The best Anti-Aliasing for v-ray renderings

V-Ray V-Ray Common Render Theory
1

Lately I have been obsessed with anti-aliasing in vray and have been trying to find the best quality vs. time AA.

For my animations I usually use Adaptive Sub. AA (-1,2) because I have found it to be so much faster for an acceptable quality. It does great with glossy surfaces and smooths them out beautifully. But I don’t like how it renders detail, it makes small light fixtures flicker, and garbles up detailed texture patterns such as carpets!

Everything that I’ve read says that QMC AA is the best to use.

So I tried to use QMC AA, and this is what I noticed. The details render better and the lights don’t flicker anymore, but all of my glossy textures are quite noisy. But if I increase the QMC to something more than (1,6), and increase the shader subdiv higher than 16…well it’s just too costly for the time that I have to render animations and I end up with a 640x480 frame that takes 30 min to render. Does anyone have a fast/high-quality animation solution?

The below image was rendered with the following settings:

640x400 @ 10 min.
AA: Adaptive subdiv (-1,2)
rQMC sampler: (0.85 adaptive, 0.01 noise thres) Global mult:1
GI: primary(irr map: 50, 50), secondary (light cache: 1000)

I have found these settings to work well in general for interior spaces. If you want to view the animation of this image go to:
http://www.ramyhanna.com/timebased/strakejesuit.html

Any fast/quality Adaptive QMC settings would be greatly appreciated!

Thanks,

ramy

2

if you haven’t already, I’d read the manual, it gives some guidelines as to when its best to use which antialiasing method.

3

Ramy,
The flickering is completely due to a lack of sub pixel sampling…it’s just that simple and using adapt subdiv with a -1/2 is not going to give you good sampling at all. I’d recommend using that configuration with test renders only. When you are rendering finals, and you have lots of smoothly shaded surfaces and not a lot of detail, you can take advantage of this sampler but you will need better settings. People say use QMC because it’s results are more predictable and you’re certain to get what you need if you use decent values. I would recommend (for production) a min/max of 2/5 or 3/6 if you really need to bring out detail. You should also disable ‘Use QMC thresh’ and carefully test different Clr thresh values because if it’s not low enough, increasing the min/max rate will do you absolutely no good (they act as a check for each other).

Above all, definitely loose that 50 value for HSph Subdivs, you simply dont need it for this see as far as i can see. Start with 20. That could easily cut your rendering time in half. I recommend no more than a Medium preset for the IR map also…you simply dont seem to have the detail that requires more than that.

4

Comparison between AA’s

Thanks for the tip. that could be helpful.

Below is a comparison of what I’m finding in my results. In these images all of the vray shaders have a glossiness of 8 (no interp. for animation). They are all using the same irr. map and light cache from file:


engine: adaptive sub
min/max: -1/2
clr thresh: 0.1
rQMC: 0.85/0.005


engine: AQMC
min/max: 1/5
clr thresh: qmc samp.
rQMC: 1.0/0.01


engine: AQMC
min/max: 1/16
clr thresh: qmc samp.
rQMC: 0.85/0.005


engine: adaptive sub.
min/max: 0/3
clr thresh: 0.1
rQMC: 0.85/0.2

I’m still playing with differnt values and settings for this shot.

5

Here is a little tip:
For adaptive qmc, to get optimal amaunt of samples per pixel in optimal amount of time, you basicly need to specify a range of sampling to perform for almost every pixel. For us, since we deal with animations/moving deforming objects we need to take advantage of this feature all the time, to avoid flickering.
What I am really talking about is vlado’s universal settings. Exept with few alterations:
you may set qmc sampling to be 1/12. 1/24, 1/32, 1/64 and 1/100. You may set theshold for qmc sampler to 0.001.
What this will do is, it will apply a specified range of samples to each pixel. Which will ensure of the noise free image in best time possible.
The difference of this method with the adaptive subdivision, is usually you can apply a lot more samples per pixel then adaptive subdivision will ever let you without running out of memory. :slight_smile:

6

Dmitry,
I don’t mean to dish what your saying, but I would discourage anyone from using such high max ranges in their image sampling. First of all, when the ‘Use QMC thresh’ option is enabled (which it is by default), the Adapt QMC sampler will look to the QMC Noise thresh value as authorization to actually do the sampling. But, if you don’t have the noise threshold set below 0.01, you’ll never get more than a Max of 5 or 6 even applied. If you do set the noise thresh down to 0.001, it will go after those ridiculously high subdivisions…when it does need to. I can’t imagine a reason why you would ever need higher than 8 subdivisions…it would be like printing an image on paper higher than 600dpi…which is the accepted limit to your eye’s ability to see detail. You could bring the fastest computer to a grinding halt with a high max value and a low noise threshold value. Just my 2 cents.

