I think we both agree that the mean average of sample n and sample n-1 method is computationally cheap. (n added to n-1 then divided by two) No need for FFT and IFT. Lot's of early sample rate conversion chips worked this way IIRC. You can improve on the results by using more samples, taking averages and weighting those averages to produce an intermediate value. This is all less demanding than using a FFT and IFT.
Hmmm ... obviously I did understand you correctly the first time. Taking the mean average of only 2 samples is a horrible way of upsampling and should be avoided at all costs ... as I said earlier, only the really bad ones use that kind of method.