The implemented algorithm is the following: for each pixel on the detector, a curve is plotted of the median plane value against the individual pixel value in this plane. This curve shows the pixel response from which a robust linear regression provides a pixel gain. The image showing all pixel gains (i.e. the flat-field) is normalized to have an average value of 1.
By-products of this routine are: a map of the zero-intercepts and an error map of the fit. Both can be used to visually check the validity of the fit. A bad pixel map can also be produced by declaring all pixels whose value after normalization is below 0.5 or above 2.0 as bad.
The eclipse recipe is called twflat. Check out the manual page (isaacp man twflat) for more details about how to use the command.