More calculation in shadow areas & edges

[Designerman77]

Hi Corona team,

What do you think about applying stronger calculation power to shadow areas and edges of high contrast?

I noticed that most of the noise and fireflies occur in shadow areas.

And very often one has to render longer, just because of poor antialiasing on edges with strong contrast.

The rest of the scenes are already clean much earlier.

Quite frustrating to accept long render times just because of antialiasing and noisy shadows.

Instead of less denoising on bump maps (core 3), I suggest a stronger & intelligent calculation on those two mentioned areas.

[houska]

Hi Designerman77,

thanks for your feedback.

As for the noise, it's really hard to tell what the issue is without seeing the specific scene or the render at least.

But the aliasing on bright edges is a well-known and well-understood issue and there's not much to do about it. It's well described in this thread in an old discussion about another renderer, but it's very precisely written and applies to Corona as well. (You're looking for MasterZap's answer): https://forums.cgsociety.org/t/render-to-exr-equals-bad-aliasing/1100220/4

Otherwise, I suggest you look at this article on our Freshdesk help: https://coronarenderer.freshdesk.com/support/solutions/articles/5000515615

[Designerman77]

Hey Houska,

thanks a lot for your reply.

I actually know the techniques how to reduce antialising on hard contrasting edges.

Been playing around with combination of highlight compression, reducing GI/AA to 4-8, highlight clamping on 4-10, changing image filter,
higher glare intensity... all combined... :)))

Antialiasing gets better...

A pitty that one loses the dynamic in the render.



The other point, which also other users start to notice: somehow the sharpness and details got worse in Corona 4.
Renders look washed out... ( AI denoiser on about 0.5 - 0.7 )


Please have a look at those two renders:

First one rendered with Corona 1.7 - nice clear details & antialiasing.
Rendered around 35 minutes on a 7 year old !!! 4-core i7 3,4 GHZ

Second one rendered with Core 4 Beta... all seems blurry and the metal looks washed out (as mentioned by other users, too)
Rendered around 20 minutes ! on a brand new 8-core Xeon 3,2 GHZ !!! ???

Super odd... since the Xeon is actually 3-4 times faster. And the PC has incomparably better cooling than the old one.


Some of us users have the impression that renders in Core 4 Beta became washed out and AA got much worse, even after lots of passes.


Bad results with same shader & lighting techniques and 3-4 times faster Computer ???


Hard to get my client work done nicely at the moment...

[houska]

Renders look washed out... ( AI denoiser on about 0.5 - 0.7 )

Thanks for the images, but the images are uncomparable. They are from different points of view.

From what I can see, the second image uses AI denoising, whereas the first one doesn't? If so, then splotches are expected. AI denoisers are meant for preview only and can never be used as final render solution unless you want to compromise image quality! Also note that longer rendering won't get rid of the splotches - even if you apply an AI denoiser on a perfectly clear image, it will produce splotches! That's the nature of the AI denoisers.

As for the speed, it's nearly impossible to compare two different viewpoints on two different PCs with two different versions of Corona.

Edit: That said, there's one feature that can make images appear more blurred (though very subtly, hardly perceivable actually) - the HQ image filtering, as can be seen here: https://corona-renderer.com/comparer/52qGKY But again, you have to compare two exact same renders

[Designerman77]

Hi Houska,

thank you for your quick reply.

Just because of a bit different angles - to call two pictures "incomparable"... I don't completely agree.
I guess you see the HUGE difference !

Especially as the second one was rendered on a 3-4 times faster PC. And never the less with much worse result.

Also a fact is that Corona Core 4 Beta delivers clearly worse results than Core 1.7

Regarding the AI denoiser I didn't know that its "only good for previews".
Thanks for that info.

The native "hi quality" denoiser still applies less calculation on bumps and - very bad - on round edges.
This is why I don't use it.

Although me and also several other users confirmed that this "selective reduced denoising" is absolutely bad, your programmers still haven't improved it since 3 months or so.

I would love to use the native denoiser... but not with this totally useless feature which makes renders "fall apart".


Edit: rendering the same scene right now...
Corona "high quality" denoiser seems to apply much more uniformly than in Core 3... so rounded edges don't remain visibly noisy.
My impression ? Or has it been changed / improved?

 




Since you doubt in the comparison of the two renders, I am going to set up a scene in Core 1.7 and open it in Core 4Beta to render both as A/B-comparison.


But I'm quite sure the points I mentioned do exist...

