Now that it's becoming clear what's doable, what do we pretend the 20/20 goal was all along?

1. Get rid of the grid (out of the box?) and the same-size restrictions on the tile images.
2. Subtler semantic meaning -- if any -- just steal textures, hint of subject, sometimes background -- not: we're tiling this Stat of Lib with these AMERICANS.
3. Often -- in existing image-tiling works -- it works out that the foreground of the tile carries the semantic information and its BG the useful HSV contribution. That's a somewhat unnatural separation it would be nice to avoid. One solution is to lighten the semantic load of the foreground. Another is to use less-decipherable tiles (ideally, but not this semester, only parts of the images).
4. If one treats the wavelet-transforms of images as the found objects, supposing e.g. that they came stored that way, they come with a different set of primitive image operations. In the frequency domain, rescaling a potential tile image is a task on the order of just selecting a region of pixels in the space domain. That images can be stored in a manner that caters to multiresolution techniques, that people use words like "pyramid," makes you immediately want to hang little pyramids on the nodes of big pyramids like tinsel on trees.