interior test [Distributed Path integrator explained]

[NiZu]

Hi everybody,
My first post here , i've been following luxrender for about 6 months :downloading builds looking at forums etc.. and this is my first test image .
I'm very fascinated by the Luxrender project .. it's an unique blend of very ambitious technical targets (spectral render ? how sci-fi is that !) and yet, there's a great attention to practical aspects : gui , usability . This combination really impressed me.

Not easy to get started though, Lux is mainly an unbiased physically accurate render .. for overnight renders is easy to setup ... but what about keeping the advantages of a physical based engine and sacrifice the physical exactness to favor speed ? it sure requires knowledge and study , but I'm starting to be very optimistic about that

This picture was rendered with hilbert / distributed path .
on a q6600 4gb ram ,ubuntu, Lux rc4 64bit. 1600x1200 pixels , The hilbert bucket cycle took 20min (8spp) then i stopped after 1h20min at 24spp. Bloom and glare in luxgui .
No particular story : just an interior test room, 200k quads , statues from Stanford scans and the3dstudio.com

here are some considerations, let me know what you think ...(and correct me if i'm wrong!)

- high number of glossy samples : that made a real difference in speed !
glossy reflections require much more rays and samples to converge (than diffuse) so : 1 glossy / 1diffuse samples means that, by the time you have 10 glossy samples, the image is still noisy because 10 samples for glossy isn't enough ..but you're wasting samples on diffuse rays that are already noiseless.
i used 1 diffuse / 4 glossy samples . Rendertime was 1/3 of using 1 glossy / 1 diffuse.

- samples per pixel for a clean image : depend a lot on the sampler and integrator :
Metropolis might be still noisy at 500spp , lowdiscrepancy + bidir converges at about 100-200spp .
This pic used hilbert / distributed path : with 8spp was decent (!) and almost noiseless at 24spp.

-sunsky is soo slow : not just in lux , it's a nasty light type : a huge arealight going thru relatively small openings ..
and it's supposed to light the room evenly (=lots of bounces + the ones needed to reach the room )
Not to be underestimated : unless the engine is heavily biased and interpolates samples .. sunsky in interiors requires a lot of samples.

In this pic i used arealights on windows with a gradient (brown to blue ) ..cheap trick i know, but it's another 2x speed gain.
Maybe a sample multiplier per lightgroup (not possible yet , right?) so to calc. 2x samples for sky and 1x for artificial lights ..could be fast and more accurate.

Thanks for viewing and for any comment, i'll post more versions with other setups.
EDIT: here is a pic of the setup :

note the the number of bounces is very low , just the bare minimum .
i highlighted with a yellow box the parameters i understand (more or less ) and tweaked with a reason , the rest is from preset or random testing .
(there isn't much documentation on those but i guess it's because lux is in costant developement , and those might change)
Bounces and samples depend -of course- on the specific scene ..i'd say mostly on the amount of direct / indirect light .
EDIT2: in further tests i reverted the settings outside the yellow box to the presets, ignore those in the pic (or, suggest me how to use them )

[fboyer]

great work ! i spent a lot of time to study ld/dp rendering, i believe radiance is working hard in this area, to get cleaner images in lower render times.
you achieved a great quality in 1h1/2, but it's still noisy for animation purposes (which is actually the main goal of distributed path/low discrepancy rendering).

anyway, stunning work !

[Dimitri]

Nice render indeed... the information in the accompanying text too precious...

[NiZu]

Thanks !,
The noise is subtle , but in animations i guess it'd be panning around the screen and that's really annoying.
I never tested animations so i wonder : is distributed path suggested for anim. just because it's fast or there are other reasons?
By the way, i didn't use noise removal on that pic.

Also i'm rendering the same with expphoton ..50 samples in 40 mins , but much stronger noise and some bright pixels. doesn't seem better.
Will be interesting to compare to bidir and real sunsky for quality (time could be 5-10x though)

[Atom]

How are you getting the light flares on those ceiling bulbs in the alcoves?

[NiZu]

I used the glare in luxgui with 8 blades , i also used gaussian bloom (which does a circular halo , glare does the rays/flares ) glare took 10 mins to compute , gaussian bloom 1/2 hour (almost same times with render stopped or going on)
I noticed some odd behaviour with those effects : if computed at same time, it takes forever , render goes on but the glare-bloom doesn't appear and tonemapping won't update.

[NiZu]

Hi,
i've rendered some more tests , with interesting results, comparing to the 1st post (distributed path , arealights on window ) after same time :

bidir 6 bounces : not a big difference , there's just some more noise in corners (areas of very indirect light ) to get same quality of 1st test about 1,5-2x time.
No real difference or improvement in lighting and photorealism (pic not uploaded )

Using sunsky with distributed path or bidir : quite a difference, seems that sunsky impacts rendertime much more than the integrator (at least for bidir vs. distributed , inthis particular scene)
Physical sunsky does improve photorealism and quality but could take 5x time or more.

using light layers, you could see that artificial light is already noiseless , while sunsky is still a long way from clearing up. in the final pic the noise appears where sunsky contributes more than artificial.
As i said , it makes sense that sunsky is so expensive, it generates lot of noise even in biased renderers with ibl if used for interiors , i have no idea what kind of sampling strategy could be used to 'balance' this..

[Atom]

Yeah, I kind of wish those glare and bloom features has a progress bar that shows you what it going on with each one. I have experienced crashes before from turning the features on and off because they did not seem to do anything. In reality, they were busy calculating away but there was no GUI feedback (even a text message) to let me know whether I really clicked that button or not.

[zsouthboy]

Great test!

This is why Lux is the future of rendering.

[NiZu]

Thanks,
i'm very confident in the future of lux (in due time of course, render engines are such a huge work)

Atom : there is a text feedback in the bottom left saying "computing lens effects and tonemapping" , but no progressbar (so i'm not sure if doing bloom and glare at same hangs or just takes very long, like >1h)

A note on materials : i didn't use any matte , walls are glossy with exp 25 and specular intensity 0.01 or similar ...which imho makes a more realistic matte, isn't much slower, and more important avoids fireflies. I got the idea that bright matte can cause fireflies reading somewhere on blenderartist and it proved right.
(here i think :http://blenderartists.org/forum/showthread.php?t=159535 )

a note on skylight : the door in my scene is quite big , but outside there are columns and a ceiling (a covered porch) this makes the worst case for skylight : you expect a certain amount of daylight to come trhu, but it takes a few bounces. (removing the porch makes it 2-3 times faster .. but you can't easily do that, in everyday archviz jobs )