NVIDIAGameWorks
What is NEE-AT?
Hi!
I'm curious about how NEE-AT works. It seems to perform quite well even in many light scenarios.
What's the overall idea? Is it based on a paper?
Hi Tom,
It's something we're working on - no paper yet unfortunately but we're working on one. The idea is to build dynamic sampling distributions for lights based on feedback of their actual contribution to the image, so it's a learning technique like ReSTIR but simpler and more coarse. In theory it should be able to work as standalone light sampling approach, and/or together with ReSTIR (where it would provide better candidate samples to ReSTIR).
We're still playing with various components and changing things - hope I'll be able to share more info soon!
I see, thanks for your reply. So this is based on some world space spatial data structure to hold the learnt distribution?
Also, practically speaking, since NEE-AT needs to learn from past frames, how can I render 1SPP images with RTXPT in "Reference mode" that use NEE-AT to its full potential (with temporally improved distribution)?
So this is based on some world space spatial data structure to hold the learnt distribution?
It's currently using screen space to store the learnt distribution(s), which is in some ways similar to ReSTIR DI, but I'm working on the world space part for indirect lighting, so in the future both will be used. This is why I don't want to go into any detail for now - it is still changing a lot.
Also, practically speaking, since NEE-AT needs to learn from past frames, how can I render 1SPP images with RTXPT in "Reference mode" that use NEE-AT to its full potential (with temporally improved distribution)?
You can't at the moment, it's intentionally being reset for deterministic behavior reasons. But you can use realtime mode (with all denoising and TAA disabled and with settings matching reference mode) and you should get 1spp with lighting history.
I could add something like a "pre-warm number of frames" for reference mode that would allow update for any temporal technique before starting accumulating frames?
I could add something like a "pre-warm number of frames" for reference mode that would allow update for any temporal technique before starting accumulating frames?
That would be great, thank you!