Upload 2 files

generator.py

@@ -550,10 +550,14 @@ class RetroArtConverter:
         # Set LORA scale
         if hasattr(self.pipe, 'set_adapters') and self.models_loaded['lora']:
             try:
-                # For SDXL with LORA, use set_adapters with proper names
-                adapter_names = ["retroart"]  # The adapter name from loading
-                self.pipe.set_adapters(adapter_names, adapter_weights=[lora_scale])
-                print(f"[LORA] Set adapter 'retroart' with scale: {lora_scale}")
+                # Try retroart first (if loaded with that name), then default_0
+                for adapter_name in ["retroart", "default_0"]:
+                    try:
+                        self.pipe.set_adapters([adapter_name], adapter_weights=[lora_scale])
+                        print(f"[LORA] Set adapter '{adapter_name}' with scale: {lora_scale}")
+                        break
+                    except:
+                        continue
                         
             except Exception as e:
                 print(f"[WARNING] LORA set_adapters failed: {e}")

models.py

@@ -164,12 +164,19 @@ def load_lora(pipe):
     print("Loading LORA (retroart) from HuggingFace Hub...")
     try:
         lora_path = download_model_with_retry(MODEL_REPO, MODEL_FILES['lora'])
-        pipe.load_lora_weights(lora_path)
-        print(f"  [OK] LORA loaded successfully")
+        # Load with explicit adapter name to avoid default_0
+        pipe.load_lora_weights(lora_path, adapter_name="retroart")
+        print(f"  [OK] LORA loaded successfully as 'retroart' adapter")
         return True
     except Exception as e:
-        print(f"  [WARNING] Could not load LORA: {e}")
-        return False
+        # Fallback to default loading
+        try:
+            pipe.load_lora_weights(lora_path)
+            print(f"  [OK] LORA loaded successfully (default adapter)")
+            return True
+        except Exception as e2:
+            print(f"  [WARNING] Could not load LORA: {e2}")
+            return False
 
 
 def setup_ip_adapter(pipe, image_encoder):
@@ -191,15 +198,29 @@ def setup_ip_adapter(pipe, image_encoder):
         # Load full state dict
         state_dict = torch.load(ip_adapter_path, map_location="cpu")
         
-        # Extract image_proj and ip_adapter weights
+        # Debug: Print available keys
+        print(f"[DEBUG] State dict keys sample: {list(state_dict.keys())[:5]}")
+        
+        # Extract image_proj and ip_adapter weights with flexible key matching
         image_proj_state_dict = {}
         ip_adapter_state_dict = {}
         
         for key, value in state_dict.items():
-            if key.startswith("image_proj."):
-                image_proj_state_dict[key.replace("image_proj.", "")] = value
-            elif key.startswith("ip_adapter."):
-                ip_adapter_state_dict[key.replace("ip_adapter.", "")] = value
+            # Handle different possible key formats
+            if "image_proj" in key:
+                # Remove any prefix before image_proj
+                clean_key = key.split("image_proj.")[-1] if "image_proj." in key else key
+                image_proj_state_dict[clean_key] = value
+            elif "ip_adapter" in key or "to_k_ip" in key or "to_v_ip" in key:
+                # IP adapter weights might not have prefix
+                if "ip_adapter." in key:
+                    clean_key = key.replace("ip_adapter.", "")
+                else:
+                    clean_key = key
+                ip_adapter_state_dict[clean_key] = value
+        
+        print(f"[DEBUG] Found {len(image_proj_state_dict)} image_proj weights")
+        print(f"[DEBUG] Found {len(ip_adapter_state_dict)} ip_adapter weights")
         
         # Create Resampler (image projection model) with CORRECT parameters from reference
         print("Creating Resampler (Perceiver architecture)...")
@@ -220,13 +241,24 @@ def setup_ip_adapter(pipe, image_encoder):
         # Load image_proj weights
         if image_proj_state_dict:
             try:
+                # Try strict loading first
                 image_proj_model.load_state_dict(image_proj_state_dict, strict=True)
-                print("  [OK] Resampler loaded with pretrained weights")
+                print("  [OK] Resampler loaded with pretrained weights (strict)")
             except Exception as e:
-                print(f"  [WARNING] Could not load Resampler weights: {e}")
-                print("  Using randomly initialized Resampler")
+                # Try non-strict if strict fails
+                try:
+                    missing, unexpected = image_proj_model.load_state_dict(image_proj_state_dict, strict=False)
+                    print(f"  [OK] Resampler loaded with pretrained weights (non-strict)")
+                    if missing:
+                        print(f"    Missing keys: {missing[:5]}...")  # Show first 5
+                    if unexpected:
+                        print(f"    Unexpected keys: {unexpected[:5]}...")  # Show first 5
+                except Exception as e2:
+                    print(f"  [WARNING] Could not load Resampler weights: {e2}")
+                    print("  Using randomly initialized Resampler")
         else:
-            print("  [WARNING] No image_proj weights found, using random initialization")
+            print("  [WARNING] No image_proj weights found in state dict")
+            print("  Using randomly initialized Resampler")
         
         # Setup IP-Adapter attention processors
         print("Setting up IP-Adapter attention processors...")