From the 2.5 docs -> "What's new/changed in LLBLGen Pro v2.5?"
The performance of fetching entities has been improved with as much as 30% compared to v2.0 speeds. Also the memory footprint has been made smaller and less objects are allocated behind the scenes which makes it more efficient for the Garbage collector.
Another added feature is execluding fields from being fetched:
It's now possible for an entity fetch to exclude fields or specify the subset of fields to fetch for an entity, a set of entities or a prefetch path node. This offers the ability to for example not load a large data field for an entity or set of entities to speed up performance. You can then load the data for the excluded fields later on into the same entity objects using specialized methods on (SelfServicing) the entity or entity collection or (Adapter) the DataAccessAdapter classes. This also gives the ability to specify that for example only field A, B and C should be fetched for entity E.
I strongely recommend upgrading to v.2.6, try out the demo version first and do some benchmarking yourself, you'll see the difference.
