Think of it this way. It generates exactly the code you write. It's nearly a one-to-one translation between the input forms and output code. But that's why it's fast.
If you want to see just how fast, try `time LUMEN_HOST=luajit make -B`
On my box, LuaJIT is able to recompile the entirety of Lumen and run all the tests in 1.3 seconds flat.
As for Urn, I'd be happy to answer specific comparison questions. I've been putting together a little intro guide to Lumen, but for now the best way to jump in is to study test.l and learn by example: https://github.com/sctb/lumen/blob/master/test.l
EDIT: By the way, Lumen runs on Torch too. If you want to do some hardcore machine learning, it's as easy as installing torch and running `LUMEN_HOST=th bin/lumen`.
Try evaluating `(require 'torch)` and `(require 'nn)`, then dive into the Torch guides. You can bring Lumen's full macro system to bear on any ML problem, which greatly simplifies the boilerplate generation involved with standardizing datasets and such.