A "Google Deep Dream" style filter?

Something similar could probably be made. Obviously not using the entire google image database, but something that creates an image fr om two (or more images) combining them based on similarities in pattern and luminosity.

I'm thinking that the simple case would be using two source images, call them image A and B, wh ere A is the overall image that is being matched and B is the image that is being ripped into pieces and rearranged to match A as well as possible.

Image B would be put through the Bomber at different scales and a cumulative Difference calculation would be made, picking the pieces that are closest to image A in any area. This could be looped for offsets, or even distortions. This would likely need another channels being used for keeping track of what the current closest match is. Fortunately, there's the alpha channel that can be utilized.

At some point you'll hit the barrier of how many loops and Bombers you can put an image through, but maybe you'll have an image of some kind built up by that time.