GI comparisons by using GI caches

Started by Tangled-Universe, July 30, 2012, 12:52:09 PM

Previous topic - Next topic

Tangled-Universe

Hey everybody,

Since the introduction of GI caches you can do some very useful things. Not only can you have flicker free animations more easily, or make skyboxes, you can also make the contribution of GI to your scene more easily visible.

First I calculated the GI cache 4 times at GI 1/2/4, GI 2/2/4, GI 3/2/4 and GI 4/2/4. I kept the blur radius a bit lower to make the differences a bit more clear, hopefully.
Then after that I switched off the sunlight and rendered the image 4 times using the GI caches respectively.
The result is a render only lit by the calculated GI.

In the image below you can see an older scene of mine rendered 4 times.
Top left = GI 1/2/4
Top right = GI 2/2/4
Bottom right = GI 3/2/4
Bottom left = GI 4/2/4

As you can see in general there's quite a lot being lit by the GI. The surfaces are relatively near each other so there's quite some bouncing possible here.
If you look at GI 1/2/4 you'll see that the overall lighting is quite flat. There isn't much detail if you look at the crevices and especially in the plants and fine shadows of the sand.
If you look at GI 2/2/4 or at higher relative detail settings these factors improve.
However, going beyond GI 2/2/4 isn't of much use in this particular case.

Later on I will dig up some other older scenes and see how this works out.

I think in overall, if you have the time, this may be a nice way to see if increasing GI in your scene is really helpful or not.

Dune

Interesting approach, Martin. It is indeed obvious that 2/2/4 is the minimum to be used, but if you look at the lower right render you can see quite a bit more  shadow between the rock and sand than in the 2/2/4, which seems a lot nicer, even than the lower left render. How about the differences in render time?

Tangled-Universe

Yep GI 3/2/4 is actually a bit better indeed...I don't know rendertimes by heart, only know that cache generation was something like this:

GI 1/2/4 = ~2 minutes
GI 2/2/4 = ~8 minutes
GI 3/2/4 = ~14 minutes
GI 4/2/4 = ~22 minutes

Something along that line and I believe that after that the rendertimes weren't that much different.
I need to check that again.

Later I will see if I can find nice examples of vegetation and clouds. Those 2 elements generelly benefit the most from good GI settings, so let's have a look what's sane and what's not in those situations.

TheBadger

Highly useful info here! Thank you, Martin.
It has been eaten.

dandelO

I think I prefer the bottom-right, 3/2/4, if you look into the distance at the very far plateau, the distant lighting in 4/2/4 seems flatter, over-riding more of the slightly shadowed areas of the cliff and seemingly covering the entire surface in GI, regardless of primary shading/displacements on the rock face. Maybe 4/2/4 is more 'correct' but I agree that 2/2/4 is pretty damn fine. Is it worth the nearly 50% extra render time to go for 3/2/4? I'd say, in this case, no.

I absolutely love the GI options now, we can render out final scenes at respectable, sensible, render quality but with as high(or low) GI as you like, mega-sampling the prepass at render detail/1 and then rendering at 0.5 - 0.75 detail for the final pass using that GI cache. Or, go the opposite way to save on time, using a lower GI cache for speed. And the ability to use cache blending for animation is fantastic.

Also, on upping GI detail further, maxing the 'ray detail multiplier' removes the triangulation of the initial GI pass, giving a more accurate spread of light to eventually use in rendering the last pass, before tuning back the renderer(and RDM) for the final pass, giving much better results, even at GI relative detail/1.

Cool tests, Martin. :)

Tangled-Universe

Thanks Martin,

The RDM is something I may go into at a later stage. It's an advanced setting, not really well known yet, but GI surface details is for example.
So likely I will do GISD first.

I just started with this so first of all I also need good scene examples where GI is really required.
Then I probably can automate these tests using render elements.

rcallicotte

Thanks TU.  As always, depths of help.
So this is Disney World.  Can we live here?

FrankB

I am not sure if I am seeing differences between image 2,3 and 4 at all.
If anything, I can say that the shadows are deeper in 2,3,4 compared to 1.
In 1 I see the least shadow, so more detail.

Thanks for sharing, though!
Frank

Oshyan

It depends on what you think of as "detail" and whether realism is the goal. The darker shadows are probably more realistic, they're a (theoretically) better calculation of the limitation of light transmission to those narrower spaces.

- Oshyan

Tangled-Universe

Then render without shadows Frank ;)
Indeed Oshyan. Detail also accounts for shadows and they are indeed more detailed in the GI relative detail 2 and 3 renders.

I hope I can find some time coming week to do some more (better) tests and show them.

FrankB

No, you misunderstand me. The reason why there isn't so much difference between the last 3 images is because the scene needs increased sample QUALITY to bring out more detail in the shadows.
I don't have a way to demonstrate this, but try GI 3/5 and compare with the other renders.

Regards
Frank

Tangled-Universe

Can you explain Frank?
In your first post you explain there's the least shadow in GI 1 and "thus has more detail". I don't understand that conclusion, because I see GI different I think.

The detail in GI is still mainly determined by GI relative detail, rather than sample quality. Why? The relative detail determines the amount of points where the GI is calculated and sample quality is how many times those points are assessed.
So ultimately the GI relative detail is the key to denser, thus more detailed, lighting in/out of shadows.
The sample quality is rather a measure for how accurate the values for each GI measurement is.
This would result in "true" values for every GI point and can give the "true" depth of a shadow. Yet, still you don't have many GI points so despite the accuracy there won't be much detail.

So I think sample quality is not as determining as relative detail.

FrankB

You are generally right, Martin, but as far as I know, the higher the sample quality, the more of the weaker light sources are taken into account. For example, years ago I had tested a scene with no sunlight, but a sunlit moon (no fake sun, just a sunlit moon), and I wanted that light reflected by the moon to light my scene foreground.
Increasing the sample quality helped make that faint light "moonlight" visible on the terrain.
Anyway, might be that you're right, but there still isn't a lot of difference between the last three images ;-)
Perhaps the scene isn't optimal to showcase the differences?

cheers,
Frank

Tangled-Universe

Yes you're right it's perhaps not an optimal example, although there are definitely visible differences.
I have to think about your sample quality examples. I think I know what you mean, but I can't put it into my own words yet.
There seems to be more about it then than I initially assumed.