Atmosphere sampling and sky color

[Eneen]

Hi,

I've checked something and want to share it with you as I don't know answer why it is so.
Please take a look at those images.









Sampling is set to 32,256,1024 and 4096 and 4096 sampling looks best. Is there any way to get 4096 look (colors are nicer) on lower sampling? Is there any parameter that controls it?

Best regs.

[Tangled-Universe]

I'm pretty sure this hasn't much to do with sampling.
I think the differences are inter-frame differences of GI calculations.
Everytime you render the same image the GI solution is slightly different. Simply explained because of its nature.
Try repeating the test and disable GI and you'll see there's hardly difference.

[Eneen]

But without GI I will not get skylight on clouds.
I've checked and it's not interframe calculations, lowering sampling = less yellow on horizon and brighter blue on sky, please check yourself.

[Tangled-Universe]

Without GI you don't get GI (skylight, as you name it) on your clouds, logical. However, that wasn't your initial question. You asked whether it had to do with samples and for that you need to rule out GI contributions, so disable it.
Then you'll find out if it has to do with GI or not and then deal accordingly. It was to test, not as solution

Regarding the yellowness I can imagine it has to do with sampling. If you have too few samples in space then the renderer won't pick up the light decay.

However, for sky 'blueness' I think samples doesn't matter that much (since you probably have plenty samples) and that it really is a GI thing you're seeing.

[Eneen]

Ok, increasing samples with GI off makes less noise and no change in look as you have said. But why lower sample count makes difference when rendering with GI when increasing GI setting (relative detail and sample quality) makes no difference - I mean it doesn't make sky more yellowish like with setting more samples in atmosphere settings?

[Tangled-Universe]

For consistent results with GI you need very high GI sampling and also quite high GI relative detail.
As I explained the calculation for GI is slightly different everytime you render your frame. That's the nature of how it is calculated.

So, you need to question yourself: what do I want/need?

Is this a problem or only an observation? I think the latter.

If you want more yellow at the bottom the increase bluesky density (and its decay consequently). The sky turns more brighter blue, but if you increase fake darkness then it will become darker again, but preserving the decay.

As you can see there's a fix for these kind of things and no way you really have to use >32 samples for situations like these.

[Eneen]

My goal is to get most realistic result without having to set everything and tweak.  Long time ago I switched from biased (mental) to unbiased (maxwell) and my life is a lot easier now. It's no problem to set something high (and get high render times). Problem is what settings should be set high.
Moreover making GI setting high (sampling quality to 100 and supersample prepass on) and even relative detail to 10 make sky look really same, it has yellow cast on horizon on GI prepass preview but it disappear on final render. That's why I'm asking.

[Oshyan]

If you want an unbiased render, you'll need to use an unbiased renderer, which TG is not. TG's renderer was not designed to be used with the kind of "quality" settings you're talking about here, atmosphere samples at 1000 is frankly pretty crazy, as is GI sample quality 100, relative detail 10. T-U's suggested approach for getting the *result* you want is very, very sensible in light of the astronomical render time you will end up with using the settings you're testing with. If those kinds of tweaks aren't something you want to deal with, perhaps you can find a way to create an atmosphere model in Maxwell?

- Oshyan

[Eneen]

From what I've tested it seems, that to get most consistent result I don't need crazy GI settings, 1 and 1 on GI rd and GI sq is enough, and making them 3 and 3 is over any need (I mean only atmosphere and cloud quality) of course with detail set to 1. As T-U have said sampling may influence sky color and my test showed that it does. But question is why increasing sampling makes difference in color only when GI is ON (I've checked it as T-U suggested, my bad on using non-pro "skylight" word). Shouldn't it be like that when GI is on I make GI settings high and I get consistent result. Moreover I wonder about this prepass color difference:





Upper image is 32s atmosphere calculated on GI settings 10/1. Lower is just a prepass. Image from prepass is very similar to what I get when I render image without GI. Lack of knowledge in area of what parameter is responsible for what and their relationship is pain here.
In Maxwell there is no such thing like haze level and so on, so you can get natural result from it but it is limited, not talking about clouds. It will give nice colored sky eg via "water" setting, but still it is not *that* if you know what I mean.

[Tangled-Universe]

Ghehe, you're special
So for the 3rd and last time I'll try to explain to you that the differences of the result has little to do with the sampling. Forget about that, period.
It's the usage of GI.

The GI is being sampled at points (the number of points is determined by GI relative detail) and the accuracy of the value of that sample is determined by GI sample quality.
The sample positions in screen space are random, so everytime you render the same image the sampling pattern isn't exactly the same
If they aren't random then rounding errors in calculations result in tad tad differences between previous sampled values.
So 2 factors affecting the differences why GI almost never gives the same solution between renders. The only way to avoid this is using higher GI settings.

