL FIX: Use Compel's built-in padding method to handle length mismatches
+                # This is the recommended way from Compel documentation for SDXL
+                [conditioning, negative_conditioning] = self.compel.pad_conditioning_tensors_to_same_length([
+                    conditioning, negative_conditioning
+                ])
                 
-                pipe_kwargs["prompt_embeds"] = prompt_embeds
-                pipe_kwargs["pooled_prompt_embeds"] = pooled_prompt_embeds
-                pipe_kwargs["negative_prompt_embeds"] = negative_prompt_embeds
-                pipe_kwargs["negative_pooled_prompt_embeds"] = negative_pooled_prompt_embeds
+                pipe_kwargs["prompt_embeds"] = conditioning
+                pipe_kwargs["pooled_prompt_embeds"] = pooled
+                pipe_kwargs["negative_prompt_embeds"] = negative_conditioning
+                pipe_kwargs["negative_pooled_prompt_embeds"] = negative_pooled
                 
                 compel_success = True
-                print(f"[OK] Compel encoded: pos={prompt_embeds.shape}, neg={negative_prompt_embeds.shape}")
+                print(f"[OK] Compel encoded: pos={conditioning.shape}, neg={negative_conditioning.shape}")
                 
             except Exception as e:
                 # Compel encoding failed - fall back to standard encoding