TopSpoiler
Regular
Yes, it's a starting point, not a turnkey solution. Apparently, they did some research into this area and then, for whatever reason, abandoned it.
Also AMD did not take this approach for FSR. I guess it has good reasons, not going that way for the hardware they are developing it for. Who know, maybe coming down from 60 to 16 ms was already much of the optimization possible when running this without specizalized hardware? Maybe we'll be seeing a revival of this approach with DirectML, once more hardware has the TFlops to spare.
I don't think single image based ML reconstruction would considered again because,
- it needs time and energy consuming training per game.
- result are not as accurate or detailed as temporal reconstruction.