An amalgram mosaic is similar to a photomosaic, in that a number of unrelated source images are used to match a target image, using an RGB errror-minimization function. Unlike photomosaics, the image isn't broken up into little tiles. Instead, the images are averaged together and then normalized.
Finding the optimal subset of images, from a larger set, for an amalgam mosaic is a much harder problem than finding the best matches for a photomosaic, and it is very computationally expensive to make good ones.
I like the amorphous quality of the final result, however.