Thanks for the quick response - much appreciated!
I’ve kept a close eye on Lumen for a while and even before 's presentation I went through his Siggraph document. From my understanding Lumen takes a hybrid approach about irradiance cache where long distance traces use a more traditional fixed grid probe layout in world space and then the second one uses layout that’s adaptive to surfaces on the screen. My question was more relating to the former.
The idea is that if there’s part of the computation that is both more expensive in terms of memory and/or compute but also changes less infrequently and produces smaller amount of data as it’s output, then that portion of the work can be offloaded and delta-streamed.
And with that in mind, the world-space cache seems like a good fit. I think of it more as a 3D grid of discrete data points than a monolithic 2D atlas (though I know that’s how it typically gets stored), so most areas in the 3D grid may change very infrequently, so they could be streamed individually such that the whole atlas updates incrementally.
Here’s a timed-link to OTOY’s 2021 presentation that shows what “seems” to be what I’m talking about:
They don’t go into details, but it suggests that there’s an irradiance cache offloaded to their RNDR distributed network but then the device uses that cache for lighting when rendering locally. I may have misinterpreted what they said they’re doing though, given how light on details they were. I did reach out to OTOY as well to clarify.
For that kind of a use-case, screen-traces would be very limited as there’s only a single object in view that would need them, and it’s screen coverage would be relatively small. If screen-traces density and/or accuracy could be tuned down that could also help for rendering on the device.
Then for the lighting from world-space probes, that would update only as the scene representation around the object changes, like in the video when an occluding object comes in (the hand) or when emitting surfaces change (the laptop screen).
Given the small 3D area of interest in this use case, the 3D grid may even be able to be tuned to a higher density while still keeping the incremental update stream bandwidth relatively low.