Actually, I cannot say too much about the algorithm right now, mainly because :
1. This is still highly experimental, not finalized, and subject to changes.
2. We are in the process of writing a scientific publication about it and we surely don't want to inspire other people (faster than us:) ) too much.
This is work still in progress. When our paper will be published somewhere I will tell more, of course, probably with a detailed blog post.
I'm sorry about that, but writing scientific papers is one of my main activity (I'm evaluated on that particular aspect of my work), so I can't let other people re-use what I've done and publish the same idea quicker than me.
What I can tell is that the algorithm is far from being trivial (that's why it takes quite a lot of time to render), and it already competes with state-of-the-art automatic color transfer techniques published by others (except for single colored layer of course

).