About noise in general

[BBB3]

Thanks Illab. Definitely one thing to try.

I'm not a friend of solutions that involve tweaking the material subdivisions on objects, though, as it make it nearly impossible to work with a library of stock objects if you have to go through each material and adjust the subdivisions depending on which GI solution you are using.

[greysheep5]

AA 1x 9x doesnt give necessarily more samples than 1x5x, see the interstation dmc tutorial, then you should understand how and why high AA max values reduced the overall taken material and light samples:
http://www.interstation3d.com/tutori...yfing_dmc.html

the AA threshold decides about the number of samples used in end.
for me 1x100x is way faster than 1x4x for me, but also more grainy.
so one lowers the AA threshold to 0.005 or lower p.e (dmc threshold stays at 0.01),
1x100 /0.005 is still faster to me than 1x4x, -both noise free solutions.

this is irritating at first, as this was not so in older versions,
but actually i like the way, and it helps getting more speed in end. important is only to use in different new way.
p.e.: for br-lc still/animation i use 1x100x, with AA threshold around 0.005-0,003, and i use pretty high subdivisions everywhere (glossy, lights,etc), as they are divided by the number of max AA samples. also in br i use 256.
the nice is i have then only one value i need to control for the image quality, thats the AA threshold,
rarely i adjust the dmc adaptivness, usually about 0.7-0.95 (similar to a new type of universal setup for me at least)

i am not saying that this might help for Bertrands scenes or similar cases, but maybe worth a try?

cheers
stefan

so vray for c4d works just like vray for max i suppose?

why do you unlink the clr trshold of the adaptive dmc sampler settings from the general dmc settings if i may ask, did i understand that correct? can´t you just let the adaptive dmc sampler linked to the general dmc sampler settings an adjust the noise treshold there?

anyway... i don´t quite get your saying that it´s a kind of new universal workflow stefan, when you have to push and adjust all the individual light subdivs and the glossy subdivs in the materials? isn´t that tedious work and against the idea of the universal workflow?

so could one raise just the global subdivs multiplier in the dmc sampler settings to up all subdivs instead of tuning the individual values? of course that would be bad for the total rendering time but much less time consuming.

best regards,

christoph.

[greysheep5]

one more thing...

with our scenes the min/max of the adaptive dmc sampler with 1x and 100x plus a noise treshold of lets say about 0,005 is not only fast but has good quality.

but only if the lights and material subdivs stay at their default values, otherwise we get even better results but also massive rendertimes! it makes a huge difference if the light subdivs with the 1/100 workflow stay at 8 subdivs or if we push them to for example 128... of course, that should be this way i guess, but most of the time we don´t need that extra noisefree quality which in our case makes vray very slow. but maybe you have more rendering power than we have, stefan.

i´ve been studying the http://www.interstation3d.com/tutori...yfing_dmc.html infos here over and over but i can´t get any real conclusion out of it.

so maybe it´s time for a new universal workflow tutorial vlado, i second that wish!!!

best regards,

christoph.

[simmsimaging]

I've been testing on a current image project and what I found was that neither material and light subdivisions, nor AA sampling levels had a significant impact on grain, or render time for that matter. The only thing that really reduced fine grain, and had the big impact on time, was the noise threshold (I use the DMC sampler).

I ran tests at 1/20 up to 1/100 and with global subdivisions at x1 and x10 (material and lights are all basically default at 8 samples). Increasing either of those made a nominal difference to my render time (like 10 secs on a 4-5 minute renderr) but also made almost no difference to noise. Lowering the noise threshold from .01 (which is my usual default) to .006 nearly doubled the render times - but did visibly reduce the noise.

Might be scene specific, but my experience has always been similar. All that said, I actually don't mind a bit of noise as I find it helps me to integrate into photography, which is often what I'm doing.

/b

[PIXELBOX_SRO]

Glad to hear I am not the only one... thought I was going crazy these last few days... more sample = more noise... wha?! very strange. In my case it's the skydome providing dark night lighting with no GI. 1/5 gives a clean result, 1/9 brings noise.
Did you get to the bottom of it Martin?

lol...nope...i ve been poited towards the DMC tutorial to address the issue...apparently when i come to understanding that tutorial, i will find out that it is actually quite normal

NOW

what i dont understand is the actual value of the noise threshold....what does 0.1 mean? what does 0.002 mean? whats the mathematics behind it?

