1. Duplicity works with a "full plus increments" model, whereas Tarsnap's archives are deduplicated snapshots. If you've been doing daily backups for a year, getting your data back using Duplicity means downloading a full backup followed by 365 increments; to get your data back using Tarsnap you directly download whichever archive you want.

2. Related to the above, Tarsnap's snapshot model allows you to delete archives independently; with Duplicity, you can't delete an archive if you want to keep increments which depend on it.

3. Tarsnap has built-in support for restricted keys; this allows you to set up a server which can create new archives but can't read or delete them. (With a bit of effort you can create "user" keys in AWS which can do S3 PUTs but not GETs or DELETEs; but if your server is compromised that wouldn't stop someone from overwriting your backups.)

4. Duplicity uses librsync, which is considerably more memory-hungry. It's possible that they worked around this by limiting its matching-block search to individual files -- I haven't looked at their code -- but that would result in it failing to recognize moved or copied files.

5. I haven't run benchmarks, but algorithmically speaking Tarsnap should be significantly faster at creating archives.