Cycles Development Updates

(doublebishop) #970

Just did a test in one of our production scenes, including compositing and preprocessing times, the total render time was reduced by 5% (1:54 to 1:48)

Nice work!

1 Like
(Simon Storl-Schulke) #971

Amazing how 6 lines of code can make such a difference.

1 Like
(Kumaran sip) #972

Still waiting for Intel open image denoiser.

(mathieu) #973

I really like that :slight_smile: 2 lines of code actually for 5-20% speedup, that’s 2.5 to 10% speedup per line, nice. Another good example that benefit to the end user can be inversely proportional to the quantity of code.

1 Like
(doublebishop) #974

OIDN (intels denoiser) is waiting on a fix from intel before it can be in the main branch of blender –

(DDB) #975

anybody knows if its for the viewport too ?

(Ace Dragon) #976

Something is moving on the denoising side, another fix for black dots.

It also looks like Brecht is doing some Cycles work again, mainly in the area of bugfixing (as 2.8’s feature set is likely frozen now).

(LordOdin) #977

You provided him a link that says nothing about his question lol.

Basically you can enable gpu in your preferences so it has the option to render on gpu with any files that are set to render with gpu. Other wise you will need a script to force any blend file you try to render to render on gpu

(Lsscpp) #978

About Cycles developement, can anyone confirm, or tell something more about what @bliblubli says here?

(Indy_logic) #979

Yeah, I can confirm this too. I did a job a few years ago that used Blender on a render farm. Some of the machines had 1080ti’s but most didn’t. As long as I went to those machines and selected the GPU in the Blender’s preferences, when the renderer encounter an environment that had no GPU, it just fall back to CPU. The other trick was to make sure that you selected to render with GPU in the scene you’re rendering.

(<== Lost? Click Me) #980

Only bug fixes will go to 2.7 branch from now:

(lolwel21) #981

Any updates on either double-sided Fresnel or Principled thin transmission?

(Ace Dragon) #982

Stefen Werner has taken another step towards a production-ready OIDN patch (in this case, it being included as a static library).

Still, I do think this will be for 2.81 at the earliest (as 2.8’s release is too close to commit big new things unless they make an exception).

(English is not my native language) #983

I tried to test the patch on Linux, but my limited knowledge about building things prevented me from doing it. I have no idea what to do with mkl-dnn dependencies :tired_face:

Ok, I left in blank the lines related to mkl-dnn from cmake-gui and it seems that this works?
Sponza 300 samples:

Using the Denoise node under Filters (I guess that’s how OIDN is used):

(mib2berlin) #984

Hi YAFU, how is it compare quality/render time to Blender denoiser?

CHeers, mib

(English is not my native language) #985

Hi mib. I have not done many tests yet, I’m not even sure I’m using OIDN correctly :stuck_out_tongue:
Anyway, at first glance, OIDN is doing an excellent job. The advantage of OIDN is that at low samples rates can achieve good results without visible patches and without losing much detail (Considering that a very noisy image does not even have much detail). But we know that Blender denoiser is mainly designed to eliminate residual noise from high sample rates renders. It will be necessary to test if for high sample rates and for production, which of the two is more stable for animation.
What I do not understand is if in OIDN things like Strength, Radius, etc. can not be configured, or is just that in Blender node it has not yet been implemented.
Here a simple comparison at low sample rate. You see blue sphere, OIDN is really having problems with those fireflies:

Blender denoiser (default)


OIDN + Blender Despeckle node

By the way, @LordOdin has shared Theory Studio build for Ubuntu 16.04 / 18.04 distro based.

When you extract the file you will see in the folder the three files related to OIDN libraries in there (libOpenImageDenoise). You copy those three files to your /usr/lib. There is also Windows Build there. It also includes Scrambling Distance and Dithered Sobol Lukas patches.

1 Like
(Stefan Werner) #986

All the options that the current OIDN library has are exposed in the Blender node.
Regarding the blue sphere in your particular example, try toggling the “HDR” checkbox, that should make a difference.

1 Like
(English is not my native language) #987

Apparently it does not get better (HDR off is worse). Here the 2.79 original file that I used to test Lukas denoiser:
test file.blend (1.8 MB)

I am tetsing in 2.79 because in 2.8 hair thickness has changed. Anyway I am using simple nodes configuration just with denoising passes enabled and Denoising Normal and Albedo connected (and Despeckle in one of the test). I have seen in your patch entry that people use more complex node configurations.


The OIDN looks suspiciously bad. It is highly unlikely that it has such a hard time dealing with fireflies, even if they are quite dense. It looks more like an implementation issue to me.

(English is not my native language) #989

I remember that Lukas Stockner also had some problems with those fireflies in this same test file. Lukas did an excellent job correcting the problem.
I’m not sure if it’s worth clarifying that Lamp has disabled Multiple Importance Sampling. But that’s the way it’s interesting to test with those fireflies.