Stochastic Progressive Photon Mapping demo

[lycium]

I seem to recall in the original GPUSPPM demo that unrolling an object intersection loop gave like a 4x speedup on nvidia cards.

[StompinTom]

I seem to recall in the original GPUSPPM demo that unrolling an object intersection loop gave like a 4x speedup on nvidia cards.

Easy on my poor brain, Thomas! If it was as easy as unrolling crepes then I'd do it in a second.

[Borgleader]

I seem to recall in the original GPUSPPM demo that unrolling an object intersection loop gave like a 4x speedup on nvidia cards.

Easy on my poor brain, Thomas! If it was as easy as unrolling crepes then I'd do it in a second.

I dont even know wth you guys are talking about

[lycium]

from ompf: http://ompf.org/forum/viewtopic.php?f=10&t=1492#p16550

thachisu is toshiya hachisuka, principal author (i think) of sppm.

[CTZn]

You are missing the point lycium, it's all about frenchy pancakes.



Talking about wich, it's good to read that you are tweaking MLT !

[lycium]

oh i've never had a crepe!

here's a further improved mlt result, 20 mins on a 3.2ghz i7:

[StompinTom]

oh i've never had a crepe!

here's a further improved mlt result, 20 mins on a 3.2ghz i7:

Water_Pillars_Light better mlt.png

This is just MLT on the CPU, right? Looking great!

[PureSpider]

what has changed? I noticed a more multi-coloured look

It does full spectral rendering and motion blur now!

[Zom-B]

We've been improving MLT in Indigo quite a lot recently

Quite hard to tell whats the difference.... maybe you can explain a little more

[CoolColJ]

The way SPPM is used in the GPU Luxrender offshoot (I use it with CPU only), you start off with large round photons, and then progressively gets smaller and more defined over time. You set the inital size, amount of photons and then how quickly the photon size decreases each pass

I think Indigo could use a similar format somehow with MLT
ie start with large maxchange value and have it go down progressively?
Maybe have it as an optional switch

That would get around the splotchy look and slow convergence of having small max change values, but also having more accurate details that the small max change values give

[CoolColJ]

We've been improving MLT in Indigo quite a lot recently

Quite hard to tell whats the difference.... maybe you can explain a little more

getting the caustics to show in the refracted parts of the glass sphere this early - very complex path
Light through sphere - water - floor - water - refraction through sphere etc

[CTZn]

[...]and then progressively gets smaller and more defined over time. You set the inital size, amount of photons and then how quickly the photon size decreases each pass

I think Indigo could use a similar format somehow with MLT
ie start with large maxchange value and have it go down progressively?
Maybe have it as an optional switch[...]

That's one of my early suggestions to Nick, IIRC I was invited to investigate further on the potential. Estimating time frame seems to be the issue here.

Adding a maxchange delta perhaps ? Each mutation would use an arbitrary value within the fixed range, so to discard time concerns... That kind of tweak sounds as experimental as promising.

The way to test its efficiency would be:

A: Render one scene for a fixed duration using a relevant parameter value.
B: Render n times the scene with a variating parameter for duration/n, then merge.

And C: compare visually the merged image in B with the original in A.

[burnin]

Decade later... 9:38 AM · Jul 7, 2021
Post from Otoy, rendered w/ Octane (WIP)

First WIP results of PPM kernel (left) vs. Pathtracing (right) - both equal render times!

[pixie]

I love these back to the future kind of posts! Thanks burnin!

[lycium]

Hope we're all doing ok, nice to see these old topics again