Optimize OpenCV FFmpeg on NVIDIA GPU

I’m trying to build a Windows 11 GPU‑accelerated video‑stitching pipeline in Python and have hit a wall with OpenCV’s CUDA modules. My environment is:

Visual Studio 2022, CUDA 12.9, Python 3.11

OpenCV 4.x from source + opencv_contrib, with all WITH_CUDA flags enabled

Repeated compile failures in cudaarithm and cudaimgproc (Thrust “unimplemented for this system,” constexpr errors, linker conflicts)

Conda and pip installs only provide CPU‑only builds, and community CUDA wheels won’t load

I need a reliable way to decode, undistort/warp, detect/match, blend and re‑encode fully on the GPU—either a working OpenCV+CUDA build recipe for Win64 or an alternative GPU pipeline (VPF, PyNvCodec, NPP, etc.). Any guidance or sample configs would be greatly appreciated!