7

Good two cents man! IMO
I think that Dmitry haven’t low time limit for production, so he takes all these heavy parameters and may be oversample sometime, may be everytime but the result is sure to be here. No risk method in heavy production.
My two cents :roll:

8

Dmitry,
I don’t mean to dish what your saying, but I would discourage anyone from using such high max ranges in their image sampling. First of all, when the ‘Use QMC thresh’ option is enabled (which it is by default), the Adapt QMC sampler will look to the QMC Noise thresh value as authorization to actually do the sampling. But, if you don’t have the noise threshold set below 0.01, you’ll never get more than a Max of 5 or 6 even applied. If you do set the noise thresh down to 0.001, it will go after those ridiculously high subdivisions…when it does need to. I can’t imagine a reason why you would ever need higher than 8 subdivisions…it would be like printing an image on paper higher than 600dpi…which is the accepted limit to your eye’s ability to see detail. You could bring the fastest computer to a grinding halt with a high max value and a low noise threshold value. Just my 2 cents.

No worries.
Here is why: Im not sure how things work in arch viz, so I cant really say for sure. However, often in film we run into situations where there is so much happening on screen, and often in dark scenes (night settings) where the only bright parts are reflections, that gets to be a trouble.
Trust me when I say that, I have done some extensive testing.
And in some specific cases, these are the mesures we have to resort to. I would love for sampling to be low, and render times mega fast. But for us often thats not the case.
About production times, priad, yes and no. At the moment we are working on series, where a show delivers ever 2/3 weeks. And often we end up rendering 10,000 to 50,000 frames.
Another note is that, vray in maya is a bit diferent from vray in max. So sampling wise (though Im not too sure) in max vray solves better sampling then in maya. This is to be tested though so dont quote me on it.

9

Well it’s certainly cool to see how others work.

10

after tweeking

After some playing and tweeking this is what I discovered for Adaptive QMC. You can still keep your min/max low and get good results as long as your shader reflective subdivs are much higher.

In the previous examples I used shaders with reflective subdivs of 8. Well in this next example I upped the reflective subdiv to 45.

I also noticed that reducing my bucket size from 64x64 to 48x48 gave me a couple of seconds, as well as a better solution.


engine: AQMC
min/max: 1/6
clr thresh: rQMC sampler
rQMC: 1.0/0.03
bucket size: 64x64


engine: AQMC
min/max: 1/6
clr thresh: rQMC samper
rQMC: 1.0/0.03
bucket size: 48x48

I’m going to continue to edit these settings, I’m sure that I can get even better and faster results!

11

i suppose ur storing the glossies in the lc allready, right?

12

Correct

13

Re: after tweeking

well the increasing the subdivs as you say wont do much good when noise threshold is so high…when importance sampling is conducted (which you have set to 100%), one cant work well without the other. any reason why you set it to 0.03? that’s really high

also, from the white dots, it doesnt look like you’re using subpixel mapping and/or clamp out…though that might because of the high noise value.

14

Re: after tweeking

Those two were not using either subpixel mapping or clamp out. The reason the noise is set to 0.03 is because that was an acceptable quality vs. time for me. I know it’s high. But for this scene it seemed to work well.

15

I don’t kow if it can help you :sweat_smile: for my final render I use

Adaptive QMC
mini 1
max 20
Clr Thresh 0.005 to 0.001

Antialiasing filter
Mitchel Natravelli.

It work very well and It’s not too slow…

Laurent

16

[deleted nonsense…]

My settings are almost always these:
AQMC min/max=1/4, clr th linked to QMC sampler
QMC sampler: noise th=0.01 and the rest default.
And material subdivs, area light subdivs etc always higher than default.

I just did some tests on glossy reflections and AA.

