Page 3 of 4 FirstFirst 1234 LastLast
Results 31 to 45 of 53

Thread: Open Image Denoise node for DP Filter

  1. #31
    Open Image Denoise, is an excellent solution for LW.
    I'm not a programer, not even in mac.
    How could I use OIDN in LW 2019 in mac os system?

  2. #32
    areté adk's Avatar
    Join Date
    Feb 2003
    Location
    Melbourne
    Posts
    1,505
    Thanks for yet another great addition to LW Denis, and peebeearr for those interesting test. The DOF one is very interesting.
    I'll need to have a go at this when I get a chance.

  3. #33
    Registered User
    Join Date
    Mar 2018
    Location
    Perth, Australia
    Posts
    64
    Did some more tests with a more or less benchmarked scene....

    I found out that OIDN has trouble retaining small details when using Brute Force GI. When not using GI OIDN retains the small details. I have no clue why that is.

    What I also discovered is that the more AA you use, the smoother the image. This has nothing to do with how AA'd the buffer passes are (Raw RGB and Normals). Only the beauty pass apparently needs as many AA samples as you can get. This is quite ridiculous as in the sample scenes on the OIDN websites gallery, you can get amazing results with very low samples indeed. (Just like in LW only when NOT using GI). In particular the Sponza scene.... very low samples but the OIDN version retains details that are not even visible from the source!

    Here are some more samples.

    Raw input image [no GI 1m22s]
    Click image for larger version. 

Name:	Raw-No_GI-1m22s.jpg 
Views:	64 
Size:	2.01 MB 
ID:	144594

    OIDN result [no GI 1m.43s] - pretty smooth and retains details (look at the ceiling rosette)
    Click image for larger version. 

Name:	OIDN-No_GI-1m43s.jpg 
Views:	66 
Size:	1.14 MB 
ID:	144595

    OIDN result [Interpolated GI 9m16s] - Very low settings (PR-50, SR-32, MinPS-2, MaxPS-50, AngTol-20, Interpl-50). OIDN did not cope with this....
    Click image for larger version. 

Name:	OIDN-Intrp_GI-9m16s.jpg 
Views:	63 
Size:	1.58 MB 
ID:	144596

    OIDN result [Brute Force GI 4m17s] - Looks pretty smooth but most small details are totally gone. The only way to get this back is to crank up AA's a LOT!!
    Click image for larger version. 

Name:	OIDN-BF_GI-4m17s.jpg 
Views:	73 
Size:	1.38 MB 
ID:	144597

    All images are done with a minimum samples AA of 8. No Adaptive, Box filter and Filter Radius set to 0.

    I do not know why details are lost when using GI. The filter should work with Monte Carlo GI but for some reason it doesn't... I don't have Octane/Kray... perhaps someone could do some tests with Octane or Kray??

    Mind you that this scene is not really optimised.... I used one of the scenes that floated around when LW2018 just came out and everyone was trying to get the lowest render times possible. Check the thread here.

    P.S. Thanks Denis, didn't know about the presets... works like a charm!
    P.P.S. all renders are done with OIDN 0.8.2

  4. #34
    Registered User
    Join Date
    Apr 2016
    Location
    Barcelona, Spain
    Posts
    526
    Quote Originally Posted by peebeearr View Post
    Did some more tests with a more or less benchmarked scene....

    I found out that OIDN has trouble retaining small details when using Brute Force GI. When not using GI OIDN retains the small details. I have no clue why that is.

    What I also discovered is that the more AA you use, the smoother the image. This has nothing to do with how AA'd the buffer passes are (Raw RGB and Normals). Only the beauty pass apparently needs as many AA samples as you can get. This is quite ridiculous as in the sample scenes on the OIDN websites gallery, you can get amazing results with very low samples indeed. (Just like in LW only when NOT using GI). In particular the Sponza scene.... very low samples but the OIDN version retains details that are not even visible from the source!

    Here are some more samples.

    Raw input image [no GI 1m22s]
    Click image for larger version. 

Name:	Raw-No_GI-1m22s.jpg 
Views:	64 
Size:	2.01 MB 
ID:	144594

    OIDN result [no GI 1m.43s] - pretty smooth and retains details (look at the ceiling rosette)
    Click image for larger version. 

Name:	OIDN-No_GI-1m43s.jpg 
Views:	66 
Size:	1.14 MB 
ID:	144595

    OIDN result [Interpolated GI 9m16s] - Very low settings (PR-50, SR-32, MinPS-2, MaxPS-50, AngTol-20, Interpl-50). OIDN did not cope with this....
    Click image for larger version. 

