Cleaning up motion blurred reflections

[Wilyman]

Please could I grab the scene for a look too, Oliver?

I'm still struggling with subdivs set at 4-1000.
This render took 4 hours at this size, but I'm still getting some fairly red bits in the SampleRate RE, and also some totally black buckets (right hand side of the two highlights on the right)...



It'd be good to see where I'm going wrong.

Cheers,

John

[^Lele^]

That's for Ollie to decide.
To be clear, however, on what is problematic in these kind of scenes: if you render it without motion blur, you'll notice the streak is a *very* bright highlight.
That alone will need -currently- a lot of sampling (other engines downright clamp. and good night.), because lowering noise in high energy pixels is a challenge.
The motion blurring makes the statistical chance of finding that highlight smaller (depending on moblur and motion parameters), requiring more samples to ensure it's found, and that shoots the rendertimes through the roof.
Further, this can be compounded by color mapping on either side (either not LWF, or the image has grading in the VFB showing different values than they really are.).
I speak of this because the contrast ratio of your shot looks off (on two different calibrated monitors, at that), with the blacks too black, and the highlights seemingly too bright (i can only judge them by the core-to-tail ratio, the image being LDR).

I could suggest to leave max AA unbounded (set it as high as it's able to go), raise the min samples still (6/, and to make buckets bigger to avoid those cutoffs.
But it sounds way too extreme for the kind of shot that is.

Without that scene, or any other which is misbehaving, i can't tell you for sure what it is which isn't working properly.
FYI, we are bound by implicit secrecy, so no data will be made public (in the case of Olli, i assumed he wanted -at least a little crop- for public consumption.)

[Wilyman]

That's for Ollie to decide.

Of course. My question's aimed at him.

Further, this can be compounded by color mapping on either side (either not LWF, or the image has grading in the VFB showing different values than they really are.).
I speak of this because the contrast ratio of your shot looks off (on two different calibrated monitors, at that), with the blacks too black, and the highlights seemingly too bright (i can only judge them by the core-to-tail ratio, the image being LDR).

It is LWF and there are no adjustments in the VFB. It's just underexposed as it's lit by a single small light, with the intention of adding it over the top of the HDRI-lit hero render.
To give you an idea of where the brightness should be, it's a white car.

Without that scene, or any other which is misbehaving, i can't tell you for sure what it is which isn't working properly.
FYI, we are bound by implicit secrecy, so no data will be made public (in the case of Olli, i assumed he wanted -at least a little crop- for public consumption.)

Sure. I'll try and chop this one up and send it over shortly. Thanks Lele.

Cheers,

John

[^Lele^]

It is LWF and there are no adjustments in the VFB. It's just underexposed as it's lit by a single small light, with the intention of adding it over the top of the HDRI-lit hero render.
To give you an idea of where the brightness should be, it's a white car.

See? i was so out of kilter assuming stuff.
Yes yes please, a scene!

[kosso_olli]

Please could I grab the scene for a look too, Oliver?

Sure, John. Please tell me where I should send it to.

The motion blurring makes the statistical chance of finding that highlight smaller (depending on moblur and motion parameters), requiring more samples to ensure it's found, and that shoots the rendertimes through the roof.

Are there any guidelines or rules to follow regarding the motion blur settings, just to make sure the sampler is finding these important pixels? I imagine one single revolution of the wheel would be sufficient, if the camera exposure is long enough to make on spoke a "blurred streak" of exactly 360°, no? I hope this makes sense...

I am rendering a test right now, but I have to say its sticking to the blurred highlight for nearly 6 hours. But it looks clean I have to say. Will post the result later.

[Wilyman]

Sure, John. Please tell me where I should send it to.

Thanks Oliver. My e-mail's john@circle-media.com

Cheers,

John

[^Lele^]

There are no sure-fire solutions that i know of (and not just for V-Ray, mind you. I may just be that ignorant, of course, in which case i stand to be corrected.), both cases are children of the same kind of statistically arduous set of circumstances.
Small, distant, and very intense lights, tiny geo to catch the highlight, moving fast, with a sharp specular, and maybe some DoF thrown in.
Anything that can be done to counter any of the above points, without sacrificing looks too much, is going to help.

