For a number of projects where very deep integrations and accurate flat fields are required, and the targets are point sources or much smaller than the frame size, the sky background itself can be used to calculate the flat field. When an object is dithered on the chip, so that it exposes pixels at different positions in different integrations, the resulting 'stack' of images may be combined to produce a very accurate flat field. Since a reference image, such as the ``LAMP OFF'' frame, is not subtracted, a bias frame of the same integration time must first be subtracted from all these images. They should then all be median combined, with suitable scaling introduced to match the medians of each image to account for the variability of the sky background level. The resulting image is then normalised to obtain the flat field.
To subtract the bias from the science images, use COMPUTE/IMAGE. Then use COMBINE/CCD to median all the frames together with multiplicative or additive scaling to match individual frame medians (you should examine the results of both options to determine which is working best for your data). Use STATISTICS/IMAGE to calculate the image median, and COMPUTE/IMAGE to divide by this to get the normalised flat field.