Using an model that was trained with random data with different licences is in my opinion more legal dangerous than an library that represents the current status of image processing and you can't patent the current state of the art.