Good LC subdivs number?

[thomthom]

In the VfSU manual it says that the LC subdivs should be set so that at the end there is few black spots. If the screen if filled up when LC is only half you can half the subdivs.

This has been a thing that's had me pondering as I often use a low amount, usually 250, for small test renders, but then when it comes to produce the final render I'm not sure how many subdivs I need. And sitting watching the LC calculate only on a trial and error basis hasn't been ideal. All though, experience had given me numbers that usually work.

Today it occurred to me that I can calculate the exact number of subdivs needed to make LC fill the entire image and no more. The Square Root of (Width * Height of the image)

But usually I find that I can get away with less subdivs. 80% seems to be fine. What's your experience with LC and how do you work out the numbers for final images? Is there any times where bumping up LC very high gives any benefits?

[Jackson_cg]

Re: Good LC subdivs number?

The weird thing with LC subdivs is that I've tried as low as 100 and as high as 2000 (before I started reading the manual!) and apart from 2000 yeilding insanely slow LC passes, I've never been able to see any difference in the final render so I stopped changing it over a year ago: I keep it set at 500 for all my renders, most of which are 1400 x 800 px (although according to the calculation it ought to be 1058 for that resolution). Maybe it's time for me to run some more tests. :-\

[thomthom]

Re: Good LC subdivs number?

Doing some reading to find out how LC really works. Noticed an interesting note from the manual: http://www.vray.us/vray_documentatio...ht_cache.shtml

Do not apply perfectly white or very close to white materials to a majority of the objects in the scene, as this will cause excessive render times. This is because the amount of reflected light in the scene will decrease very gradually and the light cache will have to trace longer paths. Also avoid materials that have one of their RGB components set to maximum (255) or above.

And on this page, http://www.vray.us/vray_documentatio...examples.shtml you clearly see the difference with different values, but that's just the pure LC. I don't see that much difference in the full render...

[thomthom]

Re: Good LC subdivs number?

And, do I understand it right that the optimal number of passes relates to the number of CPU cores on your computer?

[dalomar_cg]

Re: Good LC subdivs number?

Alright, I guess I've got to debunk a couple of LC myths today...strap in boys, we're goin' for a ride ;D

So the way LC works it by tracing rays from the camera through the scene. Those rays essentially bounce around, and as they bounce more and more, they gather more information about the lighting of the scene. The actual "placement" of the initial starting point of those rays deals with the "image plane" which is divided by the number of pixels that you've set your final rendering to. THE VISUALIZATION OF THE LC IS NOT AN INDICATION OF ITS QUALITY. Why? its still the same scene on the other side of camera, so the only thing that dictates whether 300 LC subdivisions looks good or not is the size of your image. The reality is that those 300 subdivisions are traced the same "depth" throughout the scene REGARDLESS of image size. 300 subdivisions is going to lead to the same level of quality whether the image size is 200x100 or 20,000x10,000. Not to mention on the other end of the spectrum that you'd have to have upwards of 5,000 subdivisions to fill in all of the pixels in a hi res render.

The other reason why using the visualization of an LC solution is rather useless is that all of your subdivisions will be compiled into samples that have a certain area. What really matters is that each SAMPLE has enough subdivisions to have accurate illumination info. Unless you've changed something, those LC samples will be THE SAME regardless of the size of you're image. Therefore, nothing about resolution changes the accuracy of what is actually used to reconstruct you LC solution.

Therefore, that leaves us with two main ways of determining the actual quality of the LC solution. First is the subdivisions themselves. Generally, I'd say that LC subdivisions follow a "times two" kinda mentality. Basically, if we say that 500 subdivisions is a Medium Low quality, to get to a Medium quality, we'd times that by to and get 1000 subdivisions. If we want to go to a Medium High quality, then times 1000 by 2 and get 2000. Times that by two and get 4000 for High quality. That all being said you'll almost NEVER need high quality...one of the main reasons is next.

The second aspect that determines LC quality is Sample Size and is typically the most overlooked aspect of LC settings. The reason why I hold this setting in such high regard is that this is the "funnel" for LC. All of your 200 or 10,000 subdivisions have to be compiled into these samples, and your level of achievable quality is dictated by those samples. In the end, I could throw a million subdivisions at a bad sample size and still get an image that I'm not happy with. So if we are using a ton of subdivisions, we're not going to get the most from them until we give them the sample size with enough "breathing room" to show their stuff.

