Are combinatory logic terms always larger?

Question

So there is an algorithm to convert lambda calculus terms to combinatory logic using SK combinators. It produces things that explode in size. I would like to know more about this explosion in size. I can't seem to think of a better algorithm however. I have heard of functional languages being practically compiled to combinators so it seems that a better algorithm must exist. I looked up David Turner's paper on the topic and he basically just says to apply a few optimizations and that they cause a "considerable improvement".

Does "considerable improvement" mean that the size drops to only a polynomial increase? Is there a known way to convert lambda terms to combinatory logic with only a polynomial (or less?) increase in size? If such an algorithm exists is it practical?