Page 1 of 2

sobol and metropolis

Posted: Fri Aug 10, 2018 11:57 am
by lacilaci
So I haven't seen too much differences in my test when testing sobol and random. One is better in one case but worse in other and vice versa.

However I'm seeing pretty big differences when using metropolis in combination with direct light cache and tweaked settings.
I know that it introduces some bias and it's also visible in a bit different lighting, but it seems (in those limited cases at least) that it is a good tradeoff.
Also not only I get better noise performance using metropolis at 90% mutation and 4 rejects. But also no fireflies and in a moderately clean render also no denoiser artifacts(which show a lot using default 40% mutation probability)

In the results shown here, the interior shows more shitty denoising using metropolis. But these renders would still need more time than 5 minutes for both sobol and metropolis to make denoiser usable.

It was important though to use direct light cache. Cause log power with metropolis will still show some noise in bright areas.

So my question is, are there some known downsides using my settings here? Other than a bit different lighting? Cause if not. It could be pretty impressive to see directcache+indirectcache+metropolis performance against some very complex situations in the future.

Re: sobol and metropolis

Posted: Fri Aug 10, 2018 3:31 pm
by Dade
lacilaci wrote:
Fri Aug 10, 2018 11:57 am
Also not only I get better noise performance using metropolis at 90% mutation and 4 rejects. But also no fireflies and in a moderately clean render also no denoiser artifacts(which show a lot using default 40% mutation probability)
A 90% of large mutation probability means Metropolis works like a pure Random sampler 9 times out of 10 so the difference with a 100% Random sampler should be nearly negligible. If you add the fact that Metropolis is a bit slower than pure Random sampler, you should get about the same results or better by using just a pure Random sampler.

Re: sobol and metropolis

Posted: Fri Aug 10, 2018 3:43 pm
by lacilaci
Dade wrote:
Fri Aug 10, 2018 3:31 pm
lacilaci wrote:
Fri Aug 10, 2018 11:57 am
Also not only I get better noise performance using metropolis at 90% mutation and 4 rejects. But also no fireflies and in a moderately clean render also no denoiser artifacts(which show a lot using default 40% mutation probability)
A 90% of large mutation probability means Metropolis works like a pure Random sampler 9 times out of 10 so the difference with a 100% Random sampler should be nearly negligible. If you add the fact that Metropolis is a bit slower than pure Random sampler, you should get about the same results or better by using just a pure Random sampler.
maybe, if i used the default 512 I'd get the same but slower results than random sampler. But I used 4 and tried also 1. Which does give a bit different look but no fireflies and a lot less noise.

That said, what does the reject value do? :D

Re: sobol and metropolis

Posted: Fri Aug 10, 2018 4:26 pm
by Dade
lacilaci wrote:
Fri Aug 10, 2018 3:43 pm
That said, what does the reject value do? :D
The max. mutation rejection value is the max. number of consecutive small mutation rejected.

Intuitively, it stops Metropolis to be stuck on small mutation of very bright paths: the default value allows up to 512 consecutive small mutation of a very bright path. With 4, it will be "stuck" only for 4 times.

P.S. as written before, Metropolis with the parameters you are using is pretty much working like a random sampler.

Re: sobol and metropolis

Posted: Fri Aug 10, 2018 4:40 pm
by lacilaci
Dade wrote:
Fri Aug 10, 2018 4:26 pm
lacilaci wrote:
Fri Aug 10, 2018 3:43 pm
That said, what does the reject value do? :D
The max. mutation rejection value is the max. number of consecutive small mutation rejected.

Intuitively, it stops Metropolis to be stuck on small mutation of very bright paths: the default value allows up to 512 consecutive small mutation of a very bright path. With 4, it will be "stuck" only for 4 times.

P.S. as written before, Metropolis with the parameters you are using is pretty much working like a random sampler.
Thank you for the explanation.
I'll play with these settings cause so far I'm either confused or something doesn't add up. I'd expect random sampler to be a lot closer to sobol in terms of performance than metropolis with these settings...

To clarify, using low reject values and mutation probability of 100 should give same result as random sampler?(with the exception of random sampler being maybe faster?)

Re: sobol and metropolis

Posted: Fri Aug 10, 2018 4:53 pm
by B.Y.O.B.
lacilaci wrote:
Fri Aug 10, 2018 4:40 pm
To clarify, using low reject values and mutation probability of 100 should give same result as random sampler?(with the exception of random sampler being maybe faster?)
If the large mutation probability is 100% there are no small mutations, so the max. rejection number doesn't matter anymore.

Re: sobol and metropolis

Posted: Fri Aug 10, 2018 5:12 pm
by lacilaci
B.Y.O.B. wrote:
Fri Aug 10, 2018 4:53 pm
lacilaci wrote:
Fri Aug 10, 2018 4:40 pm
To clarify, using low reject values and mutation probability of 100 should give same result as random sampler?(with the exception of random sampler being maybe faster?)
If the large mutation probability is 100% there are no small mutations, so the max. rejection number doesn't matter anymore.
Now I'm confused a lot, cause I'm sure I saw noticeable differences even between values of 1 and 4 for reject, while having mutation probability at 100.
Gonna do some more testing as soon as I have some time, hopefuly tomorrow morning.

Re: sobol and metropolis

Posted: Sat Aug 11, 2018 6:19 am
by lacilaci
So here are my results:
All settings except metropolis are at default. using cpu+path+dlc and no clamping

There is clearly a big difference in performance between metropolis(mutation probability:100,max rejects:1) and random sampler.

It does give a different look, but in cases where time is more important than accuracy I think it's a great option to boost rendering(Unless there are other issues with these settings I'm not aware of yet) as it is not only considerably faster but also no fireflies!

Also, even with mutation probability of 100, the max rejects value does make big difference. Where the default value of 512 is giving similar results to random but weaker performance.

Re: sobol and metropolis

Posted: Sat Aug 11, 2018 6:28 pm
by Sharlybg
For my experience for testing purpose including metropolis sampler and extrem setting you need a bit more complexe scene ( at least specular shader some emissive light and some textured surface ) to better highlight each method strenght and weakness.

Normally fireflies shouldn't be a problem. When Luxcore is set the right way for a given scene fireflies just disappear.
Most of the time clamping should easily solve it when lighting setup is correct.

Edit : also DLCS is not the best light strategy for your interior scene. I think it is king when it come to deal with many small light source case.

Re: sobol and metropolis

Posted: Sat Aug 11, 2018 7:16 pm
by lacilaci
Sharlybg wrote:
Sat Aug 11, 2018 6:28 pm
For my experience for testing purpose including metropolis sampler and extrem setting you need a bit more complexe scene ( at least specular shader some emissive light and some textured surface ) to better highlight each method strenght and weakness.

Normally fireflies shouldn't be a problem. When Luxcore is set the right way for a given scene fireflies just disappear.
Most of the time clamping should easily solve it when lighting setup is correct.

Edit : also DLCS is not the best light strategy for your interior scene. I think it is king when it come to deal with many small light source case.
Hm, so far I haven't seen any dificulties using metropolis set up as I did. Random and sobol give fireflies until clamping is so strong it affects total contrast and flattens highlights.

I agree that in that particular view dlcs doesn't do much. but in general dlcs seems to help metropolis with areas where direct light hits. I noticed that in exterior views.