SimonButtes Closeup

[DannyG]

Close up is impressive to say the least, rock faces are topshelf

[Henry Blewer]

This will have to be put aside for a month or so. I don't have enough memory to render it.

Also, the new machine is delayed some more. I can't afford the RAM and the HD at the same time. I'll buy the HD now. Then I can order any cables I need with the RAM later.

[Oshyan]

Unfortunately hard drives have gone up massively in price recently. You might want to wait on the HD, they're not likely to get any more expensive, and they just might get cheaper as the impact of the flooding becomes more clear.

- Oshyan

[Henry Blewer]

New render of this using the new computer. Detail 0.5375, AA 24 with 1/16 samples for the First sampling level and Pixel noise threshold of 0.15. Supersample prepass and GI details checked. GI 2/4/6, with a 4 hr 22 min render time.

http://www.flickr.com/photos/njeneb/6349710165/sizes/o/in/photostream/

http://njeneb.deviantart.com/#/d4gcmux

[dandelO]

Looking very nice, Henry. Congrat's on the new machine!

I bet there wouldn't be much noticeable visual difference(if any at all actually) and with a much faster render time by using AA=6 at max samples/non-adaptive.
36 is a very generous amount of samples per pixel and 36 all round would probably give you an equally nice result in about 1/4 of the render time than the massive amount of max' samples(576) taken at AA=24 with first sampling at 1/16, which gives you a minimum of 36 samples. I think that really is overkill and wasting unnecessary time, time that could be used making your next scene!

I'm not taking into account the GI surface details as that'll be what's adding most of the render time here in this scene, but try out both methods of AA sampling on a different test scene that has 'GISD' unchecked, save both results and flick between them in your picture viewer, I'd bet the only real difference will be much less render time.

I might be completely wrong but I don't think the visual difference will even be noticeable in each image back-to-back.

EDIT:

I ran a test on one of Walli's trees at; AA=6 at max samples/AA=12 at 1/4 samples. Both give a minimum sample count of '36'.

Ignoring the ridiculous total render times for both because of my slow computer, the one at AA=12 took more than double the time and there's pretty much no gain in quality(and AA=24 at 1/4 sampling would have be lots longer still for a very minimal gain, I quit the render as it was just too much for a quick test on my computer but the area that did render wasn't much better).

It is very slightly softer but I think the extra render time far outweighs any visual gain that comes from using such extreme sampling levels. Maybe it's just me but I don't think it's worth it at all to be using values like 576 samples per pixel when even 144 samples took double the time and turns out pretty identical. I understand there will be special cases where it would work very well but in general I don't think I'd waste the time. Thoughts, anyone?

[attachimg=#]

[Tangled-Universe]

I bet there wouldn't be much noticeable visual difference(if any at all actually) and with a much faster render time by using AA=6 at max samples/non-adaptive.
36 is a very generous amount of samples per pixel and 36 all round would probably give you an equally nice result in about 1/4 of the render time than the massive amount of max' samples(576) taken at AA=24 with first sampling at 1/16, which gives you a minimum of 36 samples. I think that really is overkill and wasting unnecessary time, time that could be used making your next scene!

There's a little but significant error in your calculation of AA samples when rendering with AA24 with 1/16th first samples.
This setting does not mean the minimum samples is 36 per pixel, but that 36 is the average number of samples.
What the pixel noise threshold does is setting a threshold for noise upon which the renderer decides to apply more or less samples to the pixel.
It depends on the initial level of noise too and I think Matt assumed that the noise distribution is equal, like gaussian, when he decided to show the setting of the average number of samples being used.

So effectively you'll never know if the average number of samples, 36 in this case, is applied everywhere.
Also, you'll never know the minimum samples used.
All you know is that when you use higher pixel noise threshold levels that the number of samples being used tends to be skewed towards less and that it is unlikely you'll use a lot more than the average samples and VERY unlikely you'll use the max number of samples.
When you render with "AA6 full" every pixel will have 36 AA samples applied, regardless whether necessary or not.
However, if you render with AA16 or higher with 1/16th first, then in some noise-regions more AA might be applied compared to AA6 full because the threshold hasn't been reached yet.
This is the reason why these high AA levels can give very good results without having crazy rendertimes.
The pixel noise threshold is an important factor for this.

In all cases I think the rendertime is (somewhat) longer and it's up to the user to decide whether it's worth it or not.
I bet the difference IS there at all and am pretty sure that it is noticeable, though pretty subtle.
For the B-spline filter the difference is the easiest to notice.

I'm not taking into account the GI surface details as that'll be what's adding most of the render time here in this scene, but try out both methods of AA sampling on a different test scene that has 'GISD' unchecked, save both results and flick between them in your picture viewer, I'd bet the only real difference will be much less render time.
I might be completely wrong but I don't think the visual difference will even be noticeable in each image back-to-back.

GISD is mostly unnecessary to use.
I recall Matt only recommended it when there's a lot of small scale dark shadows in the image, like dense vegetation with not much direct light on it.
Not many scenes have this kind of lighting so don't require it and if they do it helps a little, but surely isn't a magical setting worth the at least 2x longer render. I haven't found many cases for GISD in 6 years, so I think the most noticeable effect indeed is on rendertime

[dandelO]

Hi, Martin. I edited my above post with a comparison, I'll render out one at AA=24 at 1/16 to see just what the difference is there too.

