I think the main problem is that it seems like you primarily trained on photographs. If I put in a photo, or something like a movie poster, it does a reasonably good job; but with digital art the software fails to identify the correct subject more often than not.