unified sampling 2012

[gustavoeb]

is unified sampling any good?

[origin]

is unified sampling any good?

There's something off in the way unified sampling works (or how it is quessing there are enough samples for pixel). I'm a bit used to vray ADMC, the way it works is pretty straightforward...I'm doing tests right now, will post results later.

[Maximus]

i'm running tests with Unified Sampling and it seems pretty promising, i must admit i'm still confused on how exactly works, but one thing i've noticed is it samples glossy reflections and area shadows while the normal antialias method doesnt.
Unified Sampling is a bit slower than traditional method but as i said it avoids you going to use samples in the shaders for glossy for example or into area lights.
Gonna test it further!

Also i've noticed that if you have enabled the Unified Sampling the Diagnostics "View Sampling" and "View Final Gather Points" dont work, wonder if it is a bug or if MR really dont let you use the diagnostic while running in Unified mode.

still didnt try ICE, but so far is awesome release!

Max

[Kzin]

did also some tests with unified. sofar, its not that fast as progressive rendering with motion blur. it renders faster then normal aa, but i have to say the mb performance is now way better in generel. will do some more test with glossys and area shadows, also ambient occlusion, things that needs alot of sampling.

i had high hopes that mi worked more on ip's, but nothing changed, lets hope for the next release for improvements or a new gi engine, dont know.

but again, no bsdf shader (which renders alot faster for brute force things then the brdf mia_mat) and also no ibl shader in softimage. would be interesting to know why it is not in.

[origin]

It seems quality = 1 doesn't reach max samples? Or I didn't understand how it works...

[grahamef]

It seems quality = 1 doesn't reach max samples? Or I didn't understand how it works...

Your cutoff settings are stopping the sampling before it reaches the max samples. Basically cutoff is there to bail out once you're "good enough", i.e., once the difference between successive sample levels is less than cutoff, it doesn't bother sampling anymore. So by increasing cutoff from 0.01 to 0.1 you're guaranteeing shorter render times even if you increase quality.

[origin]

Yes I expected it to work like you described but in this example: min 1 max 1000 (!) quality=1 and cutoff = 0,001, and image is still noisy while cutoff is very small.

[Kzin]

Yes I expected it to work like you described but in this example: min 1 max 1000 (!) quality=1 and cutoff = 0,001, and image is still noisy while cutoff is very small.

because your quakity setting is to low to reach 1000 sample.


i am testing and testing but its really hard to get good rendertimes with good quality compared to normal aa, more hard with motion blur.
you need alot of samples if you want good quality for high raycount stuff, there is no way to get around this (brute force wise). ;)

[Maximus]

i found this Unified sampling really better if you are dealing with glossy and area lights, its way faster in cleaning especially on glossy where you would need 128 glossy samples in the shader to reach the same quality.
Same goes for shadows samples, i'll test it further and come back

[origin]

Yes I expected it to work like you described but in this example: min 1 max 1000 (!) quality=1 and cutoff = 0,001, and image is still noisy while cutoff is very small.

because your quakity setting is to low to reach 1000 sample.

the doc says something diffrent "You can also get more samples than Maximum by setting Quality higher than 1". For my understanding quality is something like multipler for maximum...but It doesnt work as doc says.

[Kzin]

i found this Unified sampling really better if you are dealing with glossy and area lights, its way faster in cleaning especially on glossy where you would need 128 glossy samples in the shader to reach the same quality.
Same goes for shadows samples, i'll test it further and come back

yes, you are right. my mistake, sometime i think "around some corners". ;)
its because i used progressive rendering and you have the same benefits with this rendering method, glossys, areas and motion blur renders way faster with progressive then with normal aa. with unified, its easier to get better quality for sampling things, but its not faster in generel. but i have to do some more tests.

@origin:
thats right. but i think with quality 1 you cant reach 1000 samples. if you use quality like 4 or something, you will reach 1000 samples (you used quality 1 and max 1000 samples in this example).
your tests shows exactly (good one btw) that, the quality value ist the most importent one, like its mentioned in the docs. from my tests its the same. you can use low max settings like 64 and quality like 1, if you increase it now to 2, the quality will be better. if you increase it to 3, you will see there is no difference in quality anymore. if you want a cleaner image, you have to increase the max sample value.

[Gabba]

So, rise the minimun samples instead of reduce the cutoff. Didn't tried to rise quality, but reducing the minimum you get better results. Also, another thing to consider. If it's like "progressive rendering" in the "old" ctrl-ghost the number of samples you use in the material will affect the final result. I have to search the old thread and see if I posted the tests I did.

[Saturn]

A few comments coming from mental ray 3.9 beta phase that explain a bit more unified sampling.
Theses are from Bart from mental images :

This is to help you understand the new Unified Sampling feature in order to test it.

There are several new string options, which you must add manually to the option block to test.

