Ugh, I think that GPULM cannot derive vertex normals from the mesh and requires normal map to be present, to avoid showing UV seams. CPULM doesn’t seem to have same limitation.
Both are GPULM:
Material without flat normal map (and with Lightmap Directionality setting on):
I think GPULM fcuked me over again!
For comparison (took only whole evening to build Preview Lighting) CPULM:
Is anybody still experiencing overocclusion issues in 5.1.1? I am using NVIDIA driver 536.23
Here is an example of GPULM (8k samples), Loushang and Path Tracer in an exterior scene, with only this mesh and the default atmosphere setup (directional+skylight on static).
GPULM
Anything before 5.4 is cursed 
What makes you assume that 5.4 won’t be cursed either? GPULM has a track record of not working in some deal-breaking way in pretty much every release since it was added to main engine branch. Given all the data points from previous release and extrapolating trajectory from them, it’s pretty much certain GPULM won’t be usable in 5.4 either.
It does feel like it is getting marginally better, but I agree, it seems like a case of a feature becoming usable Soon™. I tested this also in 5.3, same issue, no clue why.
I am guessing that dealing with the quirky Loushang or the slow CPULM are still the only viable production options?
Because someone did test bakes for me using 5.4 branch and ue5-main and everything seemed to work as expected.
I really hope you are right, but I remain skeptical 
So, imagine that - if I disable Skylight (static), VLM works with GPULM:
WTF Epic?! How hard is it to have a brainstorming session where you think of as many possible scenarios as you could for baked lighting and then test it with GPULM ?!
I hope GPULM is not a next-to-be-thrown away plugin, because it’s no fun to wait for ~2 hrs to bake lighting in that scene with CPULM on 20 cores/28 threads CPU.
I am hopeful myself, however, to be fair, I’m pretty sure GPULM worked in 5.3EA and was broken in the final release.
We’ll add another bug to the list.
Any of my meshes that don’t have collision models I use a zero quality fallback mesh, since UE seems to want to keep fallback meshes fully in memory even if they aren’t really being used for anything.
I also use the cvar r.RayTracing.Nanite.Mode=1 which should theoretically load all the nanite meshes as full quality for raytracing and GPULM, but it doesn’t quite seem to work perfectly. I always get some meshes that don’t bake properly.
Usually what I do to resolve this is temporarily change the fallback quality to max on all meshes, then bake, then turn it back down. If I don’t save any levels when I do that, it doesn’t break lighting.
This looks similar to the issue I was dealing with a while ago. See Epic's GPUlightmass - #479 by Ozykz
I had to create a custom engine version with fixes backported from 5.4 (ue5-main branch on github) in order to resolve it. anything newer than 5.1 is essentially a no go for GPULM.
In 5.4 (github from last week), you do see details / variations when switching to detail lighting.
we’ll probably have Meta’s fork of 5.4.0 by Summer, if we are lucky
It’s been awhile and still no fix for static skylight messing up VLMs when lighting is baked with GPULM 
@Matthew.Ivey I see a lot of refactor commits for GPULM, but none of the commits addresses VLM/Skylight issue I keep posting about here 
So far EA5.4 seems to have a functioning GPULM. I have done several test builds and have not seen any major problems.
Hey 
I’ve been periodically checking the thread so don’t worry, I appreciate all the issue reports and feedback! We’re doing what we can to mitigate issues in the future and fix issues from previous regressions however it is quite time consuming.
All I can ask is that everyone continue to keep the issue reporting and feedback/suggestions coming. I can’t promise anything of course, but seriously, we really appreciate it!
Why on Earth does GPULM not support baked AO, when CPULM does?!
Added secondary UV map and baked AO in Substance Painter:
I definitely would love not to have to do this outside of UE.
GPULM isn’t at feature parity, it’s missing a number of things CPULM can do sadly.
The big one I’m waiting for is for GPULM to respect the per-light indirect lighting intensity setting. (I want it specifically for skylights)