Tried these settings:

noise=0.01 and min/max=1/4
Variable subdivs for the material:
4:15.2s
8:19.8s
16:27.5s
24:34.2s
32:41.5s
(noise levels in the image: good from 16 subs)

min/max=1/4 and material subs=8
variable noise th
0.05: 5.4s
0.02: 13.6s
0.01:19.8s
0.005:21.0s
0.001:22.5s
(noise levels in the image: all crappy)

min/max=1/4 and material subs=16
variable noise th
0.05: 5.4s
0.02: 12.9s
0.01:27.6s
0.005:51.8s
0.001:53.0s
(noise levels in the image: good from 0.01 and lower, but all exactly the same, so no improvement from 0.01 to 0.001)

material subs=8 and noise=0.01
min/max variable
1/4:20.0s
1/8:39.6s
1/16:50.0s
1/32:52.6s
(noise levels in the image: good for 16 and 32 and looking exactly the same)

material subs=8 and noise=0.002
min/max variable
1/4:22.2s
1/8:58.3s
1/16:208.4s
1/32:333.6s
(noise levels in the image: good for 16 and 32, but rendertimes are way off…)

From the best images (bold ones), the best image is the one with min/max=1/32, noise=0.002 and 8 subdivs of course, but rendertime is super high.
All of the other images where material subdivs=8, have the worst quality.
The second best quality is found in the image with noise=0.01, min/max=1/4 and subdivs=32. I also did a test with noise=0.005, min/max=1/4 and subdivs=32. Rendertime doubled (88s) and now it is nearly identical to the best image (333s…)

Here are these images:
01)noise=0.002, min/max=1/32 and subdivs=8 (333s)

02)noise=0.005, min/max=1/4 and subdivs=32 (88s)

03)noise=0.01, min/max=1/4 and subdivs=32 (41.5s)

Of course this only tests glossies and not fine detail for example in your scene. But since your question originally was about the glossies, this is a good test. You can do similar tests on a fine detail scene only (for example with a white non reflective material) and see what gives you good enough results. Then combine it with the previous test and choose the best quality vs speed settings.

What you can also see is that when using default min/max rate and default material subdivs, the noise treshold below 0.01 doesn’t make much of a difference.

So my conclusion is that if you want fast rendertimes and good glossies, you have to increase the material subdivs instead of going crazy on the min/max rate and noise treshold.

17

i found this helpful:

http://www.cgarchitect.com/upclose/VI/Week17/image\_sampling.pdf

18

flipside:
Though ur test seem pretty damn exaustive imo the quality of the glossys in an non studio set will be directly connected to the environment it self and its lighting quality. Therefore an interior arch viz shot is a completelly diferent story, me thinks…
i will try to post some of my own tests when i get 15 min! :frowning:

:wink:

19

Yeah could be… But usually if you have a problem like this, it’s not a bad idea to do some small test where only one thing changes at a time instead of trying a bunch of settings at once. It’s also best if you want to test AA that you leave out the GI calculations etc out of your test as this will be affected too when for example changing the QMC noise treshold.

Some years ago when I still had loads of time, I did a lot of these tests on every vray setting, creating graphs out of it, that was very very usefull!! Then you see how each value behaves, you would be surprised how some of these curves looked like, rendertimes going sky high once you passed a certain value or sometimes no effect at all in a range of values etc…

For the few interiors that I have done already, I usually got away with the settings from my previous post (increasing the material subdivs instead of all the rest). For animation it’s different of course, but also there if you only see the noise when viewing it frame by frame, why bother increasing quality. For example this animation:
http://www.aversis.be/3dviz\_portfolio/23.htm
Look at the render at the bottom, very noisy, but nobody complained as you didn’t see that noise when the animation played (also not in the high res version, the one on my site is very low quality :slight_smile: )

I’m sure this is different for every scene and client… But these settings are so critical and I see many posts with people using some very high quality values that I’m sure in many cases you could just alter one setting which can cut rendertimes in half without losing quality in the image.

20

quoted for agreement! :smile:
and i do remember your graphs pretty well!very usefull stuff bu the way!
:wink:

nevertheless i feel quite usefull to tweak either aa and gi quality with just a few clicks, instead of tweaking through materials and lights subdivisions…but that s just me! :smile: