Fog performance

[anan_vozak]

Hi, it would be really stunning to have similar speed in Vantage when ReSTIR is on, cause now a light tree with restir only decreases the quality and quite frequently, aside from low sampling, gives some blocky artifacts (which disappear when disabling the light tree). Currently, fog doesn't properly work for final renders (it is impossible to render it clean), nor for realtime preview (DLSS doesn't know how to properly denoise scenes with fog, but Optix denoiser does it mid)

[npg]

I was able to reproduce some splotches when using reservoir resampling and scattering fog in Vantage. But you said light tree is the issue, which I could not reproduce. I've seen some unusual noise patterns developing when using more than 1 sample with the light tree, but it wasn't related to fog and it generally cleans up over time. I'm not sure if we can agree that fog doesn't work for final renders - we've been using it for that. Maybe there is an issue with a specific combination of settings and/or scene. DLSS seems to have the same problem as reservoir resampling - the scattering fog leads to noise in the inputs it uses like normals, positions, etc. and it's not meant to receive such noisy g-buffers.

[anan_vozak]

I was able to reproduce some splotches when using reservoir resampling and scattering fog in Vantage. But you said light tree is the issue, which I could not reproduce. I've seen some unusual noise patterns developing when using more than 1 sample with the light tree, but it wasn't related to fog and it generally cleans up over time. I'm not sure if we can agree that fog doesn't work for final renders - we've been using it for that. Maybe there is an issue with a specific combination of settings and/or scene. DLSS seems to have the same problem as reservoir resampling - the scattering fog leads to noise in the inputs it uses like normals, positions, etc. and it's not meant to receive such noisy g-buffers.

I apologize if it sounded rude. I didnt mean that fog works really bad, or some of my complaints are issues by themselves, but that fog currently is quite unreliable and hard to deal with and light tree and restir only making it worse. Maybe it's a user error, but in scenes with tiny lights, i always get results like this, where even 2400 passes are not enough to achieve clean renders (examples done in fullhd, but it's still noticeable in more hires renders). All these examples were done without fog gi scattering. Some blocky artifacts occur very rarely, and i cant reproduce it now, so I am attaching a render from a few years ago, but it still sometimes happens (but I am unsure, cause i dont render fog with light tree and restir usually). Since I cant reproduce it, maybe it was fixed or something, so lets ignore it. So I understand why dlss and restir fails with fog, but just wish that fog sampling could be improved in future, to be able to achieve clearer results, like in the link above. Ofc example from the link is just a demo of an unreleased pathtraced engine and I aint a programmer to understand why its so fast compared to Vantage

[npg]

I don't think I've seen these rectangular artifacts, though it could be a really bad case of the pattern that sometimes shows up when using >1 light tree sample.

I hope we can improve the sampling at some point, be it with restir or something else. But it's not a high priority currently.

[anan_vozak]

Hi again! I forgot to ask about Area ReSTIR: are there any plans on updating ReSTIR PT to a new one or is this paper not ready for implementation yet?

[npg]

We're currently using Restir only for DI and we currently don't have specific plans for implementing more "flavors".