Octane AI denoiser - initial tests, results and next steps

[Goldorak]

can't wait for its release, hopefully it can be integrated well into plugins? will it work with networked GPU's?

We plan to just make it part of Octane itself (both local/cloud) once it's done. It will work through the Octane imager module system, which will include 1st party modules like this one shipping in all OR plug-ins.

[pegot]

Is that a different denoiser than what we currently have or an enhancement to the existing system?

[Goldorak]

Is that a different denoiser than what we currently have or an enhancement to the existing system?

Totally novel. Used machine learning.

[calus]

We plan to just make it part of Octane itself (both local/cloud) once it's done. It will work through the Octane imager module system, which will include 1st party modules like this one shipping in all OR plug-ins.

Did Octane devs also investigated on the AI driven Adaptive sampling ?
Post-process denoising is not good for all production.
Can't the AI denoising feature also be used to produce a better noise map to drive adaptive sampling (instead of post process denoising) ?
It could bring to Octane better adaptive sampling a lot more efficient and less user input dependent...

Maybe the ultimate solution is to combine both AI driven Adaptive sampling and AI de-noising.
It seems to be the way Nvidia's Ai denoiser was implemented in IRAY:
"This isn’t just for interactive rendering at artist workstations, as a fully converged noiseless final frame renders is about four times faster; if a final render took an hour before, with the new AI it will take approximately 15 minutes."
https://www.fxphd.com/blog/whats-the-st ... w-iray-ai/

[Goldorak]

We plan to just make it part of Octane itself (both local/cloud) once it's done. It will work through the Octane imager module system, which will include 1st party modules like this one shipping in all OR plug-ins.

Did Octane devs also investigated on the AI driven Adaptive sampling ?
Post-process denoising is not good for all production.
Can't the AI denoising feature also be used to produce a better noise map to drive adaptive sampling (instead of post process denoising) ?
It could bring to Octane better adaptive sampling a lot more efficient and less user input dependent...

Maybe the ultimate solution is to combine both AI driven Adaptive sampling and AI de-noising.
It seems to be the way Nvidia's Ai denoiser was implemented in IRAY:
"This isn’t just for interactive rendering at artist workstations, as a fully converged noiseless final frame renders is about four times faster; if a final render took an hour before, with the new AI it will take approximately 15 minutes."
https://www.fxphd.com/blog/whats-the-st ... w-iray-ai/

We're choosing to start with AI denoising in this way first for several reasons, but it is all possible on top of this and we're looking into all of it.

[swipswap]

Will this have an increased rendering speed on motion blurred areas or areas with heavy light scattering?

[Goldorak]

Will this have an increased rendering speed on motion blurred areas or areas with heavy light scattering?

We are training it on DOF, MB and more, but it's very hard to measure real world success until we release a test build and see what user feedback tells us in aggregate (tuning will continue for many months after that I would expect).

So far it's doing well on random scenes we throw at it (10 - 50 samples) which are not in way related to the training samples. This is a critical test that has been passed. We're probably over the 50% mark towards a preliminary experimental release which users can help validate further.

[ivankio]

How is progress on the denoiser? Any estimates?

[ivankio]

I guess I can only hope your are not planning to push denoiser to 4.0.

[Goldorak]

I guess I can only hope your are not planning to push denoiser to 4.0.

Unless something unexpected happens during the remaining integration work, we're planning to have the initial test release of the AI denoiser in 3.09.

As a caveat, that first release should be considered experimental tech. Getting it out to everyone by 3.09 (basically as soon as it's embedded in Octane), means we will not have had much time to do exhaustive testing prior to sharing it with the community for deeper QA.

The goal of getting this feature to users as soon as possible in 3.09 is to gather feedback on any corner cases ASAP (e.g. where denoising is too destructive), so we can further tune and improve the system's performance subsequent updates.

While the AI denoiser can do pretty awesome work, as in the tests and scenes we've shown, we may run into situations in real world usage where the AI isn't denoising correctly or usefully, or worse, goes Skynet on us once it discovers humanity is the root source of all unpredictable scene noise