Kottke has a neat post about dancing twigs.
It strikes me you could generalize this into a method for generating animation using any type of found object.
- Generate an animation using standard methodology. For example, you might use a stick figure animator to create a sequence of dancing image.
- Search within a body of found images to find the one most similar to the generated image.
- Replace the generated image with the found image.
- Assemble the found images into a movie.
I’ll (lazily) skip steps 1 and 2, because they are hard and you can easily visualize the output, but step 3 might lead to this sequence of twigs:
And step 4 would lead to this movie:
As an extension which allows for filling in gaps where there was no suitable found image, you could train a GAN on the pairs. The GAN would be rewarded for generating the found image from the source image.