Oh I see now, the difference is just in scaling, which is arbitrary. So these differences you see shouldn't cause any downstream issues, and they are intentional. I just forgot the origin of this.

We have this following code in estimateSizeFactorsForMatrix. It is not strictly necessary (the scaling of the size factors just changes the intercept), but ensures the size factors have geometric mean of 1. In order to not have any changes with respect to the original DESeq and its size factors, the default size factor correction in DESeq2 is identical to the original 2010 paper's implementation and the geometric mean is not equal to 1.

  if (incomingGeoMeans) {
    # stabilize size factors to have geometric mean of 1
    sf <- sf/exp(mean(log(sf)))
  }