diff --git a/tools/sam_masks.py b/tools/sam_masks.py
new file mode 100644
index 0000000..93cb4af
--- /dev/null
+++ b/tools/sam_masks.py
@@ -0,0 +1,83 @@
+"""SAM2 mask generation script.
+
+Usage:
+    python tools/sam_masks.py --input video.mp4 --output masks_dir/
+
+Outputs one binary PNG per frame: frame_0000.png, frame_0001.png, …
+Uses center of first frame as positive point prompt, propagates across all frames.
+Requires: torch, segment-anything-2, opencv-python
+"""
+import argparse
+import os
+import sys
+import tempfile
+
+import cv2
+import numpy as np
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--input", required=True)
+    parser.add_argument("--output", required=True)
+    args = parser.parse_args()
+
+    os.makedirs(args.output, exist_ok=True)
+
+    import torch
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    print(f"Using device: {device}", flush=True)
+
+    # Extract frames to temp directory (SAM2 video predictor needs image files)
+    with tempfile.TemporaryDirectory() as frame_dir:
+        cap = cv2.VideoCapture(args.input)
+        if not cap.isOpened():
+            print(f"ERROR: cannot open {args.input}", file=sys.stderr)
+            sys.exit(1)
+
+        total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        idx = 0
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+            cv2.imwrite(os.path.join(frame_dir, f"{idx:04d}.jpg"), frame)
+            idx += 1
+        cap.release()
+
+        print(f"Extracted {idx} frames", flush=True)
+
+        from sam2.build_sam import build_sam2_video_predictor
+
+        predictor = build_sam2_video_predictor(
+            "facebook/sam2-hiera-large",
+            device=device,
+        )
+
+        with torch.inference_mode():
+            state = predictor.init_state(video_path=frame_dir)
+
+            # Center of first frame as positive point prompt
+            cx, cy = width // 2, height // 2
+            _, _, _ = predictor.add_new_points_or_box(
+                inference_state=state,
+                frame_idx=0,
+                obj_id=1,
+                points=np.array([[cx, cy]], dtype=np.float32),
+                labels=np.array([1], dtype=np.int32),
+            )
+
+            for frame_idx, obj_ids, masks in predictor.propagate_in_video(state):
+                # masks shape: (N_objects, H, W) bool tensor
+                mask = masks[0].cpu().numpy().astype(np.uint8) * 255
+                out_path = os.path.join(args.output, f"frame_{frame_idx:04d}.png")
+                cv2.imwrite(out_path, mask)
+                print(f"frame {frame_idx + 1}/{total}", flush=True)
+
+    print("done", flush=True)
+
+
+if __name__ == "__main__":
+    main()