Watching the reflex video annoyed me cause they invented some new measure called PC latency and somehow claim there's only 1ms of latency now. I'm assuming that means latency added by the engine?
2 frames are needed to calculate the data to feed the Optical Flow. At least according to their Frame Generation press release 2 years ago. Considering your image and how DLSS3 was described in the same presentation, I'd say it is pretty much interpolation. It wouldn't make much sense otherwise.
You could use frame N and frame N-1 to compute optical flow, and use that to forward predict to frame N+1. Definitely not as straightforward (eg disocclusion artifacts) but 4 frames of interpolation seems a bit silly
Comments
Watching the reflex video annoyed me cause they invented some new measure called PC latency and somehow claim there's only 1ms of latency now. I'm assuming that means latency added by the engine?
I wouldn't put deceptive marketing beyond them.