The inference results of the mlflow model and the onnx model created with Azure Machine Learning's AutoML feature are different

Question

I built an instance segmentation model using the AutoML feature of Azure Machine Learning. As a result of the training job, both an mlflow model and an onnx model were generated. When performing inference with these models, the results were significantly different. Could you explain the cause?

If possible, I would like to create an onnx model that produces the same inference results as the mlflow model and deploy it locally.