Please see the following draft BIP which should decrease the amount of
bytes needed per transaction. This is very much a draft BIP, as the design
space for this type of improvement is large.

This BIP can be rolled out by a soft fork.

Improvements are around 12% for  standard "one in two out" txn, and even
more with more inputs hashes.

https://gist.github.com/JeremyRubin/e175662d2b8bf814a688