Update models.py

models.py

@@ -218,20 +218,64 @@ def load_image_encoder():
 def load_sdxl_pipeline(controlnets):
     """Load SDXL checkpoint from HuggingFace Hub."""
     print("Loading SDXL checkpoint (horizon) with bundled VAE from HuggingFace Hub...")
+
+    # --- START FIX ---
+    # Load tokenizers and text encoders from the base model first
+    # This guarantees they exist, even if the single file doesn't have them
+    print("  Loading base tokenizers and text encoders...")
+    BASE_MODEL = "stabilityai/stable-diffusion-xl-base-1.0"
+    
+    try:
+        tokenizer = CLIPTokenizer.from_pretrained(BASE_MODEL, subfolder="tokenizer")
+        tokenizer_2 = CLIPTokenizer.from_pretrained(BASE_MODEL, subfolder="tokenizer_2")
+        
+        text_encoder = CLIPTextModel.from_pretrained(
+            BASE_MODEL, subfolder="text_encoder", torch_dtype=dtype
+        ).to(device)
+        
+        text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(
+            BASE_MODEL, subfolder="text_encoder_2", torch_dtype=dtype
+        ).to(device)
+        print("  [OK] Base text/token models loaded")
+
+    except Exception as e:
+        print(f"  [ERROR] Could not load base text models: {e}")
+        print("  Pipeline will likely fail. Check HF connection/model access.")
+        # Allow it to continue, but it will likely fail below
+        tokenizer = None
+        tokenizer_2 = None
+        text_encoder = None
+        text_encoder_2 = None
+    # --- END FIX ---
+
     try:
         model_path = download_model_with_retry(MODEL_REPO, MODEL_FILES['checkpoint'], repo_type="model")
         
+        # --- START FIX ---
+        # Pass the pre-loaded models to from_single_file
         pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_single_file(
             model_path,
             controlnet=controlnets,
             torch_dtype=dtype,
-            use_safetensors=True
+            use_safetensors=True,
+            
+            # Explicitly provide the models
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            
         ).to(device) # This main pipe MUST be on device
+        # --- END FIX ---
+
         print("  [OK] Custom checkpoint loaded successfully (VAE bundled)")
         return pipe, True
+        
     except Exception as e:
         print(f"  [WARNING] Could not load custom checkpoint: {e}")
         print("  Using default SDXL base model")
+        
+        # The fallback logic is already correct
         pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_pretrained(
             "stabilityai/stable-diffusion-xl-base-1.0",
             controlnet=controlnets,
@@ -240,7 +284,6 @@ def load_sdxl_pipeline(controlnets):
         ).to(device) # This main pipe MUST be on device
         return pipe, False
 
-
 def load_loras(pipe):
     """Load all LORAs from HuggingFace Hub."""
     print("Loading all LORAs from HuggingFace Hub...")