[joconnell]

In general terms any of the noise thresholds set how much difference there is allowed between samples for vray to stop increasing quality. Say you're doing gi light sampling and a you've got 5 rays being bounced off a surface. The first ray hits a light and returns a bright white colour (lets say rgb 255,255,255), the second hits off a wall and returns a red colour (255,0,0), the third goes out a window into the skylight and gets a blue value (0,0,255), another ray hits some random green thing (0,255,0) and the last hits a black object (0,0,0). All of these samples get fed back into vray and it starts to compare the results of each ray. Lets say you've set a noise threshold of 50 for arguments sake. What vray will do is look at our 5 samples of white, red, green, blue and black and see how close their RGB numbers are to each other. Since all of our numbers are nearly 255 apart and our noise threshold is 50, vray will say that the samples aren't close enough to each other to give us a smooth result, so it'll go to another level of sampling. It'll fire another set of rays which will hit surfaces between our white, black, red, green and blue surfaces and lets say for convenience that it manages to return values that are half way between each colour so our rgb numbers are now 128 apart. Again since our noise threshold is 50, vray will do another level of sampling again and if we conveniently manage to get colour samples in the middle of our new values, we're down to the difference between each sample being 64. Vray will have to do one last round of sampling to get us below our threshold of 50 between all of the samples, and then it'll stop.

The threshold is pretty much an accuracy value. If you've got samples that are really far apart in numbers, this will be visually represented as grain in your render. If you get samples that are close together, it'll be smooth. The noise threshold is telling vray how close together you want the samples to be before it stops refining. In some ways it's similar to what maxwell is doing when you give it a certain amount of subdivions it can use at a maximum, or a certain amount of time that it's allowed to render for. With every scene being slightly different in terms of it's range of colour, contrast and difference in lighting you might find that you can use bigger or smaller thresholds to get a clean result - for a simple pure white scene with a dome light you might be able to use low thresholds since all of the rays in the scene are going to be hitting white or shades of grey, all of the samples will average out very very quickly. If you've a complicated scene with loads of different lights, materials and parts that are dark and parts that are bright, there's a good chance that the rays boundnig around the scene are going to have a really wide range of values. To try and bring those values close enough to get a smooth result is going to be far harder than our white scene so it'll make vray use lots more samples to achieve the same level of smoothness.

[PIXELBOX_SRO]

Hi Joconell,
thanks for a very detailed insight!

[greysheep5]

I've been testing on a current image project and what I found was that neither material and light subdivisions, nor AA sampling levels had a significant impact on grain, or render time for that matter. The only thing that really reduced fine grain, and had the big impact on time, was the noise threshold (I use the DMC sampler).

I ran tests at 1/20 up to 1/100 and with global subdivisions at x1 and x10 (material and lights are all basically default at 8 samples). Increasing either of those made a nominal difference to my render time (like 10 secs on a 4-5 minute renderr) but also made almost no difference to noise. Lowering the noise threshold from .01 (which is my usual default) to .006 nearly doubled the render times - but did visibly reduce the noise.

Might be scene specific, but my experience has always been similar. All that said, I actually don't mind a bit of noise as I find it helps me to integrate into photography, which is often what I'm doing.

/b

for one of our scenes it made a quite huge difference. we pumped the light subdivs of all the lights from the default 8 up to 128 and let the rest of the settings untouched (irrad map + lc, dmc sampler min=1 max=100) and the rendertimes went through the roof. but it was a mixed indoor/outdoor scene (entrance of a building lit with peter guthrie hdr on the outside and a suspendes glossy ceiling with area lights above it) which was kind of complex renderwise. the time went up to like 4h for a 2500x2500 pixel sized render on 3 modern quadcore i7 machines with 16gb ram, times we are not used to as normally everything renders in max 1h for our use, except very highres special renders. but it was not a very scientific test, i should look into this scene once more with other settings to have a good comparison i guess.

like you brett we don´t mind a bit of noise so we normally tend to stay with a 1/100 setting and the standard subdivs in light and material. and yes, the value of the noise treshold makes quite a difference. we tune it from 0,2 for very rough previews to 0,1 for normal renderings up to 0,008 or even 0,005 for very fine and not so grainy renders... just about as you do from what i read above. but 1/4 can be slower and produce better quality than 1/100, that is also a thing we observed, like stefan said. so the lc+irrad workflow with 1/100 is quite perfect with us in terms of quality vs speed.

best regards,

christoph.

[joconnell]

Hi Joconell,
thanks for a very detailed insight!

