Rendering speeds between the different modes...

General questions about Indigo, the scene format, rendering etc...
User avatar
CoolColJ
Posts: 1738
Joined: Mon Jun 25, 2007 1:47 pm

Post by CoolColJ » Tue Aug 14, 2007 12:47 pm

I have been having a lot of splotch problems with that scene with the Alien Artifacts I recently did, lots of reflective metal surface and heavy bump mapping probably makes it worse.
I have already rendered it twice, once to over 4000samples and it took days, and still splotchy in MLT BiD with max consecutive rejections of 2000. Then redid it again in MLT mode much faster right now

anyway I saw these quotes from Ono - which made me think a little, so I did some testing to figure out the right settings....
Reducing the max number of consecutive rejections reduces noise, especially fireflies, at the cost of introducing some bias into the render.

It is the maximum number of consecutive rejections allowed in the Metropolis-Hastings algorithm before a mutation accept is forced.

------General tracing parameters------
"max_depth" "10000"
"bidirectional" "false"
"metropolis" "true"
"hybrid" "false"


------Metropolis light transport settings------
"large_mutation_prob" "0.4"
"max_change" "0.03"

all rendered to 6 mins 320x240 image size


MNCR 0
Done 6020000.00000 samples (78.38542 samples per pixel)
16646.48344 samples / second (60.07275 micro-seconds / sample)

MNCR 25
Done 11060000.00000 samples (144.01042 samples per pixel)
30426.43564 samples / second (32.86616 micro-seconds / sample)

MNCR 50
Done 10840000.00000 samples (141.14583 samples per pixel)
29920.80770 samples / second (33.42156 micro-seconds / sample)

MNCR 100
Done 10990000.00000 samples (143.09896 samples per pixel)
30385.82236 samples / second (32.91009 micro-seconds / sample)

MNCR 250
Done 10990000.00000 samples (143.09896 samples per pixel)
30314.52187 samples / second (32.98749 micro-seconds / sample)

MNCR 500
Done 10930000.00000 samples (142.31771 samples per pixel)
30181.85997 samples / second (33.13248 micro-seconds / sample)

MNCR 750
Done 10940000.00000 samples (142.44792 samples per pixel)
30254.40539 samples / second (33.05304 micro-seconds / sample)

MNCR 1000
Done 11010000.00000 samples (143.35938 samples per pixel)
30291.60746 samples / second (33.01244 micro-seconds / sample)

Doesn't seem to be any difference in rendering speed like many have believed, but setting it to zero and under 10 definitely slows it way down...?!

You can see with the lower settings, there much less firefly noise spots, but some of shading and reflections on the metal objects are missing.
250 is the lowest I'd go, while 500 seems to the sweet spot, with diminishing returns above that.

and just for kicks I did a render with 10000 setting, no difference in rendering speed, but way splotchy :)
usually most of those spots do go away with increased rendering time, but only up to a point.

Will 600 as my all round setting :o
Attachments
CCJ_indigo_MNCR.JPG
CCJ_indigo_MNCR.JPG (276.88 KiB) Viewed 3348 times

mrCarnivore
Posts: 517
Joined: Sun Mar 04, 2007 6:20 am
Location: Stuttgart, Germany

Post by mrCarnivore » Tue Aug 14, 2007 7:37 pm

@CoolColJ: You are a sick, sick man! ;-)

Thx for the comparison. I think that one should go into the manual, so that everyone can see directly in which way the render is influenced by this parameter...

User avatar
CoolColJ
Posts: 1738
Joined: Mon Jun 25, 2007 1:47 pm

Post by CoolColJ » Tue Aug 14, 2007 8:49 pm

another thing is that with higher MNCR values, caustics show up much earlier and denser with MLT, almost as well as with Bidrectional on

In that glass rendering on the first page, with an MNCR of 2000 I was getting this curved caustic on the left wall. Now with a MNCR of 600 it's barely there at all, will probbaly show up much later on

So I guess it's always trade off between 2 things....

User avatar
Kram1032
Posts: 6649
Joined: Tue Jan 23, 2007 3:55 am
Location: Austria near Vienna

Post by Kram1032 » Wed Aug 15, 2007 1:06 am

250 is the lastest, that has no fireflies,
1000 is quite ok,
10000 has quite many, already - but still not too much :)

mrCarnivore
Posts: 517
Joined: Sun Mar 04, 2007 6:20 am
Location: Stuttgart, Germany

Post by mrCarnivore » Wed Aug 15, 2007 2:10 am

Hey Kram, try to resist translating every single word from german to english. There is no lastest... last is already last. You can be more last than last. :-)

It's similar to the German word "Einzigste" which doesn't exist either, but lastest definitely sounds worse!

Anyway, back to the topic: Do the fireflies in the renders with MNCR at 1000 and above clear eventually or are they there for at least days?

User avatar
Kram1032
Posts: 6649
Joined: Tue Jan 23, 2007 3:55 am
Location: Austria near Vienna

Post by Kram1032 » Wed Aug 15, 2007 2:16 am

-.-
I don't translate from german to english.
I use english, directly.

I often heard "lastest news", though... (even if it's wrong)

mrCarnivore
Posts: 517
Joined: Sun Mar 04, 2007 6:20 am
Location: Stuttgart, Germany

Post by mrCarnivore » Wed Aug 15, 2007 2:20 am

You heard latest news which is kind of the the newest news.

lastest (with 2 s) does not exist in english...