[houska]

Since you doubt in the comparison of the two renders, I am going to set up a scene in Core 1.7 and open it in Core 4Beta to render both as A/B-comparison.

If your only point is the reduced denoising on rounded edges, then don't do the comparisons. We're aware of that and it's intentional, because it brings more detail into surfaces with bump-mapping. We might change that in the future, but I cannot make any promises, sorry :-(

If you want to compare the images to compare the speed of rendering, then you can do it, but you need to do it on the same PC, exact same scene and ideally with noise limit (because the newer versions can be slower, but better quality)

[romullus]

If you want to compare the images to compare the speed of rendering, then you can do it, but you need to do it on the same PC, exact same scene and ideally with noise limit (because the newer versions can be slower, but better quality)

Hmm, would rendering with time limit, actually be more fair comparison? Noise level can be interpreted significantly different between various Corona versions and that could lead to false conclusions. If one would render 1 hour in version A and then one hour in version B, then there should be little room for misinterpretation of results.

[Designerman77]

Hey Houska,

the reduced denoising in Core 3 is - of course - not the actual point I talked about.

Neither rendering speed is my topic.

It is antialiasing and the blurry look in Core 4 Beta.

Im doing an A / B comparison right now, using the same amount of passes.
And of course the exact same setting.

I guess this should be a quite fair comparison.

[Designerman77]

Test done.

In my opinion the best result was delivered by Core 4 Beta - with AI denoiser.
AI denoiser looks less noisy than the Corona Hi quality in shadows and round edges. And it even preserved more details.
Last point (more details) is astonishing. Didn´t expect it.

I'm extremely surprised... especially because of my impression from other renders, where parts of the scene looked totally blurry.
And I was "freaking out"... :)


Now I'm literally confused.


In terms of antialiasing, there is literally (almost) no difference between Core 1.7 and Core 4 Beta.
Also contrary to my observations in other scenes. Didn't expect this outcome.



I guess there is no other way than rendering enough passes, applying a bit of highlight clamping / lowering GI/AA, etc... if you want clean antialiased edges and low noise.


Dang... I already see myself as a render-farmer with 8 x 32 cores... :)))
Seems the only way to client-ready renders in a super short time.


BTW, Houska, this rises the question why the AI denoiser is "not for final render quality" ?
To me it looks like the contrary.


Greetings!

[houska]

Please do note that more passes will not help you when you want to solve aliasing on the bright edges! Read that first link that I linked in the first reply above to see why.

As for the AI denoisers, they can be surprisingly good in some cases, yet fail in others. The key point is that no-one (not even the developers themselves) understands why they denoise a single given image the way they do. They are just trained on a huge set of reference pictures and then they use that information to denoise new images. And often they "hallucinate" detail where there shouldn't be any, so I'd be careful here. That's the only reason why we don't recommend the AI denoisers for production - their inherent unpredictability.

[Designerman77]

Hey Houska,

thanks a lot for this specific info!

Yes, the "hallucination of details" in AI denoiser... I can confirm it in some cases... as well as blurring out entire areas without any obvious reason... Areas that stay blurry even after doubling the passes.
If not even the developers totally understand it... then its "artificial intelligence". :)

Yep... also noticed that more passes don't really improve AA.
In older versions, though, I had the impression to observe the contrary.

However... "observations" can be misleading.


Thanks again !



P.S. to end with the thought from the beginning: in my opinion, "smarter" calculation in shadows (noise) and on AA would be a huge benefit.
Because in many cases, jaggy AA is the thing that makes a render look not 100% realistic.

[Designerman77]

P.P.S.

so, the only way - at the moment - to improve AA is by tweaking in the settings...
And rendering big size (like 4000 pix), so fine details / edges get a proper resolution ?

[houska]

No, unfortunately, tweaking the AA vs. GI setting won't help with the bright jagged edges either, because they are an inherent feature of the physically-based rendering. In the second link that I posted in my first post, you can see what can be done to help with the issue. It's written from the point of view of 3dsMax, but it should be applicable to C4D too.

[Designerman77]

Hi Houska,

thanks a lot. I had studied the tips in the link from you.

Highlight compression, bloom & glare, etc. I use as standard anyway.
Clamping mostly not... but lately I did, since it seems to help with AA.
Also sharpen&blur sometimes help... you are right.

Do you agree with the assumption that rendering in big size also reduces AA problems?
Seems logical to me...since the picture has more pixels for details.

[houska]

The problem with aliasing on bright edges will be still there, if you render in bigger resolution. But, what you can do is render in higher resolution and then scale the image down in post.