PhotonGI cache

[Dade]

I tried the B.Y.O.B.'s idea of having a mix over glossy between brute force and cache. This is a test with a sharp transition (just immagine to add a blend between the 2 instead of a sharp one):

and this the same scene after having moved the camera around:

As you see, the size of brute force area changes because changes the view angle. It is quite worrying for animations. I will try if it can work with a veeeeery large transition area.

[lacilaci]

Do you think it could give nice performance boost with scenes full of glossy materials?

[andreymd87]

yep!! last build works perfectly.
no dark area. and black walls

looks really nice

[Dade]

Do you think it could give nice performance boost with scenes full of glossy materials?

A bit, it allows to lower the "glossiness" threshold from 0.2 to 0.025 in my current tests so the cache will be used quite more often on glossy materials.

To recap, this is brute force only:

this is cache only:

this is with mix with a sharp transition:

and this is mix with a blend transition:

The above test have been done with a glossiness" threshold of 0.0 and a glossy roughness of 0.001 (pretty much the most nearly specular surface you can have).

My current problem is more with the little fireflies caused by brute force over glossy:

This is a test with roughness of 0.025 and works well from the reflection point of view.

[lacilaci]

wow, that looks great

[provisory]

My current problem is more with the little fireflies caused by brute force over glossy

Don't you think, an improved, more noise based adaptivity could handle this (too)?

[Sharlybg]

I remenber when we were discussing about cache feature:

No because of GPUs. It will be built during a pre-processing stage so it can be transferred to the GPUs.

excuse me if i'm insisting like this but i've more question

If we take CPU only as reference (no GPU acceleration).

1) Do we lose a lot of performance compared to online method ?

On-line training is better because the quality of the light cache improve over the rendering time. The quality of light cache built during a pre-processing phase is constant over the rendering time. CPU Vs. CPU, a light cache with on-line training is better than one without.

About a comparison between photon based caching vray like and denoise algorithme cache "Corona paper". corona online was slower and wasn't really interesting for GPUs but was efficient for denoising any kind of noise ( specular/diffuse/reflection) and could help speed up Bidir engine. finally the current PGI is the best bet anyway .

But just wonder if we cann't pick something fron this to add on top of PGI ? a kind of hybrid ?

[Dade]

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

[B.Y.O.B.]

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

By the way, have you seen this article in Matt Pharr's blog? https://pharr.org/matt/blog/2019/02/27/ ... ing-1.html

[Dade]

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

By the way, have you seen this article in Matt Pharr's blog? https://pharr.org/matt/blog/2019/02/27/ ... ing-1.html

I remember also some stuff about spherical light source sampling (probably written by Iliyan Georgiev). I have also some memory of papers about the topic of sampling triangles by Solid Angle people. All topics that may be worth exploring after PGI.