Chaos Corona Forum
Chaos Corona for 3ds Max => [Max] General Discussion => Topic started by: Bormax on 2016-05-24, 09:06:46
-
Hello,
Don't remember from wich daily I start to fill that Corona 1.4 builds clean picture slower, but I was too attracted with denoising, so I didn't concentrate on this. But now I decided to make some tests and compare render speed of Corona 1.3 and 1.4. Results made me a bit sad... Here are comparison pictures rendered with no denoising:
pictures 1, 2 - project made with Corona 1.3 and rendered 13 min with 1.3 and 1.4 (sorry for blured area - names....)
picture 3 - crops zoomed 200%
plus statistics and render setup UIs
picture 4 - crop of the project I just start to work on, made with 1.4 and rendered with both 1.4 and 1.3, zoomed 200%
As you can see Corona 1.4 gives much noisier result with the same render time. Also parsing time is longer. Beside that CoronaMultiMap gives different color combination in 1.3 and 1.4 version (look cars colors).
I'd like to ask - is it the price for posibility to denoise picture or something wrong with my render settings or what ever from my side? Did somebody make the same kind of tests?
Win 7, dual Xeon E5-2660, 32Gb of RAM, Max 2014
-
Can you show us the sampling focus render element?
-
Here it is.
-
Oversampling trees....reminds me how Grant Warvick requested sample multiplier so he could lower it per object via Vray attributes so that they get less sampling since noise isn't visible.
Not sure if there can be intelligent solution that would guess correctly when noise isn't obvious for human eye. Good cause to keep adaptivity optional or adjust stregth perhaps.
-
The noisy splotches on the wall of the building look very similar to the noise I have in one image that I can't recall seeing in earlier versions - does the wall material use low level reflection with low glossiness?
Makes me wonder if there was thorough internal testing done of older versions vs 1.4 - I think this would have certainly became visible in some cases.
-
The noisy splotches on the wall of the building look very similar to the noise I have in one image that I can't recall seeing in earlier versions - does the wall material use low level reflection with low glossiness?
Makes me wonder if there was thorough internal testing done of older versions vs 1.4 - I think this would have certainly became visible in some cases.
Here is outside wall material structure. Inner brownish walls have reflectivity set to 0 :)
-
Thanks - similar to my setup: BlendMtl using two slightly reflective materials with low glossiness. Maybe blending materials is problematic in 1.4.
Ondra, I guess you have several machines with different versions of Corona for internal tests anyway. Could you test this kind of materials in 1.3 and 1.4? Or anyone else? I am super sorry but I really have no time to do any testing right now...
-
So, I still don't understand - are these results of my tests normal for Corona 1.4, or there really something wrong? If there something wrong, I'd like to know can I do something to make Corona work as it should? Can somebady else make this kind of tests and share results, please?
-
How about render settings on both? Did you change something or used default settings?
-
How about render settings on both? Did you change something or used default settings?
Settings are default. If you want you can check them, I atached UI screens in ferst post here
-
Hm, no one else has comparison ?
-
Hello Bormax
you can connect the scene files for verification ?
-
Hello Bormax
you can connect the scene files for verification ?
Hello Rafpug,
If this situation would happen with one or two scenes, I would decide that problem somewhere inside those scenes, but I have the same picture with all scenes - new or old.
Here are couple of new tests, all pictures rendered after reset of render settings, 5 minutes of render. White color of trees - RGB 150, empty cube - RGB 128 with no reflection.
More complex scene, more complex materials - more visible difference of noisiness
-
I don't think that those results are unexpected. Afterall, adaptivness isn't meant to speed up rendering, it just helps to distribute noise more evenly over entire image. So, if you have situation where there's good mix of areas which tends to hide noise pretty well, with areas where's noise is very evident even in low quantities, then adaptivity might be just counterproductive. So if you have rendering of house with big flat walls in the forest or in meadows, or nicely lit interior scene with big windows and even bigger hairy carpets, in that situation adaptivity may do more harm than good. I guess it isn't so universal thing, as we hoped it to be. Maybe decision to hide its controls, wasn't the best one?
What puzzles me more, is how adaptivity reacts to brigtness in diffuse texture. It assigns much more samples to darker textures than to brighter ones. You can clearly see that in first post in the striped wall part.
-
Hello Bormax
Denoise works well , you have to understand the disturbing element in the scene.
-
Hi Team,
Before renewing my license, it seems important that someone from the team clarifies these very interesting tests.
For me, the denoising function is an interesting gimmick, but a gimmick.
What is important to me, it is the raw rendering results
Thank you in advance
-
Hi, as already mentioned, adaptivity is not intended to speed up rendering. It just makes the noise distribution more uniform.
Here are some examples: https://coronarenderer.freshdesk.com/support/solutions/articles/12000004975
In the first example with the glasses you can see that the DOF areas got much better, but that's "at the cost of" some uniform areas which were slightly less noisy with adaptivity disabled.
In the last example you can see how much adaptivity changes when using denoising with it. Disabling adaptivity would sometimes lead to situations like this.
1.4 adaptivity is not infallible, but it is definitely better than no adaptivity at all.