Name:	OIDN-Intrp_GI-9m16s.jpg 
Views:	63 
Size:	1.58 MB 
ID:	144596

    OIDN result [Brute Force GI 4m17s] - Looks pretty smooth but most small details are totally gone. The only way to get this back is to crank up AA's a LOT!!
    Click image for larger version. 

Name:	OIDN-BF_GI-4m17s.jpg 
Views:	73 
Size:	1.38 MB 
ID:	144597

    All images are done with a minimum samples AA of 8. No Adaptive, Box filter and Filter Radius set to 0.

    I do not know why details are lost when using GI. The filter should work with Monte Carlo GI but for some reason it doesn't... I don't have Octane/Kray... perhaps someone could do some tests with Octane or Kray??

    Mind you that this scene is not really optimised.... I used one of the scenes that floated around when LW2018 just came out and everyone was trying to get the lowest render times possible. Check the thread here.

    P.S. Thanks Denis, didn't know about the presets... works like a charm!
    P.P.S. all renders are done with OIDN 0.8.2
    Did you use the normal and albedo pass? I'm getting excellent results with details in my tests.
    English is not my native language so please be patient.

    Salvador Ureña
    http://urenasalvador.wixsite.com/portfolio

  5. #35
    Registered User
    Join Date
    Mar 2018
    Location
    Perth, Australia
    Posts
    64
    Quote Originally Posted by Asticles View Post
    Did you use the normal and albedo pass? I'm getting excellent results with details in my tests.
    Yes I do..... This is my node setup. I do clamp my Albedo pass to max 100% (or 1) with the DP Limit node although I have not seen any difference if I don't... I just follow Intel's instructions here.

    Node setup
    Click image for larger version. 

Name:	Nodes.JPG 
Views:	79 
Size:	59.6 KB 
ID:	144598

    I can see significant degradation when I either not use Albedo or Normal (You can't use normal without Albedo though).

    Can you please post your results? Also your setup? I'm quite curious why I get the results I get.

    Cheers,

    Arthur

  6. #36
    Member
    Join Date
    May 2006
    Location
    France
    Posts
    3,965
    Just a note to say that the Histogram node in DP Filter may help you
    to see if your albedo input is correct (not going beyond 1.0).

    Gerardo Estrada tested the case when you have transparency and an albedo input,
    you need an albedo/raw with at least a 1.0 value (white)
    to get the denoiser working in this region, not zero (black)
    even if you don't have defined détails here.

    Applying the denoiser on pre-rendered images is a different thing
    compared to post-processing directly on rendering,
    the filter radius (aka oversampling) may inhibit the denoiser.

    Denis.

  7. #37
    Registered User
    Join Date
    Mar 2018
    Location
    Perth, Australia
    Posts
    64
    Quote Originally Posted by dpont View Post
    Just a note to say that the Histogram node in DP Filter may help you
    to see if your albedo input is correct (not going beyond 1.0).

    Gerardo Estrada tested the case when you have transparency and an albedo input,
    you need an albedo/raw with at least a 1.0 value (white)
    to get the denoiser working in this region, not zero (black)
    even if you don't have defined détails here.

    Applying the denoiser on pre-rendered images is a different thing
    compared to post-processing directly on rendering,
    the filter radius (aka oversampling) may inhibit the denoiser.

    Denis.
    Hi Denis,

    I use the LW image viewer to see if my values are above 100%. I verified that when I use your clamp/limit node the values will not exceed 1 or 100%.

  8. #38
    Registered User
    Join Date
    Mar 2018
    Location
    Perth, Australia
    Posts
    64
    Quote Originally Posted by peebeearr View Post
    Hi Denis,

    I use the LW image viewer to see if my values are above 100%. I verified that when I use your clamp/limit node the values will not exceed 1 or 100%.
    But anyway, that is not really the issue I am facing here... it is the lack of detail while using MC GI.

  9. #39
    Registered User
    Join Date
    Apr 2016
    Location
    Barcelona, Spain
    Posts
    526
    Quote Originally Posted by peebeearr View Post
    But anyway, that is not really the issue I am facing here... it is the lack of detail while using MC GI.
    How many samples do you need to recover the details?
    English is not my native language so please be patient.