By default, the sample size is dictate by the percentage of the screen and is set to 2% (displayed as .02). This seams like pretty small, but in actuality, that generally leaves us with only 250 samples for a scene. This is setting can give a generally good approximation of the illumination (which is really more what LC is for), but a lot of the finer illumination details gets lost. The thing is though, that sample size, with a subdivisions amount around 1000, produces a pretty good balance of subdivisions for each sample. However, thats looking at the samples 1 on 1...for the final rendering, there's some interpolation between samples, so that balance is not really as needed. As we decrease the sample size from 2% to 1%(or even 0.5%), we'll have a better quality potential (because are samples are smaller), but we'll have fewer subdivisions per sample leading to a "noisier" solution. Don't worry about this though because A) there some interpolation of the LC solution when rendering the final result, so it would have to be REALLY noise for us to see a big difference... B) once you kick up the subdivisions for your final, a good deal of that noisiness will go away.

A little homework for you guys if you want to test this out first hand. 1) Set LC to both Primary and Secondaries... 2) Set the Interp Samples to 1...this will remove any interpolation in the final result, and allow you to really see the samples. 3) Test varying LC subdivisions first at the default sample size...notice how at a certain point, it won't really look that much better. 4) Now test some different sample sizes (.01, .005, .0025) and see how the result looks at a single SubD level. The larger sizes should look a little smoother, and the smaller ones will look noiser. 5) lastly, try testing a larger number of subdivision with the default sample size and a much smaller one. Notice some of the details that you'll be able to see.

Thats seams it for me right now...Yes Thomas, passes should be set to the number of available threads.

[thomthom]

Re: Good LC subdivs number?

dalomar, if you ever write a v-ray book I'm signing up for pre-orders. This made much more sense to how LC works than the V-Ray manual or any other LC articles I've read.

I was trying with LC + LC earlier and also tried if the Sample size did much. But due to the interpolation, I didn't see the major effect. The only time I touched the interpolation value was increasing it.

Very interesting, and very useful to know that the sample size is the value that makes the biggest difference.

I did some additional tests with different resolutions, that answered a few questions now that I can see the actual samples.

While I've previously just adjusted these values by "messing about" and still get where I wanted, it's good to know how this actually works so I know what needs to be changed when and what does what.

Reading the V-Ray manual again on LC makes more sense now, but that's solely thanks to you dalomar.

...so, how about that book..? ;D

[dalomar_cg]

Re: Good LC subdivs number?

dalomar, if you ever write a v-ray book I'm signing up for pre-orders. This made much more sense to how LC works than the V-Ray manual or any other LC articles I've read.

My name isn't on the manual you're reading??? ;D


Yea, the interpolation is key to really see what's going on. Without that, the solution will look really blurry and there won't be much distinction there. For actual rendering though, an interpolation number less than 5 is not advised (either 10 or 15 is our default). Most of the time you shouldn't have to worry about interpolation settings. LC gets a "bad wrap" for being a really blurry solution. Its the sample size issue that's really the cause for that because LC can be just as accurate as anything else. In fact. The Progressive Path Tracing mode (ie. Maxwell render mode) will produce extremely accurate images...just think of the sample size being so small that its the size of one pixel.

[thomthom]

Re: Good LC subdivs number?

I want the dalomar bible! ;D

[dalomar_cg]

Re: Good LC subdivs number?

That's a slippery slope...;D

[cat4fish]

Re: Good LC subdivs number?

I want the dalomar bible! ;D

Me too. First time deep touch of LC setting.
That will really be an classic vray LC tutorial, with test picture!

[Jackson_cg]

Re: Good LC subdivs number?

LOL... and to think I almost never change ANY settings in LC!

Thanks for the info Damien.

[thomthom]

Re: Good LC subdivs number?

For those of you that like to see pictures rather than text. Here's some of my test renders while experimenting with this.

Rendered with LC+LC no interpolation.

First is a series of tests with sample size of 2% with subdivs at 100, 200, 400 and 800.






It's also interesting to see the effect in the thumbnails:




Then a series with sample size of 0.5%










Notice that the first series are smoother.

[dalomar_cg]

Re: Good LC subdivs number?

Great Tests...Thanks Thomas for posting these...Its interesting how by the time you got to 800 subDs with the samples @ 2% that the image was largely smoothed, yet those same 800 subDs with the samples @ .5% still had some more room to get smoother. Proves the point that a higher number of subDs really benefit from smaller sample sizes. The thing I think is possibly more interesting though are some of the details present in the .5% tests that are practically nonexistent within the 2% tests. The most striking one is the shadows from the person standing out side.