FWIW, the scene does use LC retrace, and LC prefilter is at 50.
i think llab has sorted it…at least according to some of my tests
yes, I already tested what llab suggests but it’s seems to be totally different from what Brett is saying
no it is still all based on this paper/tutorial i think and not contradicting, up to a certain subdivision the AA threshold is the only really counting. i think thats also in the dmc tutorial. additional this special case only considering some extra related things like glossy ray interpolation and retrace value which is not covered by this tutorial
stefan.
confusion is rising… ![]()
lllab method is similar/derived to/from interstation 3d tutorial (high subds because of high AA …well, more or less), ok?
Brett said that he uses default values for subds, that means 8 (not an high value)
Actually my opinion is slightly different:
tweaking the DMC sampler setting allows me only to solve antialiasing problem, no more noise coming from materials or lighting; for that I always end to raise their subdivision on local basis (very time consuming), I don’t like use global multi because it always lead to rendertime waste.
I have also made some test using high AA setting like you suggest (100) but for my scenes is slower; as an example right now I’m working with max 32 AA and 0,03 thershold (subs are various starting from standard 8 up till 256).
To summarize:
glossy and light/shadows are solved by their subds
antialias is solved by AA
crl threshold controls general noise/speed ratio
…pretty basic workflow but actually for me is the only way to go
Well, I have a few test I could upload if there is any interest…
They are from a working scene, so are not totally complete.
They are full materials applied, multiple Vraylights, VrayDomelight+HDR, vray DOF, opacity maps, vraylightmaterials, glass, vrayblendmtrl, tons of maps/textures, overall a good mix…
Oh, to clarify an earlier comment I made, I did not mean to say the Chicago scene is not a “real” one, I just meant one with materials, etc. working scene with client’s specs. ![]()
Just to clarify: my workflow is very similar to yours. I use default Sub-d’s in *lights*, but often crank up material sub-d’s as you do. I tend to use 32 as a base - not from stringent testing for optimization, but just a number I have found tends to be a solid minimum given that I use tons of glossy reflections.
However, in the last small test I did neither light nor material sub-d’s made as much difference to the existing noise (based on using the global sub-d to crank them up) as the DMC noise threshold value did. That part is definitely not in-line with llab’s results from what i understand. It might be due to the fact that most of my renders are single or few objects rather than interiors, but usually with fairly complex materials and lots of lights and VraylightMtl geometry lights.
“However, in the last small test I did neither light nor material sub-d’s made as much difference to the existing noise (based on using the global sub-d to crank them up) as the DMC noise threshold value did. That part is definitely not in-line with llab’s results from what i understand”
it is exactly what i say if i understand you right.
it doesnt matter if you do it via global subdivision or direct in the settings, important is they are only high, so they dont play any matter anymore(due the high dmc AA, which takes overall, as also in the dmc tutorial)
stefan
p.s overall i am sure there are many ways and variants which leads to noise free setups, not only one method.
No, i should have been more clear: there was almost no real difference in the noise or the render times when I raised the material and light subdivisions, which I did by cranking up the global multiplier from 1 to 10. The noise was the same till I lowered the DMC threshold. That cleaned it up, but also increased the render times as expected.
that’s was exactly my question, how do you solve the noise coming from light with only 8 subds? with my scenes (actually interiors with quite a lot of objects and lights) I’ve tested that using DMC for that kind of task is slower that crank their subds
I hadn’t, until now, really tested it too thoroughly. I was trusting to the direction of the Universal method so if it was really noisy I would just raise the AA levels (up to 1/100, but usually around 2/50 would be all I could see a difference at) and lower the noise threshold down to .006 or .007 (rarely, if ever, went below that). Raising the light sub-d in my recent test didn’t seem to make any difference at all though.
for me is fine if you could share a test scene, hoping someone else to follow
yep..i am in too ![]()
um, I am sorry, but I meant test renders, not the actual scene…sorry for any let downs ![]()
edit: But I will put up the detailed settings, at least…
all lights = 256
1st bounce: BF = 256
2nd bounce: LC = 1000 retrace checked, use for glossy checked, prefilter 50
ADMC = 1/100
all material glossy subd = 256
phys. camera DOF = default 6
AA = none
Envir override = no
color mapping = reinhard 1.0, .25 burn, 2.2 (first 3 boxes checked, clamp level default)
DMC sampler = default
region division = 32
Should be basically what Stefan had posted in his tests.
I should also mention that for these I had the Vray lens effect, but not turned on, in case it makes much difference…
Obviously, render time changed drastically as aDMC clr thresh changed…
originally rendered at 1200 pixels (HDTV) but forum resized them…so keep that in mind when looking at the times…
Also, these are straight from the vFB with no post whatsoever, except watermarks ![]()
Thanks for the test Voltron. It is a good baseilne and I would say this is pretty much the expected result: that that raising the noise threshold that much would be both faster and far grainier. I’d be very curious about the results of picking one noise threshold values and doing some comparisons at various sub-d levels for the mats/lights/gi, and also at fixed sub-d’s and various AA levels? I think that would really help narrow down the time/quality contributions. It’s pretty long render times there though - so maybe even just a section could be done for speed?
(Colin’s Tweaker script might be really useful for these sorts of tests…)
/b
its strange that you can still see some noise in the sample where DMC is at 0.004…but i guess its because of DOF set at 6
my tests here render noiseless
currently renderign image at 11.500px wide and almost 4.000px high with this method.
rendertimes ar skyhigh ![]()
rendering on 7 comps through DR (various GHz and core count comps) and it seems like the rendertime will be between 24 -30 hours :)))))
Just wanted to test it
will post some samples when its done
BTW this method really “LOVES” opacity mapped leaves :)))) …crawls when it hits such geometry.
now i think i coould have picked a smaller size image to really test the setup ![]()
may I ask if you have any elements already available? …reflection, lighting, shadow and sample rate
interesting:). make sure to use subpixel mapping on( i guess you have, just to make sure), otherwise it will be very slow. retrace i use normally at only 0.2 or highest 0.5.
for opacity maps make sure to use no filters which blur the maps. they are fast here too:-)
voltron7, what computer do you use? it seem very slow for the rather simple scene you tested?
specially the simple wall material seems so grainy in your tests?
maybe the “candle lights” in the glass behind the opacity mapped cone are the reason for slow down?
These were rendered on a single core 2 quad Q9550 2.8ghz with 16GB RAM, windowns vista 64bit.
**edit - also involved was a single, old, slave (core2duo 6700 2.6 ghz 4GB RAM)
Nothing fancy or up to date.
Scene is no really that simple, just view shows small area.
There are 70 lights and about 3.5 mil polygons…
Floor tiles all modeled with floor generator.
Walls material is actually a vrayblend with glossies, bump, etc.
Actually, all materials in the scene have some sort of reflective value.
Interestingly enough, subsequent test with the Dome light turned off have reduced render times a lot. Still as grainy, but at clr thresh of .01 without dome it was looking at 30 minutes vs. 1h:21m
I didn’t render any elements…
LC retrace was at 1.0, although LC calculation was the fast part!
Still will try more tests asap, but new work is pressing again…
the high rendering times for this level of noise are the reason I di not use BF/LC for that project.
I was getting much cleaner results with IM/LC with render times no longer than 30 minutes for that scene.
So, for now, I am still sticking with IM/LC for production.
edit: The wall material started as blend (like varied venetian plaster, but client wanted it to read even, so I just turned down the blend, but now I think maybe I should have just used a simple material…


