I agree about the offsets--clarified that language.

For the helper functions, what do you think about something like this (which would avoid the business of passing around the output of each intermediate step with a variable like rejected_pixels, which is what I'm currently doing):

combiner = ccdproc.Combiner()
combiner.threshold_reject(ccddata1, ccddata2, ccddata3,...) # options for setting min/max
                                                                # omitted for brevity
combiner.clip(ccddata1, ccddata2, ccddata3,...)   #options omitted for brevity
combiner.calc_weights(cddata1, ccddata2, ccddata3,...)
combined_image = combiner.combine(cddata1, ccddata2, ccddata3,...)

Actually, if we went down this road, then the images could be passed into the initializer for the Combiner and then no need to keep specifying which images need to be combined:

combiner = ccdproc.Combiner(ccddata1, ccddata2, ccddata3)
combiner.threshold_reject(...) # ONLY options for setting min/max
                                                # go in here because the combiner instance
                                                # knows what images will be combined.
combiner.clip(...)   # again, only clip options need to be specified.
combiner.calc_weights(...)  # and scale/offset options..
combined_image = combiner.combine(...)

Although, using masked arrays may help with np.median or np.mean, we may want to run our own tasks there to properly account for bad pixels, weighting, and propagation of errors. So for right now, I'm make method just a string 'median' or 'mean' that calls something like combine_median.

My thought had been to construct, behind the scenes, a "weighting array" for each image that incorporates the original pixel mask, the threshold rejection, clipping and incorporate scaling/zero offsets/weights, then multiply each image by its "weighting array", then feed the result into median or mean to avoid re-implementing those.

Either way, I suppose the option can accept a string and how we handle that behind the scenes doesn't really matter.

+1 to the combiner class and that the list of images is initialized on input. that looks good to me.

I've put in the combiner class and rebased on the most recent version of the logging branch to ease merging--let me know if you see further improvements to be made here, though!