Lytro has the founder's PhD thesis on their website, which explains the process and is worth a quick read[1]. It relies on an array of microlenses in the camera to produce an image on the sensor that you won't really recognise as the scene in question, and then they apply a bunch of maths to extract the scene information at any particular DOF or focus point.
You can't apply it to an older photograph that wasn't taken with the array of microlenses.
It's also worth noting that the approach is a tradeoff between planar image resolution and depth resolution. The prototype had a microlens size of 12x12 pixels, which means total image resolution was 144x lower than that of the sensor, in exchange for 12 steps of depth information. There's a reason all the demonstration images on their website are only a couple hundred pixels across and still blurry; it's all the technology can manage.
I suspect that application of this microlens technology and processing to a large-format image (probably with a scanning back) could yield enough information to make it worthwhile. Of course, the huge sensor means natural DOF is even smaller, so you probably have a commensurately greater resolution-loss factor to obtain the same depth information... but I ain't sure of that.
[1] in fact, please read and understand it before making uneducated guesses about what they're doing. This really is a new technique that really does work, but it has tradeoffs.