Author Topic: Sampling Worklfows  (Read 14707 times)

2012-09-09, 19:43:43

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
I've been testing a really heavy scene the last couple of days. 56.000.000 instanced polygons, about 7.000.000 unique. It's a cathedral with lots and lots of very small openings (windows) everywhere. I had big problems to get acceptable rendertimes with normal engines like vray etc. Now, after serveral renderings and different tests of various settings I'm getting better rendertimes with less noise.
Currently I'm using (besides the defaults):
Pathtracing as primary GI solver
HD Cache as secondary GI solver
Subpixel AA: 2
Display AA: 5
GI samples: 4
Lights group, samples:16

I've noticed that most of the noise is in the direct lighting pass, so i decreased GI samples and increased Lights samples. I think i'm getting a good balance now - more or less the comparable amount of noise in direct light and GI passes.
Now I'm wondering if there are some techniques to make it better. For example, I don't know what subpixel AA exactly does (since it's not explained somewhere).

Thanks for any suggestions ;)
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-09, 20:02:58
Reply #1

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
subpixel AA is only for antialiasing bright highlights. Display AA is just how well will zoomed-out image look in Corona VFB, it does not affect rendering itself.

Next corona version will come with manual, I hope I'll have some time for it next week.

You can post the render with stats and complete settings screenshot, I'll look into it.
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2012-09-10, 13:58:13
Reply #2

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
Ok, I will prepare some images this evening.
I noticed that much higher values for GI and Light samples work much better here. 20 for GI and 64 for Lights give a pretty clean result.
I've been testing several combinations of high sampling values which result in less and less passes (in the same time). The lower those sampling values are the more passes it will reach - but Rays/s will decrease (from about 2.600.000 to 1.600.00). As far as I understand the technique behind those passes and samples relationship it means that a pass is the "time it takes to go through every pixel in the image" and samples are of course the number of rays shot through every pixel.

1) My question is: Is it correct that it is faster to have high sampling values (and reaching less passes) than having a lower sampling setting and reaching more passes (which I think would result in a better antialiasing result but more noise in lighting and GI)

2) I've been using HD cache for secondaries now but I noticed slight lumincance differences between several renderings, which will result in flickering animations. I have set "Max records" to the highest possible value but I suppose I will need to increase "Pt samples" for a more stable solution. What do you think is needed to get a clean solution?

Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-10, 14:29:40
Reply #3

Javadevil

  • Active Users
  • **
  • Posts: 399
    • View Profile

Hi Deadclown, you mention 64 light samples, the light sample multiplier only goes to 10.
Which one are you turning up ?

cheers

2012-09-10, 14:45:23
Reply #4

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
Are you using the alpha v2 ? I'm using a later version so that may have changed or be a completely new setting :-/ . I have a "Lights" group here with a "samples multiplier" spinner.
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-10, 17:48:50
Reply #5

Javadevil

  • Active Users
  • **
  • Posts: 399
    • View Profile

I've got the latest build ( 27th Aug ), the samples multiplier spinner only goes to 10.

2012-09-10, 18:26:09
Reply #6

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
I fear it's not the latest, I've got 30th August here.
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-10, 18:27:10
Reply #7

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
Hi,

this all should be improved in the next release:

1) GI should render with the same speed (maximal) regardless of the PT samples setting
2) Area light multiplier will go up to 999
3) It will be possible to crank up the quality setting of HD cache records, resulting in less flickering
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2012-09-10, 18:29:10
Reply #8

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
Awesome! That's great news, I'm looking forward to test that ;)
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-10, 23:43:37
Reply #9

michaltimko

  • Active Users
  • **
  • Posts: 305
  • Coronaut (c)
    • View Profile
me 2 !
Coronaut!(c)2011

Supporting Corona in commercial projects since pre-alpha

2012-09-11, 00:06:34
Reply #10

Javadevil

  • Active Users
  • **
  • Posts: 399
    • View Profile


Great news !!!

