Corona v1.3 vs 1.7

[cecofuli]

As in the title, a small test FOR FUN of different Corona version.

What it's strange is the v1.7 NL. In both v1.6 and v1.7  it's 8.14.
But the v1.7 has, clearly, less noise, especially on the ceiling, left /right walls and in the DOF.
Is it a bug?

[Ondra]

the noise estimation is not perfect, so when we improved the sampling in 1.7, the measured noise remained the same, even though the picture is much clearer

[maru]

Great tests Cecofuli!
The 1.3 shots appear more "smooth" because of different filter type or size, right? It could be probably easily fixed by enabling blur+sharpen in 1.6+
I am also curious how this test scene would behave with "render until it's clear" scenario, rather than setting a fixed time limit. That could be however tricky to setup, especially now that the noise level requires different values across different versions.

[Ludvik Koutny]

1.3 one doesn't have just different image filtering, it is simply more clean - less noisy, that's it.

My guess would be it's caused by adaptivity. Would be worth running one more test with adaptivity off.

[Ondra]

things I can think of that could cause this:
- adaptivity: different part of the same image could be more noisier in 1.3, post-1.4 the noise is more or less uniform across the image. How to test this: look at the rest of the image, isnt the noise pattern reversed somewhere?
- image filtering: we were previously blurring the image during rendering to reduce noise, but that was not compatible with adaptivity. 1.6+ corona has the option to apply the same blurring as post-effect interactively. How to test this: disable image filtering in 1.3 and rerender, is it more noisy?
- GGX vs. older, more primitive BRDF model. How to test this: does it happen also when the scene is not glossy?

[maru]

1.3 one doesn't have just different image filtering, it is simply more clean - less noisy, that's it.
My guess would be it's caused by adaptivity. Would be worth running one more test with adaptivity off.

Take a look at sharp edges on Cecofuli's tests. 1.3 is simply more blurry. With that in mind, you cannot really judge how noisy it is.

I am attaching an example of 1.6 hf 2 with and without adaptivity. If anything, it becomes more noisy with adaptivity off, so I don't think re-running the tests with adaptivity off makes sense. ;)

[Ludvik Koutny]

Well, I did try to take it into photoshop, and blur the 1.7 result about as much as 1.3 one, and it still appeared significantly noisier, even with the slight blur. Blur is not a denoising magic bullet, especially in such a small amount. Of course, in majority of cases, adaptivity helps, that's why it's on by default, but there still may be some cases, where it focuses sampling more on different places. It can often produce images that are noisier on certain places, to ensure noise level is even across the image.

The most obvious step in noise level is between 1.3 and 1.4. Adaptivity was introduced in 1.4, so it would not be unreasonable to assume adaptivity could have impact on it, and re-run the test with it disabled, if Cecofuli has the time to do so.

And lastly, I think it's actually easy to judge how noisy the image is - by just looking at it :) Ultimately, we deliver our images to clients, who don't really care about what the analytic noise level is, but how it looks to their eyes, and I bet if you shown this image to random group of non-cg people, majority of them would consider 1.3 to be the cleanest.

So, if there's any chance of performance regression since 1.3, even if it was just in very specific scenes, it's really worth investigating it :)

[cecofuli]

Hi guys..

I didn't expect so much interest.

v1.3 has default Filter setting, with Width 2.0.
I think that Filter Type isn't so important.

But, it looks like, with the same Filter width, the v1.3 add more "blurry" to the pixels.
So, it was better make the test with Filter width=1, in all version (Sorry)

In attachment the scene (Max 2012 - Corona v1.3 )

[cecofuli]

Ok, new set of test, with all rendering Filter width = 1

[romullus]

I like how you have to walk an extra mile to make 1.3 look worse than latter releases :] It looks like adaptivity indeed had played major role here.

[cecofuli]

Maybe with the v1.7 we can use a stronger Filter Width to blur a bit the noise, like in the v1.3 =)

[Ludvik Koutny]

If you want to take filtering out of equation, it's better to set filter type to none rather than width to 1. There is a very slight possibility that filter set to 1 in 1.3 may actually be adding to the noise :)

But other than that, it's what I suspected. You can see on the new test that adaptivity makes some places more noisy, while other places cleaner, generally making the noise level more even across the image. It looks like corona's adaptivity tends to prioritize darker places a bit more. Maybe that could be somehow tweaked in the future :)

[cecofuli]

Ohh.. I'm so stupid! Ahaha.. you are right! But, today I had a very crazy day (really crazy)!
But, I agree with you. Adaptivity should be improved a bit in my opinion.

[Njen]

A classic trick we sometimes use in VFX when dealing with dark images in modern renderers is to overlight the scene (expose up every light relatively) to make the render go faster, then in comp to expose down the image back to 'normal'.

[lacilaci]

Is this a fair comparison? Low res render with object/shader/texture and lighting complexity more or less equal across the screen?

AFAIK the purpose of adaptivity is to clear up complex objects in a simple scene equaly so that you don't end up with clean render and couple of noisy objects taking ages to get at acceptable level. I tested adaptivity on/off on couple of scenes yesterday and it worked as expected...

Or are we talking about adaptivity being broken? So that it slows down rendering in the long run?