The scan approach seems very simple. Pick a size, find some good matches, move on to the next size, repeat.
However, there are many questions involved with the process.

  1. Should we allow one source tile to be used multiple times?
  2. How many tiles should be picked at each size?
  3. How many tile sizes should be used?
    4. and
  4. What the heck does a good final image look like?
Obviously, the answers to questions 1 through 3 depend on the answer to question 4, and vice versa. All of these revolve around a balance between
meaningful and interesting tile images and a good representation of the target image

The following links show some experiments in each of these domains:

Important variables: Less important:

Sections

  1. Intro
  2. The Approach
  3. First Method Results
  4. Second Method Results
  5. Backing away to 20 meters...
  6. Meta Results