diff --git a/scripts/gpu_video_demo.py b/scripts/gpu_video_demo.py
new file mode 100644
index 00000000..366465ee
--- /dev/null
+++ b/scripts/gpu_video_demo.py
@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+import os
+import time
+import argparse
+
+import cv2
+import torch
+import numpy as np
+
+# --- Replace this import with your repo’s actual model API ---
+# e.g. from sam2.model import SamuraiModel
+from sam2.samurai import SamuraiModel
+
+def parse_args():
+    p = argparse.ArgumentParser(
+        description="Interactive GPU/CPU demo: draw a box, then segment next 5 frames."
+    )
+    p.add_argument(
+        "--video",
+        type=str,
+        default=r"C:\Users\vansh\OneDrive\Documents\app proj 1\app project 11\samurai2\myvid1.mp4",
+        help="Path to input video file",
+    )
+    p.add_argument(
+        "--outdir",
+        type=str,
+        default="outputs",
+        help="Directory to save mask images",
+    )
+    return p.parse_args()
+
+def main():
+    args = parse_args()
+
+    # 1) Open video
+    cap = cv2.VideoCapture(args.video)
+    if not cap.isOpened():
+        raise FileNotFoundError(f"Cannot open video: {args.video}")
+
+    # 2) Read first frame & let user draw a box
+    ret, first = cap.read()
+    if not ret:
+        raise RuntimeError("Failed to read first frame")
+    bbox = cv2.selectROI("Draw bounding box", first, False, False)
+    cv2.destroyWindow("Draw bounding box")
+
+    # 3) Prepare device
+    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    print(f"[INFO] Using device: {device}")
+
+    # 4) Load model and move to device
+    model = SamuraiModel()            # adjust constructor as needed
+    model.to(device)
+    model.eval()
+
+    # 5) Ensure output dir exists
+    os.makedirs(args.outdir, exist_ok=True)
+
+    # 6) Segment next 5 frames
+    for i in range(1, 6):
+        ret, frame = cap.read()
+        if not ret:
+            print(f"[WARN] Could not read frame {i}; stopping early.")
+            break
+
+        start = time.time()
+
+        # --- preprocess: convert BGR→RGB, HWC→CHW, normalize if needed ---
+        img = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+        tensor = (
+            torch.from_numpy(img)
+            .permute(2, 0, 1)
+            .unsqueeze(0)
+            .to(device)
+            .float() / 255.0
+        )
+
+        # 7) Run inference with the initial box as prompt
+        with torch.no_grad():
+            # adjust call signature to your repo’s API:
+            mask = model.predict(tensor, box=bbox)  
+            # expect `mask` as a torch tensor [1,H,W] or numpy array
+
+        # 8) Post-process & save mask
+        if isinstance(mask, torch.Tensor):
+            mask = mask.squeeze(0).cpu().numpy()
+        mask_img = (mask * 255).astype(np.uint8)
+        out_path = os.path.join(args.outdir, f"frame_{i}.png")
+        cv2.imwrite(out_path, mask_img)
+
+        elapsed = (time.time() - start) * 1000
+        print(f"[INFO] Frame {i} → saved to {out_path}  ({elapsed:.1f} ms)")
+
+    cap.release()
+    print("[DONE] Segmentation complete.")
+
+if __name__ == "__main__":
+    main()