[BTS] How much time does it take to render these encounters?

While Bindr is faster on updates than most games of this kind, I'd still like to provide some insight on how long encounters take to make. Pretty much the biggest immovable constant that can only be offset by more GPU hardware is render times. As I've reinvested all of the stuff into Bindr, we have 2x4090s (because 5090s cost more than a car at this point). Two are also the maximum that can be pushed into a regular PC box without a threadripper processor and board with a custom cooling solution, which alone is like $10,000+ without the GPUs.

Networked rendering is another solution, but my networking is ass, so not feasible to transmit scene data that crosses 20 GB on the regular.

Depending on scene complexity, in the most optimized case with all the bells and whistles, like guided sampling, setting nominal luminance, overpowering lights slightly to converge faster, it still takes quite a bit of time. Especially on very fancy scenes. Average time for an exterior shot with proper participating media (atmospherics) is around 10 minutes to get noise down to an acceptable level:

Here we have MC's ugly mug holding a fake ass phone. After 12 minutes:

He took a photo of Jen, which will be relevant as well:

That also took plus 10 minutes.

Now, interior shots are easier to manage, especially when well lit, in super ideal situations, a 4K render can take 2 minutes. But those are rare. Usually it's 5-ish minutes. For a time average over an encounter that lasts 110-ish frames, it's 6 minutes per frame, I imagine. 7.5 on the high end.

So the rendering alone takes around 12-13 hours per encounter on average. On top of that, there is at least 2-3x more time taken by scene setup, posing / "animating", clothing, facial posing, camera setup, storyboarding, writing, it all adds up.

Eventually, we'll be able to nab 2x5090s, at this rate within a year. Then render times will halve to around 6-7 hours, which will enable us to use our time better.

Generally, a 4090 is twice the pathtracing performance of a 3090. So those exterior shots would take 20+ minutes on 2x3090s, 40+ minutes on a single 3090 with atmospherics and all the bells and whistles. Allegedly, PT has jumped two fold with the 5090 again (this is good for us), even though in games it's merely an uplift of 30% wrt 4090.

On other GPUs, you'd get an OOM failure due to memory constraints and it would fallback to the CPU. You can then write, and I did back in the day when Bindr started with a 3080 10G, scripts that downsample all the textures, reducing texture memory pressure to 1/4 of what it was.

But that creates seams issues and all kinds of other stuff that needs filtering. Also an obvious reduction in skin quality in close shots etc.

Can you do it without participating media simulation?

Yes, but you don't want me to. Here's an example of GTA V with participating media simulation:

And here it is without:

Historically, in art, after the "discovery" of perspective... The intuitive discovery of the "ethereal blue fog", the blueish hue as you go further away from the canvas/camera, was the second most important in achieving convincing art. Later it was understood it's actually atmosphere scattering sunlight, specifically combination of Rayleigh scattering and Mie scattering (and other smaller contributions).