Dumb Question: PMC difference from path tracing

[xxdanbrowne]

So now that I have a GPU capable of running a different kernel than direct lighting without seizing up I have a question:

Is PMC like a *better* version of path tracing?

[matej]

Is PMC like a *better* version of path tracing?

Only in certain situations. Some artists doing interirors swear that PMC is the way to go. In my experience while doing mostly what could be considered productviz, you only get a slowdown factor of 4 for the exact same quality as PT.

[glimpse]

My favourite so far is PT, as in some (majority) cases where are lot of glossy reflections it seems that this kernel is workin' faster (then PMC). Again it very much depends on the scene (lighing, material) & that's only my personal preference.

At best (if You have some time) You could do tests on Your own to see what works best for You.

[gordonrobb]

What is the actual difference though? Or is the answer too complex?

[xxdanbrowne]

@matej/glimpse:

Thanks guys. I see the same symptoms as you: PT gives awesome renders (as does PMC), but PMC is 4x slower. On my new machine I seem to be only 2-3 times slower for most scenes using PT as DL. That said I haven't done anything brutal like mesh emitters or caustics or anything so that might slow things down significantly.

But I honestly can't see the difference between PMC and PT. Though it might be that I don't have 100% trained up artists eyes (because some scenes that look awesome to me apparently look shit to some high end professionals!)

Anyways: same question as Gordon; Is there anyone who is a heavyweight kernel guy (or one of the genius devs?) who could give an overview of the difference between PMC and PT?

[riggles]

Well, I'm no kernel guru, but here is my understanding.

Both are Monte Carlo path tracers, but PMC (which stands for Population Monte Carlo) is adaptive, or as the Octane manual says a "mutating" algorithm. Instead of equally distributing samples across the image plane until you get a converged image, PMC is adaptive. It looks at the variation between pixel samples to determine which parts of the image plane need more sample concentration (high variation areas) and which are ok with less (low variation areas).

The reason PMC can seem slower is because it's sampling, analyzing, adapting, resampling, etc., so it has that extra overhead. Where it can come in handy is scenes that have lots of variation, such as lots of small mesh emitters or with some caustics, in which the regular path tracing kernel may never give you clean results without going to absurd max sample levels.

That's my $0.02

[xxdanbrowne]

@riggles

Thanks man, that made sense.

[gordonrobb]

Not to me wouldn't adaptive be quicker?

[grimm]

The more branching (decision making) that the GPU does the slower it is. The advantage with PMC is that even though you are getting fewer samples per second, the noise will clear faster in certain cases. It can help with interior scenes or like riggles said with caustics.

Jason

[matej]

Not to me wouldn't adaptive be quicker?

It is in special cases like small light emitters, SSS & caustics, afaik. But this better sampling achieved by PMC doesn't show up immediately, only after the render is nearing completion, you'll start to notice PT doesn't improve your scene anymore (eg. fireflies don't clear, or clumps of noise stay the same), while PMC (maybe) would. But this makes testing quite time consuming and usually not worth it - just go with PT and do noise clear in post.

(admittedly I don't have experience with interior rendering, so maybe in such cases PMC shines a lot better)