diff --git a/vr180_matting/sam2_wrapper.py b/vr180_matting/sam2_wrapper.py
index d2134ad..7cc2c41 100644
--- a/vr180_matting/sam2_wrapper.py
+++ b/vr180_matting/sam2_wrapper.py
@@ -5,6 +5,8 @@ import cv2
 from pathlib import Path
 import warnings
 import os
+import tempfile
+import shutil
 
 try:
     from sam2.build_sam import build_sam2_video_predictor
@@ -33,6 +35,7 @@ class SAM2VideoMatting:
         self.predictor = None
         self.inference_state = None
         self.video_segments = {}
+        self.temp_video_path = None
         
         self._load_model(model_cfg, checkpoint_path)
     
@@ -68,18 +71,46 @@ class SAM2VideoMatting:
         except Exception as e:
             raise RuntimeError(f"Failed to load SAM2 model: {e}")
     
-    def init_video_state(self, video_frames: List[np.ndarray]) -> None:
+    def init_video_state(self, video_frames: List[np.ndarray] = None, video_path: str = None) -> None:
         """Initialize video inference state"""
         if self.predictor is None:
             raise RuntimeError("SAM2 model not loaded")
         
-        # Create temporary directory for frames if needed
-        self.inference_state = self.predictor.init_state(
-            video_path=None,
-            video_frames=video_frames,
-            offload_video_to_cpu=self.memory_offload,
-            async_loading_frames=True
-        )
+        if video_path is not None:
+            # Use video path directly (SAM2's preferred method)
+            self.inference_state = self.predictor.init_state(
+                video_path=video_path,
+                offload_video_to_cpu=self.memory_offload,
+                async_loading_frames=True
+            )
+        else:
+            # For frame arrays, we need to save them as a temporary video first
+            
+            if video_frames is None or len(video_frames) == 0:
+                raise ValueError("Either video_path or video_frames must be provided")
+            
+            # Create temporary video file
+            temp_dir = tempfile.mkdtemp()
+            temp_video_path = Path(temp_dir) / "temp_video.mp4"
+            
+            # Write frames to temporary video
+            height, width = video_frames[0].shape[:2]
+            fourcc = cv2.VideoWriter_fourcc(*'mp4v')
+            writer = cv2.VideoWriter(str(temp_video_path), fourcc, 30.0, (width, height))
+            
+            for frame in video_frames:
+                writer.write(frame)
+            writer.release()
+            
+            # Initialize with temporary video
+            self.inference_state = self.predictor.init_state(
+                video_path=str(temp_video_path),
+                offload_video_to_cpu=self.memory_offload,
+                async_loading_frames=True
+            )
+            
+            # Store temp path for cleanup
+            self.temp_video_path = temp_video_path
     
     def add_person_prompts(self, 
                           frame_idx: int, 
@@ -231,6 +262,16 @@ class SAM2VideoMatting:
             
             self.inference_state = None
         
+        # Clean up temporary video file
+        if self.temp_video_path is not None:
+            try:
+                if self.temp_video_path.exists():
+                    # Remove the temporary directory
+                    shutil.rmtree(self.temp_video_path.parent)
+                self.temp_video_path = None
+            except Exception as e:
+                warnings.warn(f"Failed to cleanup temp video: {e}")
+        
         # Clear CUDA cache
         if torch.cuda.is_available():
             torch.cuda.empty_cache()