Off to grab the latest build

2012-09-15, 21:57:49
Reply #11

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
hmm.. it's kind of fitting into this topic so I'll ask that here: I've noticed that (big interior scene) distant objects will need way more time to get clean. Near objects are already clean in a short time, but walls which are relatively small in the rendering (far away) are still very noisy. Is there anything that can be done about that?
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-15, 22:00:41
Reply #12

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
Can you give and example (image or scene) of such situation? Is it when using PT+PT, or PT+HD?
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2012-09-15, 22:11:40
Reply #13

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
It happens mainly with pt+hd, pt+pt seems to be a bit more consistent here. I sent you an image via pm.
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-17, 13:29:34
Reply #14

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
mhh... Another thing I'm struggling with. Lets say I have a room with several little CoronaLight spheres. I'm using pt+pt. The lighting pass is clean in seconds, but in the GI pass I've a lot of little fireflies, which are of course samples hitting the light-spheres. So is there a way to exclude the lights from GI calculation (GI pass would be clean very fast then). I'm using the latest build.
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-17, 18:20:47
Reply #15

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
I am really sure that the fireflies are not from hitting lights directly, because Multiple Importance Sampling automatically accounts for that (unless you have lights sampling set to something other than MIS).

Also, you cannot exclude lights from GI calculation, GI is induced solely by direct lighting, you wouldnt have any GI without the lights.

So, the possible reasons for fireflies are either
- caustics (light behind some solid glass/mirror)
- bright spots (light very close to some geometry)
- general inefficiency of path tracing in interior scenes

First two causes can be easily fixed at the cost of some small bias by setting "max sample intensity" to about 5-10. The last is intristic deficiency of path tracing, I would recommend switching to PT+HD in that case.

Anyways, it would be much better, if you could post renders and screenshot of your render settings ;)
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2012-09-17, 19:19:32
Reply #16

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
Ok, I can't show you the scene I'm talking about but i made an example which is pretty similar. So, it will likely be the second point of your list. I didn't think about this... but it's pretty obvious now

My question about excluding light was meant in that way: "Do GI rays directly hit the lightspheres?" I thought direct lighting would be the light emitted by the spheres and GI would be bounced light from the surface the direct light illuminates.

Anyways, It seems the fireflies have their origin at the surface beneath the lightsources.


P.S. Is max sample intensity for GI only or for both directlight and Gi samples?
« Last Edit: 2012-09-17, 19:29:01 by DeadClown »
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-17, 19:52:34
Reply #17

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
Its only for GI; its not needed for direct light. Just set it to some reasonable value to balance the bias/variance ratio ;).

About the direct/indirect: GI rays can randomly hit lights. But because Corona uses MIS, such cases are properly weighted, and do not cause any fireflies or bias. They, in fact, increase the robustness when using large area light sources and when sampling very close lights. When using MIS the distinction between direct lighting/GI partially vanishes.
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2012-09-17, 20:13:32
Reply #18

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
whoaaa... that's fixing several problems at once! Awesome! :) You should consider moving the spinner to the Global illumination group - that would make those things clear. I didn't use the max sample intensity spinner a lot since i didn't know exactly which values it will clamp (like direct light for example).
MIS is always active here :)
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-17, 20:36:17
Reply #19

Ludvik Koutny

  • VIP
  • Active Users
  • ***
  • Posts: 2557
  • Just another user
    • View Profile
    • My Portfolio
You should always press production defaults button located in "About corona renderer" rollout. It sets render settings to correct defaults (initial defaults are for debug purposes only). And it, among other settings, sets also max sample intensity to 10, which is recommended value ;)

Press this button everytime you create a new scene ;)

2012-09-17, 20:46:29
Reply #20

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
I do - everytime. But 10 was simply to high for this scene.
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-17, 20:52:10
Reply #21

Ludvik Koutny

  • VIP
  • Active Users
  • ***
  • Posts: 2557
  • Just another user
    • View Profile
    • My Portfolio
Value of 10 should not be very noticeable... ;) It could introduce some weaker secondary reflections and things like that, but it should not affect scene such as this one that much :)

2012-09-17, 20:56:19
Reply #22

racoonart

  • Active Users
  • **
  • Posts: 1446
    • View Profile
    • racoon-artworks
I can only say that lowering this value helped a lot...
Why does GI sample intensity have an influence on reflection (not considering the intensity of reflected surfaces of course) ?
Any sufficiently advanced bug is indistinguishable from a feature.

2012-09-17, 20:58:28
Reply #23

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
thats great suggestion (moving it to GI rollout), I've already done it ;)

I think Rawalanche means diffuse reflections
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)