Hi,
Same sample count in karma. But the Thin-Walled material has more noise.
What is the reason? How to reduce them?

Thanks!

jerry7
Hi,
Same sample count in karma. But the Thin-Walled material has more noise.
What is the reason? How to reduce them?

Thanks!

It looks like thin wall removes the direct refract skip optimization that you get with thick refraction since you can assume thick will always encounter an exit surface before a light source. Thin can't make this assumption that the surface is closed since it's meant to be used for open surfaces such as windows. Perhaps a checkbox or render property could be added for cases where you know it's a closed surface.

Ideally, it would just use non-probabilistic sampling and always sample both sides of the bsdf instead of sampling the reflection 4% of the time and the refraction 96% of the time--which is why the reflection is so noisy, it almost never samples it.

This problem has been present since the era of Mantra.
https://www.sidefx.com/forum/topic/62083/ [www.sidefx.com]

I have always been baffled by the fact that direct glossy reflections ever have any noise

In my mind they require only a single ray per pixel not even pixel sample as a pretty clean base and yet if based on probability they hardly ever get even that in many cases

It's even worse with low probability but strong light sources that are being reflected
I have always viewed this as a step back from more traditional ray tracing where such cases were matter of seconds on much older hardware

tamte
It's even worse with low probability but strong light sources that are being reflected
I have always viewed this as a step back from more traditional ray tracing where such cases were matter of seconds on much older hardware

The problem with not using probability is that once you get a few ray recursions in, the render time starts increasing exponentially if you always continue every branch. Ideally probability would be weighted on throughput so that bright reflections are properly sampled. Also for being able to reduce noise for the first few branches, it would be nice if you could define a depth to switch to probabilistic, so that the first refracted reflection could also be not noisy.

jerry7
This problem has been present since the era of Mantra.
https://www.sidefx.com/forum/topic/62083/ [www.sidefx.com]

In the world of non-pbr mantra, the glass shader was 100% noise free, however each bounce increased render time exponentially.

The thin wall in Redshift has not this problem. Karma really needs some optimization about sample algorithm.

jsmack
Also for being able to reduce noise for the first few branches, it would be nice if you could define a depth to switch to probabilistic, so that the first refracted reflection could also be not noisy.

This is sort of what I meant, If just just camera ray hits make sure they sent at least one path per brdf lobe per pixel (not even pixel sample), then as the path continues it can be probabilistic So essentially just depth 0 would make sure all lobes are covered
Especially direct samples that evaluate lights are a low hanging fruit, there is really no reason why 0 depth direct glossy samples should be noisy

Whenever someone compare renders like these I just don't get it, what is the point you are trying to prove here?
That an almost 10 years old render-engine is faster than a not even 2 years old one?
Karma is a baby compared to how long RS has been around, plus Redshift sucked at retractions for a long time not sure why you aren't willing to give Karma the same patience. Should we compare Redshift 1 to current Karma XPU to see what happens?

I'm not saying that Karma shouldn't improve, it stills lacks a lot of common-place functionality but it's getting there and fast! also we should celebrate all the good things that's doing better than other render-engines, crypto mattes? incredibly fast to render, volumes? a breeze to setup and render, Volumetric lighting? a dream come true on how easy is it to use, Supporting USD features? better than anyone, I could go on...

fusion9
Whenever someone compare renders like these I just don't get it, what is the point you are trying to prove here?
That an almost 10 years old render-engine is faster than a not even 2 years old one?
Karma is a baby compared to how long RS has been around, plus Redshift sucked at retractions for a long time not sure why you aren't willing to give Karma the same patience. Should we compare Redshift 1 to current Karma XPU to see what happens?

I'm not saying that Karma shouldn't improve, it stills lacks a lot of common-place functionality but it's getting there and fast! also we should celebrate all the good things that's doing better than other render-engines, crypto mattes? incredibly fast to render, volumes? a breeze to setup and render, Volumetric lighting? a dream come true on how easy is it to use, Supporting USD features? better than anyone, I could go on...

I don't have no patience. The reason of I recalling this problem is that the same problem has been present throughout mantra's entire lifecycle.
I don't want to see it happen to Karma again.

jerry7
Same sample count in karma. But the Thin-Walled material has more noise.
What is the reason?

We assume...

• thin-walled = a thin object that can be viewed from both sides (eg a window)
  
• refractive = a "think" object which cannot have a viewpoint looking at a backface (eg a crystal ball)
  

Refractive objects have many more noise optimizations applied to them because of the manifold assumption (eg we reduce noise by clipping away lighting from backfaces etc...)

jerry7
How to reduce them?

What if you leave it as refractive (ie not thin walled), but set the IOR to 1?

brians
What if you leave it as refractive (ie not thin walled), but set the IOR to 1?

Wouldn't that also kill all reflections?

brians

jerry7
Same sample count in karma. But the Thin-Walled material has more noise.
What is the reason?

We assume...

• thin-walled = a thin object that can be viewed from both sides (eg a window)
  
• refractive = a "think" object which cannot have a viewpoint looking at a backface (eg a crystal ball)
  

Refractive objects have many more noise optimizations applied to them because of the manifold assumption (eg we reduce noise by clipping away lighting from backfaces etc...)

jerry7
How to reduce them?

What if you leave it as refractive (ie not thin walled), but set the IOR to 1?

They are not the same thing.