experiments with using variance ontop of lowdiscrepancy

Discussion related to the implementation of new features & algorithms to the Core Engine.

Moderators: jromang, tomb, zcott, coordinators

experiments with using variance ontop of lowdiscrepancy

Postby Radiance » Tue May 27, 2008 10:28 pm

Hey,

Last night i did some experiments with a modification to the lowdiscrepancy sampler.

My aim is to try to be able to render a decent resolution complex animation frame from a professional animation,
in an acceptable time for production.

I added a usevariance param to lowdiscrepancy, which resamples pixels based on variance.
This seems to do a good job.
I've still got some normalization issues as i'm opting to use per screen based normalization and therefore
need to prescale the contributions before splatting to the film.

I've managed to get down to 2.5 hours for a 1440x810 frame.

I think once my normalization is correct, and some more optimization/tweaks, together with an ICC build i should be able
to get this frame rendered to an acceptable output quality in 2-3 hours.

Once this is done i'd like to render a scene from 'elephants dream' as a test.

I've still to tweak the scene a bit, and adding some better variance construction estimate is also something to investigate.
the scene is not yet luxified so there's still no decent materials and some normal issues.

greetz,
radiance
Attachments
emo1.jpg
User avatar
Radiance
 
Posts: 3968
Joined: Wed Sep 19, 2007 2:13 am

Re: experiments with using variance ontop of lowdiscrepancy

Postby droid » Tue May 27, 2008 11:21 pm

FYI... I've also dabbled with this concept and it does give EXCELLENT results for some scenes. Areas where the variance changes slowly and suddenly leaps up can be tricky (e.g. low probability specularly-viewed caustics) but I guess such effects are the last thing on people's minds when rendering animations.

Ian.
droid
 
Posts: 23
Joined: Mon Oct 15, 2007 5:48 am

Re: experiments with using variance ontop of lowdiscrepancy

Postby zsouthboy » Wed May 28, 2008 12:09 am

I think you should shoot for a scene of BBB instead of elephant's dream, but I like the idea :)
zsouthboy
 
Posts: 327
Joined: Sun Oct 14, 2007 9:28 pm

Re: experiments with using variance ontop of lowdiscrepancy

Postby dougal2 » Wed May 28, 2008 12:16 am

hmm, a scene is like a minute ? 30 frames/sec for 3 hrs? Plus export time ? hmm.... it might be ready for christmas ;)
User avatar
dougal2
Developer
 
Posts: 3074
Joined: Mon Jan 14, 2008 7:21 am

Re: experiments with using variance ontop of lowdiscrepancy

Postby Radiance » Wed May 28, 2008 1:03 am

hmmm, yeah well it's just a test eh,
we can render for a couple of weeks together with dade's machines or something. :)

i gues 20-30 seconds should be enough :)

if we render it a PAL resolution it should be 3x faster :)

Radiance
User avatar
Radiance
 
Posts: 3968
Joined: Wed Sep 19, 2007 2:13 am

Re: experiments with using variance ontop of lowdiscrepancy

Postby Dade » Wed May 28, 2008 8:20 am

Radiance wrote:we can render for a couple of weeks together with dade's machines or something. :)


I'm working on the project of a deskside cluster built of diskless-headless nodes (only q6600+motherboard+2GB ram) with an hand-built case. The price of the single node is quite low (under 250 euro + VAT) so I'm playing with the idea to build a 4xNodes box (16 cores). I'm going to post some more information in the off topic forum soon :mrgreen:
User avatar
Dade
Developer
 
Posts: 4800
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

Re: experiments with using variance ontop of lowdiscrepancy

Postby Karl vom Berge » Wed May 28, 2008 9:28 am

Radiance wrote:Hey,
I added a usevariance param to lowdiscrepancy, which resamples pixels based on variance.
This seems to do a good job.


Great!

This is a classical adaptive sampling technique in Monte Carlo integration. However, adaptive sampling can introduce bias, depending on how you have implemented this. I just want to point that out.

Ref. Veach PhD thesis, chapter 2, section 2.7.1.

Adaptive sampling just based on variance is a great technique, but it is not perfect in the sense that on some areas of scenes you actually can allow for quite a lot of noise (e.g. highly textured objects, or very complex shadows (e.g. from tree canopies), generally: on high frequent signals in the image). Perceptual heuristics can be used here in order to determine if the renderer needs to refine pixels or not. These techniques avoid computing stuff a human can't see.

See my post in this thread: http://www.luxrender.net/forum/viewtopic.php?f=8&t=451

Karl
Last edited by Karl vom Berge on Wed May 28, 2008 4:38 pm, edited 1 time in total.
User avatar
Karl vom Berge
 
Posts: 81
Joined: Mon Jan 14, 2008 7:11 am
Location: Antwerp, Belgium

Re: experiments with using variance ontop of lowdiscrepancy

Postby Dade » Wed May 28, 2008 1:41 pm

Dade wrote:I'm working on the project of a deskside cluster built of diskless-headless nodes (only q6600+motherboard+2GB ram) with an hand-built case. The price of the single node is quite low (under 250 euro + VAT) so I'm playing with the idea to build a 4xNodes box (16 cores). I'm going to post some more information in the off topic forum soon :mrgreen:


Ahah, found someone brave enough to have already built what I have in mind (however his system is not diskless): http://helmer.sfe.se/

It is funny, his exotic "case" it is available from my local Ikea store too: http://www.ikea.com/it/it/catalog/products/00107874
User avatar
Dade
Developer
 
Posts: 4800
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

Re: experiments with using variance ontop of lowdiscrepancy

Postby dougal2 » Wed May 28, 2008 1:48 pm

That's a very neat casing solution! (not so hot on cable management though)
User avatar
dougal2
Developer
 
Posts: 3074
Joined: Mon Jan 14, 2008 7:21 am

Re: experiments with using variance ontop of lowdiscrepancy

Postby Radiance » Thu May 29, 2008 9:00 am

Ha,

perfect for winter dade,
you can grab a pair of nice, warm socks in the morning :)

Radiance
User avatar
Radiance
 
Posts: 3968
Joined: Wed Sep 19, 2007 2:13 am

Next

Return to Architecture & Design

Who is online

Users browsing this forum: No registered users and 0 guests