    Salvador Ureña
    http://urenasalvador.wixsite.com/portfolio

  10. #40
    pass:sword OFF's Avatar
    Join Date
    Feb 2003
    Location
    Russia
    Posts
    994
    I made several tests on a simple scene and came to the conclusion that if you use luminescent polygons as a light source, the noise reduction in this case will kill small details (in random order). But if you use a conventional light source - the amount of noise decreases in direct proportion to the number of samples in the light source. More or less balanced values - from 16 to 32. If you raise the number of samples higher - the rendering time greatly increases, as with the increase of the rays Brute Force.

    Poly light

    Click image for larger version. 

Name:	LumeObj_001.png 
Views:	80 
Size:	1.39 MB 
ID:	144627

    LW Light (16 samples)

    Click image for larger version. 

Name:	Spherical light_001.png 
Views:	78 
Size:	1.34 MB 
ID:	144626

    GI 2 rays
    Camera settings 1/32, 001, 003

  11. #41
    Registered User
    Join Date
    Mar 2018
    Location
    Perth, Australia
    Posts
    64
    Quote Originally Posted by OFF View Post
    I made several tests on a simple scene and came to the conclusion that if you use luminescent polygons as a light source, the noise reduction in this case will kill small details (in random order). But if you use a conventional light source - the amount of noise decreases in direct proportion to the number of samples in the light source. More or less balanced values - from 16 to 32. If you raise the number of samples higher - the rendering time greatly increases, as with the increase of the rays Brute Force.

    Poly light

    Click image for larger version. 

Name:	LumeObj_001.png 
Views:	80 
Size:	1.39 MB 
ID:	144627

    LW Light (16 samples)

    Click image for larger version. 

Name:	Spherical light_001.png 
Views:	78 
Size:	1.34 MB 
ID:	144626

    GI 2 rays
    Camera settings 1/32, 001, 003
    In my examples the main light comes from a conventional light source --> 2 Portal lights. The rest is indirect lighting from bounces... That makes the whole thing a bit less usable for interior shots of this kind. I have another interior whith much more portal lights on more different angles (ie. more modern type of interior) and that comes out quite nicely.

    So the conclusion we can draw here is that whenever indirect lighting is used, the filter can't handle it properly. The question is --> WHY???? And how to solve it!

    Damn, I so wished this would work properly!! I will still use it and in many cases it will give you a better end result...

  12. #42
    Member
    Join Date
    May 2006
    Location
    France
    Posts
    3,965
    OpenImageDenoiser has been updated to version 0.9,
    you can download it here,
    https://github.com/OpenImageDenoise/...14.windows.zip

    If you installed previous version you just need to overwrite the "OpenImageDenoise.dll"
    library in the plugin directory of the node.

    The DP Filter node itself hasn't been modified, except the readme file,
    including much more documentation about the denoiser behavior
    and different ways for setting auxiliary maps.


    Denis.

  13. #43
    Registered User
    Join Date
    Mar 2018
    Location
    Perth, Australia
    Posts
    64
    Quote Originally Posted by dpont View Post
    OpenImageDenoiser has been updated to version 0.9,
    you can download it here,
    https://github.com/OpenImageDenoise/...14.windows.zip

    If you installed previous version you just need to overwrite the "OpenImageDenoise.dll"
    library in the plugin directory of the node.

    The DP Filter node itself hasn't been modified, except the readme file,
    including much more documentation about the denoiser behavior
    and different ways for setting auxiliary maps.


    Denis.
    Thanks for the heads up Denis. I'll give it a run for the money tomorrow!

  14. #44
    Registered User
    Join Date
    Mar 2018
    Location
    Perth, Australia
    Posts
    64
    Quote Originally Posted by dpont View Post
    ...
    The DP Filter node itself hasn't been modified, except the readme file,
    including much more documentation about the denoiser behavior
    and different ways for setting auxiliary maps.

    Denis.
    Hi Denis,

    Couldn't find much difference in the readme file though....
    Just downloaded it....
    File name: readme_OpenImageDenoise.txt
    File size: 2075
    Date: 27-03-2019

    Arthur

  15. #45
    Member
    Join Date
    May 2006
    Location
    France
    Posts
    3,965
    Quote Originally Posted by peebeearr View Post
    ...Couldn't find much difference in the readme file though....
    Just downloaded it....
    Yes I missed the upload…totally,
    it is 05/12/19 dated now.

    Thanks for report,
    Denis.

Page 3 of 4 FirstFirst 1234 LastLast

Bookmarks

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •