Plus I don't see how having a frame inserted that has no CPU generated game data attached to it cannot be perceived as lag if it occurs during a button press. I don't think this technology will be particularly desirable in all cases. Especially if it interprets the frame incorrectly for some reason.
The following is my educated guess about how latency is impacted:
My first assumption is that every other frame is virtual. Eg. it is not possible to have two base frames in a row, or two virtual frames in a row. If this assumption is not correct, then the manner in which virtual frames are inserted would determine the impact to latency.
My second assumption is that the mechanism of adding a virtual frame in and of itself does not increase latency since, unless Nvidia engineers are lying on twitter, virtual frames are not interpolated but are rather forward projected.
Virtual frames do not help with (input) latency of course. They are totally disconnected from the game update loop. If every other frame is virtual, the engine is effectively running at half rate.
Based on the fact that frame insertion does not yield a 100% speedup, it stands to reason that, at least if my first assumption is correct, there is a performance hit to "base framerate," which would in turn be the main driver of added latency.
My question is what happens when you have a capped frame rate. If I run a game with a 60 FPS cap and have enough power to hit 60 FPS without frame insertion, will DLSS 3 insert new frames anyway? In other words, am I stuck with a "base framerate" of only 30 FPS? On a laptop or power constrained environment, this might be preferable behavior, but not on a desktop.
Ideally DLSS 3 would have a "prefer latency" setting and a "prefer efficiency" setting for how to behave when running up against a framerate cap.
All in all, I wouldn't be surprised if DLSS 3 ended up a lot like DLSS 1 in being a great idea that just needed more time in the oven. Of course it's possible DLSS 3 might be great from the get-go, but if it isn't, I'm sure Nvidia will have it figured out for DLSS 4. Virtual frame insertion is the future, obviously, and in time all vendors will have their own solutions.