In the case of John, it even clips sampling at 1k subdivs, requiring more (his highlight is in the 1000s float, f.e., for max intensity.), and so even more time.
And even more min-AA padding, because of the speed and size of the highlight, to be caught everywhere it should, and not leave gaps.
24-10000 to reach a 0.01 noise threshold isn't really very feasible. Although it will work.

Clamping may help some bits, but not a lot: some more tech may be needed to solve these kind of rendering problems automagically, under any kind of condition and setup, methinks.

EDIT: Rendering with an unbounded max AA (say, 24-10000 subdivs. ugh.) and a higher N.T. (say, 0.1) would likely allow to render a properly shaped, but slightly noisy, specular pass which could be denoised efficiently enough for it to be usable in post.

[Wilyman]

Are there any guidelines or rules to follow regarding the motion blur settings, just to make sure the sampler is finding these important pixels? I imagine one single revolution of the wheel would be sufficient, if the camera exposure is long enough to make on spoke a "blurred streak" of exactly 360°, no? I hope this makes sense...

I'd imagine for most wheels, a rotation of 360/n degrees, where n is the number of spokes, would be sufficient to give you the longest possible highlight. I guess the amount of blur on the badge would be the only noticeable difference.
Whether that would make a difference to the sampling is not for me to say though.

I'm also making a bunch of tests at the moment, based on Lele's feedback, so will update when I can.

Cheers,

John

[kosso_olli]

This is the result from today. The render settings are described as per Lele, thanks again!
It took 100 minutes on 3 nodes at 1k square.
The settings for the motion blur are: Two revolutions of the rim (720°) over 100 frames, rendered at frame 50. Motion blur duration is 50 frames, geometry samples for the wheel are 24. The highlight effect turned out great.
John, I would send the sene for this image. Can you live with that?

[Wilyman]

John, I would send the sene for this image. Can you live with that?

Sounds good to me.

[^Lele^]

This is the result from today. The render settings are described as per Lele, thanks again!
It took 100 minutes on 3 nodes at 1k square.
The settings for the motion blur are: Two revolutions of the rim (720°) over 100 frames, rendered at frame 50. Motion blur duration is 50 frames, geometry samples for the wheel are 24. The highlight effect turned out great.
John, I would send the sene for this image. Can you live with that?

Well, wow.
I am very impressed by the results.
Question, could you also share the EXR with, perhaps, the specular/reflection channels?
I'd be interested to have a play in nuke and see if with all the clamping and stuff i can still tighten it and keep it looking this good.

edit: i just noticed the tiny highlights from the brake disk holes' rims. :drool:

[kosso_olli]

Lele, I sent the renders to you.
One question: Might these settings also help with noisy bokeh shapes in DOF, caused by extremely bright highlights?

[^Lele^]

They might, yes.
Two main sides to this story: probabilities, and intensities.
DoF will make probabilities low, just like for moblur, requiring many AA rays (min, the fixed part, chiefly) to clean as a result.
The second part is what is indeed found, by all those rays: if it's a small, intense highlight, maybe even creating bokhes, you will want to throw many camera rays at it so it gets anti-aliased (the max AA here.), the more intense, the more rays, effectively.

Thanks for the images, and yes, they can be pulled and tightened a bit as they are already, and even more if they were denoised.
Very glad to have been of some help, while we wait for new tech to help more.

[Wilyman]

I've had a chance to do a few more tests here.

The main thing I've realised is that I was making it much harder for myself by testing on a wheel with such a tight chamfer. As soon as I tried it with Oliver's wheel instead, things started looking nicer.

So this is a test with just a single lightsource hitting the wheel, which could then be overlaid onto the hero render, in order to get something like the reference pic in my original post.
It doesn't look that pretty on it's own, but it's starting to get the effect I'm looking for...



Similar render times to the one lit by sun/sky above. This was 4 hours on a single workstation, and to be honest, it's still running out of samples on the brightest parts at 16/500 subdivs, NT 0.25, but I felt it was clean enough if it's going to be going on top of other layers.

Thanks for all your help on this Lele. It's been really interesting and informative.

Cheers,

John

[^Lele^]

Super stuff!
It is definitely on the expensive side of things, but man doesn't it look utterly interesting...