I haved converted your model too onnx and implement with ONNXRuntime C++. I seems can get good performance ~ * [fsanet.cpp](https://github.com/DefTruth/lite.ai.toolkit/blob/main/ort/cv/fsanet.cpp) * [fsanet-cpp-demo](https://github.com/DefTruth/lite.ai.toolkit/blob/main/examples/lite/cv/test_lite_fsanet.cpp)