How to deal with using non training data for inference, and the inference results are not realistic enough to restore