Realized the 2nd screenshot i sent was just the lumens were off for some reason… set to unitless messed things up, but the other one with the street light… not baking at all… so still issues in that level… level does have sublevels if that matters at all.
Probably best to report it as a bug and submit project files with the bug ticket
ive been told that 1) GPU Lightmass shares code with the path tracer, and we’re limited to 128 lights now 2) Due to the same reason, non inversed square falloff lights are not supported
test in 4096 ! detail seem not very good , but all scene build can work fast in 15-20 min in 4096/4096 sample ! 2048 can do preview!
That’s good to know, but it still doesn’t explain all the issues with spotlights, as I’m still seeing vanishing lights in scenes with 1 spotlight that is set to use inverse square falloff.
Try turning lightmap compression off in the world panel. I don’t know that this is directly related to GPU Lightmass but the compression seems to be absolutely brutal on them in this release. I suspect it might have more to do with compression as related to virtual texture lightmaps, which weren’t really testable in 4.25 because they never worked correctly.
edit: Also are you using irradiance caching?
CPU/GPU comparison, definitely seems to be some significant missing features here.
Scene Setup: Stationary directional light (area shadows enabled), static skylight, static spotlight with 25 outer cone angle. Using raytraced AO. All settings are otherwise default, no lightmass portals
Honestly not really clear on what the irradiance cache settings do so I may have picked some ****** ones. Some immediately noticeable things:
- Temperature of the scene is completely different, not sure of the cause but I suspect GPU LM doesn’t sample the color of the skylight correctly.
- Static spotlight completely disappears when GPU rendered, I believe its intensity is being scaled by its outer cone angle.
- Behavior of area shadows is completely different, not sure it does anything with GPU LM.
WARNING: Beware of using HRDI with very directional light for SkyLight cause I found that since very very old days ( 4.17+ or maybe form the beginning ) SkyLight object has a bug that causes that the HDRI light is X-MIRRORED according to the HRDI image !
Somehow, using the new GPU Lightmass in 4.26, I got much worse results than with Luoshuang’s solution in 4.25… In the 4.26 lighting is not that bad overall, but I got very bad and visible light seams and light leaks. I’m testing with the same geometry and the same hdri… I’m a bit confused and frankly, disappointed.
I can accept a little noise but seams & leaks are disqualifying factors.
i cant even get 90% of my lights to work in the new gpu lightmass…
Seems like it is better to just turn the denoiser off right now (this is a wall of modular tiles)
Comparison without lightmap compression:[SPOILER]
Good catch Arkiras, seems that denoising “trick” exaggerates UE4 problems with lightmap seams - which is logical.
Unfortunately, I must work with datasmith CAD data and I must rely on automatic lightmap UV’s. With the normal CPU baking I can crank up quality settings pretty crazy and have tiny, barely noticeable seam. Out of curiosity I must try 4.26 GPU Ligthmass with GI samples like 32 768 - but this is not a reasonable solution to the problem because GPU had to make things faster.
Another thing, that in Luoshuang’s solution lightmap seams was unnoticeable, so it was promising.
just use this lightmass setting in GPU lightmass bake!
Enjoy and cheers!
tc22, are those Lightmass settings doing anything to GPU bake? It doesn’t seem to be doing anything, however, I tried your settings and this is what I received in my test scene.
https://i.ibb.co/Z6XfprB/1.jpg https://i.ibb.co/1Mhhfzz/2.jpg https://i.ibb.co/Ychk0Dn/3.jpg https://i.ibb.co/q5hgLxg/4.jpg
found another bug… Gpu lightmass on alpha geometry casting inverted shadows. As in the transparent parts block light… and the the opaque parts on the texture let light pass through instead of casting shadows.
Doesnt look like inner cone angle works in gpu lightmass.
source radius on spot light doesnt work in gpu light mass.