Long term, the result was a 45-minute test suite which spent most of its time setting up records in the DB, and which could not be parallelized except by adding more instances of the real DB (otherwise the tests would interfere with one-another's assertions about the DB state).
To provide a counterpoint though, an advantage touted by advocates of complete decoupling from the DB is that your UT suite can run in O(one minute), rather than O(ten minutes). E.g. see https://www.youtube.com/watch?v=tg5RFeSfBM4.
I'm not 100% sold on this approach yet (separating your domain objects entirely from the ORM wrapper is an uphill struggle), but it's interesting.
Also since we strive to make our databases upgradable, it's important that the actual schema update scripts themselves are tested and used directly.