Harmy, when you set a curve in the 'R/G_diff' layer turn it on temorarily to check the effect. Set a curve which looks like an upside-down 'U' cut in half. That is, hug the left of the histogram to about one-quarter of the way up then gradually roll it off to finish at the midpoint of the top line in the grid.
This curve will then make sure changes only take effect in areas of extreme difference (ie, clipped areas).