Denoiser pass Like Disney Research paper and implemented in Renderman

[jaco graaff]

This can greatly reduce render times and maybe even create better previews. I am going to try it out in the non-commercial version of Renderman - since I am skeptical on how well it works in animation - but then it is not simply denoising - the algorithm uses motion vector data in it's calculation



http://www.disneyresearch.com/wp-content/uploads/Path-space-Motion-Estimation-and-Decomposition-for-Robust-Animation-Filtering-Paper.pdf

http://www.disneyresearch.com/publication/pathspace-decomposition/
http://renderman.pixar.com/view/denoiser

[agentdark45]

Damn that's cool! I wonder if this tech could be used to upscale / denoise already rendered footage (dvd's e.t.c).

[fb01]

not sure if this is the same as the one that appears in render man 20? - recently released -
that needs separate pass(es) in order to do its magic

is there any chance of this sort of thing happening in corona or other renders - or is patented etc to stop that happening any time soon?

 

[Jarda]

Hi,
work on adaptive image sampling as well as denoising in Corona is currently under way. The plan is to have a preliminary version by the beginning of the fall, and a stable one by the end of the year.

Some comments on the Disney paper: As far as I know (unofficial), the spatiotemporal filter described in this paper has not been implemented in RenderMan. It's a fairly involved tech that does not easily blend with many renderer.  I believe that denoising in RenderMan 20 is really just a spatial filter (working independently on individual images).
Cheers!

[stock]

 Great to hear that.
Do you know if this is a related technique? Could it be implemented in Corona?

[jaco graaff]

Posts like...

http://cgpress.org/archives/renderman-20-now-available.html
"This release incorporates noise reduction technology developed by Disney researchers...."

http://renderman.pixar.com/view/renderman20
"...but RenderMan's new denoising technology, originally developed at Disney,..."

does claim the Dysney Research Noise Filter algorithm is used - although I do not know to what extend. It might be a simplified subset.

Renderman also only use the "noise-filter" part of the total technology developed by Dysney Research in Zurich which also includes:

• up-scaling,
• motion blur,
• dof,
• frame-interpolation (render only every 3rd frame)
• and capabilities and some new ways of handling volumetrics: http://www.disneyresearch.com/project/joint-importance-sampling/

I think there will be some cases where this could be useful - like a controlled camera move in a architectural rendering or some of the supporting background fx.

Some of this functionality might fall apart in a more dynamic scenes with complex textures and fast articulate movements like character animation.

There is a huge difference between a purely post denoiser algorithm and one that uses "deep-sceen data" like z-depth, motion vectors and object-ids integrated into the rendering process.

As 4K is going to be the standard in the near future it makes sense. Our processors power is growing and make render faster (who tried to render with maxwell in 2006???), the quality and size of output also grows exponentially. There is also only so much detail the human eye needs to perceive quality.

http://www.cambridgeincolour.com/tutorials/cameras-vs-human-eye.htm
http://www.red.com/learn/red-101/eyesight-4k-resolution-viewing
https://plus.google.com/+YesIKnowThat/posts/LToL7NFHU4r


Just having fun...

[Ludvik Koutny]

Yeah, i agree that if Pixar released Renderman 20 with new denoising options and at the same time released paper about denoising techniques, it would be really silly to assume those two things are unrelated. What's in Renderman 20 is most likely (at least a part of) what's in that paper.

[rfletchr]

The did an interesting trick on the lego movie to denoise their shots. They used motion vectors to sample a pixels value in then previous/upcoming frames.
It would be interesting to know if corona could(perhaps already is) using caching between frames to reduce noise.

[Ondra]

The did an interesting trick on the lego movie to denoise their shots. They used motion vectors to sample a pixels value in then previous/upcoming frames.
It would be interesting to know if corona could(perhaps already is) using caching between frames to reduce noise.

It is not using this.

In general I would like to state this:

there are a lot of ways to reduce noise/render times. Some of them are more "clean" in that they do not produce bias/artifacts, do not require user input, tedious tweaking, ... We are looking into the "clean" options first (2 people are working on it at the moment). Only after we do reasonable clean optimizations, we will explore those more unreliable options, such as post-processing denoising.

[Juraj]

Reminded me of this paper from Siggraph '15 (my brother sent it to me I don't browse this stuff much) :- )

A Machine Learning Approach for Filtering Monte Carlo Noise
http://cvc.ucsb.edu/graphics/Papers/SIGGRAPH2015_LBF/

[pdmenon]

check out www.innobright.com for denoising of MC images.

BTW the rights to the patents for this, including the Learning Based Filtering,  is held by innobright and not disney and this feature has been integrated into a number of leading renderers such as Arnold, V-Ray etc., and hopefully shortly into Corona also.

[rombo]

Yeah, I'd like to see a solution like this being implemented as well in Corona - Especially for Archviz imagery - I'm finishing an animation and the noise is a frustrating element for those of us who don't have access to hundreds of thousands of euro renderfarms - Even with the corona portals optimization, absence of dof & motion blur it's taking between 30 min and 1h to clean up the massive noise from the animation, between 50 to 100 passes on a 3930k.
Unless I'm doing something wrong, with this engine I get twice as much rendertime to get the same look I get in VRay in half the time (for this particular type of work - It's my 1st corona job).

But then again, the light distribution is better, that's true... :>

[Juraj]

Unless I'm doing something wrong, with this engine I get twice as much rendertime to get the same look I get in VRay in half the time (for this particular type of work - It's my 1st corona job).

No, that is correct. You're comparing speed of Vray when using IR/LC with precomputation, which is just fantastic stuff for animation speeds of static flythroughs. Even Vray would be unusable for animation archviz if you had to use BF/LC in interiors for example.

Same look...well, sort of, depends on scenario. But yes, Corona at the moment is not a good solution for archviz anims. Costs too much money to be viable in our budgets.

[pdmenon]

Please check innobright.com/documentation/ for a How to guide on using Altus Standalone denoiser with Corona Renderer.

[JCdeBlok]

Speaking about nice..  Sometime ago I wrote a firefly post filter that does a good job without any blurring what so ever.

Have a good look at the examples here and notice how good it works even in the MB/DOF areas of the children''s room and that it doesn't take away any of the 'life' out of the image: http://blog.irayrender.com/post/91961451241/firefly-removal-max-script

How it works is as follows:

Stage one, detecting fireflies:  For each pixel it's surrounding pixels are checked for contrast, if a certain  number of surrounding pixels have more then a given threshold of contrast it's marked as a firefly.

Stage2: fixing it: there are 255 possibilities for the 8 surrounding pixels being over/under the contrast limit. For each of these possible combinations a definition is made for which pixels to use for reconstruction.

for example (1=over limit, 0 is under)

111
0P0
000

In this case one the bottom two rows will be averaged to generate a color value for P.

Feel free to steal it for Corona! :)