In HINT, only a few pixels are transmitted in the first level. Then, when the next level is being transmitted, pixels in four directions (N,E,S,W or NE, SE, SW, NW) can be used to predict the current pixel, so a better prediction is possible. Also the content of the whole image becomes apparent when only a fraction of the data has been transmitted, thus allowing the user to abort (either because the image is not of interest or because a low-quality image is sufficient).
My project was to improve on this. One of these days, I'll get around to putting the abstract of my thesis on-line...