Author Topic: HD Success rate and noise in corona 1.4  (Read 6737 times)

2016-05-17, 18:56:37

valwizard

  • Active Users
  • **
  • Posts: 55
  • "We get up in the morning. We do our best. Nothing
    • View Profile
Hi, Guys! Congrats with new version of Corona renderer. I am rendering an animation in Corona. It is an interior but very heavy one with heavy proxy the polygon amount is almost 40 mln in already optimized scene. Trying to cut render time down to a reasonable one, something like 1hour on 2Xeon 32 cores. Still there is noise mostly in the GI pass. For the moment I have AA/GI on 20/4 MSI 11
The HD success rate is very low. What could cause that? What is the relation between success rate and amount of geometry/lights/reflections? The number of rays is ok IMHO 6.5 mln is ok isn't it?
Whatever I will really appreciate for some advise /help. I don't want to use denoiser as this is bluring the image and washing the clean sharp details, after testing I did not find any difference between Corona denoiser and Red Giant denoiser or Neat video plugins, except is fully integrated in Corona and doesn't need to buy additional expensive plugins. Sorry I can not attach the full image and the scene due to copyrights and NDA.
« Last Edit: 2016-05-17, 19:32:09 by valwizard »
"We get up in the morning. We do our best. Nothing else matters! "

2016-05-17, 20:20:13
Reply #1

Juraj Talcik

  • Active Users
  • **
  • Posts: 4050
    • View Profile
    • studio website
I pondered upon this too, as this applies to lot of my (surprisingly interior) scenes as well. It's usually glass and metallics. Previously HDCache (deprecated) had threshold up to which it would work.
talcikdemovicova.com  Website and blog
be.net/jurajtalcik   Our studio Behance portfolio
Instagram   Our studio Instagram, managed by Veronika

2016-05-17, 20:57:06
Reply #2

cecofuli

  • Active Users
  • **
  • Posts: 1507
    • View Profile
Usually, for a modern interior, simple scene, HD rate must be between 50-80.
With the UHD you cannot control nothing about Cache, except fo the two preset

Still: bigger sample with less subdivision
Animation: smaller samples with more subdivision

I suggest you to use Still, if you are working for single renders.

My suggestion is to try to render with in clay and see if the problems are in the shaders.
Then hide some heavy geometry etc...
You can also install the good old A6, with the HD and see if the result is better.
Honestly, I would like to have more control for the UHD ;-)

2016-05-18, 12:06:02
Reply #3

valwizard

  • Active Users
  • **
  • Posts: 55
  • "We get up in the morning. We do our best. Nothing
    • View Profile
I pondered upon this too, as this applies to lot of my (surprisingly interior) scenes as well. It's usually glass and metallics. Previously HDCache (deprecated) had threshold up to which it would work.

Yeh, There is a hell lot of Golden heavy details and glass chandeliers.

I just realised that the soft furniture using faloff in diffuse is clean after 10 passes, but I still have noise on the wall which is only diffuse with sight reflection/ glossiness  , just semi matt  paint. Any suggestions regarding this? Is this a normal behaviour? Actually this noise still persist in GI pass while the other passes are pretty clean. Increasing the GI it will decrease AA and I have a lot of reflective edges which at certain point start flickering.

Thanks guys

"We get up in the morning. We do our best. Nothing else matters! "

2016-05-18, 12:28:47
Reply #4

Juraj Talcik

  • Active Users
  • **
  • Posts: 4050
    • View Profile
    • studio website
Low-glossy flat surfaces are worst offenders. I am not sure if they didn't receive less sampling previously because of lower reflectivity but perhaps Adaptivity should of have solved this in 1.4. Didn't do extensive testing yet.
talcikdemovicova.com  Website and blog
be.net/jurajtalcik   Our studio Behance portfolio
Instagram   Our studio Instagram, managed by Veronika

2016-05-18, 13:43:04
Reply #5

valwizard

  • Active Users
  • **
  • Posts: 55
  • "We get up in the morning. We do our best. Nothing
    • View Profile
Just  a test - taking down the reflection in the material did nothing to the HD success rate   still tough the level of noise has not change much. I am not really sure how the Adaptivity works - is it similar to Vray, like raising more samples in the difficult areas? As long I understood from the actual post regarding Adaptivity/ Denoising the samples are brought from the other areas, weakening them( pouring the quantity) Is that correct?  What this figure 5 in the box  means ? 5 samples? 5subdivisions? a relative number?
"We get up in the morning. We do our best. Nothing else matters! "

2016-05-18, 15:42:43
Reply #6

cecofuli

  • Active Users
  • **
  • Posts: 1507
    • View Profile
And, about the Corona Denoise vs Neat Image.
Neat image cannot do this great job, I think.
Only two passes, 12 seconds...


