Are Neural Networks the future of sampling distribution?

[Khazmadu]

I think the DMC sampler is doing a great job at this time.I read a lot about neural networks in the last weeks and found it highly interesting.
Using a Neural Network for deciding how the samples are distributed in the image should be a way to enhance the results even more.
Users could have the option to submit their own sampling datasets to enhance the results of the Network even more.

What do you think about that? Is that something that might be possible in future or is it just too complicated and not worth the effort?

[vlado]

There are a number of instances in which some sort of AI could be useful in rendering, so we'll see...

Best regards,
Vlado

[luma91]

This concept is rather intriguing. I've been very interested in the way neutral networks work, and would be very interested in seeing how they could be employed in optimising renders. I imagine each scene/shot would be very different, so over time as you keep re-rendering Vray could learn how to optimize for that particular shot or scenario and so thus gradually your times reduce little by little. How it could be implemented however is far beyond me, but the idea is definitely something else!

[Junglejuz]

So there's this:
https://blogs.nvidia.com/blog/2017/0...r-ray-tracing/

[Neilg]

So there's this:
https://blogs.nvidia.com/blog/2017/0...r-ray-tracing/

my friend sold his nvidia stock the day before they announced this... solid 25% jump the same day.

[francomanko]

i think im going to need a Neural Network to keep up with this stuff, I'm starting to feel like my Gran trying to use a VHS recorder.

[vlado]

i think im going to need a Neural Network to keep up with this stuff, I'm starting to feel like my Gran trying to use a VHS recorder.

There is a bit of that, isn't there

Best regards,
Vlado

[RotemS]

i think im going to need a Neural Network to keep up with this stuff, I'm starting to feel like my Gran trying to use a VHS recorder.

Luckily you are already equipped with one!

The NVidia blog post is interesting, the results sure do look impressive, but they are really demonstrating a post-process denoiser rather than "Bringing AI to Ray Tracing", aren't they? The network does not affect the actual 3d sampling from what I understood.

[luma91]

Luckily you are already equipped with one!

The NVidia blog post is interesting, the results sure do look impressive, but they are really demonstrating a post-process denoiser rather than "Bringing AI to Ray Tracing", aren't they? The network does not affect the actual 3d sampling from what I understood.

That's it, yes. The denoising was not what I was curious about, although the results are still impressive in their own way.
I'm interested in the concept of using neural networks to replace DMC style sampling.

[RotemS]

Another paper with similar goals

http://cvc.ucsb.edu/graphics/Papers/SIGGRAPH2017_KPCN/

[vlado]

I'm waiting for the nVidia paper too so that I can compare the approaches...

Best regards,
Vlado