Octane4 and AI adaptive sampling Tweet

[calus]

Octane4 is an amazing improvement and the roadmap looks like most of the major production issues will be adressed in a clever way.

But what About "AI adaptive sampling" to replace our clunky adaptive sampling system ?
or maybe the current system can be feed with a noise map based on the difference between the beauty and the denoise pass ?

There is a talk about that at GTC 2018 :
https://2018gputechconf.smarteventsclou ... tleSort&p=

S8788 - Adaptive Ray Tracing Rendering Powered by Deep Learning
Session Speakers

Andrew Tao - Distinguished Engineer, Director for Deep Learning Applied research, NVIDIA
Carsten Waechter - Ray Tracing Software Architect, NVIDIA

Session Description

This session will present a proof of concept where a deep neural network was trained with pairs of Iray ray traced images (one arbitrary ray tracing iteration number and one fully converged image) and theirs structural similarity index (SSIM). Originally thought as a method for measuring the similarity between two images, SSIM index can also be viewed as a quality measure versus a reference image or, in our case, as a ray tracing rendering progress. The DNN can now from any render iteration of arbitrary scene infer a rendering progress estimator but also provides heat map pictures of the scenes that can be used for adaptive rendering, focusing ray tracing engine power on appropriate zones.

Or is it already the same technique used in AI lights ?

[milanm]

Bumping this thread with a recording of the presentation, and a big +1 for AI driven Adaptive sampling (and AI lights too).

http://on-demand.gputechconf.com/gtc/2018/video/S8788/

There is an inherent flaw with Ai denoiser and that is the fact that it's trained on final renders from the past. That means that there will always be a scenario where a user who bought Octane because of the Ai denoiser might get disappointed later on in the middle of the real world project. AI denoiser will always lag behind the current trends/styles/looks.

Also, as opposed to adaptive sampling, importance sampling, coherence etc., denoising as a post process doesn't improve Octane as a render engine in any way.

More intelligent sampling might be the 'holy grail' for a brute force pathtracer such as Octane Render.

As artists, we know what works and what doesn't. I know that I'll get a ton of 'fireflies' if I put the sun in front of the camera and I have reflective surfaces in the frame. I know that area light in the fog will have the most noise right next to the surface of the area light. If Octane was somehow aware of those cases, that would be a significant improvement over the current adaptive sampling method. And then AI denoiser could be used to get the image from 95% to 100% because that's a scenario where it actually performs the best.

Regards
Milan