diff --git a/tensorrt_llm/inputs/utils.py b/tensorrt_llm/inputs/utils.py
index 79804a4115a..6cd97a91b72 100644
--- a/tensorrt_llm/inputs/utils.py
+++ b/tensorrt_llm/inputs/utils.py
@@ -6,7 +6,6 @@
 from urllib.parse import urlparse
 
 import aiohttp
-import cv2
 import numpy as np
 import requests
 import torch
@@ -88,6 +87,9 @@ def load_video(
         format: str = "pt",
         device: str = "cuda") -> Union[List[Image.Image], List[torch.Tensor]]:
 
+    # Keep this import local to avoid importing cv2 if not needed
+    import cv2
+
     assert format in ["pt", "pil"], "format must be either Pytorch or PIL"
 
     # Load video frames from a video file