All good - I wrote that a bit tired and hungover, I reckon I can redo it more clearly! I really want to figure out the DMC, but I've got a few gaps in my knowledge to work out.

[lllab]

hi Bertrand,

what i meant is i use the same high subdivision all time, so nothing to tweak and very fine for premade libs, we use all presets too with that settings.
what is an alternative maybe is to use the global multiplier but, i havent tested on this in detail yet (from pure logic should be the same).

the only value i adapt is threshold in AA, sometimes to 0.003, (but not changing the dmc threshold, it is not the same as the AA threshold and i dont need it has low as AA thr. in 2.0 core, DMC thrs. i keep usually at 0.01 or lower not more than 0.005), in my works it renders faster noise free in that setup than with low subdivisions. it seem then to get easier noise free in shorter time.
as some already stated this system seems not to suit all work flows, for me it is definitely faster, for others not though, they seem to have better results keeping it at 8.
so only trying which is better for yourself might bring something.

cheers
stefan


of course it might be not for all type of work flows, so i can say only from my working.

cheers
stefan

[pengo]

Sorry lllab do u use high vraylight subds too like 64 or 128 or more combined with the adaptive image sampler min/max 1/00 and clr threshold to 0.003 (noise threshold if i correct understood u leave it by default to 0,01 to get not too high rendertimes) ? How many subds in materials generally?
Sorry my bad english
Thanks in advance

[lllab]

normally all 256, lights, mats, dof etc.
from pure calculation 1x32 woudl be optimal dmc(256/40=8 ) then according to the DMC tutorial, but in my tests higher max AA makes it faster.

[pengo]

Sorry surely I'm stupid but testing all the possible combination trying to reach at the same time with global subds multiplier the hsph subds in irrad map to 50 as default (because I suppose we are talking irrad+LC) mats and shadows subds to 256 (leaving them by default to , I get anyway noise GI, so I ask is needed to cranck up the irrad map hsph subds too until 256 with your method?
And do u use with those high parameters 1 as adaptive amount? Or u leave it to 0,85?
In few words have I to cranck up even irrad maps hsph subds (leaving global subds multiplier to 32 to get 256 (default mats and light subds 8 x 32 Global subds multiplier= 256subds)
And doesn't all be divided by 100 that is the max value in AA image sampler ?
If so I should use in hsph something like 157 x32 Global sample multiplier I'd get=5024 subds /100 max AA value I'd get 50
Is it correct?
And to be clear I specify that the example scene is a little room with only one door-window lighted by a dome light with guthrie's map
Sorry maybe are stupid questions but I'd like to understand as clear as possible with my bad english .
Thanks again

[olika]

Hello to all DMC Scientists
Couldn`t really remember that i`m ever staring on a thread like this. Since i found this thread i look at it several times a day if some new posts and conclusions coming up. So i really feel a need to contribute *something* to this thread to keep it on, cause i`m really looking forward to find the solution in this DMC science. The DMC prevent me to sleep enough for a long time now. I have the feeling that most things i thought i discovered over the years are mostly crushed now (again oh no!). What i learned over the years is when i thought i know something, this was wrong... Think that`s CG in a way. Saturday i run also some tests on my own.



Realized to interesting things.
First thing, in the past i hardly dared to increase the brute force subdivs cause when i remember right rendertimes go insane?! Now it`s faster with higher subdivs. I guess maybe it`s a cleaner GI solution (less noise from the beginning) what the AA DMC sampler had to deal with and even with more work for the GI, the AA DMC had less work and therefore it`s faster?

Second thing was, i could be wrong but, i had to increase the Global Multiplier to 32 to produce the same result with the rendering before. So all three pictures before i increased subdivs in Light and BF to at last 32 sudivs. In the last Image i decreased the Light and BF back to 8 subdivs and increased the Global multi to 32. But this is not multiply in my opinion. Multiply would be a global of 4, so 4 x 8 subdivs are 32 subdivs, right? Instead it`s more like a override when you have to put in 32 to reach the same result. ?? When i`m right the word multiply mislead me in the last years. Ok, this post is hardly about DMC thing in the end, i realize now, but all those things operate together right? Will try to continue the tests (especially DMC behavior in connection with this) with this super simple setup, cause this super simple thing is even enough to get me crazy in fiddling around for hours!
Hope we`ll find peace in the end!

Grüße
Oli Kentner

[lllab]

hi pengo, i didnt mean IR hem subdivision (this is also not affected by the global multiplier),
only light, glossy, brute force etc get the high subdiv. combines with high AA max (which divides the subdiv),