I don't mind compressing the video like DSR. Besides, I don't need hours of footage. Even a few minutes will be a good enough to test to see if I want to play at higher resolution or fps. For instance, I think in a blind test, I can tell the difference upto 720 fps. (On a youtube video) Airforce pilots are known to be able to do this for up 600. (read in a random forum, source not available)
Here is another methodology of doing the same thing. Instead of saving pixel data, the game could save the command you input and the way AI interacted with you. After you are done playing the game, the game engine would then slowly render and recreate that scenario in a much higher quality. I'm thinking, the bottleneck happens because the burden isn't only generating that amount of graphics, but it is generating it immediately. So by allowing the computer to take all the time it needs, it should in theory generate a better quality.