[sillysaurus3]

This is actually touching on a deep observation I've spent many years thinking about.

A hypothetical perfectly-realistic CG rendering would need to be perfectly realistic to animals as well as humans. Then you "capture" the realistic rendering using a virtual camera, which happens to look real for humans.

It's computationally infeasible, but the strategy seems correct.

(No one has ever achieved a perfectly realistic 100% synthetic video of complex scenes. It's good to keep such mountains on your radar -- someone will conquer it one day, but how? Could it be you?)

[yathern]

> A hypothetical perfectly-realistic CG rendering would need to be perfectly realistic to animals as well as humans.

That's not necessarily a true assumption - unless you mean realistic as in an exact light-propagation simulator, not realistic as in fooling humans. On the nanometer-scale there are things that happen with refraction that would probably be more accurately done if using wavelengths and photon energy as the units, start-to-finish. But at the end of the day, the models we use for rendering are already extremely effective - the uncanny valley arises with how we use them.

> No one has ever achieved a perfectly realistic 100% synthetic video of complex scenes

I get what you're saying, but that's a little bit subjective, like the turing test. Some people see some CGI scenes in movies and are shocked to find out they weren't real. Some people see a real scene and complain about how bad the CGI is! An interesting litmus test is Moff Tarkin. Some people were shocked they found a perfect lookalike, and had no clue it was CG. Some people were apalled at how poorly it was done.

[sillysaurus3]

It's important to qualify the claim:

- 100% CG: no mixing in real life footage.

- video. No still frames.

- reasonably complex scenes. Think nature documentary.

- reasonably long: 30 seconds or more.

- not a degenerate scene: it would be bogus if the camera was fixed, staring at a tree stump.

The overall thrust of these criteria is that no one, anywhere, has achieved fully realistic CG video. And this is easy to demonstrate. If we had that capability, you'd be able to fool people in a double blind test so that they wouldn't be able to pick out real video from fakes any better than random chance. In a scientific setting, a nature video will wipe the floor with our best CG video ten out of ten times.

[yathern]

Thanks for qualifying more. I'm surprised you didn't mention having humans in the video as a point - since that's pretty much what we're best at detecting fakes.

I believe the technology already is at a point for us to do this. The technology isn't the barrier to meet your criteria as much as the attention to detail. For a rendered video scene to look good, the main issue is the animation.

Still image renders are already at the point of fooling humans if enough time went into it. Then the remaining issue is the animation. Animation requires incredible attention to detail - some of which can be improved by a more nuanced understanding of the physics involved (eg water simulation or bird flight).

> no one, anywhere, has achieved fully realistic CG video

According to your specific criteria for realistic CG video.

But I still would argue the technology is already there. They just haven't made your video yet because realistic CG is done in movies. And movies aren't prioritizing meeting that specific criteria. But the talent and technology is there.

What about scenes like this from the Jungle Book?

https://www.youtube.com/watch?v=e6yKWQ24rF8

Of course it breaks your rubric for having the boy be real. But get rid of the boy, and less stylized setting (high saturation and fog and smooth tracking camera), people would absolutely be fooled in at least some of the scenes in that clip.

> In a scientific setting, a nature video will wipe the floor with our best CG video ten out of ten times.

Okay

[mbroncano]

Spectral rendering [1] could possibly fulfill that role, given a proper mapping to the specific color receptors of those animals.

[1] https://en.m.wikipedia.org/wiki/Spectral_rendering