Page 21 of 32

Re: OpenImageDenoise

Posted: Wed Feb 06, 2019 12:26 pm
by Sharlybg
lacilaci wrote: Wed Feb 06, 2019 12:19 pm This is incredible... I might actually give bidir a try in some cases!
Bidir will be less painfull for your standar and the kind of complexity you deal with in your daily workflow with something like a AMD Threadripper 1950X. Without such kind of hardware i fear you will hardly find any advantage to use it even with OIDN on.

No ?
Anyway it is always a pleasure to see your work so seeing your art on BIDIR could also be magic ;)

Re: OpenImageDenoise

Posted: Wed Feb 06, 2019 12:39 pm
by lacilaci
Sharlybg wrote: Wed Feb 06, 2019 12:26 pm
lacilaci wrote: Wed Feb 06, 2019 12:19 pm This is incredible... I might actually give bidir a try in some cases!
Bidir will be less painfull for your standar and the kind of complexity you deal with in your daily workflow with something like a AMD Threadripper 1950X. Without such kind of hardware i fear you will hardly find any advantage to use it even with OIDN on.

No ?
Anyway it is always a pleasure to see your work so seeing your art on BIDIR could also be magic ;)
Well, I'll see how and if I can use bidir, but until now since BCD is trash I couldn't use bidir for anything cause it would render for 10 hours, and if I needed changes after that I'd rather quit my job.

With oidn you can get 1 minute previews that clearly show the scene and lighting etc, and then let render for ~2 hours for hi-res and if blendluxcore gets on a decent renderfarm then I basicaly only need fast previews on my HW! That's how I've been working with corona for a very long time (most of work was on rebusfarm).. There is a lot that goes into this, to me this denoiser is just as important improvement as gi caching!

Re: OpenImageDenoise

Posted: Wed Feb 06, 2019 12:41 pm
by Dade
BiDir+Metropolis is pretty much the ultimate "Fire & Forget" rendering solution: the problem has always been it does a lot of work that contribute very little to the rendering (this is the big untold true about MIS aka Multiple Importance Sampling).

Re: OpenImageDenoise

Posted: Wed Feb 06, 2019 1:57 pm
by lacilaci
Is there something that could be done on the side of luxcore about memory consumption? I'm doing some 5K renders and denoising just kills my 32gb ram.

Re: OpenImageDenoise

Posted: Wed Feb 06, 2019 2:01 pm
by epilectrolytics
Dade wrote: Mon Feb 04, 2019 12:37 pm
epilectrolytics wrote: Mon Feb 04, 2019 12:14 pm
B.Y.O.B. wrote: Mon Feb 04, 2019 11:10 am I wonder how good OIDN fares vs. BCD when objects are behind glass or seen through a mirror.
Why not render a "deep albedo" like I suggested earlier with this example (right part is the albedo image):
I can do it but would be useful ? The problem we is we don't really know what Oidn needs. For instance, glass is a pure specular material but it has ray split (Reflect Vs Transmit) so it will have quite some noise and may work better with current Albedo while mirror is pure specular with no split and no noise so it may work better with your method. We are going to need some test and input from Intel guys.
I made a test:
.
specularnoisetest.jpg
.
Good news: OIDN denoises through arch glass!
Bad news: OIDN does not denoise efficiently what's in a mirror.
Request: Option for mirror material to render albedo AOV with specular depth.
I don't think we need this for metal, real glass or glossy.

Re: OpenImageDenoise

Posted: Wed Feb 06, 2019 4:23 pm
by Dade
epilectrolytics wrote: Wed Feb 06, 2019 2:01 pm Bad news: OIDN does not denoise efficiently what's in a mirror.
Request: Option for mirror material to render albedo AOV with specular depth.
I modified the code so all delta materials (mirror, glass, archglass and null) show the reflected/refracted albedo instead of just black. It seems to work well.

Re: OpenImageDenoise

Posted: Wed Feb 06, 2019 5:45 pm
by B.Y.O.B.
Seems like it doesn't work for Bidir - is it possible there or are there technical limitations?

Re: OpenImageDenoise

Posted: Wed Feb 06, 2019 7:49 pm
by Dade
B.Y.O.B. wrote: Wed Feb 06, 2019 5:45 pm Seems like it doesn't work for Bidir - is it possible there or are there technical limitations?
I forgot to update BiDir code. I extended the new Albedo AOV handling of specular surfaces to BIDIRCPU and LIGHTCPU (and now using the path throughput for the albedo).

Re: OpenImageDenoise

Posted: Wed Feb 06, 2019 7:53 pm
by epilectrolytics
Dade wrote: Wed Feb 06, 2019 4:23 pm I modified the code so all delta materials (mirror, glass, archglass and null) show the reflected/refracted albedo instead of just black. It seems to work well.
Awesome!
BiDir and mirror compatibility, these are two major improvements put into builds in just half a day :shock:

I waited for the last build to finish and loaded it into BlendLuxCore on Windows but there it crashes when trying to render.
Standalone with Cornell works well, also the earlier BiDir build, but not the last (delta material albedo) in Blender.

Re: OpenImageDenoise

Posted: Wed Feb 06, 2019 8:06 pm
by B.Y.O.B.
epilectrolytics wrote: Wed Feb 06, 2019 7:53 pm I waited for the last build to finish and loaded it into BlendLuxCore on Windows but there it crashes when trying to render.
Standalone with Cornell works well, also the earlier BiDir build, but not the last (delta material albedo) in Blender.
Can you post a scene that crashes?