Pixar/Disney denoising by deep learning paper, is this already known by the dev team?

Hi there.

I just saw this in cgpress.org

And here is the link to the paper:


Is this known already or is this useful?
Also there was someone here trying to develop some kind of deep learning app for denoising, this could help too :slight_smile:


Also, I may be wrong, but aren´t the architecture pictures in the paper down with blender? again… I may be completely wrong, just asking here…

Deep learning seems the way to go for the future in regards of denoising. Looking forward to what the future brings in that area.

And you spotted “blender” for some reason:

We alsothank the following Blendswap artists for creating the scenes inboth Fig. 7 and the training set: Jay-Artist, Mareck, MrChimp2313,nacimus, NovaZeeke, SlykDrako, thecali, and Wig42

I saw Nvidia show demos of deep learning AI denoising probably using their Iray? render. It was doing an incredible job, although it was video and not high res pictures so i could not compare the accuracy of the end result.
But yea machine learning AI is without a doubt going to play a big role in renders in the near future i think!.

Here you go. https://www.youtube.com/watch?v=yPJaWvxnYrg

Not that long ago a guy Razorblade here in the forum was training a neural network on denoising blender images.
He did show some results here, apparently as a developer he is able to create neural networks, and train it with images.
But the results where not yet of a high quality standard, though he seamed convinced it could work.

He gave frequent updates about his progress, and as for blender development, he went pretty fast, in terms of a few weeks he had something working. After showing the first results (who where not yet high quality, but as for sure something neural was going on there) he told he took a rest for a week on the project. Several weeks have passed since, and i haven’t seen him around on the forum here. He seamed to be pretty active on the forum before, and he was very active on that project, So I hope he’s in good health.

I lost the link to his thread, in which he asked for help for some sample images to train his neural network with.
If you look around or search on neural network i’m sure you find it.

That’s the mentioned thread:

Ive been looking at many different noise removal systems, Im building a test branch with a few different versions.

A machine learning approach to adaptive sampling for monte carlo rendering, This looks very interesting. It runs on PBRT so update to Cycles shouldn’t be that hard for a dev. Not tried this yet as still working on my SURE based error adaptive sampler first.

But if your interested here’s the code for PBRT, you could always just download and compile PBRT with it to see if the results you get are good enough for Cycles integration.