Understanding DMC Sampler

[peterguthrie]

Nice one Akin.

I'm not entirely clear on how the min shading rate is different from min samples?

[RockinAkin]

Simple explanation:
The Min Samples value is limited (clipped) by the the Max Secondary Samples (Per Primary Sample) value.
The Min Shading rate value is not limited and will in fact raise the Max Secondary Samples (Per Primary Sample) value.

More complicated explanation:
Min Samples is better thought of as 'Min Adaptive Samples' (as it is named in V-Ray for Maya, and which I'm hoping vlado also renames in V-Ray 3.0 for max).
This value basically says: 'of X number of Adaptive Samples possible, this value will set the Minimum number of Adaptive Samples'.

When your AA settings become higher (like 10 Max / 1 Min), the max number of secondary samples gets lower and lower until it eventually becomes 1 Max / 1 Min Secondary Samples (Per Primary Sample).
In this scenario, there's no more room for adaptive secondary samples anymore, and the Min Samples value has no effect at this point.


The Min Shading Rate value however isn't bound by any other values... it sets the floor for the maximum number of secondary samples to be taken per primary sample... regardless of any other settings.

Makes sense?

[ARTECONI-CGI]

Hi All,

Thanks for this Jo and Grant, it is one of the most exciting improvement I had experienced in vray in the last few years, I just hope it is consistent and usable in all the scenarios.
I had the chance to test this new method at work today on a pretty simple scene and after 5 minutes I had the whole studio behind me staring at the unbelievable speed and cleanliness of the render.

I have to admit that I have gone through this thread 5 times and read the article from Toni Bratincevic again and again but I still struggle to grasp all the concepts at 100%.
Anyway driven by the instinct I have followed your workflow John and noticed that it is very efficient.
What I don't get is what you are trying to balance: why don't you just start with the max amount of subdivision everywhere instead of raising them gradually?

To me it seems straightforward to act like this:
A) Set min to 1 and max to 100
B) Use high subdivisions everywhere, as a default.
I have put 512 pretty much everywhere, Brute force, Glossiness, Lighting and it seems to work beautifully.
C) All I have to do is decide with the color threshold how much noise I want.

I didn't notice big slow down passing from 256 to 512 subdivisions or 1000.
In my opinion it just make sense to have as many subdivisions as we can. Leaving them low at 64 we just risk to have not enough samples to clean the noise.
The max value will anyway subdivide the value of those subdivision by 100 and start to shoot rays until he hits the color threshold.

I have the feeling that I am missing something from your workflow. Maybe my antialiasing is not good?
What do you look for in the vraysamplerinfo? Mine is pretty blue. Most of it? just a slight lighter blue on the edges. Does it mean it is good?

Why do you increase your subdivision gradually and then reduce the max amount? You're talking about finding a balance?
Do you mean that an antialiasing 1/100 is not as good as 1/50? or less... 1/32?

Kind Regards,

Giacomo.

[RockinAkin]

Giacomo:

1 min / 100 max is simply overkill... only in scenes with REALLY small geometric detail like leaves or hair will MAYBE need something 1 min / 50 max for a high quality result.
Even your samplerate pass backs this up - since it is nearly all blue - v-ray isn't even getting anywhere close to the 100 max aa subdivisions that you've set.
But a drawback of using such a high maxiumum value for antialiasing is that almost all other secondary values in your scene (gi, refl, refr, sss, etc) become meaningless unless your also set their values ridiculously high. Refer to the calculator to see for yourself: http://www.cggallery.com/tutorials/dmc_calculator/

The tutorials where 1 / 100 max or something similar is recommended is similar to setting V-Ray on autopilot... it will eventually get you a nice final image with minimal understanding of what's happening under the hood. But don't be fooled into thinking that it's the best way to get a scene rendered. Using some of the methods described in this thread will get you to a cleaner image, with faster rendertimes - if you just understand where you need your samples allocated to.

[ARTECONI-CGI]

Hi Akin,

I have just seen your portfolio: Impressive! Great lighting and great characters.

