chore: add simple image resizing to prevent OOMs

app_local.py

@@ -254,8 +254,59 @@ def infer(
     if device == "cuda":
         torch.cuda.empty_cache()
         gc.collect()
+    
+    # Resize image to max 1024px on longest side
+    def resize_image(pil_image, max_size=1024):
+        """Resize image to maximum dimension of 1024px while maintaining aspect ratio"""
+        try:
+            if pil_image is None:
+                return pil_image
+                
+            width, height = pil_image.size
+            max_dimension = max(width, height)
+            
+            if max_dimension <= max_size:
+                return pil_image  # No resize needed
+                
+            # Calculate new dimensions maintaining aspect ratio
+            scale = max_size / max_dimension
+            new_width = int(width * scale)
+            new_height = int(height * scale)
+            
+            # Resize image
+            resized_image = pil_image.resize((new_width, new_height), Image.LANCZOS)
+            print(f"📝 Image resized from {width}x{height} to {new_width}x{new_height}")
+            return resized_image
+            
+        except Exception as e:
+            print(f"⚠️ Image resize failed: {e}")
+            return pil_image  # Return original if resize fails
+    
+    # Add noise function for batch variation
+    def add_noise_to_image(pil_image, noise_level=0.02):
+        """Add slight noise to image to create variation in outputs"""
+        try:
+            if pil_image is None:
+                return pil_image
+                
+            img_array = np.array(pil_image).astype(np.float32) / 255.0
+            noise = np.random.normal(0, noise_level, img_array.shape)
+            noisy_array = img_array + noise
+            # Clip values to valid range
+            noisy_array = np.clip(noisy_array, 0, 1)
+            # Convert back to PIL
+            noisy_array = (noisy_array * 255).astype(np.uint8)
+            return Image.fromarray(noisy_array)
+        except Exception as e:
+            print(f"Warning: Could not add noise to image: {e}")
+            return pil_image  # Return original if noise addition fails
+    
+    # Resize input image first
+    image = resize_image(image, max_size=1024)
+    
     original_prompt = prompt
     prompt_info = ""
+    
     # Handle prompt rewriting
     if rewrite_prompt:
         try:
@@ -293,22 +344,6 @@ def infer(
             f"</div>"
         )
     
-    # Add noise function
-    def add_noise_to_image(pil_image, noise_level=0.02):
-        """Add slight noise to image to create variation in outputs"""
-        try:
-            img_array = np.array(pil_image).astype(np.float32) / 255.0
-            noise = np.random.normal(0, noise_level, img_array.shape)
-            noisy_array = img_array + noise
-            # Clip values to valid range
-            noisy_array = np.clip(noisy_array, 0, 1)
-            # Convert back to PIL
-            noisy_array = (noisy_array * 255).astype(np.uint8)
-            return Image.fromarray(noisy_array)
-        except Exception as e:
-            print(f"Warning: Could not add noise to image: {e}")
-            return pil_image  # Return original if noise addition fails
-    
     # Set base seed for reproducibility
     base_seed = seed if not randomize_seed else random.randint(0, MAX_SEED)