Interesting Idea... "VRay AI Scene Optimizer"

[joelly3d]

This is a big request, but is also an interesting idea.
It crossed my mind that since all parameters are known and certain optimizations are common... it should theoretically be possible to use trained AI to automatically optimize a scene. Image based training is already being widely used, why not use it for this?

The way I was thinking about an approach would be as follows...

1. Create your scene.
2. Get it to render how you want it to look... even if very slow.
3. Then feed the VRay scene and the target output render into a new "VRay AI Scene Optimizer"
4. The VRAISO would then tweak parameters for materials, lights, GI, Sampler, DMC, and other scene settings to find and discard what is wasteful as far as landing at the target result faster.
   1. I realize this may take hours of trial and error computations and render tests to create comparisons to base tweaks off of.... but if you let it go and go to sleep while it works, who cares right?
5. Then it could prompt you a readout of what it will tweak, so you can see what it is doing, and then prompt you to automatically apply the tweaks.

Something like this would be a good use of machine power, and also would be a great way to teach you efficiency.

Correlation coefficients for the win!

It's a different idea, so I wanted to throw it out there. What do you all think of something like this?

[joconnell]

As you say image based training is being used on final results, the under the hood light cache prepass is doing what you describe and there's apparently more things in the works to cut down the quality of soft / undetailed areas!

[fixeighted]

imo it's always a 'wait and see' situation. Most of the situational issues we have been fighting with over decades have been mainly addressed and the remainder will eventually disappear, as the software/hardware solutions become more focused.
As an example, I recall saying to friends, in the early 90's, that humans would be able to be rendered realistically; they all laughed at me.
These things you mention will 'eventually' be sorted,; I have zero doubt about that

[glorybound]

We've come a long way baby! A couple of short years ago I was spending 70% of my time-fighting V-Ray. Now, today, I don't even touch it. I am not sure how much better it can get.

[joconnell]

One way of looking at it was the previous idea of should you waste human time trying to get a few more percent off your render times? Are you putting in more hours tuning render settings than you're going to gain back? If you're someone like bobby where you're doing fast turnaround and putting out a scene in a day, there's no point in running ai to try and get a 20 minute speed up on three hour long still image renders. If you're doing long sequences which are similar - lets say vfx or animation gigs, then it might be worth your while since the gains ripple over multiple shots.

[glorybound]

Yeah, diminishing rate of return. I rarely do scenes that take more than an hour to render and that's all the bells and whistles at 4K.