I'm looking for an explanation and/or example of why using 4 components vs 2 components is important for EVSM. In my limited tests, using 2 component 32 bit gives much nicer results overall vs 4 component 16bit. Dropping from 4 to 2 components does result in a bit more light-leaking, but it doesn't seem bad in the test scenes I'm looking at. It's still hugely improved vs straight VSM. Going down to 16bit however (even 4 components) causes pretty noticeable aliasing, and the required bias seems to result in peter-panning as well.

I remember specifically reading a presentation from Ready at Dawn, where they had to fall back to using 16bit for memory reasons on The Order. Is there a good reason they chose this trade-off, vs dropping down to 2 components?

That's very helpful, thank you.

What kind of values did you end up needing for leak reduction? Is that a value that you think makes sense to expose to artists per light, or did you get away with a hard-coded global value?