Can you summarize the different experiments you've done?

Right now my cache does 32 sample RIS for direct lighting with 1 selected sample to trace, and 1 cosine hemisphere random sample for indirect lighting.

Have you found any improvements?

vMF distributions might interest you dl.acm.org/doi/pdf/10.1...

Or you could do what Lumen did, and build a small octahedral probe at each cell, and then compute the CDF.

* If you're ok only setting it once per frame, you can use a counter per cell initialized to 1, and atomicMac(cell, 1), and then if the result is 0 you write to it
* If you really need multiple updates per frame, you can use a linked list link.springer.com/content/pdf/... or list of buckets per cell

Couple of different solutions:
* If you're ok only setting it once, you can do so once you're the one to initialize the cell github.com/bevyengine/b...

Once again, thank you so much for all the tips!

I tried out the luminance tracking, works great! github.com/bevyengine/b...

Now I might be able to get ReSTIR GI validation rays working and further decrease the lag (before the cache was not stable enough to validate against)

Oh, gotcha! I didn't end up having that problem because I don't store albedo per cache cell, I always just use the per-ray albedo. Gives pretty much the same result (no denoiser because I'm on my laptop atm)

Do you mind expanding on the albedo bit as well? Not quite sure I understand how that works.

Thanks for sharing! Luminance tracking should be fairly easy to implement, I'll give that a go tomorrow. And I'll also look into the cache append scheme.

Also glad you enjoyed my article! I've already made a bunch of improvements, so expect a sequel when Bevy 0.18 comes out in ~a month!

Also making a list of cache requests and then updating cells using a random request per cell is interesting. Currently I have a fixed point per cell (position of first request to the cell), which can lead to some ugly artifacts if it poorly represents the cell.

jms55.github.io/posts/2025-0...

2. How do you sample direct and indirect lighting per cell? For direct lighting, is it just RIS? And BRDF sampling for indirect lighting?

3. Do you treat cache cells as always diffuse surfaces, or do you account for the specular BRDF when querying the cache?

Great post! Some questions on your world cache:
1. Assuming you have a temporal blend factor to accumulate radiance per cell, how do you get both stability, and adapting to light changes quickly? Also, do you separate direct and indirect accumulation?

With any luck, I should have some more time to work on Bevy soon, and we should be able to catch up and process Zorah. My PC has "only" 16 GB of RAM. Lots of work to do to get to the point we can process, load, and render it!

Great post! Bevy is on a fairly old version of meshoptimizer at this point, we're overdue in updating our from_mesh() algorithm to take advantage of the new simplification and possibly clusterization code (would have to see how it compares to METIS).

That said, please do try it out and let me know what the limitations are! (keeping in mind it's experimental and not production ready).

I'm a developer, not a designer, so I've only been testing it on a few static scenes. If you try it, please let me know how it fares for real use cases!

* No skinned meshes or morph targets
* No alpha masked or transparent material
* No support for specular materials
* You need a NVIDIA RTX GPU (for DLSS-RR)

All things I hope to solve eventually, but not in this release.