That time of year again I guess... This image, like most of my images is 95% models and 5% lucky positioning. Included are:
Oak from Klas
Norway Maple, Sycamore Maple, Honey Locust and Smooth Elm from Xfrog
Dry grass, fresh grass, shrubs and dirt objects from Walli
and one population of the internal rock object.
A smaller compressed version is attached, but you can view the full-size uncompressed image here: http://www.archer-designs.com/zp/albums/digital-art/terragen-2/autumn-tree-line-full.jpg (http://www.archer-designs.com/zp/albums/digital-art/terragen-2/autumn-tree-line-full.jpg)
Very nice Ryan. Loads of contrast there. It looks like it wouldn't flicker too much if you'd animated the camera, out of interest, what were you AA settings and render time?
Cheers
Jon
Thanks Jon,
Rendered at 1920 x 1080 - Detail 1, AA 6 (default sampling), GI 1/6 - 11 hours on my system, so likely around 5 hours on one of the new fancy computers I would guess.
Thanks fella. Out of interest, why did you go detail 1 when models are influences by AA?
Unless I am quite mistaken (possible), detail level still has quite a bit of influence on the entire scene even when using ray-traced models. I will have to do some tests soon.
As I understand the Ray Trace side of things, AA defines the quality of objects, detail defines displacements. So AA includes atmosphere as well as objects, high AA means you'll need less quality in atmosphere/cloud settings. With the new ability to define the sampling, we have a lot of power on how things render/AA/speed.
The detail setting does influence GI, shadows and other things not related to the objects themselves. Quick test here using all the same setting except detail. One image the detail is at 1 the other the detail is at 0.1.
From my posted image using a detail of 1 was probably totally not useful, 0.5 or so would have likely given good enough GI and shadow detail.
Thanks Ryan. Interesting comparison. I tend to render at 0.5 with model heavy scenes. I don't see much difference with 0.6 or 0.7. The GI side of things is probably something else. Something I need to do some tests with.
As an aside, I've been experimenting with insane AA settings and playing with the Edit Sampling. My work tends to have moving cameras, so flicker control between frames is paramount. In some scenes, I've had to knock up AA to 32 to deal with render inaccuracies, which was a lead that a post from Oshyan had lead me to. More 'mean samples' gives a more accurate initial pixel. But you need to dial up the pixel noise threshold to compensate for render times. 0.2 has been a common value.
Try the same scene with....
Detail 0.5
AA 32 (Edit Sampling / First Sampling Level 1/16, Pixel Noise Threshold 0.2)
GI 2/4
At 720p it should take about 2hrs to render tops
Sorry Ryan, it's not going to make a difference to how your scene looks. It's just interesting to see if we can knock down your render times.
No problem Jon, I'm always interested in finding ways to speed things up. Doing a test at 1280 x720 with your suggested settings.
The compressed image really does not do this justice. Awesome scene. As for luck, I think your skill has much more involvement.
Hi Jon,
Tried your settings, the render took 12h 09m for a smaller image (1280 x 720) than my original render (1920 x 1080). Quality as expected was very similar.
Wow. I didn't expect that. Hrm. I was getting 2 hrs on some of my plant heavy scenes.
I was told that model quality in the node's drop down, didn't effect quality anymore, so I've kept them low. I wonder if this is the reason. I'll have to do some tests when I get back from holiday.
Sorry to have led you up a garden path. Thanks for giving it a go though.
Cheers
Jon
Quote from: Hetzen on October 10, 2011, 03:41:23 PM
Wow. I didn't expect that. Hrm. I was getting 2 hrs on some of my plant heavy scenes.
I was told that model quality in the node's drop down, didn't effect quality anymore, so I've kept them low. I wonder if this is the reason. I'll have to do some tests when I get back from holiday.
Sorry to have led you up a garden path. Thanks for giving it a go though.
Cheers
Jon
Hi Jon,
Model quality settings only apply when you're NOT using raytrace objects.
Enjoy your vacation!
Cheers,
Martin
Can anyone explain the term pixel noise threshold to me, preferably in non technical, or at least not too technical terms? This is such a geek term, I can't make any sense of it ;)
Hi Ryan, sorry for hi-jacking! Regarding your render: I think this render suffers from the low object geometry detail. If that weren't the case it would be a much nicer image overall. Also, I don't know, the xfrog models look so small to me, maybe that's because there should be more and smaller leaves? Could be it's just me...
Cheers,
Frank
I like a lot the rocks surface.. is it a fully procedural surface or image projection ?
luc
Quote from: FrankB on October 10, 2011, 05:55:05 PM
Can anyone explain the term pixel noise threshold to me, preferably in non technical, or at least not too technical terms? This is such a geek term, I can't make any sense of it ;)
Hi Ryan, sorry for hi-jacking! Regarding your render: I think this render suffers from the low object geometry detail. If that weren't the case it would be a much nicer image overall. Also, I don't know, the xfrog models look so small to me, maybe that's because there should be more and smaller leaves? Could be it's just me...
