Fix GPU inference freezing after first frame

src/models/Model.h

@@ -85,8 +85,8 @@ class Model {
 private:
 	/* data */
 public:
-	Model(/* args */) {};
-	virtual ~Model() {};
+	Model(/* args */){};
+	virtual ~Model(){};
 
 	const char *name;
 
@@ -259,11 +259,12 @@ class Model {
 					 const std::vector<Ort::AllocatedStringPtr> &inputNames,
 					 const std::vector<Ort::AllocatedStringPtr> &outputNames,
 					 const std::vector<Ort::Value> &inputTensor,
-					 std::vector<Ort::Value> &outputTensor)
+					 std::vector<Ort::Value> &outputTensor,
+					 const std::vector<std::vector<int64_t>> &outputDims,
+					 std::vector<std::vector<float>> &outputTensorValues, const std::string &useGPU)
 	{
-		if (inputNames.size() == 0 || outputNames.size() == 0 || inputTensor.size() == 0 ||
-		    outputTensor.size() == 0) {
-			obs_log(LOG_INFO, "Skip network inference. Inputs or outputs are null.");
+		if (inputNames.size() == 0 || outputNames.size() == 0 || inputTensor.size() == 0) {
+			obs_log(LOG_INFO, "Skip network inference. Inputs are null.");
 			return;
 		}
 
@@ -277,6 +278,40 @@ class Model {
 			rawOutputNames.push_back(outputName.get());
 		}
 
+		// For GPU execution providers, we need to recreate output tensors for each inference
+		// to avoid tensor reuse issues that cause processing to stop after one frame
+		// Explicitly check for known GPU providers: cuda, rocm, migraphx, tensorrt, coreml
+		// Note: DirectML is not currently supported in this plugin despite documentation mentions
+		bool requiresTensorRecreation = (useGPU == "cuda" || useGPU == "rocm" || useGPU == "migraphx" ||
+						 useGPU == "tensorrt" || useGPU == "coreml");
+
+		if (requiresTensorRecreation) {
+			// Clear and recreate output tensors for GPU inference
+			outputTensor.clear();
+
+			Ort::MemoryInfo cpuMemoryInfo = Ort::MemoryInfo::CreateCpu(OrtAllocatorType::OrtDeviceAllocator,
+										   OrtMemType::OrtMemTypeDefault);
+
+			for (size_t i = 0; i < outputDims.size(); i++) {
+				outputTensor.push_back(Ort::Value::CreateTensor<float>(
+					cpuMemoryInfo, outputTensorValues[i].data(), outputTensorValues[i].size(),
+					outputDims[i].data(), outputDims[i].size()));
+			}
+		} else {
+			// For CPU mode, output tensors should have been pre-allocated during initialization
+			if (outputTensor.size() == 0) {
+				obs_log(LOG_ERROR,
+					"Output tensor is empty for CPU mode. Tensor allocation may have failed.");
+				return;
+			}
+		}
+
+		// Final defensive check that outputTensor is valid before calling inference
+		if (outputTensor.size() == 0) {
+			obs_log(LOG_ERROR, "Output tensor is empty. Cannot proceed with inference.");
+			return;
+		}
+
 		session->Run(Ort::RunOptions{nullptr}, rawInputNames.data(), inputTensor.data(), inputNames.size(),
 			     rawOutputNames.data(), outputTensor.data(), outputNames.size());
 	}

src/ort-utils/ort-session-utils.cpp

@@ -151,7 +151,8 @@ bool runFilterModelInference(filter_data *tf, const cv::Mat &imageBGRA, cv::Mat
 	tf->model->loadInputToTensor(preprocessedImage, inputWidth, inputHeight, tf->inputTensorValues);
 
 	// Run network inference
-	tf->model->runNetworkInference(tf->session, tf->inputNames, tf->outputNames, tf->inputTensor, tf->outputTensor);
+	tf->model->runNetworkInference(tf->session, tf->inputNames, tf->outputNames, tf->inputTensor, tf->outputTensor,
+				       tf->outputDims, tf->outputTensorValues, tf->useGPU);
 
 	// Get output
 	// Map network output to cv::Mat