V-Ray Scene Optimisation Script

[^Lele^]

Sorry for the late reply; there is no need to change the light cache settings when changing the resolution.

Vlado, I made a test of this, and found it's not quite as you say, if you excuse the terrible presumption (/me hides).
I did the same type of tests ages ago and gave it for granted since, but so many things changed that i thought i did indeed lose grip on the way the LC works.
Yet, these tests seem to confirm the idea of a shading rate is the one providing for better consistency.

I took the scene we're testing, and slapped a gray diffuse shader as override.
Then rendered it with the default LC settings (1k subdivs, 0.02 sample size, nearest filter of 5, retrace on) at the original resolution and at double resolution.
Then i changed JUST the subdivs to a sample rate of 1 (768 for the original res, and 1536 for the double resolution).
Notice that by the Sample Rate concept, the original resolution image is slightly oversampled (1k subdivs, versus 768 ) while the double res is slightly undersampled (again, 1k versus 1536 ), so the case seems to be an ideal one to check the LC behaviour.
I then saved the exrs, rescaled the double res to match the original res, and subtracted one from the other.
Here are the results:


Not only (invisible here) the min and max pixel variance (curve tool) is slightly reduced for the SR 1 method (which for this scene, by my count, is actually a bit too low), but there is no sign of those brighter and darker splotches in niches and crevices, visible in the other image with default settings.
Notice that changing reformat filter makes no appreciable difference to the final result.

This result is indeed consistent with what i experienced during production, where moving objects and deforming geo were present, and rendering an oversampled LC per frame led to much stabler results across sequences (where the camera path option was not usable), with no visible "bubbling" of the GI solution.

I would very much love to hear your thoughts on this, as i feel i still do not quite understand the LC ins and outs as i ought to.

p.s.: i also have the same comparison done without retrace on, and with direct LC visualisation, although i find those a bit to the side, and not so indicative of real production setups.

[^Lele^]

To compound on my misery, i rendered the default res shot with an SR of 4 (ie. 1536 subdivs, those used for the double res), and the results show a much better shadows definition in the delicate transitional areas, a generally better contrast across the image, and better self-occlusion from geo.
It's also marginally darker (as the difference, SR1-SR4 is bright-ish).

[eyepiz]

Hello,

Interesting thread, good stuff!...I just though I would throw the scene in RT GPU and see what the outcome would be...Beside the unsupported features, Ie: displacement, Fast SS2 it did not look to bad for 20 minutes unbiased.

[^Lele^]

Very nice!
Care to test render the scene I linked on google drive a few posts earlier? it's a wee bit cleaner than the GPU one, which looks more in the ballpark of the 22mins render i did last, but i'd be curious to see what's the difference.
Speaking of which, what's the hardware you're rendering with?

[eyepiz]

Very nice!
Care to test render the scene I linked on google drive a few posts earlier? it's a wee bit cleaner than the GPU one, which looks more in the ballpark of the 22mins render i did last, but i'd be curious to see what's the difference.
Speaking of which, what's the hardware you're rendering with?

Sure...I'll test it when I get home tonight...Here is the production version of the same render with my production settings, obviously not optimized for speed.



Dual 8-Core Xeons.
(4) GTX 6Gig Titans

[^Lele^]

I am DROOLING (on all counts.). ^^

[Morne]

Nice script.
If I'm using the archviz preset and don't need super clean elements as not much compositing will happen afterwards, whats the 1st thing I should look at to bring render times down - possibly even half or a 3rd of the time (for argument sake lets assume some noise is "acceptable"/render time)

[^Lele^]

Well, the time spent rendering is proportional to the numbers shown in the info panel at the bottom (SPP, or samples per pixel), to the noise threshold sensitivity, and to the adaptive amount.
Aim for 512 samples across the board, and for a noise sensitivity of 8 bits, to begin with.
You can accomplish that through the GSM multiplier alone (it supports alt+drag for very fine increments. 0.529 will do the trick, BUT it will keep the balance so that direct lighting is samples less, and GI is sampled more than shaders.), or by setting all the base subdivisions to 8, and the GSM to 1.
You may also want to set the max AA to 4 (so the image is rendered a fixed-4 AA) if your geometry isn't all that needy (ie. large flat areas, and no pixel/sub-pixel geometry).

EDIT: if that's still too slow for you, feel free to go lower: 256 SPP is decent for a preview without noise poking your eyes out, but of course the harder-to-sample areas WILL look well noisy.
EDIT2 : result for 256 SPP and a 4/4 AA (a tenth or so of the rendertime for the ArchViz preset):

[Morne]

EDIT: if that's still too slow for you, feel free to go lower: 256 SPP is decent for a preview without noise poking your eyes out, but of course the harder-to-sample areas WILL look well noisy.
EDIT2 : result for 256 SPP and a 4/4 AA (a tenth or so of the rendertime for the ArchViz preset):

[ATTACH=CONFIG]20554[/ATTACH]

So then 512 SPP on your pc will take 14 min for this frame size and also be a little cleaner?

[Morne]

What am I looking for in the readout for SPP?
Theres a bunch of numbers, if I choose archviz preset, I don't see 512 anywhere, just my brain staring at all the numbers.

...And when would an Archviz animation with moving objects (cars driving, people walking, fountains) not work with the Archviz preset and instead I should use something like VFX-std?

[^Lele^]

So then 512 SPP on your pc will take 14 min for this frame size and also be a little cleaner?

That was the whole aim for me as a TD: to be able to predict final render times quite accurately.
Here's the results of 512 SPP across the board:


as the image name says

edit: any slack from exactly double the render time is due to the first core being used for everything else as well. it's Spa-Francorchamps weekend, you see. i can't close my f1 news pages... :P

[^Lele^]

What am I looking for in the readout for SPP?
Theres a bunch of numbers, if I choose archviz preset, I don't see 512 anywhere, just my brain staring at all the numbers.

...And when would an Archviz animation with moving objects (cars driving, people walking, fountains) not work with the Archviz preset and instead I should use something like VFX-std?

SPP is indeed written nowhere.
You want to look at the info panel where it tells you the Total (shader, sss, lights, ibl/gi) Samples.
The Base samples show only how many samples (per pixel: SPP) will be cast per AA ray.
Move the Global Subdivs Mult. spinner until one of the TOTAL samples lines reads 512. then you'll understand.

[AlexP]

Im getting an error trying to make scene changes. It seems to be trying to access a render I saved (and have since deleted) for some reason?

[^Lele^]

Im getting an error trying to make scene changes. It seems to be trying to access a render I saved (and have since deleted) for some reason?

Umh, i have a feeling the path is set into the "separate render channels" before the script runs.
That being a bitmap construct, not a simple string, not finding the file on disk errors out on either save or restore of the properties.
I could fix that with a try/catch.
But if you could try and wipe the path to confirm it's indeed what i think it is, it'd be helpful.

[Morne]

Move the Global Subdivs Mult. spinner until one of the TOTAL samples lines reads 512. then you'll understand.

OK if I bring GSM down to 1, then Total Lights Samples is the 1st thing that comes down to 512 (I think)
So now only place I see 512 is here and it says 512|(1024) / 800|(1600)

Does that mean it is now 512 SPP? Not sure I understand the min max etc. What is the 512|(1024) / 800|(1600) numbers in bold then?

And somewhere in the thread you mentioned I should aim for less camera samples vs the rest. So if everything is 512, what should camera samples be then and what part of the info panels tells me that?