Yes, of course you are correct about the averaging of samples used, I just referred to how it is labelled in the pixel sampler node. At AA=6/max, then the mean number of samples shouldn't drop below 36.
There will definitely be some quality gain with higher sampling, my point is that I just don't think that the extra render time justifies the task of doing so, in most cases.

[Tangled-Universe]

You made the same mistake again in your comparison
It's not minimum, but average, the render pixel sampler doesn't mention minimum as far as I know?
edit: I see your image now, thanks it's indeed the minimum samples, with an average (the average number of minimum samples. So it's not 36, but could also be less, or more. Depends on the pixel noise threshold, for instance.
Is the pixel noise threshold the same in the comparison?
It's not the best comparison anyway, since it's perhaps a bit too simple.
In more complex and dense scenes the subtle differences are definitely there.

I do agree with you that it does not always gives an improvement and that it adds rendertime.
That's something I have been saying from the beginning that it "might" look better, but not guaranteed.
Seems general perception is shifting towards "TU says we should always do this". Definitely not.

I started suggesting this to people to 1) find out how it might look as it can give better results, in my opinion 2) they can handle it with their machines, so rendertime is less relevant (especially for crazy guys like me who render overnight and thus care less ) and 3) most importantly to provoke/induce thinking about TG2's renderer and it's settings.
The latter I definitely succeeded in

[dandelO]

But it does.
I think that it is the average but just cannot drop below the count that says 'min samples per pixel' and when non-adaptive is used then a set 36 samples will always be taken.

[attachimg=#]

Regardless, I still think in most cases I couldn't justify using such high sampling when the extra time far outweighs the quality gain, that's all I'm saying.

Cheers, Martin.

[Tangled-Universe]

But it does.
I think that it is the average but just cannot drop below the count that says 'min samples per pixel' and when non-adaptive is used then a set 36 samples will always be taken.

Yes it can drop below that count, otherwise it wouldn't say "mean" behind the value.
So a mean number of minimum samples. With 36 samples being the mean and 50% of the samples being <36 or >36 for all the pixels.
When you average the mimimum of samples used for every pixel the value will be 36.
A bit strange perhaps to understand why it is depicted like that in the node, but it's a bit comparible to statistics which I use for my work.

[dandelO]

I know samples are averaged when adaptive sampling is used and I believe that is where the term 'mean' is used.

But when non-adaptive sampling is used, then you are guaranteed to use the set number of samples per pixel described in the sampler, is that not right? That's what non-adaptive sampling says to me anyway.

Cheers, man!

[dandelO]

Simply for completeness sake, and for my own curiosity, I rendered the same 'jaggy raytraced object on a plain background' scene at AA=24 at 1/16.

[attachimg=#]

Pretty much no gain in quality that is worth me spending the massive amount of extra render time on.

The quality is indeed marginally better with the higher AA but it still doesn't justify using it in *most cases to me.
This is, of course, only one instance and only RTO was used here, I'm not saying that it won't be beneficial in some cases but in a scene like Henry has posted here I think the extra time is really just wasted, comparable results are achievable in less than 1/4 of the time.

Sorry to deviate so much from discussing your image here, Henry, but I think you could really get the render time down on such scenes(without the GI surface details, of course, that really isn't necessary here at all in my opinion) without using such extreme AA settings.

[FrankB]

I too think there is nothing to gain from this strategy, that would be visually noticable. I think your comparison demonstrates that pretty obviously. Maybe there are situations where this strategy would produce a visually noticable difference, but nobody seems to know what these situations could be. It's definitely not required for a "Walli Pine before standard background" render, that's for sure

[Tangled-Universe]

Again guys, I'm only saying it's interesting to explore it as it sometimes can give a subtle and nice difference (especially in heavy object scenes and not examples like the above as Frank pointed out + the response to his attempt to try this out seem to be positive, or not?) and with fast machines it's not too much to handle.
Hence I only recommended it to Zaai and Henry so far as they have enough brute render power to try this without bleeding too much for lost time.

Of course this should never be the common way to do it, I never said that. Still it seems some think I say to everybody that they should do it this way(?).
So everybody should continue with the common way to do it, by keeping everything reasonable within the standards which most of us experienced users recommend to everyone.
If I opt for something different to try, it does not mean I want to change "the standard"
But I'll be careful next time with suggesting something out of the ordinary

Martin is perfectly right to say it does not justify the rendertime for him. Mostly it indeed doesn't, agreed to that before.
For the future it can be very useful, as Jon figured out this approach is one of the very few, possibly the only one!, to render out an animation with lots of trees with small (specular) leaves without flickering and/or popping.

I think I've explained myself enough now why I suggested to try this

Cheers,
Martin

[dandelO]

I didn't think you were suggesting that this be the common way to do things from now on, not at all, Martin.
Your ideas are always well educated and greatly appreciated and I knew there must be some instances that this option would be beneficial, I just didn't think it would be very beneficial for Henry in this scene and that's why I said so.

I decided to do a test just to see if there was any general quality gain from adding huge amounts of possible samples. I found pretty much none for a general setting of jaggy objects on plain background.
It certainly doesn't justify me to do it this way, not on this piece of crap pc, anyway! I just thought Henry could get a much faster render by using a more standard method in this instance.
Thanks also for the information above about when it would be very useful to use the AA this way, for popping speculars etc, I would probably never have worked that one out and I shudder to think what I'd do if I was facing such a situation and had to rely on doing this! I'd likely just throw in the towel and say, nope, this isn't the scene for me!

Cheers, dude!