You translated "der letzte" to "the lastest" even if you did it unconsciously. :-) Seems obvious because of the two t's...

But don't worry. Happens to everybody every now and then!

User avatar
Kram1032
Posts: 6649
Joined: Tue Jan 23, 2007 3:55 am
Location: Austria near Vienna

Post by Kram1032 » Wed Aug 15, 2007 2:23 am

I DON'T TRANSLATE
I even think that stuff in english.
Might be saved wrong, though

If that should be discussed further, use PM, pls :) (Don't guess so, though)
___Off topic, sorry___

User avatar
CoolColJ
Posts: 1738
Joined: Mon Jun 25, 2007 1:47 pm

Post by CoolColJ » Wed Aug 15, 2007 10:07 am

mrCarnivore wrote: Anyway, back to the topic: Do the fireflies in the renders with MNCR at 1000 and above clear eventually or are they there for at least days?

the default value is 1000, and if you look at the renders I did in my Test pics thread in the Work in progress section, the ones where I let them render up to 3000+ samples, they do eventually get nice and clean.
With darker indirect lit areas though it can be a problem. I find with a value of 1000, the render will start noisey then clean up a bit when you hit the 1000-1500 mark and then more fire flies pop up and then they eventually get softer and then mostly dissappear.

Now last week I rendered that Alien Artifacts scene at 800x600 with MNCR of 2000, 3x Supersampling, MLT Bidirectional and even after 4000 samples it was still full of fireflies etc.... plus it was very slow, took days...
I am redoing it now with MNCR of 600, MLT, 2xSupersampling and rendering much faster and a lot cleaner!!!! Looks just as good too shading wise so far. Already up to 1700 samples now in a few hours already. Still has some spots though. Will post it in my Test pics thread when done

With MLT the render will look clean and smooth usually when you hit the 4000-5000 sample mark if you have no textures, just bare diffuse surfaces
Caustics may take longer though, if you have lower MNCR values I think.

I think I should put all this info in the Wiki 8)

User avatar
CoolColJ
Posts: 1738
Joined: Mon Jun 25, 2007 1:47 pm

Post by CoolColJ » Fri Aug 17, 2007 2:40 pm

I'm back with another mega test -
8)
this is part 1 - using default settings for bounces, max rays and MLT settings etc


all rendered for 10mins.

Man MLT looks horrible :shock:
MLT is not bad outdoors in my renders so far.
Anyway it looks like the Bidirectional modes fair the best with this kind off interior artificial type lighting. Even though with Bidrection on, you don't reach as high a sample count the difference in noise is pretty obvious.
Hybrid is the slowest of the bunch though

some modes don't even show the caustics that should exist, or may take much longer to show, if at all.

IMO Bidrectional MLT is the best of the bunch, especially with my settings - see part 2 coming up :)

I'll do a larger and longer render compare between Bidrectional MLT and PT, to see which is the top dog :D
So far there are these white dots you get with Path Tracing modes, that don't show in MLT type modes

also there is some bug with Bidrectional MLT where sometimes caustics like that reflection on the bottom of purple sphere appears somewhere else. It has rotated upwards in this case....
Maybe related to the bug which also cause Bidrectional MLT to move the Sun position
Attachments
CCJ_Indigo_DefaultRenderCompare.JPG
10 min renders, default settings
CCJ_Indigo_DefaultRenderCompare.JPG (236.54 KiB) Viewed 3184 times

User avatar
CoolColJ
Posts: 1738
Joined: Mon Jun 25, 2007 1:47 pm

Post by CoolColJ » Fri Aug 17, 2007 4:08 pm

Now here are the MLT modes with my settings of 0.03 for max change, which I got after extensive testing.

MLT now looks better and more accurate too in a shorter time than the default 0.01 as shown above 8)
Attachments
CCJ_MLT_My Settings.JPG
MLT modes with my settings
CCJ_MLT_My Settings.JPG (159.21 KiB) Viewed 3170 times

mrCarnivore
Posts: 517
Joined: Sun Mar 04, 2007 6:20 am
Location: Stuttgart, Germany

Post by mrCarnivore » Fri Aug 17, 2007 7:33 pm

WOW. You test-o-maniac. ;-)

Great to finally see somebody do some extensive tests and share them with others!

Seems 0.03 max change is really to better alternative. It's faster and leads to better results...

User avatar
manitwo
Posts: 1029
Joined: Wed Jul 05, 2006 4:50 am
Location: Tirol - Austria

Post by manitwo » Fri Aug 17, 2007 11:59 pm

thanks for that test coolcolj!
the noise is much better - the dirty-patchy look is gone.
i vote for "0.03" as the new default.

User avatar
Kram1032
Posts: 6649
Joined: Tue Jan 23, 2007 3:55 am
Location: Austria near Vienna

Post by Kram1032 » Sat Aug 18, 2007 1:27 am

VERY GREAT TESTS!

I'd say, Hybrid BiD and MTL BiD look quite simmilar...
You should add something SSSish, to that test, though - and one simmilarish outdoor test :)

Wedge
Posts: 441
Joined: Sun Jan 14, 2007 11:33 am
Location: East Coast, USA

Post by Wedge » Sat Aug 18, 2007 4:36 pm

If any new defaults are confirmed by Ono I will be sure to use them. Thanks to everyone for testing. :)
Content contained in my posts is for informational purposes only and is used at your own risk.

Post Reply
95 posts

Who is online

Users browsing this forum: No registered users and 25 guests