Update app.py

app.py

@@ -1,3 +1,4 @@
+import spaces  # MUST be first, before any CUDA-related imports
 import gradio as gr
 import torch
 from diffusers import (
@@ -15,35 +16,19 @@ import cv2
 from transformers import pipeline as transformers_pipeline
 from huggingface_hub import hf_hub_download
 import os
-import spaces
 
 # Configuration
 MODEL_REPO = "primerz/pixagram"
+device = "cuda" if torch.cuda.is_available() else "cpu"
+dtype = torch.float16 if device == "cuda" else torch.float32
 
-# Note: For ZeroGPU, device detection happens dynamically
-# We'll set device inside GPU-decorated functions
-print("Using ZeroGPU - GPU will be allocated on-demand")
+print(f"Using device: {device}")
+print(f"Loading models from: {MODEL_REPO}")
 
 class RetroArtConverter:
     def __init__(self):
-        self.models_loaded = False
-        self.device = None
-        self.dtype = None
-        self.face_detection_enabled = False
-        print("RetroArtConverter initialized - models will load on first generation")
-    
-    def _initialize_models(self):
-        """Lazy model initialization - called on first generation when GPU is available"""
-        if self.models_loaded:
-            return
-        
-        print("Initializing models...")
-        print(f"Loading models from: {MODEL_REPO}")
-        
-        # Set device (will be cuda when called from GPU-decorated function)
-        self.device = "cuda" if torch.cuda.is_available() else "cpu"
-        self.dtype = torch.float16 if self.device == "cuda" else torch.float32
-        print(f"Using device: {self.device}")
+        self.device = device
+        self.dtype = dtype
         
         # Initialize face analysis for InstantID (optional)
         print("Loading face analysis model...")
@@ -144,7 +129,7 @@ class RetroArtConverter:
             self.pipe.scheduler.config
         )
         
-        # For ZeroGPU, we don't use model_cpu_offload
+        # For ZeroGPU, don't use model_cpu_offload
         # self.pipe.enable_model_cpu_offload()
         
         self.pipe.enable_vae_slicing()
@@ -160,8 +145,7 @@ class RetroArtConverter:
             except Exception as e:
                 print(f"⚠️ xformers not available: {e}")
         
-        self.models_loaded = True
-        print("✓ Model initialization complete!")
+        print("Model initialization complete!")
     
     def get_depth_map(self, image):
         """Generate depth map from input image"""
@@ -222,9 +206,6 @@ class RetroArtConverter:
     ):
         """Main generation function"""
         
-        # Initialize models on first run (lazy loading for ZeroGPU)
-        self._initialize_models()
-        
         # Resize image maintaining aspect ratio
         original_width, original_height = input_image.size
         target_width, target_height = self.calculate_target_size(original_width, original_height)