Experimental 2.77 Cycles Denoising build

[Grzesiek]

I'll have to start playing with the de-noising part. I wonder if you animate the sample seed, will it still be as noticeable since each frame would have a total different sample.

Is the de-nosier part of the nightly build by now? or separate download?

[mib2berlin]

Hi, it is in experimental builds, links in https://blenderartists.org/forum/sho...=1#post3127350

Cheers, mib

[1pepe]

There are any 32bit build of the experimental build?, i tried

https://blenderartists.org/forum/sho...=1#post3127350

links and while it apear to have builds when you try it the LWR denoising option isnt there.

[Mobiledeveloper]

There are any 32bit build of the experimental build?, i tried links and while it apear to have builds when you try it the LWR denoising option isnt there.

you can download denoising build on graphicall

http://graphicall.org/1190

[Ace Dragon]

Another heads up, Lukas has committed another batch of commits.

They include a lot of code cleanup and improvements to NLM (the former in part being due to obsolete/failed algorithms being removed like the original WLR).

[Razorblade]

ehm, i did a rendering with objects behind glass, and the objects almost got blurred away.
And i wonder if this is how denoise should behave ?, i think bluring should be done based upon similar object origin.
And in the case of glass, (then asside reflections), it should look beyond it and then smooth on object level again.
Where smoothing is mostly light based, texture's dont need to be smoothed

But i wonder if it currently indeed works like that, or that it rather is some shape recognition (ea edges) with smooth..

[lsscpp]

As I posted earlier, there's a paper that describes the use of "virtual flash", which is a very quick&clean (few samples i guess) image, that is used to drive a denoising algorithm. Now, in Blender AO pass set to distance=0 gives a similar result, a quick&clean image that has all the materials properties intact, like trasmissions, reflections, textures, bump etc...
Wouldn't these infos be useful to avoid over-smoothing behind glass or reflected areas?

[Razorblade]

enabling world AO isnt that hack that kinda emulates global radiosity of light, as a global illumination ignoring the true light setup.
(as way for getting interiours rendered without getting on windows to lit the scene up for a daylight interior).
not sure how it could help.. (i'm thinking the model known surfaces textures shaders etc, the unknown part is the light distribution).
So a correct light distribution model is calculated..and hm thats the part that's need denoising, besides the part that light can bounce randomly and thus be more pressent/reflected un-even distrubuted.. perhaps only the light 'flow' needs to be denoised based on pixels with similair surface ... but hm thats me thinking, i dont know current interpretation.
..

[zeauro]

It looks like cross-denoising option is not working with GPU.

[aliasguru]

Basically, I'm getting a distorted render when using multiple GPUs. See this screenshot of a viewport render here:

Denoiser_Issue.JPG

I've got three TitanX CUDA GPUs working on the image. The first thrid of the image comes out fine, the rest is distorted as you can see. This only happens to me when using the denoiser build. I've tested many of them, in the last 4 months none of them (not even self compiled ones) worked for me. CPU denoising works fine though.

Deactivating two of three GPUs and only rendering on one of them works fine too.

Has anyone else experienced this issue?

[burnin]

Yes, noticed it quite a time ago... working on one GPU at once. Using two, doesn't work at all. Anyway, it is not advisable to use multi GPUs on a single frame.
Tho, working good with animations (splitting frames per GPU1, GPU2 & CPU).

[B-Rae]

Tho, working good with animations (splitting frames per GPU1, GPU2 & CPU).

I would do this. If you are only rendering a still image, use 3 instances of blender, use 1/3 the samples you were going to use with a different seed value for each, then render each on a different GPU and merge all 3 - you might actually end up with a clearer image this way anyhow. Slightly more clunky, but a good solution for now.

[lukasstockner97]

Just a quick update:
- I'm working on a update that will improve quality a lot for hair/fur/grass/fine detail in general. Right now it makes DoF worse, so I'm still working on fixing that. Downside is that memory use is going up, but that's only ~50MB for 256x256 tiles.
- The next experimental build will include some render pass changes, meaning unlimited light groups, an AOV output node and the Cryptomatte ID pass system.
- The Virtual Flash image could work well, I'll definitely try it.
- MultiGPU should also be fixed soon, along with an OpenCL version of the denoiser.

[ejnaren]

Lukas, that sounds amazing. You are a star. Will the light groups be able to be saved and adjusted again after closing blender?

[lukasstockner97]

Lukas, that sounds amazing. You are a star. Will the light groups be able to be saved and adjusted again after closing blender?

They'll work just like any other pass that's in Blender right now, except that you can name it - so, you can save them individually, or save all at once in a Multilayer EXR. Later, you can then open them in the Compositor and tweak and recombine them as you like.

Also, glad to hear that it will be useful

[Indy_logic]

Lucas,
This sounds amazing! Excited to see what you come up with.

What about the AOV node? I can only imagine how it works. Does it allow you to send the output of any node in the shader graph to it's own pass? If so, that's going to be a god send! Any information you can give us is greatly appreciated.

[lsscpp]

AOV and Lightgroups are a HUGE contribution to Cycles. Period

[lukasstockner97]

Lucas,
This sounds amazing! Excited to see what you come up with.

Well, it's not exactly new, other renderers had this stuff for ages

What about the AOV node? I can only imagine how it works. Does it allow you to send the output of any node in the shader graph to it's own pass? If so, that's going to be a god send! Any information you can give us is greatly appreciated.

Yes, that's the idea. A basic example can be seen here: http://pasteall.org/pic/show.php?id=110778

[Ace Dragon]

New NLM kernels have been committed (there's another commit stating they work on CUDA as well).
https://lists.blender.org/pipermail/...ry/092599.html

I assume these could be the ones Lukas was talking about when mentioning fine detail improvements.

[lukasstockner97]

New NLM kernels have been committed (there's another commit stating they work on CUDA as well).
https://lists.blender.org/pipermail/...ry/092599.html

I assume these could be the ones Lukas was talking about when mentioning fine detail improvements.

No, not yet - the new NLM kernels should get significantly reduce the noisy outlines next to edges with strong brightness differences. In technical terms, the new kernels use the Condat algorithm and the patchwise NLM extension for faster filtering with better quality.