Thanks for the reply, I will have another go tomorrow because the scene I have tested was pretty simple.
But if I remember correctly 1/50 was taking longer of 1/100. The more I was lowering the MAX value and the more the rendering time were increasing.

So 1/100 was sensibly faster than 1/50. 1/50 faster than 1/32. 1/32 faster than 1/16 and on and on.
I know what you mean: using 1/100 sounds exactly like using the universal settings.

The main difference now is that my materials and lights have huge subdivision values. 512 as average.

That seems to have lead the rendering engine to a much faster solution.

Much faster than the usual universal settings.

I will run some more tests tomorrow to check if my scene is a special case and try to reproduce the behaviour. I have no Max at home at the moment. I couldn't have waited for tomorrow otherwise...

Giacomo.

[ARTECONI-CGI]

The feeling I had is has follow:

it seems like that with high subdivisions on materials, brute force, DOF and lights Vray clean the render very fast, while not having enough subdivision on them, Vray rely on some other less efficient method to clean the noise. A much slower one.

I will confirm it tomorrow.

[RockinAkin]

Yes, it is completely expected that 1/50, or 1/32, or 1/16 will take longer to render than 1/100 - only because your secondary subdivision settings (lights, materials, etc) are being set so high!

If you're using a secondary subdivision value of 500 subdivisions all over your scene (lights, materials, etc):
With 1/100 AA - VRay may take a max of 25 secondary samples per AA sample.
With 1/50 AA - VRay may take a max of 100 secondary samples per AA sample.
With 1/25 AA - VRay may take a max of 400 secondary samples per AA sample.

So when you bring your AA settings down, you need to bring your secondary samples to reasonable levels as well. So for example:

If you bring your AA settings to 1/50 - then bring your secondary subdivisions to 250. and you'll have a V-Ray take a max of 25 secondary samples per AA sample. (Like it previously was doing at 1/100 AA)
If you bring your AA settings to 1/25 - then bring your secondary subdivisions to 125. and you'll have a V-Ray take a max of 25 secondary samples per AA sample. (Like it previously was doing at 1/100 AA)

See what I'm getting at? Use the calculator to check if you have any confusion.
You want to set the AA according to your scene's requirements, then tweak secondary sampling to meet the perfect balance of speed / minimal noise.

[grantwarwick]

Good to see others getting massive speed improvements as well.

Although it does feel a little sad KNOWING your renders are as fast as they can get :P

[joconnell]

Yep - that said though we're asking a lot of vray - imagine doing the same in another renderer - I'm working on some full frame cg stuff at hd at the minute and occasionally I'll complain about a 45 minute render time as being too slow, but then again I'm kind of using everything I possibly can save for 3d motion blur

[snivlem]

If you bring your AA settings to 1/50 - then bring your secondary samples to 250. and you'll have a V-Ray take a max of 25 secondary samples per AA sample. (Like it previously was doing at 1/100 AA)
If you bring your AA settings to 1/25 - then bring your secondary samples to 125. and you'll have a V-Ray take a max of 25 secondary samples per AA sample. (Like it previously was doing at 1/100 AA)

Then why not leave it at 1/100 if it is just a linear sliding scale?

What are you're "universal settings" before you start finding the sweet spot?

[joconnell]

It's the last line that's the important bit - AA first and then set all the other samplers appropriate to that. Since the AA sampler in vray is the daddy of everything else, you've really got to set that appropriately first - no point in tweaking all of your material subdivisions until then as once you change your AA all of your material settings go out of whack entirely.

In terms of production stuff I'm normally around 1/8 for simple, 1/12 for more complex and 1/16 for finer detailed stuff - anything involving really thin lines for example. I haven't dealt with hair or fine grass yet and they'd be situations that I reckon would be testing for aa sampling. The 1/100 thing was made by vlado as a catch all method but as he said himself it's not optimised at all. Every seen is different and will have different causes of noise so a single approach won't cover everything.

[Morne]

Using the calculator I did a quick test. I have about 200 or so lights in my scene inside the building and outside is an HDRI dome light. Its a street view with about 10 trees in view

Set everything to 500 and AA to 1/100. (max 25 secondary samples) A small cropped area renders in 20 min

Set everything to 80 and AA to 1/16. (max 25 secondary samples) A small cropped area renders a lot longer. I cancelled it after 20 min and it was only halfway with the render!

