I prefer to take darks separately but there are arguments both ways.
If you take a series of darks and median combine then the you will reduce the readout noise by a formula that I can't remember,
with more being better up to a point.
This leaves you with what is hopefully the consistent dark current for your particular sensor with the random readout noise reduced. This is assuming that the temperature when you take the darks is the same as when you take your lights. There are ways around this however using scaling of darks etc.
On a practical note it is also probably more time efficient to take separate darks. I tend to use 300s exposures and take about 7 darks. This will take 35mins and can be done after I have finished imaging and packing up or gone to bed. If I used ICNR and took 24 exposures in the session then my 2 hours of imaging would have expanded to 4 hours with 24 darks being taken along the way but not median combined. I also don't have access to these dark files as they don't exist as the camera internals have used and then deleted them.