-
Another examples:
-
Thanks very much Maru ! Have a good day.
Franck
-
For me adaptivity or no adaptivity question was settled after running these tests :
adaptivity OFF:
https://www.flickr.com/photos/119850875@N05/27276144062/in/datetaken-public/
VS
adaptivity ON:
https://www.flickr.com/photos/119850875@N05/26767325493/in/datetaken-public/
notice the left area atop the stairs and small wall corner in the kitchen area near the small window - adaptivity surely makes them better and imo the whole image looks a little better with adaptivity on - no portals or denoiser were harmed in making those test thus the they have more noise than such render should for a given time - strangely but when added, portals speed up things quite noticeably in this situation )
-
Thank you all guys for your responses, clarification and tests!
I just have to get used to adaptivity.
And thanks again Corona team for so intensive development and improvement of your product!
-
Hi, as already mentioned, adaptivity is not intended to speed up rendering. It just makes the noise distribution more uniform.
Here are some examples: https://coronarenderer.freshdesk.com/support/solutions/articles/12000004975
In the first example with the glasses you can see that the DOF areas got much better, but that's "at the cost of" some uniform areas which were slightly less noisy with adaptivity disabled.
In the last example you can see how much adaptivity changes when using denoising with it. Disabling adaptivity would sometimes lead to situations like this.
1.4 adaptivity is not infallible, but it is definitely better than no adaptivity at all.
Hello guys,
I would like to know what happens "under the hood" that adaptive system on Corona.
I don´t have technical knowlegement, but after watched a video that shows how adaptive works on Fstorm I thought they has different methods.
e.g fstorm´s sampling method stops to sampling areas that doesn´t need more sampling and keep focus just where it´s needed.
Corona´s sampling works that way too?
Thanks and sorry my english if it´s not clear enought.
Video:
-
That's exactly how adaptivity in Corona works too. You can add CSamplingInfo render element and see sampling distribution like in that FStorm video.
-
Yes, but the Corona adaptivity, in my opinion, isn't so good as it should be =(
It's not possible with I7 970, in 17 minutes, at 640x640, with PTS=2, to have so noise...
PS: there isn't any denoise.
(http://www.francescolegrenzi.com/Temp/Corona/0075_Bokeh.png)
-
cecofuli have you tried that scene in any other render to see how it handles those parts?
-
That's exactly how adaptivity in Corona works too. You can add CSamplingInfo render element and see sampling distribution like in that FStorm video.
Thx romullus!!
p.s: nice and informative signature ;)
-
(http://www.francescolegrenzi.com/Temp/Corona/0077_Bokeh.jpg)
-
(http://www.francescolegrenzi.com/Temp/Corona/0078_Bokeh_1.4vs1.3.jpg)
-
In attachment the original rendering
-
Corona adaptivity right now is image space adaptivity. It can only sample some pixels more or less, but it does not change the sampling itself. In some extreme cases, such as the ones you posted, no amount of samples would result in satisfactory result. If you have 0.001% chance of sampling something in pixel, sampling the pixel 10 times more will still give you only 0.01%. For these scenes to render properly, different changes in different algorithms would be necessary (which we will try to do sooner or later).
-
Ohh, ok. So, we must to wait some new release =)
-
Bormax: Try to change adaptivity interval via maxscript to higher number(maybe half number of phases or higher) as default(5):
renderers.current.adaptivity_interval = 48With this setting will be the image rendered with equal portion of samples in each pixel from pass 1 to 48.
After pass 48 will start adaptivity do its job :)
According sampling focus render element adaptivity is "drastic" in distribution of samples...so solution is let corona render image with adaptivity only last few passes. For me ideal number of interval is between half and 3/4 num. of total passes. I discovered 10% less render time and of course noise eliminated on problematic materials and dark areas :)
Someone could say, that with high interval values there will be few or only one calculation of adaptivity during rendering....but when I was watching SF render element, is it enough.
-
Thank you Bzuco!
I tried to put your script in to Corona's string option window, but got the "Invalid string option: renderers.current.adaptivity_interval = 48" massage. Did I put it to the wrong place?
-
Thank you Bzuco!
I tried to put your script in to Corona's string option window, but got the "Invalid string option: renderers.current.adaptivity_interval = 48" massage. Did I put it to the wrong place?
press F11 (maxscript listener) :)
-
press F11 (maxscript listener) :)
Thank you! :)
-
BTW: You can also change this parameter as "Adaptivity recalculation (passes)" in the "Development/Experimental Stuff" rollout (activate that rollout in "Performance settings" with "Enable devel/debug mode" first).
This does the same, no need to use a script unless you want to act automatically relatively to the render passes amount (what could get a kind of scripted adaptivity slider requested by some peple here).
Good Luck
-
I tried to put your script in to Corona's string option window, but got the "Invalid string option: renderers.current.adaptivity_interval = 48" massage. Did I put it to the wrong place?
If you want to use it as a string, then format is as follows:
int adaptivity.interval = 48
But as Frood already mentioned, this option is already exposed in UI, so no need for strings here. And let's hope it doesn't turn into another MSI200 rule :]
-
Thank you guys for your help!