I have a 2D histogram in root which I need to fit a KDE to, which I will in future sample from. I need to keep the data binned as the number of entries is high and a KDE process looking at every point would take too long.
This is correct, for multi-dimension the only implementation that we have currently is the one in RooFit but it is not very efficient.
As mentioned before, there has been some development for a new KDE class for multi-dimension. This work has not been merged yet in ROOT, but a development version is available in GitHub at this link,
You are welcomed to try it and give us eventually feedback.