Epic's GPUlightmass

that’s horrible… Do you guys report these finding to Nvidia?

Btw, I am still on 516.94 :sweat_smile:

I can confirm 5.1.1 + ERDGInitialDataFlags::None + nvidia 522.25 fixes all my issues.

Edit: The only thing I could find, which is unrelated to previous issues, is this pseudo-decal mesh causes that weird rectangular shadow there.

Just like @nSpeednSpeed I can confirm - also with ERDGInitialDataFlags::None + nvidia 522.25 - the lightmaps look great again without any strange artifacts. Really thanks for so quickly investigating this @Luoshuang.

Edit: once thing I did find while testing is what looks like some sort of lightmap-mip scaling effect on the edges of the lightmap uvs. Especially in VR its very noticable. See pic - when moving more near the dark edges disappear.
Disabling VirtualTextures makes this way less noticable but its still there.

1 Like

I also noticed that when I was testing your scene, and that is basically a ray offset issue - to avoid the decal mesh blocking light the ray origin needs to be pushed away so it starts in the front of the decal. There are two ways to do that

  1. Move the decal closer to the ground (by reducing its Z position by a fraction)
  2. Modify LightmapPathTracing.usf
Ray.Origin += max(ShadowTerminatorFix ? ShadowOffset : BaseOffset, RelativeOffset * 100) * Normal;
Ray.TMax -= length(max(ShadowTerminatorFix ? ShadowOffset : BaseOffset, RelativeOffset * 100) * Normal);

You can bump up the * 100 to 200 or something and restart the editor to see if it resolves the issue. I believe it could be good if ray bias is included in GPULM settings at some point in the future.

2 Likes

While lightmaps have mipmap issues in general, your problem is probably more specific. When testing your scene I noticed that some meshes have override lightmap res which is lower than their min lightmap resolution specified in the static mesh editor. Removing those overrides greatly mitigates the issue:


3 Likes

Unfortunately after testing another map, I think the LOD issue is still there. That driver version probably just hides the bug a bit more.

Things to note

  1. The LOD issue
  2. The road weirdness I think is related to what you said. I didn’t try that code change because I’m a bit short of time ATM, but in this case, would changing that or moving the mesh even help? Or would the blackness just appear somewhere else?
  3. After building the lights, it still says that some lights are unbuilt

Hello, has anyone tested GPULightmass with AMD RX 6xxx or even 7xxx HW how well it works, how stable it is? i wanted to get Nv 4070 Ti, but price insanity put me on hold and to think twice.

1 Like

Not sure if this is a known issue or not. I am using UE5.0.3

Enabling FBRG seems to cause strange noise, even at very high sample counts (same occurs with 35000 samples, 512 RC quality samples, and 3500 FBRG trial samples)


I actually got a response from Luoshuang on this issue a while back when I encountered it:

2 Likes

Ah gotcha. So in this instance I should either vastly increase the FBRG samples or simply disable it for this scene. I’m okay with that. It didn’t seem to vastly affect this particular sublevel anyway.

I can confirm the same artifact issue persists even up to Nvidia driver version 528.02 (WIN 10), returning to version 522.25 per your advice solved all of the GPU Lightmass baking artifacts I was having on UE 5.1, I also applied the aforementioned fix posted above

“Engine\Plugins\Experimental\GPULightmass\Source\GPULightmass\Private\LightmapRenderer.cpp”

Although the Nvidia driver issue was the primary cause of my problem.

As of this post the latest Nvidia drivers (beyond 522.25) cause black baking artifacts on Ceilings and general baking artifacts around the rest of the scene, GI Light bounces in general seem broken on the latest Nvidia drivers as of (1/20/23) and as a result ceilings and other hard to reach places receive no baked data.

(Long winded post to catch key words and phrases I tried to use when looking up this problem, it took way too long to find the fix so hopefully this posts can help with search results)

This tool is tremendously essential for my Team’s project and I’m extremely grateful for it’s development, thank you all and especially thank you for discovering the driver issue. I never would’ve thought that a driver would cause this considering on the same drivers 5.0 seems to bake fine.

I would love a bit more control over lightmap compression.

Turning it off eliminates a LOT of really nasty artifacts, but 4x data size is too much of a price to pay. Perhaps some kind of lossless compression option, or a compression quality setting so we can pick some kind of middle ground.

2 Likes

Does the just now released 5.1.1 hotfix include those improvements? Didn’t see it mentioned in notes.

4 Likes

Yes. And remember to use a driver no newer than 522.25.

3 Likes

Looks like I made a big mistake updating my project to 5.1.1,

Some of my emissive materials do not seem to work at all. I can’t figure out the reason why or a workaround.

5.0.3

5.1.1

My clear glass also doesn’t seem to let any light through

5.0.3

5.1.1

I also seem to have an issue where building reflection captures seems to destroy all the lightmaps. When that happens, I can exit without saving and try again, and then it usually works. I haven’t found the pattern yet, but it is happening to me quite a lot.

I am using Nvidia driver 522.25.

@Luoshuang I just tested with 528.49 Studio Driver - All seems to work fine. Haven’t seen any strange artifacts until now. Anyone want to try and confirm? Using launcher 5.1.1

Edit 230218 It seems there are still problems with this driver in certain cases. See here: Epic's GPUlightmass - #445 by Luoshuang

2 Likes

What about Game Ready driver ?

Haven’t tried it. Why would this be different? Whats the benefit of the Game Ready driver?

The benefit is that Game Ready driver lets you play games with latest fixes and optimizations :wink: Obviously if all you do is work in UE on your PC, it’s irrelevant.

I can confirm that on my end Nvidia Driver version 528.49 (I tested the Game Ready version) appears to work perfectly fine on my 5.1.1 project. GPU Lightmass Bake results are accurate and clean.

Thanks for your update, I have now upgraded to 528.49 and will report any issues here should I come across any.

PS: Luoshuang would know best, as you can see his post below states 528.49 still has issues so as of this moment 522.25 is still the recommended driver for now, although a fix is in the works on Nvidia’s end. I will keep my posts updated regarding this topic.

2 Likes