Help with difficult indoor lighting

zeitmeister · Post by **zeitmeister** » Thu Oct 13, 2016 6:59 am

Hey there,
I have to render an animation in 4k. CPU because of our little render farm.
Quite some frames, because the client needs 50fps and the animation is 30 seconds of duration.

The location is a parking garage only illuminated with neon lights. Glossy floor and ceiling elements, plus two cars.

What settings would you recommend?
- Pathtracing or Pathtracing with BiDir
- clamp contributions
- supersampling or render in bigger resolution with postpro denoising
- downsize filter if original resolution plus supersampling
- how many light layers (at the moment I use two; location and car lights)
- strength of the emitting materials
...

What else should I be aware of?
All my light emitters are highpoly original meshes.

I really would appreciate some hints from the experienced power users here, and of course also from the Indigo Devs to set up my scene for optimum technical speed and visual approach.

Thank you very very much in advance!!!

Gesendet von iPhone mit Tapatalk

Post by **Oscar J** » Thu Oct 13, 2016 7:34 am

Am no expert at indoor stuff of course, but BiDir does seem like the way to go with glass and glossy materials combined with small light sources (consider making bigger but invisible emitters).

Clamping is probably better than excessive SS if you get fireflies. Really low clamping values *might* cause flickering in animations according to the devs, because the threshold pixel brightness is calculated from the mean brightness of every frame. Might be worth looking into.

Render in double res+denoise is probably better if your video software can handle it. Denoising at final resolution will cause some smudging.

Light layers should not have a major effect on render times other than the time it takes to send image buffers etc.

The absolute strength of the emitters probably don't matter, it's rather the balance in power between the different emitters. You might want to overpower small/noisy/firefly prone emitters and lower them in light layers to optimise sampling.

I am sure you might know most of these things already, but anyway.

Polinalkrimizei · Post by **Polinalkrimizei** » Thu Oct 13, 2016 8:06 am

Any chance of simplifying some of your highpoly emitters? At least the ones in the distance. How many shots does the scene have?
And yes, BiDir PT is the best long-distance runner with indoor scenes to my experience.

One exception: With HIGHLY reflecting environments (I had a kitchen made out of polished stainless steel...) MLT BiDir was the way to go...
Do you have a still frame?
And any chance of seeing the finished product?

zeitmeister · Post by **zeitmeister** » Thu Oct 13, 2016 8:25 am

Oscar J wrote:Am no expert at indoor stuff of course, but BiDir does seem like the way to go with glass and glossy materials combined with small light sources (consider making bigger but invisible emitters).

So I go with PTBiDir, thanks a lot! MLT didn't work for me well.

Oscar J wrote: Clamping is probably better than excessive SS if you get fireflies. Really low clamping values *might* cause flickering in animations according to the devs, because the threshold pixel brightness is calculated from the mean brightness of every frame. Might be worth looking into.

I supposed something like that!

What value for the clamping would you recommend? 100 is set by default; my shots are mostly clean, slow and simple dolly rides with low scene or brightness change...

Oscar J wrote: Render in double res+denoise is probably better if your video software can handle it. Denoising at final resolution will cause some smudging.

Agreed! Gonna have to try to render 8k with SS of 1 and then denoising... it didn't work out well in 6k, SS1 and 1700 SSP... so maybe I really need a SS of 2?

Oscar J wrote: Light layers should not have a major effect on render times other than the time it takes to send image buffers etc.

Such large image buffers consume a lot of RAM; the fast machines do have 64GB... a SS of 4 is not possible there any more. With 2 light layers.

Oscar J wrote: The absolute strength of the emitters probably don't matter, it's rather the balance in power between the different emitters. You might want to overpower small/noisy/firefly prone emitters and lower them in light layers to optimise sampling.

You mean that I should set them very high in the material settings and lower the overall brightness of the light layer settings? Will give it a try!

Oscar J wrote: I am sure you might know most of these things already, but anyway.

Yes indeed, but there is so much more to learn...
Thank you very much!!

zeitmeister · Post by **zeitmeister** » Thu Oct 13, 2016 8:32 am

Polinalkrimizei wrote:Any chance of simplifying some of your highpoly emitters? At least the ones in the distance. How many shots does the scene have?
And yes, BiDir PT is the best long-distance runner with indoor scenes to my experience.

Will try to simplify the neon bulb emitters tremendeously; good point! The scene do have a couple of shots... 7 in general, but with variations in light and zoom.

Polinalkrimizei wrote: One exception: With HIGHLY reflecting environments (I had a kitchen made out of polished stainless steel...) MLT BiDir was the way to go...

Noticed exactly that in a former job where outside sunlight and inddor lighting was mixed. MLT with a SS of 4 or 5 did a great job there... but here it seems that MLT is definately out. It takes too long to converge.

Polinalkrimizei wrote: Do you have a still frame?
And any chance of seeing the finished product?

I can provide a still frame tomorrow; nothing secret here. Without cars for sure.
If I can manage the rendering with Indigo, the result could be seen in 4 weeks.

Post by **Oscar J** » Thu Oct 13, 2016 8:50 am

Regarding clamping: go as low as you need for the fireflies to go away, but no further.

