Author Topic: High quality image filtering playground!  (Read 10562 times)

2019-04-06, 18:19:11
Reply #15

Feodor

  • Active Users
  • **
  • Posts: 121
    • View Profile
Feodor, i was refering to Nekrobul's post. As for your results, are you comparing tent filtering with denoiser versus high quality filtering with no denoiser?

Sorry mix up the message)
So I have not checked.
I was a little puzzled by the function that degrades image quality. When using noise, "Using high quality pixel filtering with AI denoising is not supported."
At this stage of development of this function, for the time being it seems to me useless, harmful

2019-04-06, 20:19:11
Reply #16

romullus

  • Global Moderator
  • Active Users
  • ****
  • Posts: 8779
  • Let's move this topic, shall we?
    • View Profile
    • My Models
One thing i noticed, it looks that HQ filtering doesn't like refractive materials, to be more specific, not filtering itself, but its combination with denoising. Look how much denoising has smeared the pixels behind refractive lens in render with HQ filtering.

It looks that same can be said about reflections - without denoising HQ filtering gives very good results, but if denoising is enabled, it just smears details badly. I think that enabling HQF by default, is a little bit premature at this point.
I'm not Corona Team member. Everything i say, is my personal opinion only.
My Models | My Videos | My Pictures

2019-04-06, 21:46:19
Reply #17

Nekrobul

  • Primary Certified Instructor
  • Active Users
  • ***
  • Posts: 1026
    • View Profile
Ok. I made a proper test of real life sitiuation this time.

Isues i have noticed.

With new sampler reflections in the mirror (wall detail shadows) are washed out for some reason.

New sampler handles DOF areas great but only at edges of objects. Large surfaces (floor in this specific case) looks horible a lot of speckle noise.

Also reflection\refraction detail being washed out with denoiser (parfume bottle)

For both images rendertiome 40 mins.

Old sampler made 266 pases

New sampler made 230 pases

In both cases native denoiser was used.

I will purge 4k trough the night just to see what will happen.

---------------------------------------------------------------
https://www.blackbellstudio.com/
https://www.behance.net/blackbell3d
CEO at "Blackbell Studio"

2019-04-06, 22:41:55
Reply #18

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
I have an idea, can you try disabling adaptivity for this test?
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2019-04-06, 23:03:47
Reply #19

Nekrobul

  • Primary Certified Instructor
  • Active Users
  • ***
  • Posts: 1026
    • View Profile
I have an idea, can you try disabling adaptivity for this test?

Good thing i have visited forum before going to sleep, shure yes. And i will make same settings i used for previous just with 4k.


-------------------------------------------------

10hrs 588pases no adaptivity 4GI\AA as pevious

Shadow details on wall present after denoising as on image with old sampling.

And noise on the floor is much lower.

PS - The denoise error was VFB halucinacion. It is not present on saved images(denoised). I compared them using photoshops diference blending and there is only slight differenc in noise stucture. I will attach both images for it will be simplier to compare.
« Last Edit: 2019-04-07, 12:59:52 by Nekrobul »
---------------------------------------------------------------
https://www.blackbellstudio.com/
https://www.behance.net/blackbell3d
CEO at "Blackbell Studio"

2019-04-08, 08:22:03
Reply #20

Njen

  • Active Users
  • **
  • Posts: 557
    • View Profile
    • Cyan Eyed
Ok, I tried another test with limiting by passes (PT+PT, same as my first test on page 1), but I left the Corona HQ denoiser on. Details in the file name.

The main difference I see in the type of noise, is that with HQ filtering off, there is the 'classic' type of CG noise. But with it on, the noise is a little more 'smudged', which I assume is the HQ denoiser trying to deal with a different type of noise. I may try another test with all denoising off.

2019-04-08, 08:33:16
Reply #21

Njen

  • Active Users
  • **
  • Posts: 557
    • View Profile
    • Cyan Eyed
Btw, Nekrobul, if you are doing difference tests, you should not be saving as .jpg files, as they will introduce their own noise, even on 100%.

