[Bf-cycles] Monte Carlo variance denoising

Sergey Sharybin sergey.vfx at gmail.com
Mon Feb 15 20:34:02 CET 2016


There are surely different techniques how to reduce noise of Monte-Carlo
renderers, but there are number of key things we should keep in mind:

- Denoiser is not a magic which makes your noisy image look totally
awesome. Denoiser just allows to get rid of the last bit of hi-frequency
noise in the final render, without requiring to exponentially increase
number of samples.

Surely there are papers which claims to increase quality of the image after
32spp, but those techniques wouldn't be applicable for animation rendering
due to annoying low-frequency noise.

- Some papers claims to work on just a final frame. While they do indeed
provide reasonable results in certain cases, you don't have all the
information you might have from renderer. For example, it's really hard to
distinguish hair from high variance noise on the final image. Having
normal, depth and other passes will help _a lot_ here.

In a production shots you also don't have that much noise in direct light,
and will likely only want to denoise indirect light. That's where extra
information from passes would help as well.

- While you can have denoiser as a complete post-pro tool, it really makes
sense to combine it together with adaptive sampling. They are really quite
the same and doing adaptive sampling would help you make amount of noise to
be get rid of lower.

Hope this makes it more clear what direction of the subject we're
researching at this moment.

In fact, both me and Lukas Stockner were playing around with different
techniques last summer. Hopefully we can make something decent in an
upcoming months. Or at least prepare some patches so people can experiment
which approach works better.

On Fri, Feb 12, 2016 at 9:57 AM, Jean-Marie Aubry <jmma.aubry at gmail.com>

> Hi there,
> This denoising technique is most likely based on the papers by Rousselle
> et al., for instance *Robust Denoising using Feature and Color
> Information* (Pacific Graphics 2013). It works quite well (I have work on
> an implementation that is not Innobright's). It doesn't have to be tied to
> any specific renderer, just needs the beauty image do be denoised, some
> feature aovs and an estimation of their variance. So, imho it could be made
> into a node independent from the renderer.
> JM
> On Thu, 11 Feb 2016 at 18:21 x3108 at chello.at <x3108 at chello.at> wrote:
>> Hi,
>> i stumbled across IMO a great technique to denoise trough MC variance. It
>> is pretty simple. Its from Innobright.
>> Now i don’t know the legalities, but it looks pretty simple, some AOVs
>> and a MC variance filter. Could be implemented in GPL ?
>> Also would this be only a node necessarily (still cool) or could this be
>> a renderer feature (more elegant) getting it straight of the renderer ?
>> Lets say, shooting 2x 128samples and variance filter them in GPU RAM
>> without saving 2 images. Like i see it, it is pretty powerful.
>> Your thoughts on this guys?
>> Many thanks,
>> enilnacs
>> _______________________________________________
>> Bf-cycles mailing list
>> Bf-cycles at blender.org
>> http://lists.blender.org/mailman/listinfo/bf-cycles
> --
> Jean-Marie Aubry
> Please consider encrypting your emails using GPG (http://gnupg.org/)
> My current public key can be found here:
> http://pgp.mit.edu/pks/lookup?op=get&search=0xBA1DD06C03E0678C
> _______________________________________________
> Bf-cycles mailing list
> Bf-cycles at blender.org
> http://lists.blender.org/mailman/listinfo/bf-cycles

With best regards, Sergey Sharybin
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.blender.org/pipermail/bf-cycles/attachments/20160215/3ee3f4a8/attachment.htm 

More information about the Bf-cycles mailing list