Understanding DMC Sampler

[Dariusz Makowski (Dadal)]

in the past I've tested this way to go but, at the end, it seems to be almost unusable because of that ultra high material subdivision values; here the problem was material editor taking forever to update the sample sphere ...but maybe I've missed something, any ideas? how do you guys manage those high values?

Disable material editor preview

[zeronove]

just a bit less radical?

[peterguthrie]

Hi Grant, and other superior minds.

Thought I'd try your settings out on one of my scenes that I often use to test settings:



The top row is the standard universal settings, with a noise threshold of 0.005

The second row has the following settings:

Brute force GI: 128
All materials: 128 apart from the floor and the white walls which are set at 256
All lights: 128

dmc 1,50

adaptive amount 0.85
noise threshold 0.005

(I still have the noise threshold and the clr threshold linked)

So as you can see a bit of a speed gain, but I'm wondering if anyone has any suggestions to make it faster still?

This type of scene is fairly typical for one of my interiors, but more problematic are the ones where you see out to an area of grass/trees etc., although lets concentrate on this nice easy scene first!

thanks guys

PS. 1/25 took longer even than the universal method.

[3LP]

Wow guys, I'm trying out this whole new way of seeing things, and till now, it's pretty damn good!

I still need a bit of time to understand every parameter in the equation, but this is awesome!

Noiseless renders is not a unbelievable dream anymore.
I think I will have a couple of question in the following days

Stan

[Morne]

OK I'm totally confused again. When I 1st looked into DMC sampler, I understood that you should be aiming for a more blue image with light blue and dark blue and avoid red, yellow and orange.

As time went on and I became more "clever" on the subject, I "learned" that we should be aiming for a smooth blue image, with red only on edges of geometry.

But now, seems the original idea of an only blue image is the way to go again? I'm thinking in an animation environment.

Could somebody please clarify this for the sake of my sanity!

[grantwarwick]

just a bit less radical?

Works fine without lag for me....32gb ram, GTX 580, I73930k

Hi Grant, and other superior minds.

Thought I'd try your settings out on one of my scenes that I often use to test settings:

[ATTACH=CONFIG]14741[/ATTACH]

The top row is the standard universal settings, with a noise threshold of 0.005

The second row has the following settings:

Brute force GI: 128
All materials: 128 apart from the floor and the white walls which are set at 256
All lights: 128

dmc 1,50

adaptive amount 0.85
noise threshold 0.005

(I still have the noise threshold and the clr threshold linked)

So as you can see a bit of a speed gain, but I'm wondering if anyone has any suggestions to make it faster still?

This type of scene is fairly typical for one of my interiors, but more problematic are the ones where you see out to an area of grass/trees etc., although lets concentrate on this nice easy scene first!

thanks guys

PS. 1/25 took longer even than the universal method.

Hey mate. That seems about right although I don't think you need such a low threshold.
Get used to unlinking the clr threshold too. For that interior with very large glossy surfaces you can raise it to .03 or .05 (try it out)

The Noise threshold I'd keep at .01

Set your min subdivs to 2,4 or 6. (10 if you see grass as you need more primary samples)

A couple of tips-

The noise from the dome light is most apparent in the lighting and specular render element. (Vray splits reflections from strong specular highlights intelligently using an importance sampling technique)

The noise from the brute force subdivs is most visible in the Global Illumination pass.

The noise from the glossy subdivs is most visible in the reflection pass.

My advice is to have these render elements enabled during the setup of your scenes and quite simply raise the subdivs individually looking at the respective passes.
If there is no more noise in the lighting pass, move onto GI then onto reflections.

Try it out for yourself.
As you increase subdivs, the antialising has to do less of a job (and therefore becomes a darker blue)

Wow guys, I'm trying out this whole new way of seeing things, and till now, it's pretty damn good!

I still need a bit of time to understand every parameter in the equation, but this is awesome!

Noiseless renders is not a unbelievable dream anymore.
I think I will have a couple of question in the following days

Stan

Quick to convert your sorry ass :P

OK I'm totally confused again. When I 1st looked into DMC sampler, I understood that you should be aiming for a more blue image with light blue and dark blue and avoid red, yellow and orange.

As time went on and I became more "clever" on the subject, I "learned" that we should be aiming for a smooth blue image, with red only on edges of geometry.

But now, seems the original idea of an only blue image is the way to go again? I'm thinking in an animation environment.

Could somebody please clarify this for the sake of my sanity!

Ultimately there won't be a "perfect" sample rate pass.
My advice is to set your DMC to 1/100 to start. Then go through the process of raising subdivs in materials, HDR, and Brute force.
Once you have no subdiv noise, start reducing the max rate until you have an acceptable render time with nice clean AA

[grantwarwick]

Also peter, I can tell you forgot to change the material subdivs on that chrome right?
It's showing up heavily in the sample rate. Raise it to 256 also you lazy bastard

[peterguthrie]

Also peter, I can tell you forgot to change the material subdivs on that chrome right?
It's showing up heavily in the sample rate. Raise it to 256 also you lazy bastard

yup, guilty!

trying out your other suggestions too now, thanks!

[m_hinks]

Ultimately there won't be a "perfect" sample rate pass.
My advice is to set your DMC to 1/100 to start. Then go through the process of raising subdivs in materials, HDR, and Brute force.
Once you have no subdiv noise, start reducing the max rate until you have an acceptable render time with nice clean AA

Very interesting, and the method is showing good speed ups/clean ups in my tests as well as PG's, but what would you start off at with regards your DMC clr threshold? With 1/100 - maybe 0.05 until the elements are clean? Then bring it down?

Mark

[Morne]

...and I would love to know how these settings work on an animation. Anybody have time to test?

[peterguthrie]

settings for this one:

DMC Image Sampler

min 2, max 50

clr threshold 0.015

Materials

generally 128 subdivs
plasterboard: 512
flooring: 384
other glossy mats (metals etc): 256

Lights

generally 128
vraydomelight 384

GI

384 subdivs

Settings: DMC

Adaptive amount 0.85
Noise threshold 0.01

I tried a higher clr threshold but it got too noisy.

And the render time? just a shade over 10 mins which compared to 23 mins for the universal method is pretty good! (the universal one still looks marginally better, but there isnt much in it)

Any more suggestions?

[joconnell]

Tbh I'd nearly be inclined to drop your AA even further - From your sample rate I'd say what's happening is that the results your high material and light settings are feeding back to the AA sampler mean that it has pretty much no work to do - as grant says let each of the samplers do their job rather than one thing trying to solve another's problem.

[peterguthrie]

Tbh I'd nearly be inclined to drop your AA even further - From your sample rate I'd say what's happening is that the results your high material and light settings are feeding back to the AA sampler mean that it has pretty much no work to do - as grant says let each of the samplers do their job rather than one thing trying to solve another's problem.

so say I drop it to 2,32 , I'd expect the render time to go up, with better quality, assuming I keep all other subdivs the same.. Is that right? (just started a test and that seemed to be the case)

If I drop the max subdivs to 32, should I then correspondingly reduce subdivs across the board?

[peterguthrie]

Just tried renders at 2,32 and then 2,24 min/max, and I lowered the subdivs proportionally (whether or not thats the right thing to do I have no idea)

the 2,32 one was about a minute faster and the 2,24 about 45 seconds faster (both compared to the 2,50 render)

does that mean 2,32 is the sweet spot for this scene?

[joconnell]

Quite possibly but I'd say that 2, 32 is a lot of aa for the scene - sod it if it gets results you're happy with though. I'm gonna do a video of my approach to it, you can be a guinea pig and run through the method and see if it makes sense for you on that scene.