diff --git a/source/filters/filter-nv-face-tracking.cpp b/source/filters/filter-nv-face-tracking.cpp
index dfc20aa..d9b6da8 100644
--- a/source/filters/filter-nv-face-tracking.cpp
+++ b/source/filters/filter-nv-face-tracking.cpp
@@ -606,7 +606,7 @@ bool face_tracking_instance::button_profile(obs_properties_t* props, obs_propert
 face_tracking_factory::face_tracking_factory()
 {
 	// Try and load CUDA.
-	_cuda = std::make_shared<::nvidia::cuda::cuda>();
+	_cuda = ::nvidia::cuda::cuda::get();
 
 	// Try and load AR.
 	_ar = std::make_shared<::nvidia::ar::ar>();
diff --git a/source/nvidia/cuda/nvidia-cuda.cpp b/source/nvidia/cuda/nvidia-cuda.cpp
index 0eeca6e..1fc98f2 100644
--- a/source/nvidia/cuda/nvidia-cuda.cpp
+++ b/source/nvidia/cuda/nvidia-cuda.cpp
@@ -18,8 +18,8 @@
  */
 
 #include "nvidia-cuda.hpp"
-#include <stdexcept>
-#include <util/platform.h>
+#include "common.hpp"
+#include <mutex>
 
 #if defined(_WIN32) || defined(_WIN64)
 #define CUDA_NAME "nvcuda.dll"
@@ -121,3 +121,15 @@ nvidia::cuda::cuda::~cuda()
 {
 	os_dlclose(_library);
 }
+
+std::shared_ptr<nvidia::cuda::cuda> nvidia::cuda::cuda::get()
+{
+	static std::shared_ptr<nvidia::cuda::cuda> instance;
+	static std::mutex                          lock;
+
+	std::unique_lock<std::mutex> ul(lock);
+	if (!instance) {
+		instance = std::make_shared<nvidia::cuda::cuda>();
+	}
+	return instance;
+}
diff --git a/source/nvidia/cuda/nvidia-cuda.hpp b/source/nvidia/cuda/nvidia-cuda.hpp
index 5ce8022..ff3dbc2 100644
--- a/source/nvidia/cuda/nvidia-cuda.hpp
+++ b/source/nvidia/cuda/nvidia-cuda.hpp
@@ -396,6 +396,8 @@ namespace nvidia::cuda {
 		CUDA_DEFINE_FUNCTION(cuGraphicsD3D11RegisterResource, graphics_resource_t* resource,
 							 ID3D11Resource* d3dresource, std::uint32_t flags);
 #endif
+		public:
+		static std::shared_ptr<cuda> get();
 	};
 } // namespace nvidia::cuda