diff --git a/docker/Dockerfile.multi b/docker/Dockerfile.multi
index e724eb4cc629..65282c2ae406 100644
--- a/docker/Dockerfile.multi
+++ b/docker/Dockerfile.multi
@@ -49,7 +49,7 @@ RUN --mount=type=bind,source=docker/common,target=/opt/docker/common \
     NCCL_VER=${NCCL_VER} CUBLAS_VER=${CUBLAS_VER} \
     TORCH_INSTALL_TYPE=${TORCH_INSTALL_TYPE} \
     bash /opt/docker/common/install.sh --base --cmake --ccache --cuda_toolkit \
-         --tensorrt --polygraphy --mpi4py --pytorch --opencv
+         --tensorrt --polygraphy --mpi4py --pytorch
 
 # Install constraints after install.sh so cleanup() doesn't delete the file mid-RUN
 COPY constraints.txt /tmp/constraints.txt
diff --git a/docker/common/install.sh b/docker/common/install.sh
index 0de962c9afcb..0d12d812a354 100755
--- a/docker/common/install.sh
+++ b/docker/common/install.sh
@@ -15,7 +15,6 @@ tensorrt=0
 polygraphy=0
 mpi4py=0
 pytorch=0
-opencv=0
 
 while [[ $# -gt 0 ]]; do
     case $1 in
@@ -51,10 +50,6 @@ while [[ $# -gt 0 ]]; do
             pytorch=1
             shift 1
             ;;
-        --opencv)
-            opencv=1
-            shift 1
-            ;;
         --all)
             base=1
             cmake=1
@@ -64,7 +59,6 @@ while [[ $# -gt 0 ]]; do
             polygraphy=1
             mpi4py=1
             pytorch=1
-            opencv=1
             shift 1
             ;;
         *)
@@ -122,10 +116,3 @@ if [ $pytorch -eq 1 ]; then
     echo "Installing PyTorch..."
     bash $SCRIPT_DIR/install_pytorch.sh $TORCH_INSTALL_TYPE
 fi
-
-if [ $opencv -eq 1 ]; then
-    echo "Installing OpenCV..."
-    bash -c "pip3 uninstall -y opencv"
-    rm -rf /usr/local/lib/python3*/dist-packages/cv2/
-    bash -c "pip3 install opencv-python-headless --force-reinstall --no-deps --no-cache-dir"
-fi
diff --git a/docs/source/features/multi-modality.md b/docs/source/features/multi-modality.md
index e2be649bf44a..6d6512a38d17 100644
--- a/docs/source/features/multi-modality.md
+++ b/docs/source/features/multi-modality.md
@@ -26,6 +26,16 @@ Further optimizations are under development and will be updated as they become a
 
 Please refer to the latest multimodal [support matrix](../models/supported-models.md#multimodal-feature-support-matrix-pytorch-backend).
 
+## Optional dependencies
+
+Video input decoding uses OpenCV (`cv2`) and is not installed by default. Install it only for models with video inputs:
+
+```bash
+pip install opencv-python-headless
+```
+
+Image-only and text-only workflows do not require this package.
+
 ## Examples
 
 The following examples demonstrate how to use TensorRT LLM's multimodal support in various scenarios, including quick run examples, serving endpoints, and performance benchmarking.
diff --git a/jenkins/L0_Test.groovy b/jenkins/L0_Test.groovy
index 301e6f693fdf..0f7a960ee9e3 100644
--- a/jenkins/L0_Test.groovy
+++ b/jenkins/L0_Test.groovy
@@ -3591,6 +3591,7 @@ def runLLMTestlistOnPlatformImpl(pipeline, platform, testList, config=VANILLA_CO
                 sh "cd ${llmSrc} && sed -i 's#tensorrt~=.*\$#tensorrt#g' requirements.txt && cat requirements.txt"
             }
             trtllm_utils.llmExecStepWithRetry(pipeline, script: "cd ${llmSrc} && pip3 install -r requirements-dev.txt")
+            trtllm_utils.llmExecStepWithRetry(pipeline, script: "pip3 install opencv-python-headless")
             if (stageName.contains("-Ray-")) {
                 trtllm_utils.llmExecStepWithRetry(pipeline, script: "pip3 install ray[default]==2.55.1")
             }
diff --git a/jenkins/current_image_tags.properties b/jenkins/current_image_tags.properties
index b4648ceb7aed..07aa75063b77 100644
--- a/jenkins/current_image_tags.properties
+++ b/jenkins/current_image_tags.properties
@@ -13,8 +13,8 @@
 #     images are adopted from PostMerge pipelines, the abbreviated commit hash is used instead.
 IMAGE_NAME=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm
 
