I think you could fix this issue with slightly different training, say for example you label a bunch of images of containers filled to the top with liquid as "full" and "filled" it will then have full images in its database and will know to make the liquid reach that height when you ask for...