I am not sure whether there can actually be flickering, but it is a theoretical problem in worst case scenarios.

I would have though SS 1 would have been sufficient when you're denoising and resizing, but for sure go with SS 2 if it helps with noise and AA and your machines can handle it. There really isn't any reason to use SS 4 or similar anymore now that we have clamping.

For the future I'd consider a couple of GTX 1070's - the GPU renderer eats high res and SS for breakfast. And no box noise -> helps tremendously when denoising.

Polinalkrimizei · Post by **Polinalkrimizei** » Thu Oct 13, 2016 9:34 pm

What I do with bright interiors: If possible, I change the wall behind the camera to an emitter to control overall brightness and tone (used to do that with walls as exit portals + HDRI, which is prettier but slower).
This way, I don't have to wait for darker areas getting cleaned up by small emitters (some shadows take forever otherwise).
But if you need parts of your scene to be pitch black, that is no option of course!

lycium · Post by **lycium** » Fri Oct 14, 2016 12:59 am

Oscar J wrote:For the future I'd consider a couple of GTX 1070's - the GPU renderer eats high res and SS for breakfast. And no box noise -> helps tremendously when denoising.

From my POV this is the big deal. For animations you'll maybe want to do some denoising, and unfortunately the boxnoise basically makes this impossible.

It might be worth asking a buddy with a GTX 1070 or two to try rendering your scenes, if they fit in memory and render acceptably quickly, it's probably a good investment.

zeitmeister · Post by **zeitmeister** » Fri Oct 14, 2016 7:13 am

I'll upload a sample frame in a few minutes.
Why is this box noise removed with GPU rendering, but not with CPU? Is this a technical reason because of the cores or threads?
My scene actually consumes ~37 GB of RAM... maybe more. So that GTX doesn't help me there...

(

Post by **Oscar J** » Fri Oct 14, 2016 8:13 am

What render settings and rendering time did you have there? Looks quite good to me, very little denoising needed.

zeitmeister · Post by **zeitmeister** » Fri Oct 14, 2016 8:30 am

This was 6k downsampled by Photoshop.
Rendered on 8 machines; approx. 1 hour and 1700ssp; ss1.
But this won't work; I can't get rid of the noise at the front of the elevator.
So at the moment I try 4K with ss2 and 2500 ssp. Runs a bit faster.
Tomorrow we will see...
I think that a temporal denoiser will do a better job.

Gesendet von iPhone mit Tapatalk

Post by **Oscar J** » Fri Oct 14, 2016 8:39 am

Temporal denoising will probably get rid of that noise in the elevator, yeah (you are using arch glass right?). You'd be surprised by the noise levels you can get away with. Try Neat Video, I can maybe help with the settings if you run into problems (have played around with it a bit): https://www.neatvideo.com

Scene looks great by the way.

zeitmeister · Post by **zeitmeister** » Fri Oct 14, 2016 9:16 am

Thank you! Yeah; still running on AE CS4... but we do have neatvideo v3, I think.
Thanks a lot for offering help there!
At the moment 20mins per frame on 13 machines... I have to optimize that.

Gesendet von iPhone mit Tapatalk

Zom-B · Post by **Zom-B** » Sat Oct 15, 2016 6:27 am

Hey Zeiti, interesting project you have there!
50fps, slow camera movement Animation is best footage for a temporal denoiser!
In my experiance 200% of final res @ SS1 is enough, higher IGI filesize also raises the final gathering of all slave data. With so many slaves I sugest trying the "slave sends data in end" feature, but only if IGI is not to big, otherwise rendered data gets lost from slaves in halt time... But be aware that all slaves send data each after another, and once eine, they are idle until next frame starts.

For temporal denoising I sugest you do a "noisecatcher" frame for each shot, where you put a plane with the noisiest Material (without textures) in the scene at the area where the most noise is. This reference shot is then used to teach the denoiser.

Last idea is to give MLT another try raising the XYZ (forgot Name) value with default of 0.25 up to 0.85 vor higher!
With that you lose the cloudy Boise MLT generates and geht More PT like Boise Ostern, but without any Boxnoise at all.
The overall uniform noise should then be killed easy with neatvide.

Since I'm on vacation, I cant give ya the proper namens for the both options above...

arc en ciel · Post by **arc en ciel** » Sat Oct 15, 2016 7:16 am

Zom-B wrote: Last idea is to give MLT another try raising the XYZ (forgot Name) value with default of 0.25 up to 0.85 vor higher!
With that you lose the cloudy Boise MLT generates and geht More PT like Boise Ostern, but without any Boxnoise at all.
The overall uniform noise should then be killed easy with neatvide.

Since I'm in vacation, I cant give ya the proper namens for the both options above...

these options are :
- Large mutation Probe
- Max change

in my little experience, I put LMP to 0.51 - 0.7 and max change to 0.06-0.095
it's faster, but up these values the noise are more thick, bigger
maybe more easy for denoising in postpro ? I didn't try

here two images examples i ran in july, watch the name of the files for specifications, the big change is about large mutation probe from 0.78 to 1.78
the speed is multiply by 2, but the noise(s size too, the rendering's time does not change