FastSSS2 extremely slow

[AlexanderLeeCG]

I'm getting about a 10x increase in render time when I am using FastSSS2 as opposed to a standard material, and this is with Prepass at only -1.

Here is a render that took 2 minutes with standard material:


And this one took 20 minutes with Fast SSS:


Is this typical? I am using Irradiance Map GI, but it was very slow with Brute Force as well.
Here are my Fast SSS parameters:

[vlado]

What mode are you using for multiple scattering in VRayFastSSS2?

I'm wondering also about the specular options - can you show them too?

Best regards,
Vlado

[AlexanderLeeCG]

Thanks for the reply, it's really amazing that you answer questions personally!

I think you can see the multiple scattering mode settings in the screenshot.. unless there are more elsewhere that I don't know about.

I just tried turning off Back Lighting, and this seemed to improve the render time, but it's still very slow. I also tried turning the Specular intensity to 0, but that had very little effect on render time.

My character is to scale, at around 185cm tall, but I have my Scatter Radius set to 30cm for testing purposes, so maybe this excessively high setting is having some negative interaction with other settings? I tried decreasing it to 2cm but I got exactly the same render time.

Here are my complete settings (that I know of):

[AlexanderLeeCG]

also, I have noticed in testing on a simple cube that I need to set the Scatter Radius about 10 times higher than expected to get a desired scatter distance. I have a 1cm cube and I have to set Scatter Radius all the way up to 10 to get the light to travel the length of it.

[Svetlozar Draganov]

Specular Subdivisions are way too high - try to revert them to the default value of 8 and run another test render.
The Scatter Radius is not the only parameter which determines the final result for the scatter distance - another parameters that influence that appearance are Scale/Index Of Refraction/Overall,SubSurface,Scatter colors/Phase Function/Single Scatter type and things like Lights position/intensity.

[AlexanderLeeCG]

No, the Specular Subdivisions are actually just fine -- this is because of my Image Sampler settings, where I have very high Max subdivs. I did render with Specular Subdivisions at 8, but it just made the render slightly slower because the image sampler had to compensate more; but overall the effect of the change was minimal, and completely turning off the Specular Layer had a minimal effect as well.

[vlado]

Hm, there is nothing strange in those settings, so there may be something else going on. Maybe best to send a scene or a similar one to support@chaosgroup.com so that we can debug it.

Best regards,
Vlado

[AlexanderLeeCG]

I've very happy to say that I think I found the problem. Now I can render the FastSSS at a Prepass rate of 1 almost as fast as a standard material.

tl;dr: when using FastSSS2, keep Adaptive DMC Image Sampler Max Subdivs low, like 4 or 8.

It appears that the Adaptive DMC Image Sampler "Max subdivs" value has a very significant effect on the speed of the FastSSS prepass. In my original tests, I was using Max subdivs of 50, which is way higher than necessary, and was left over from my tests with Nederhorst's universal settings. However, having Max subdivs at this high value had no ill effects on my rendering with the standard material; it just required that I set the subdivs for each light, GI, specular, etc high enough to compensate for the division by 50 subdivs.

But I found that if I lowered Max subdivs to 8, or even better to the default of 4, which is all that is necessary for adequate anti-aliasing, the FastSSS prepass became immensely faster. Adjusting the "Single scatter subdivisions" value seemed to have no significant effect.

After I adjusted the subdivs of all the lights and specular and such down to match the division by 4 subdivs, it left my render time of the standard VRayMtl the same as it was before, but the calculation time of the FastSSS became probably over 40 times faster than the old tests, allowing me to increase the prepass to 1.

So a word of warning to people trying to use Nederhorst's settings: he says in his Vray for Maya primer that he recommends setting Max subdivs all the way up to 100. This would cause a massive slowdown for FastSSS. I think it is best to use the Image Sampler only for anti-aliasing, not for cleaning up noise, and when it is used for this purpose alone, a Max subdivs in the range of 4 to 8 is adequate.


VRayMtl, rendertime = 1:06


VRayFastSSS2, Prepass rate = 1, rendertime = 1:16

[Svetlozar Draganov]

Glad to hear you sorted it out.

The Image Sampler will only reach to the Max Subdivs if the Color Threshold is too low.
Universal Settings (1/100) are a general approach to tweak render settings, they will not provide optimized result but a lot of people prefer them since they can control the quality of the render with a single parameter - Color Threshold.

You will be able to get faster render times by increasing Color Threshold value accordingly.

[AlexanderLeeCG]

^ Yes, the actual image samples will only reach the max amount if there is lots of noise and the color threshold is low. And the render can be made much faster by increasing the threshold, in non-SSS situations.

The extreme slowdown of SSS prepass calculation seems to occur regardless of how high the color threshold is set. It occurs in the prepass phase of rendering, at which point the Image Sampler has not yet begun operation. There is apparently some connection between the Max Subdivs value and the internal calculation parameters for the SSS prepass - maybe when Max Subdivs is set higher, the prepass has to calculate a larger set of data to deal with the potentially higher number of Image Sampler samples.

[Svetlozar Draganov]

There might be something else involved here since according to my tests the prepass is not affected by the Image Sampler settings at all.
I have tried with 1/4 and 1/200 but the prepass was calculated for the same time.

Would it be possible to send us the scene for investigation, you can replace the character with a sphere but please keep the lighting and material setup the same.

[AlexanderLeeCG]

Thanks for testing it yourself! I see now that you are correct.

I think the real cause was probably that when I increased to 1/50, I also manually increased the subdivs for all the lights in the scene to keep it optimized, and these high subdivs were being applied to the prepass calculation. I will do some tests to make sure that this is correct.

This means that the SSS would probably actually calculate at a reasonable speed when using the Universal Settings, as long as the user doesn't mess with the default light subdivs.