-LLM_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-26.04-py3-x86_64-ubuntu24.04-trt10.16.1.11-skip-tritondevel-202607072201-15907
-LLM_SBSA_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-26.04-py3-aarch64-ubuntu24.04-trt10.16.1.11-skip-tritondevel-202607072201-15907
-LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.2.1-devel-rocky8-x86_64-rocky8-py310-trt10.16.1.11-skip-tritondevel-202607072201-15907
-LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.2.1-devel-rocky8-x86_64-rocky8-py312-trt10.16.1.11-skip-tritondevel-202607072201-15907
-LLM_SBSA_WHEEL_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.2.1-devel-ubuntu24.04-sbsa-ubuntu24.04-py312-trt10.16.1.11-skip-tritondevel-202607072201-15907
+LLM_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-26.04-py3-x86_64-ubuntu24.04-trt10.16.1.11-skip-tritondevel-202607100544-16206
+LLM_SBSA_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-26.04-py3-aarch64-ubuntu24.04-trt10.16.1.11-skip-tritondevel-202607100544-16206
+LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.2.1-devel-rocky8-x86_64-rocky8-py310-trt10.16.1.11-skip-tritondevel-202607100544-16206
+LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.2.1-devel-rocky8-x86_64-rocky8-py312-trt10.16.1.11-skip-tritondevel-202607100544-16206
+LLM_SBSA_WHEEL_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.2.1-devel-ubuntu24.04-sbsa-ubuntu24.04-py312-trt10.16.1.11-skip-tritondevel-202607100544-16206
diff --git a/jenkins/scripts/perf/local/slurm_install.sh b/jenkins/scripts/perf/local/slurm_install.sh
index 8c03385edf41..91bb7d664e62 100755
--- a/jenkins/scripts/perf/local/slurm_install.sh
+++ b/jenkins/scripts/perf/local/slurm_install.sh
@@ -46,6 +46,8 @@ slurm_install_setup() {
 
         echo "(Installing TensorRT-LLM and requirements) Install mode: ${INSTALL_MODE:-source}"
 
+        retry_command pip install --retries 10 opencv-python-headless
+
         # Support two installation modes: source (default) and wheel
         if [ "${INSTALL_MODE:-source}" = "wheel" ]; then
             # Wheel installation mode
diff --git a/jenkins/scripts/slurm_install.sh b/jenkins/scripts/slurm_install.sh
index d1302beb956a..27c591760f5c 100644
--- a/jenkins/scripts/slurm_install.sh
+++ b/jenkins/scripts/slurm_install.sh
@@ -29,6 +29,7 @@ slurm_install_setup() {
         if [[ $pytestCommand == *--run-ray* ]]; then
             retry_command --timeout 2700 pip3 install --retries 10 "ray[default]==2.55.1"
         fi
+        retry_command --timeout 2700 bash -c "pip3 install --retries 10 opencv-python-headless"
         retry_command --timeout 2700 bash -c "cd $llmSrcNode && pip3 install --retries 10 -r requirements-dev.txt"
         retry_command --timeout 2700 bash -c "cd $resourcePathNode && pip3 install --retries 10 --force-reinstall --no-deps TensorRT-LLM/tensorrt_llm-*.whl"
         gpuUuids=$(nvidia-smi -q | grep "GPU UUID" | awk '{print $4}' | tr '\n' ',' || true)
diff --git a/requirements.txt b/requirements.txt
index d14251c5c423..0b0d123d4ced 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -56,7 +56,6 @@ peft>=0.18.1,<0.19.0
 patchelf
 einops
 flashinfer-python==0.6.14
-opencv-python-headless
 xgrammar==0.1.32
 llguidance==0.7.29
 jsonschema
diff --git a/tensorrt_llm/inputs/media_io.py b/tensorrt_llm/inputs/media_io.py
index 8e93784ab650..9d456b23e4f1 100644
--- a/tensorrt_llm/inputs/media_io.py
+++ b/tensorrt_llm/inputs/media_io.py
@@ -32,7 +32,6 @@
 from urllib.parse import unquote, urljoin, urlparse
 
 import aiohttp
-import lazy_loader as lazy
 import numpy as np
 import requests
 import soundfile
@@ -43,11 +42,6 @@
 from tensorrt_llm.inputs.multimodal_data import AudioData, VideoData
 from tensorrt_llm.logger import logger
 
-# Lazy import: OpenCV is large and only needed when the cv2-backed video
-# decode path is exercised. The proxy triggers the actual `import cv2` on
-# first attribute access (e.g. `cv2.VideoCapture`).
-cv2 = lazy.load("cv2")
-
 
 def rgba_to_rgb(
     image: Image.Image, background_color: Union[tuple[int, int, int], list[int]] = (255, 255, 255)
@@ -338,6 +332,18 @@ def extract_audio_from_video(
     return audio, target_sr
 
 
+def _get_cv2():
+    """Import OpenCV on demand for the optional cv2-backed video decode path."""
+    try:
+        import cv2
+    except ImportError as exc:
+        raise ImportError(
+            "OpenCV (cv2) is required for video decoding but is not installed. "
+            "Install it with `pip install opencv-python-headless`."
+        ) from exc
+    return cv2
+
+
 def _select_cv2_stream_buffered_backend() -> Optional[int]:
     """Return a VideoCapture backend that can read from a Python `BytesIO`.
 
@@ -351,6 +357,8 @@ def _select_cv2_stream_buffered_backend() -> Optional[int]:
     open. Built-in backends (the FFMPEG path in the PyPI wheels) are always
     safe to use.
     """
+    cv2 = _get_cv2()
+
     # The stream-buffered API was introduced in OpenCV 4.13.0. Older builds
     # don't have `cv2.videoio_registry.getStreamBufferedBackends`, so signal
     # "no usable backend" and let the caller fall back to the tempfile path.
@@ -420,6 +428,8 @@ def _load_video_by_cv2(
     """
     assert format in ("pt", "np", "pil"), "format must be one of 'pt', 'np', 'pil'"
 
+    cv2 = _get_cv2()
+
     # Open the source. Two cases:
     #   (a) `video` is a file path / URL str -> hand it straight to cv2.
     #   (b) `video` is mp4 bytes -> feed cv2 from an in-memory BytesIO via