I experimented more with using cropped versions of the masked image as input. Instead of feeding the full image with the mask as input, I ended up using a cropped version of the image with a width and height both equal to twice the width and height of the mask, centered on the mask. The reason why I ended up with poor results earlier was that my crops were too large. Through experimentation, the results from the new crop dimensions matches the output from GIMP closely. Any further hints for improvements are appreciated.