"unified sampling" on
"samples min" 1
"samples max" 100
"samples quality" 0.5
"samples error cutoff" 0
"samples per object" off

It is called "unified" in part because if on, it can control sampling for regular, progressive, and rasterizer modes from the same simple set of inputs.

"samples min" and "samples max" represent the minimum and maximum number of samples per pixel. "samples min" can be a scalar (floating point number) because you can start with a sample density less than 1 per pixel. Eg, "samples min" 0.25

"samples quality" is a scalar (floating point number) that can go above 1.0 if desiring extreme high quality. Using internal error measurement criteria, the quality helps determine how many samples per pixel are actually taken. Roughly, if there are areas of high error, then more samples will be taken in that area.

You should be able to get reasonable speed/quality tradeoff with most scenes using just the above mentioned controls. We're shooting for simplicity in control, as we continue to refine it.


"samples error cutoff" allows for stopping sampling in very dark, black areas where the error is less than the given amount. The default is 0. But it could be that raising this to some small amount for some scenes could increase performance without visual quality reduction.


"samples per object" allows for overriding samples min and max at the object level. This reuses the existing samples <min> <max> flags that can be put on objects, but interprets the min and max as absolute samples per pixel, like the above samples min and max, rather than the power of 4 like it means with regular adaptive sampling. With unified for example, one might use
samples 8 100
in order to take a minimum of 8 samples per pixel for the object.

In the rasterizer mode,

"samples max" will be converted into a "samples collect" number to represent the sample density used for the rasterizer

"samples max" 64 becomes "samples collect" 8 (8x8=64 samples per pixel)
"samples max" 65 - 80 also become "samples collect" 8
"samples max" 81 becomes "samples collect" 9 (9x9=81 samples per pixel)
"samples max" 82 - 99 also become "samples collect" 9
"samples max" 100 becomes "samples collect" 10

"samples min" is ignored

"samples quality" 2.0 becomes "shading samples" 2.0
roughly indicating how many polys per pixel with a sample for each poly

The motion blur point is a key benefit to unified sampling. It will also handle depth-of-field, sampling only where extra sampling is needed to reduce error.

So you will typically need less local samples for:

area lights
glossy samples
ao samples
dof samples

time samples/contrast is ignored

One might even consider force FG

To further clarify, here's the rough steps:

1) Take samples min number of samples.
2) Calculate error for each pixel. This will vary across the image
3) Determine the number of samples desired for each pixel using the samples quality and the error. Limit that number to samples max.

That gives us the samples per pixel across the whole image, and provides the foundation for the simpler strategy for artist control. Now to refine that think of these adaptations.

4) If error less than samples error cutoff, stop taking samples.
5) Otherwise, start taking more samples for this pixel toward the desired target determined in 3)
6) After a portion of samples are taken, we re-calculate error and adjust that desired target for pixel samples and continue taking samples if we need to.

So, visually, if you want more anti-aliasing quality across your image, increase the samples quality. Now if you feel in the low error areas, like flat walls, that you don't need to increase quality anymore, you can set an error cutoff. The default is 0, so it defers to whatever your quality setting helps determine.

For example, I see I need super-high quality in some detailed part of the image, but I'm OK with 1 sample per pixel, my samples min, on my flat wall. Then, setting samples error cutoff will save some rendering time, so that it doesn't over-sample in those low error areas given my visual judgment.

You can actually look at both error (from the final error calculation step) and number of samples in layers in the diagnostic.exr written out if you use the

Code: Select all

-diagnostic samples

option at the command line.

Try "samples quality" 1.0 to get a baseline to take a look at the diagnostic.exr file. And remember that you can make quality lower or higher than 1.0.
__________________

[Saturn]

Noticed there is a new frambuffer called pixel time.
It helps to see where mental ray is sampling the most.
It remplace the view sampling diagnostics mode when using unified sampling.

[Kzin]

thx harry, in addition the maya 2012 thread for unified sampling
http://forums.cgsociety.org/showthread. ... 7&t=970700.

for future version it would be great that the sampling completly takes control over area shadows or ao, so you need only one sample for ao computing for example.
the rasterizer examples are helpful, will start to test it today with hair and ibl.

besides, ibl ist now working with mr 3.9 without spec/sampling artifacts and it renders really fast.

Moderator edit: please use the URL tags - HB

[Saturn]

I noticed the spec too it was coming from Xsi environment shader. If you used the mib_spherical_lookup it was working.
It is sad we don't have access out of the box to the IBL mode. It's damn good.

That the path they are actually taking. Removing any sample and make thing to be control globally. Have a look to the built in bsdf shader.
They don't have any sampling control and they are a lot smoother while cast less rays.

That said you can't use really bsdf shader with iradiance particle. You have to connect them into the new bsdf slot in the material shader that is not exposed.