LuxCoreRender Forums

provisory wrote: Sun Mar 03, 2019 12:29 pm Don't you think, an improved, more noise based adaptivity could handle this (too)?

I have just cut (aka biased) the main source for that fireflies, I'm getting quite good results:


with roughness 0.05. This is an interesting debug rendering:


in red are the pixels done with brute force path tracing while in blue the one with PGI cache. I'm eveluating the idea to take in acount not only the distance from the last bounce but also the angles

P.S. ray differentials would be very handy here.

Aren't high gloss (reflective) surfaces producing caustics rather than directly adding to GI (ambient lighting). At least on first couple of bounces over high gloss, then as "caustics" disperse further (more and more, up to a certain value/threshold?) they start adding to GI.

From observations, such situations can be experienced in shade near large water body areas (seaside, lake, river shore) or on film/photography sets. When sun light hits the reflective water surface at just the right angle (fresnel plays a huge roll here) tiny waves act as concave mirrors/reflectors which concentrate light into beams bouncing into shade and over glossy surfaces which propagate that energy further until "concentrated E" gets dispersed enough, adding to GI. Thus, IMO, there's no need to account bounced light from glossy surfaces with values below 0.03.

BTW (for interested) there's no pure glossy surface in existence.
& We all constantly approximate to certain level of tolerance, optimized for intended tasks.

This "cache" development looks very promising.

PS
Oh, yes i forgot about convex glossy/reflective surfaces, which directly add to GI because of light/Energy dispersion. Even, if pure glossy surface like mirror.

Dade wrote: Sun Mar 03, 2019 9:29 pm

Sharlybg wrote: Sun Mar 03, 2019 7:56 pm But just wonder if we cann't pick something fron this to add on top of PGI ? a kind of hybrid ?

The data collected by tracing photons are a gold mine: at the moment I'm in doubt between a new version of visibility map for infinite lights or guided brute force path tracing (we current sutff is done). A newer DLSC version is possible too.

It is a bit funny but now that computing indirect light is insanely fast ... we need new/better ways to accelerate direct light sampling

Yes please, current dlcs produces artifacts and is unreliable. If that would be fixed that would be amazing!

Dade wrote: Sun Mar 03, 2019 9:29 pm The data collected by tracing photons are a gold mine: at the moment I'm in doubt between a new version of visibility map for infinite lights or guided brute force path tracing

Guided brute force path tracing would be unbiased, right?
Maybe an alternative to caustic cache with its grainy<>blurry problem needing manual adjusting.
An unbiased guided render would progressively converge while the direct implementation of a caustic cache needs a proper photon count/distribution managed beforehand because it will not improve once computed.

I have just cut (aka biased) the main source for that fireflies, I'm getting quite good results:

you ____ are___ the___ man

this is art

Ok this is the final result with roughness 0.051:


I have lower the default value of glossiness threshold (i.e. path.photongi.indirect.glossinessusagethreshold) from 0.2 to 0.05. This can significantly boost the performance on some scene with large glossy surfaces.

I have increased the "retrace" parameter (i.e. path.photongi.indirect.usagethresholdscale) from 4.0 to 8.0 (the net result is exactly the same because before was 0 or 1 while now it is a blend between 0 and 1).

I have also introduced a new debug mode:
It shows in red the pixels where brute force path tracing is used and in blue where cache it is:


I pushed the patch few minutes ago. Give it a spin.

Dade, it seems a great result!

I'll check as soon as I can

I have also introduced a new debug mode:

Would be nice to switch between debug mod and normal rendering.

Looks great, what's next? Do you think an autofill and autosize cache would be doable?

lacilaci wrote: Mon Mar 04, 2019 11:22 am Looks great, what's next? Do you think an autofill and autosize cache would be doable?

Yes, it may be the next target: some type of convergence condition.