Of course, at some point the number of atmosphere samples is also of importance (in your example you're oversampling the GI compared to the sky, resulting in a mismatch between prepass and final output), but like I and Oshyan explained you need to focus on the visual result you want and not on the settings. Now you're focussing on the settings and yet complain that you have to worry about them while we try to explain you otherwise

[Matt]

T-U,

I think Eneen has successfully shown that sampling is affecting the GI. He said that 3/3 GI was sufficient to produce consistent results, and although that is usually not true in general, I believe it might be enough to demonstrate a consistent difference between high atmosphere samples and low atmosphere samples. Turning off GI is not possible because the differences he's demonstrating are in the GI.

Eneen,

I have not seen this demonstrated this well before, but your results seem consistent with some expectations I have about the way GI is undersampled when atmosphere samples are low. When there are fewer atmosphere samples, the GI is sampled more sparsely along the volume and each sample is then forced to spread over a larger area in the final image. Unfortunately this volume is larger in image space XY so you can lose some fidelity at the horizon. I have never tested at such high samples as you to see how it looks like that.

You'll never get a completely unbiased result with the current GI system because it uses an undersampled cache, but higher values will often give you a more correct result. I would suggest taking what you learn from the high-sample images and emulating that look with fewer samples. For example, you can reduce the GI lighting on just the atmosphere node without affecting the clouds. The atmosphere node has a parameter called Enviro light which affects the Enviro Light (GI) contribution to just that node. Your results show that a more realistic horizon might want less enviro light.

Matt

[Eneen]

Matt, thank you a lot for the answer.
Indeed making env light to about 0.7 makes image closer to that with very high sample count (16000). That's solution for faster rendering without tweaking colors.
I've checked also clouds and very low sample count makes clouds more transparent, and cirruses may disappear completely sometimes.
Question here: is there a rule of thumb to set number of samples for atmosphere and clouds, e.g. n. of samples per kilometer or something to be close enough? Rendering time does not matter in this experiment, more important is to be able to set similar effective quality for every element, that it would not be matter of guessing.
Last question, I don't know if correct one: is there a way to set number of samples for GI cache, I mean, is quality samples (e.g. for atmo) same number as gi cache sample number? I'm trying to go with such route: "can I get noisy picture but still keep quality of image, quality of colors, reality, on very good level? As from prepass picture above - high GI seting, here 10 for relative detail, cause prepass to be visible with similar to high sample image but it's noisy.

[Matt]

Matt, thank you a lot for the answer.
Indeed making env light to about 0.7 makes image closer to that with very high sample count (16000). That's solution for faster rendering without tweaking colors.
I've checked also clouds and very low sample count makes clouds more transparent, and cirruses may disappear completely sometimes.

Yes, there can be some problems when the samples are very low, but I'd expect you'll usually see noise the image before anything like that happens.

Question here: is there a rule of thumb to set number of samples for atmosphere and clouds, e.g. n. of samples per kilometer or something to be close enough? Rendering time does not matter in this experiment, more important is to be able to set similar effective quality for every element, that it would not be matter of guessing.

None that I can think of. The main purpose of the samples setting is to balance noise with speed. The way I look at it, the effect on the GI accuracy or detail is a secondary consequence which I allow to happen because typically when you set the samples low you are not interested in a high quality picture. Therefore it's OK if the GI loses accuracy in order to save render time. When you set the samples high enough so that no noise is visible, I'd say the GI accuracy is usually "reasonable" unless you have some higher quality image to compare with. There are some internal settings that I chose to reach what I thought was a reasonable quality/speed balance in various tests I did. In many situations the GI accuracy can get better the more samples you use, but I don't know how you could decide how many samples is "good enough" when a completely unbiased result is virtually unobtainable. It's also complicated by the fact that there is a limit to how accurate the GI can get which depends on the GI settings (particular GI relative detail and GI blur radius) will put limits on how detailed the atmospheric GI will be.

Last question, I don't know if correct one: is there a way to set number of samples for GI cache, I mean, is quality samples (e.g. for atmo) same number as gi cache sample number? I'm trying to go with such route: "can I get noisy picture but still keep quality of image, quality of colors, reality, on very good level? As from prepass picture above - high GI seting, here 10 for relative detail, cause prepass to be visible with similar to high sample image but it's noisy.

If you find settings that produce a final image you're satisfied with, but you don't want to use so many samples in the final pass, you could precompute the cache with those settings and write the cache to disk (as of v2.4, but not in the free edition). Then you can lower the atmosphere samples as far as you like for the final image. Currently that's the only way to separate the quality between the prepass and final pass.

Matt

[Eneen]

Thank you alot Matt!