ComfyStream provides it's own opencv build precompiled for nvidia gpu architectures.
When opencv installs into comfystream (docker only), the script removes any existing versions of opencv and replaces the installation with binaries in the PATH.
|
if [ "$1" = "--opencv-cuda" ]; then |
|
cd /workspace/comfystream |
|
conda activate comfystream |
|
|
|
# Check if OpenCV CUDA build already exists |
|
if [ ! -f "/workspace/comfystream/opencv-cuda-release.tar.gz" ]; then |
|
# Download and extract OpenCV CUDA build |
|
DOWNLOAD_NAME="opencv-cuda-release.tar.gz" |
|
wget -q -O "$DOWNLOAD_NAME" https://github.com/JJassonn69/ComfyUI-Stream-Pack/releases/download/v2.1/opencv-cuda-release.tar.gz |
|
tar -xzf "$DOWNLOAD_NAME" -C /workspace/comfystream/ |
|
rm "$DOWNLOAD_NAME" |
|
else |
|
echo "OpenCV CUDA build already exists, skipping download." |
|
fi |
|
|
|
# Install required libraries |
|
# Remove existing cv2 package |
|
SITE_PACKAGES_DIR="/workspace/miniconda3/envs/comfystream/lib/python3.12/site-packages" |
|
rm -rf "${SITE_PACKAGES_DIR}/cv2"* |
|
|
|
# Copy new cv2 package |
|
cp -r /workspace/comfystream/cv2 "${SITE_PACKAGES_DIR}/" |
|
|
|
# Handle library dependencies |
|
CONDA_ENV_LIB="/workspace/miniconda3/envs/comfystream/lib" |
|
|
|
# Remove existing libstdc++ and copy system one |
|
rm -f "${CONDA_ENV_LIB}/libstdc++.so"* |
|
cp /usr/lib/x86_64-linux-gnu/libstdc++.so* "${CONDA_ENV_LIB}/" |
|
|
|
# Copy OpenCV libraries |
|
cp /workspace/comfystream/opencv/build/lib/libopencv_* /usr/lib/x86_64-linux-gnu/ |
|
|
|
# remove the opencv-contrib and cv2 folders |
|
rm -rf /workspace/comfystream/opencv_contrib |
|
rm -rf /workspace/comfystream/cv2 |
|
|
|
echo "OpenCV CUDA installation completed" |
|
shift |
|
fi |
This is not flexible enough for other applications that require a valid reference to cv2
- ComfyUI-StreamPack needs this package for super resolution
- PyTrickle would benefit from an opencv build with nvidia support
Ideally we can:
- Modify the script so that it installs cv2 (opencv-cuda) using precompiled binaries defined in
pyproject.toml to remove the docker dependency and support the customized opencv install in other applications
- Consider publishing the precompiled wheels to pypip
