I noticed that---when converting unsigned integers to (signed) integers---the values of the uint are simply scaled to fit into the positive range of the int array. On the other hand, when converting from int to uint, the *full range* (i.e. negative and positive) values of the int array get scaled to fit into the uint. The result is asymmetric, and it can lead to drift when converting between types; e.g.: #~~~ In [55]: skimage.img_as_ubyte(skimage.img_as_int(np.uint8([0]))) WARNING:dtype_converter:Possible sign loss when converting negative image of type int16 to positive image of type uint8. WARNING:dtype_converter:Possible precision loss when converting from int16 to uint8 Out[55]: array([128], dtype=uint8) #~~~ I think one of two things should happen: 1) [uint -> int]: expand uint to *full* range of int. [int -> uint]: unchanged (squeeze into range of uint) 2) [uint -> int]: unchanged (fill *positive* range of int) [int -> uint]: clip negative values from int Thoughts? -Tony P.S. This change would either have to wait for PR 99<https://github.com/scikits-image/scikits-image/pull/99>, or it could be integrated into that pull request. (I haven't yet looked at what it would take to make this change.)