2019-04-11, 14:35:08
Reply #22

Ludvik Koutny

  • VIP
  • Active Users
  • ***
  • Posts: 2557
  • Just another user
    • View Profile
    • My Portfolio
And you also should not render so clean images. Sure it's good to test how the new image filter contributes to final quality image, but it will be harder for human eye to spot the differences. I'd do one more test, again with the same amount of passes, but lower amount of them (let's say a half) so you can spot the differences better.

From the current test, it appears that the HQ image filter mainly helps to resolve very small highlights/glints.

2019-04-11, 19:36:46
Reply #23

Bormax

  • Active Users
  • **
  • Posts: 568
    • View Profile
I've tried new filtering in different scenes - exteriors and interiors. I based my test on same render time for pictures rendered with new and old filtering. And have to say that all the time I get the same result - pictures are obviously cleaner with new filtering, but Corona statistic shows less passes and higher noise level. But personally I prefer cleaner picture rather than some numbers from statistic lines :) I like this new filtering. Hope to see that it starts to work with Fast preview denoising soon.
Thank you guys!

2019-04-11, 22:57:37
Reply #24

Nekrobul

  • Primary Certified Instructor
  • Active Users
  • ***
  • Posts: 1026
    • View Profile
I will get back to the testing on weekend. I just have cuple of *real life* scenes with a lot of complications on dof and highlights.
---------------------------------------------------------------
https://www.blackbellstudio.com/
https://www.behance.net/blackbell3d
CEO at "Blackbell Studio"

2019-04-12, 00:05:59
Reply #25

Njen

  • Active Users
  • **
  • Posts: 557
    • View Profile
    • Cyan Eyed
And you also should not render so clean images. Sure it's good to test how the new image filter contributes to final quality image, but it will be harder for human eye to spot the differences. I'd do one more test, again with the same amount of passes, but lower amount of them (let's say a half) so you can spot the differences better.

From the current test, it appears that the HQ image filter mainly helps to resolve very small highlights/glints.

I get what you are saying, but to me practically, the noise in both of these images (HQ and non-HQ) are the minimum amount of noise I'd be willing to accept in my renders. There isn't any reason for me to make comparisons with more noise/less passes, since I'd never use them any way. My tests are more practical for me to figure out if this feature would benefit me.

Maybe other people can focus more on the academic, less passes/more noise tests.

2019-04-12, 08:13:14
Reply #26

Nekrobul

  • Primary Certified Instructor
  • Active Users
  • ***
  • Posts: 1026
    • View Profile
And you also should not render so clean images. Sure it's good to test how the new image filter contributes to final quality image, but it will be harder for human eye to spot the differences. I'd do one more test, again with the same amount of passes, but lower amount of them (let's say a half) so you can spot the differences better.

From the current test, it appears that the HQ image filter mainly helps to resolve very small highlights/glints.

I get what you are saying, but to me practically, the noise in both of these images (HQ and non-HQ) are the minimum amount of noise I'd be willing to accept in my renders. There isn't any reason for me to make comparisons with more noise/less passes, since I'd never use them any way. My tests are more practical for me to figure out if this feature would benefit me.

Maybe other people can focus more on the academic, less passes/more noise tests.

This yes, wanted to mention that too about the logic in testing on final quality renders. It is like we are aiming for final quality in the end right? So that is what we should be comparing.

Just was to sleepy yesterday.
---------------------------------------------------------------
https://www.blackbellstudio.com/
https://www.behance.net/blackbell3d
CEO at "Blackbell Studio"

2019-04-16, 00:47:15
Reply #27

dfcorona

  • Active Users
  • **
  • Posts: 290
    • View Profile
I've tested so far and found the quality to be just about equal but the render time to go up a LOT!

2019-04-17, 18:36:46
Reply #28

lycium

  • Users
  • *
  • Posts: 2
    • View Profile
Hi, after some delay due to illness I'm happy to report that the issue with adaptivity has been fixed, and render times are pretty much the same now. So it should be a pure win :)

It's in review and testing now and should be out ASAP!

2019-06-22, 20:26:03
Reply #29

dfcorona

  • Active Users
  • **
  • Posts: 290
    • View Profile
Does intel denoiser still not work with new noise filter? In my tests it doesn't