By eye I can't see a difference in any of the render elements, except the sample rate in the 2nd option isn't as overly blue as the 1st option
So even though max secondary samples in both instances are 25, the 2nd option renders WAY SLOWER

So then will it be possible to get the 500 and 1/100 option of 20 min to render any faster?

[RockinAkin]

It's the last line that's the important bit - AA first and then set all the other samplers appropriate to that. Since the AA sampler in vray is the daddy of everything else, you've really got to set that appropriately first - no point in tweaking all of your material subdivisions until then as once you change your AA all of your material settings go out of whack entirely.

In terms of production stuff I'm normally around 1/8 for simple, 1/12 for more complex and 1/16 for finer detailed stuff - anything involving really thin lines for example. I haven't dealt with hair or fine grass yet and they'd be situations that I reckon would be testing for aa sampling. The 1/100 thing was made by vlado as a catch all method but as he said himself it's not optimised at all. Every seen is different and will have different causes of noise so a single approach won't cover everything.

Exactly right - couldn't have said it better myself.

Using the calculator I did a quick test. I have about 200 or so lights in my scene inside the building and outside is an HDRI dome light. Its a street view with about 10 trees in view

Set everything to 500 and AA to 1/100. (max 25 secondary samples) A small cropped area renders in 20 min
Set everything to 80 and AA to 1/16. (max 25 secondary samples) A small cropped area renders a lot longer. I cancelled it after 20 min and it was only halfway with the render!

By eye I can't see a difference in any of the render elements, except the sample rate in the 2nd option isn't as overly blue as the 1st option
So even though max secondary samples in both instances are 25, the 2nd option renders WAY SLOWER

So then will it be possible to get the 500 and 1/100 option of 20 min to render any faster?

This is really strange... are you sure no other settings have been messed with?

[RockinAkin]

Here's a visual example to better illustrate the what we've been saying here:

Two renders of a simple test scene containing only a sphere with glossy reflections, and an HDRI dome light.
All V-Ray settings are left at their defaults except for AA max, and the local Subidivsions value of my light and material reflection glossiness.

The first render has my anti-aliasing set to 1min / 100max. My materials and lights are set to 8 subdivisions:


The second render has my anti-aliasing set to 1min / 8max. My materials and lights are set to 100 subdivisions:


We can see that the second render is MUCH cleaner (especially noticeable in the lower half of the sphere) - even though both renders take nearly the same amount of time to render.
All I've done is swap the higher subdivisions value (100) from the Image Sampler (AA) max to the Secondary Sample (Light & Material) values.

This works because in this scene, you can clearly see that there isn't much geometric detail that vray needs to deal with - basically only the edges of the sphere, which is simple enough.
So instead of firing a bunch of extra, unnecessary AA samples at the scene hoping to clean up noise - we can better allocate those samples to attacking the noise at it's source - the glossy reflections and lighting in the case of this scene.

Now this is obviously a super simple test example, and of course different scenes will require different settings depending on what's going on in them.
But it's just to illustrates the point of how immediately setting your AA to 1min / 100max is definitely not the best way to optimize your render.

[ARTECONI-CGI]

Hi All,

I have been able to run some more tests and sadly I have to admit that the enthusiasm has gone. On a more complex scene I didn't see an improvement in terms of speed or quality.
I am used to stay at min 1 max 4 and fix the noise increasing the subdivisions. Any test with different values and subdivs has led to longer rendering times.

I am not sure what I am looking for in the samplerate pass? A good one has to be all blue but red on the edges? Or just blue and green on the edges?
It doesn't matter which scene I load, the best AA settings seems to be at 1/16 or 1/32. But then, increasing the subdivs proportionally, I didn't see such an improvement.

I wonder as well how should I behave with the LC and IRR Map subdivision when I start fiddling with the AA sampler.
Do they have to be multiplied proportionally?

Kind Regards,

Giacomo.