Cheers,
Frank
Hi Frank, I'll give it a try:
When you use non-adaptive sampling, so max samples, then the same amount of AA/samples is applied to every part of the render.
Logically, when using non-adaptive sampling, the pixel noise threshold is disabled, since everything will be AA'ed the same way.
When you use adaptive sampling, which is default set to 1/4 first samples, the algorithm kind of detects where there's noise in between pixels and decides where to apply more samples for AA and where not to:
If the noise is below a certain level the algorithm decides to take less samples and the anti-aliasing stops quicker/sooner.
If the noise is above a certain level the algorithm decides to take more samples and the anti-aliasing takes longer, since more samples are taken.
The pixel noise threshold regulates how quickly the algorithm stops with anti-aliasing.
If you use high AA levels, say 16 or 32, and use 1/16/th first samples you still have 64
mean samples.
So this could be more or less than 64 samples, depending on the pixel noise threshold.
If you set the pixel noise threshold to something around the default, like 0.025 the algorithm will apply a lot more than 64 samples even where it's likely not necessary and perhaps close to the max amount of samples where there's a lot of noise (max samples is AA^2). This is slow.
If you set the pixel noise threshold to something like 0.2, the algorithm will only apply more than 64 samples where it's really needed, thus rendering goes faster.
I'll look for the link where Matt explains this in a lot more detail, but in a nutshell this is how it works a bit.
Jon started recently experimenting with this as you may have read and it can give very interesting results in terms of quality and performance, especially animation related. In animation you need a good averaged result per frame to avoid flickering of vegetation and using adaptive high AA levels gives good results so far. You can animate lots of tiny distant leaves without noticeable flicker. Very cool.
Cheers,
Martin
Thanks for this comprehensive explanation, Martin. Very clear. But do I understand well that by noise you mean the contrast between, e.g. leaf edge and surroundings? So it's especially AA'ing the hard edges (leaves against sky), whereas areas where leaves integrate more into (~leaf colored) background, it'll take less samples?
Quote from: Dune on October 11, 2011, 03:38:07 AM
Thanks for this comprehensive explanation, Martin. Very clear. But do I understand well that by noise you mean the contrast between, e.g. leaf edge and surroundings? So it's especially AA'ing the hard edges (leaves against sky), whereas areas where leaves integrate more into (~leaf colored) background, it'll take less samples?
I don't know exactly on which basis the algorithm decides to continue or stop. You can do this solely by contrast of course if you look at it in the traditional way, but I can imagine you can also do this on a temporal-space basis, or by hue, or saturation... or a combination of some factors.
I'm not sure if you can look at adaptive AA as the classical example of a jagged black line which is smoothened by AA. I think there's more to it.
On second thought, Jo recently explained the RTO feature especially performs well on parts where leafs are against sky. Suggesting you might be right.
Thanks for the explanation Martin.
Luc, the rocks are a population of the internal TG2 rock object, distorted a bit with a powerfractal and then given texture via an imagemap.
In my head I have it that the higher the pixel noise threshold = the less the AA will be subdivided at render time, making low mean sampled AA very rough/jaggy.
If you have a high 'mean' value, raising the pixel noise threshold is wise because a lower threshold would be subdividing AA samples more than is needed, since you already have a very respectable 'mean', it isn't required that you keep subdividing and so would only waste time.
Try this by disabling the pixel noise threshold(=0) against a very high threshold value(=1). The higher the threshold the less subdividing is done and images can appear jaggy but will render very fast. Finding the balance in between time/quality is the fun part!
* I think I read in the 'guide to RT settings' thread that AA is only based on pixel luminance at the moment, so all hue/saturation values are not interpreted by the AA, the brightness/contrast of neighbouring pixels is all that defines the AA sampling.
Here it is...
QuoteMatt - Currently the adaptive sampler only keys off luminance, rather than the individual RGB components. Samples that have different hues or saturation values but give the same "luminance" will be seen as being equal, so high amounts of noise in colour and saturation may be tolerated by the sampler. This is often OK, because human vision is also less sensitive to high frequency changes in these things. It also means that the most noise is allowed in the blue channel, and the least in the green channel, due to the weightings of these values in the luminance calculation. There are many situations where this approximation isn't good enough, however, so we will add options to change this in future versions.
Hi
OK thanks for the info about rocks
Luc
put leaves on ground and on shrubs under trees, then good
There is sense behind madness.
Just another thought. Rta includes aa. So. Dialling back atmosphere quality and cloud quality settings also contributes to render times....
Super scene - feel like Autumn, looks like Autumn....maybe the off falling leaf might add something? Great Render :)
love the rocks, great mood here and light too!
Another attempt here: Pulled back the camera a little bit and changed the camera angle from 50mm (original render) to 24mm.
There's something weird about the tree line; it looks (at first glance) as if the autumn trees have strange shadows, until you see they're other trees. In that respect I preferred the first one (where I very much liked the darkness under the trees).