Best-fit texture synthesis can also be performed using an L1 or L2 norm. Using a different norm produces different results, and the difference appear analogous to the differences between autistic and normal people.
That's two big idea chunks, and i've put them together. Autism. Best-fit synthesis. Thunk. Is that creative, or a simple juxtaposition?
Here's an example. Shown to the right of the results are the boundaries of patches copied from the input. (Best-fit texture synthesis sometimes results in copying even though this is not explicit in the algorithm. I suspect this is similar to human memory recall.)
Input:
Result for L2 (normal):
Result for L1 (autistic):
As you can see, the L1 version has copied larger chunks. The L2 version has done this cool synthesis thing where a whole lot of little tiny bits form a coherent whole.
Would a normal person consider the use of such large chunks as creative? The chunks themselves are not creative, being simple feats of eidetic recall. Given that there are less chunks, the L1 result must be less creative. Real creativity should be innovative in every aspect! L1 ain't nothing but a crap-hound.
Funny thing is, despite the large chunks the L1 result took significantly longer to compute than the L2 result. There are helluvalotta ways to juxtapose even two fragments of input, and L1 can't rule out possibilities as quickly (a single outlier will not immediately rule out a particular possibility as in L2). The L2 result uses an impressive array of fragments, but it actually contains less structure, it's just this big mass of undifferentiated leafness. It might have an impressive bibliography, but it doesn't go anywhere interesting.
Addendum 19 September 04: Jiří has run my best-fit code with L0.5. (It took two days to calculate!) As might be expected, there is even more copying and the result is very disjointed. A person with a norm this low would probably be mentally retarded.
