Page 1 of 1

bidir+metropolis

Posted: Thu Feb 07, 2019 12:29 pm
by lacilaci
Any tips on speeding up? Recommended settings for interiors etc?

Re: bidir+metropolis

Posted: Thu Feb 07, 2019 1:02 pm
by Sharlybg
Try to reduce the second setting from 512 to 128. Reduce your eye path and light path below 5 if you can. set your linear tone mapper to 1 so you can work with low clamp value ( i use 10 ). for every light i use efficacity 1 for predictability.

In preview mode i set light/eye path = 2 for fast rendering.

Re: bidir+metropolis

Posted: Thu Feb 07, 2019 1:13 pm
by lacilaci
I will try, thanks.

Re: bidir+metropolis

Posted: Thu Feb 07, 2019 2:54 pm
by FarbigeWelt
Depending on your scene you may use 75%, 128 and between 1% and 10%.
If you use animation keys for the values, mouse pointer over inbox and then hit key I, the color of entry field changes color if you did it correctly. Set extrapolation of key node to linear.
Define halt condition: sample. Disable auto brightness of tonemapping. Do not clamp.

Animate, wait and select properties of best result.
Now, with selected image you can adjust clamping. Not less and not more as useful.

You can easy use low res like 25% of 1920x1080 to get a good impression for the final.

Be aware, low conj. rej. values may have influence on color depth. I often use extreme values like 1024xN, N=4 to 16.
Entry field is able to calculate for you, so enter 1024*8 and confirm with return.

Re: bidir+metropolis

Posted: Thu Feb 07, 2019 5:27 pm
by Dade
Sharlybg wrote: Thu Feb 07, 2019 1:02 pm Reduce your eye path and light path below 5 if you can.
Yup, BiDir connects every light path vertex with every eye path vertex so the complexity is O(n^2) with the path length. Reducing the max. path depth (eye + light) can lead to a good speedup.

Re: bidir+metropolis

Posted: Thu Feb 07, 2019 5:53 pm
by lacilaci
Dade wrote: Thu Feb 07, 2019 5:27 pm
Sharlybg wrote: Thu Feb 07, 2019 1:02 pm Reduce your eye path and light path below 5 if you can.
Yup, BiDir connects every light path vertex with every eye path vertex so the complexity is O(n) with the path length. Reducing the max. path depth (eye + light) can lead to a good speedup.
Yeah.. In some situation I can see a clear difference from pathtracing regarding lighting. Not sure if it is a bug in pathtracing or something.
However I also see that adaptivity does nothing for interior type of rendering, it is probably just making it slower.

Re: bidir+metropolis

Posted: Fri Feb 08, 2019 3:03 am
by Fox
About that eye path depth. Lets say i need eye path depth of 20, can i still use default russianroulette.depth 3 and cap 0.5?
russianroulette.depth 20 seems to be 3-4 times slower.

Re: bidir+metropolis

Posted: Fri Feb 08, 2019 9:29 am
by lacilaci
Fox wrote: Fri Feb 08, 2019 3:03 am About that eye path depth. Lets say i need eye path depth of 20, can i still use default russianroulette.depth 3 and cap 0.5?
russianroulette.depth 20 seems to be 3-4 times slower.
I guess, that parameter isn't exposed in blender (probably shouldn't be changed?)?

I also think... If I set lightpath to 0 do I get a pathtracer? :D

Re: bidir+metropolis

Posted: Fri Feb 08, 2019 9:46 am
by Dade
lacilaci wrote: Fri Feb 08, 2019 9:29 am I also think... If I set lightpath to 0 do I get a pathtracer? :D
Yes, BiDir is exactly a path tracer + a light tracer (aka backward and forward ray tracer) with MIS (Multiple Importance Sampling).

BiDir+MIS is good because picks the best (i.e. less noise) solution between multiple results ... but it is also slow (in term of samples/sec) because for each sample has to compute multiple results and discards the bad one (doing effectively a lot of nearly useless computations).

Re: bidir+metropolis

Posted: Fri Feb 08, 2019 9:49 am
by Dade
Fox wrote: Fri Feb 08, 2019 3:03 am About that eye path depth. Lets say i need eye path depth of 20, can i still use default russianroulette.depth 3 and cap 0.5?
russianroulette.depth 20 seems to be 3-4 times slower.
Russian Roulette is a trade off between more samples/sec and more noise. I use a very conservative RR (i.e. the cap) exactly to avoid a "bad" trade off. The default should be fine with any depth.

P.S. RR is applied to both paths: eye and light.