Given that this will probably only produce a relatively low res, high SNR image, it might be the sweet spot is to use this tech to capture a true eye-contact perspective from the glass in front of a screen, then neural net combine it with images from offset cameras, to create a 'true' through-the-screen view.

Having the screen be the camera 1) removes the need for the holepunch as well as magic island workaround 2) could clearly be used in tandem to improve the effect, even potentially reducing the required computation. Sure, Google also fixes images in post, but that's still not a substitute for better cameras. You're always information limited.

Are you sure? I remember that being present briefly in a beta several years ago, but then it was removed. I don't remember it ever coming back.

It’s a bit uncanny valley though, right?

1986 solution vs cyberpunk solution

how did we even get here?

I believe Nvidia offers something similar with Nvidia Broadcast[0]. [0] https://www.nvidia.com/en-us/geforce/broadcasting/broadcast-...

This doesn't solve the issue, because it looks like the person is always making eye contact! You don't want that—you want it to look like the person is making eye contact only when they are actually making eye contact. Constant eye contact is exhausting.

Which versions of iphones do this?

It does what??? How have I never heard about this before? Is this deployed right now? Is it always enabled? I have so many questions, don't even know where to start. That sounds amazing.