The following is a brief comparison following a very quick analysis of the PocketWorlds plugin shipped with the Lyra sample, so it could be not accurate.
Accordingly to its author:
[PocketWorlds] makes it easier to setup isolated rendering areas in the same world
This is the exact reason why we created MultiWorld in first instance,
The main feature of PocketWorlds seems to be the creation of thumbnail renders (see class UPocketCapture), apparently of single actors.
PocketWorlds apparently tries to gently implement the classic hack needed to render a 3D model needed for a UI, when you can't leave your main game world: it loads a scene somewhere inside the current game world (hoping it doesn't disturb the normal gameplay); then it plays with some render settings of a USceneCaptureComponent2D and use it to capture the render of the selected actor; then “alpha-masks” the result and returns a few textures that later can be used in the UI.
Usually similar solutions are subject to many limitations and issues (e.g. if you have a dynamic lighting setup, it common to break one of the scenes; finding a good place where to span the loaded scene is often a problem, in particular for multi-player games; audio can leak between the scenes; some physics simulations can produce unexpected results [e.g. with wind simulation]), etc. etc.). The problem is that the other scene is living inside the main game world, and so it shares and interact with many things within it.
With MultiWorld instead, you ends with really independent UWorld istances, so none of the above issues is a problem. And you can also implement custom interactive gameplay, since they're ordinary Unreal Engine maps.
Also for UI needs, MultiWorld enable artists to produce the desired results with no compromises.