I uncompressed the original file before starting, then did something similar to the following:

fpack -h -s 1 -S dark > dark.fits.1

I did this for a variety of different scales ( 1, 2, 4, 5, 6, 8, 10). The scales in the images are the following:

1	2	4
5	6	8

I chose a number of different images to test- flat, dark, qtfocus, video focus, stare, moonglint, and an md01. The md01 is just to see what happens to data we care about, in case we find out data we care about happens to be misclassified for some reason (PR images? other unknown things?) The stare I chose happened to be not too interesting, but I included it anyways (I randomly picked files).

example	rice	H (no loss)	H scale 1	2	4	5	6	8	10
dark	18449280	17596800	11831040	8838720	5797440	4953600	4193280	2862720	2113920
flat	34372800	33978240	11868480	8784000	5935680	5094720	4302720	3041280	2335680
md01	27648000	26349120	12124800	9069120	6102720	5319360	4475520	3176640	2364480
moonglint	18665280	17740800	11329920	8377920	5483520	4694400	3827520	2517120	1794240
qtfocus	23186880	22219200	11891520	8746560	5757120	4968000	4129920	2839680	2076480
stare	17498880	16525440	10618560	7467840	4861440	3830400	2934720	1877760	1434240
videodonuts	23811840	23160960	11966400	8939520	5996160	5270400	4420800	3127680	2364480

dark flat md01 moonglint qtfocus stare videodonuts