Next gen lighting technologies - voxelised, traced, and everything else *spawn*

[MfA]

NRC trains from low-res, full-length traced scene. The result(=cache) is a tiny neural network model.

Oh it tries to capture the entire scene in one model? I didn't expect that.

 

[TopSpoiler]

From my findings, current NRC doesn't seem to properly cache narrow occluded area. Hope it will be fixed.

 

[Dictator]

From my findings, current NRC doesn't seem to properly cache narrow occluded area. Hope it will be fixed.
View attachment 11037
View attachment 11038

Do they have a reference path Traced to See how it should Look in ground truth? It would be nice If they did.

Also AFAIK, NRC should also do specular bounces. You See good evidence of that?

 

[Ext3h]

Is SHaRC spatial only, or direction aware? Asking because the demo case appeared to have painfully avoided geometry where different lighting conditions would have applied to the front/back facing side a geometry thinner than the voxel size.

Simply didn't bother yet to take a look at the actual hash function, so did someone already check that?

I do suspect it doesn't, so you'd always get an unintended bleed through thin occluders for the parts of GI which had not been sampled in screen space.

 

[DavidGraham]

Alex was wondering if DDGI is still there, it is.

https://twitter.com/x/status/1770371538757272009

It has 3 techniques tight now, NRC which uses Tensor cores, SHaRC and DDGI use compute.

NVIDIA RTXGI 2.0 Now Available: The Next Frontier of Ray Traced Visuals With Neural Radiance Cache, Spatial Hash Radiance Cache & Dynamic Diffuse Global Illumination Support

NVIDIA updated its ray tracing global illumination SDK to the latest feature set in RTXGI 2.0, offering support for new technologies.

wccftech.com

 

[TopSpoiler]

Do they have a reference path Traced to See how it should Look in ground truth? It would be nice If they did.

Yes. More detailed images I took with sample accumulation on and denoiser off:

Also AFAIK, NRC should also do specular bounces. You See good evidence of that?

I think it does.

Is SHaRC spatial only, or direction aware? Asking because the demo case appeared to have painfully avoided geometry where different lighting conditions would have applied to the front/back facing side a geometry thinner than the voxel size.

Perhaps the 'voxel' in SHaRC is not a common word meaning 'volume pixel'.

 

[MfA]

I'd be surprised if they aren't using surfels for SHaRC, that will handle two sides just fine. The Neural net will find it difficult to learn surface orientation from just the sparse samples AFAICS.

I think they will have to compromise on the giant neural model approach and do something hybrid.

PS. oh, it's really voxels. This seems a lot less advanced than Dice's surfel GI.

 

[MfA]

A bit, odd. Cached radiance transfer aint exactly new, and a major point of these new GI techniques is you don't have to bake anything. Look instant feedback and dynamic gameplay! Getting rid of all that so you can add "neural" as a bullet point seems silly.

The tensor cores are there any way, though trying to capture the radiance cache in one model is a bit silly if that's still what they are doing. That feels a bit like DLSS 1, a very sloppy way to satisfy a mandate to leverage the tensor cores.

MLPs/SVMs can serve as a way to compress and interpolate between non-uniform samples, but it makes more sense to do that localized in a scene for instance for a surfel hemisphere or for prefiltered view dependent voxels.

 

[Ext3h]

I'd be surprised if they aren't using surfels for SHaRC, that will handle two sides just fine.

SHARC/shaders/Include/HashGridCommon.h at 180f825f5fbae03aaaa593727d80d1094f913f00 · NVIDIAGameWorks/SHARC

Spatially Hashed Radiance Cache (SHaRC) Library . Contribute to NVIDIAGameWorks/SHARC development by creating an account on GitHub.

github.com

Didn't use surfels. Just a subdivision of the voxel into 8 axis aligned primary normal orientations. And it's simply recording voxels in world coordinates. With no extra metadata for depth, radius or anything, nor even the option to dynamically increase resolution where potentially needed.
So it does suffer from an well known issue:

Also didn't properly account for hash collisions! So this will occasionally resolve to completely wrong hash entries. Partially self inflicted issue, as it used the same bits of entropy for both the bucket and key inside the bucket, could have further reduced the chance of collisions at the same space requirements significantly.

And don't get me started on the complete lack of any locality. It's 32bit random reads all over the place.

 

[MfA]

I've seen a paper make the argument the neural network will clean up the collissions with a little dither so you don't always hit the same bucket in a region. Though that would mean Sharc needs ray reconstruction.