PDA

View Full Version : Accurate Caustics + Importance Sampling



DrStrik9
12-10-2015, 04:42 PM
I set up a scene to render caustics without selecting Enable Caustics in the GI panel. I did this by using the same GI settings as the 11.5 Radiosity Benchmark scene, plus placed an object in the scene with extremely high luminosity value (15,000%) which in GI acts like an EXTREMELY noisy light. (I parented a 0% intensity light to this object, to be able to view the scene from this bright object's POV.) This method does render caustics — I assume far more accurately than the Enable Caustics button in the GI panel renders. But the noise factor is, well, VERY NOISY.

So I turned to Importance Sampling to see how well it reduced this extreme noise. I rendered this image with IS set to 1024, 1536, 2048, 4096, 8192, and 16384 (because anything worth doing is worth overdoing.) All other GI benchmark settings remained the same. Render times were very similar (13:42-15:48), but there was NO DIFFERENCE between renders! Conclusion: 1024 is apparently Importance Sampling’s best and “only” setting.

131421

But so far, there’s still too much noise.

So I began to raise Rays Per Evaluation (from the benchmark 4) which of course, increases render time.

131422

131423

131424

RPE 16 = 41:13. Noise is very low, but a 1080p render would take around 2:45:00. Still, it's nice to know IS is a one-setting tool (on or off), and after that, RPE is your long-rendering friend.

Ztreem
12-10-2015, 06:23 PM
You have to find a ballance between RPE and adaptive sampling to clean the noise.

DrStrik9
12-10-2015, 06:35 PM
In the 11.5 benchmark GI scene, adaptive sampling is used. What do you mean by "balance?"

Actually, I know what "balance" means. :+) So as RPE goes DOWN, so does AS threshold? As in if I choose RPE of 8 (as opposed to 12), then AS threshold might go from 0.01 to 0.001 ... ?

EDIT: ... I've tried lowering RPE below 12 in this scene, and no lowering of AS threshold delivers the same acceptable noise amount, so I think I'm already there.

spherical
12-10-2015, 08:37 PM
I'd try a high RPE. Go nuts just for bumps. I've seen recommendations of 300 or more. In my experience it didn't impact render times all that much after you go above a certain threshold. The image just gets cleaner because the solution is more accurate.

madno
12-10-2015, 08:58 PM
Regarding Importance Sampling:
To my understanding it does nothing if you don't use textured environment (don't know about Image World). If you use a light or lominous geometry or a inwards pointing world sphere with e.g. HDR map IS does not help.

Please somebody correct me if I am wrong here.

spherical
12-10-2015, 09:10 PM
Conclusion: 1024 is apparently Importance Sampling’s best and “only” setting.

According to the Docs, that is not the case; although there is zero in the way of explanation as to what changes do (which, sadly, is usually the case):

Normally this can be safely left at the default of 1024, but improvements to the quality and
speed of preprocessing can be gained by increasing or decreasing it with care.

It also does not define what is and is not a "Backdrop" that ISBG works with. I did a couple of early tests in my Dielectric Cube scene and it uses a Sky Dome. Turning ISBG on and off did influence the render quality and times.

DrStrik9
12-10-2015, 09:27 PM
Importance Sampling turned on or off DOES make a difference, but in my scene, all renders of the values in my original post make no difference. ALL pixels were EXACTLY the same in all those renders.

I'd like to know under what circumstances an IS value other than 1024 does anything at all.

spherical
12-10-2015, 11:25 PM
Um, ok. How do I see changes in the pixels between renders, then, if ALL pixels were EXACTLY the same in all those renders? Color me confused....

BTW, the reference to on/off influence from a Sky Dome was to madno's post about it being only Textured Environment.

DrStrik9
12-11-2015, 12:23 AM
BTW, the reference to on/off influence from a Sky Dome was to madno's post about it being only Textured Environment.

Oh, OK. I didn't know you were referencing another post since you only quoted my post. No problem.

So Importance Sampling only changes renders with other settings than 1024 if there's a textured environment. I'll have to test this.

kolby
12-11-2015, 02:36 AM
Regarding Importance Sampling:
To my understanding it does nothing if you don't use textured environment (don't know about Image World). If you use a light or lominous geometry or a inwards pointing world sphere with e.g. HDR map IS does not help.

Please somebody correct me if I am wrong here.

Yes, you are right.

DrStrik9
12-11-2015, 10:04 AM
Um, ok. How do I see changes in the pixels between renders, then, if ALL pixels were EXACTLY the same in all those renders? Color me confused....

Sorry, I should have been more specific. In my original post, I listed all the Importance Sampling settings I used in test renders (but did not show those renders, because they were all alike). And of course, I didn't post the images because there would be no point.

I have since found out that only Textured Environments in the scene would make any difference with varied IS settings. Since my scene used an inward-facing HDR sphere, adjusting IS settings made no difference.

I agree that the docs are a bit vague on this point, so my thanks to those who clarified.