2016-05-18, 15:59:10
Reply #7

valwizard

  • Active Users
  • **
  • Posts: 55
  • "We get up in the morning. We do our best. Nothing
    • View Profile
Well, I am glad to see these results. As I have mentioned below I am not considering the denoiser serious untill I could prove this really. For the images 600X600 or thumbnails probably it is ok but for prints like 8000 this is not a solution. There are much more powerful denoisers for still images like Imagenomic Noiseware , I just mentioned Neat for instances as it is powerful cleaning the noise in Video sequences. Again I am not going to go in comparison Corona VS Neat or REd Giant or anything, these are too different products with different purposes to be compared to.  tBut the reason I have raised this questions

I do not want to trick the result, the render should be clean without using denoisers as they are cleaning picture by blurring pixels which is not accurate( for different purposes though). Whatever regarding to my researches here is the result after turning the adaptivity off: The HD success rate raised to 41 and the noise is almost gone after 100 passes which gives about 30 % performance. Below is the result in comparison.
Render time for 150 passes Adaptivity on 1h18 min  Adaptivity off 1h 04 min,
« Last Edit: 2016-05-18, 16:08:20 by valwizard »
"We get up in the morning. We do our best. Nothing else matters! "

2016-05-18, 16:10:56
Reply #8

Juraj Talcik

  • Active Users
  • **
  • Posts: 4050
    • View Profile
    • studio website
That's quite surprising result.
talcikdemovicova.com  Website and blog
be.net/jurajtalcik   Our studio Behance portfolio
Instagram   Our studio Instagram, managed by Veronika

2016-05-18, 16:19:19
Reply #9

valwizard

  • Active Users
  • **
  • Posts: 55
  • "We get up in the morning. We do our best. Nothing
    • View Profile
I am really surprised as I am feeling that something is wrong. I hate do not understand what happens every time I do something. Again can anyone tell what is this adaptivity, how this algorithm works, why the result is opposite than you expect? Jurai, you are the most experienced guy in Corona, could you explain how this relates to samples? I am still in the dark and hate guessing even though there is a result but the most important question is Why?
I can only guess again that is making the noise more obvious and easier to detect and clean by denoiser
« Last Edit: 2016-05-18, 16:27:47 by valwizard »
"We get up in the morning. We do our best. Nothing else matters! "

2016-05-18, 16:40:29
Reply #10

Juraj Talcik

  • Active Users
  • **
  • Posts: 4050
    • View Profile
    • studio website
I am not expert on Corona's sampling at all, Maru and the devs are :- ). You can always send them a scene to test.

I have yet to test and compare how the adaptivity works (although I dislike spending time on doing such tests, I am not even systematic enough).
talcikdemovicova.com  Website and blog
be.net/jurajtalcik   Our studio Behance portfolio
Instagram   Our studio Instagram, managed by Veronika

2016-05-18, 16:42:50
Reply #11

cecofuli

  • Active Users
  • **
  • Posts: 1507
    • View Profile
I'm not expert like Juraj, but with adaptivity, Corona will spend less time on "easy" surface/shaders, and he will concentrate his energy on the most difficult areas.
That's why, with adaptivity ON, simple flat surface have more noise. But, at the same time, complex areas (DOF, MBlur, heavy glossy surface) will be much more clean (in the same rendertime)

In my opinion, if your "complex" area s aren't too much (like 10% of your image), maybe, it's better to turn off adaptivity...
But, we have no parameters for to tweak the adaptivity. Or ON, or OFF... And adaptivity recalc doesn't change a lot the final result.

I would like to hear some official Corona answer ))

2016-05-18, 16:50:30
Reply #12

Frood

  • Active Users
  • **
  • Posts: 1466
    • View Profile
    • Rakete GmbH
ON, or OFF...

... and not even that is assured in future...

Good Luck


Never underestimate the power of a well placed level one spell.

2016-05-18, 17:24:52
Reply #13

valwizard

  • Active Users
  • **
  • Posts: 55
  • "We get up in the morning. We do our best. Nothing
    • View Profile
I'm not expert like Juraj, but with adaptivity, Corona will spend less time on "easy" surface/shaders, and he will concentrate his energy on the most difficult areas.
That's why, with adaptivity ON, simple flat surface have more noise. But, at the same time, complex areas (DOF, MBlur, heavy glossy surface) will be much more clean (in the same rendertime)

In my opinion, if your "complex" area s aren't too much (like 10% of your image), maybe, it's better to turn off adaptivity...
But, we have no parameters for to tweak the adaptivity. Or ON, or OFF... And adaptivity recalc doesn't change a lot the final result.

I would like to hear some official Corona answer ))
I am not an expert neither,  but I would like to know what am I doing instead of testing. I found these new features quite interesting, for some of the guys might absolutely worthy, not for me yet, as I am not playing much with DOF and so... What I would really like to have in Corona - more flexibility and real improvements for sampling, probably material subdivisions,  better lighting behaviour. So far the guys made absolutely stunning job. By the time Corona will come to 3.3 version I am sure it will be an absolute winner.
"We get up in the morning. We do our best. Nothing else matters! "

2016-05-18, 18:47:50
Reply #14

TomG

  • Corona Team
  • Active Users
  • ****
  • Posts: 3084
    • View Profile
Might be useful for testing, you can take a look at what adaptivity does using the CInfo_SamplingFocus render element, brighter areas show where more processing is being done. Have to have enough passes for it to show up (e.g. for default of Adaptivity recalc of 5, would need to run for at least 5 passes to see anything in that render element).