I would think that would be horrifically expensive and add significant input lag.
I disagree, but even if it were expensive and laggy, it would be
just like the other kind of supersampling, no? I mean, were talking about a program that renders everything at insane resolutions just to get rid of aliasing, its the most brute force method there is, and its not exactly cheap. (But Half Life 2, for example, is over a decade old at this point and we could afford it.)
The difference would be that with temporal supersampling your still getting the latest rendered frame up on screen, just as if you were rendering normally. So since the last frame takes very little time to render, it would be much more up to date and closer to what is actually happening at the moment than a huge 4k+ downsampled rendering. Therefore it seems to me like it would actually be far less input lag. At worst it would be approximately as laggy as triple buffered vsync.
After all, were simply blending in all the frames that your GPU already rendered and would simply throw away with v-sync enabled. These are the same frames that end up being partially shown to you when you're seeing tearing. You
already pay for rendering these unseen frames in every game that has uncapped framerates, and all that temporal supersampling would add to this cost is the price of blending the frames together. So, of course, there is a bit of performance hit and it may likely take some more memory.
However, an additional benefit would be that since everything is motion blurred, you wont see jaggies or shader aliasing when in motion. Additionally, if everything is sitting still on screen, you could do what Halo Reach and Crysis 2 did; have a half pixel jitter between frames that acts like spacial supersampling.
Technically, this could give you all of the sampling benefits that GeDoSaTo currently gives, but without the necesary high minimum performance hit of rendering at high resolutions. It could be set up to only work if your getting above a certain number of frames, and when you dip below that thresh hold, it would simply skip blending inbetween frames and go back to rendering the game as it would normally. The problem being that you would probably need significantly higher than 60 fps (or whatever your refresh rate happens to be) to get enough samples for motion blur. 2x samples would probably not look very good.