Hmm I skimmed the paper on the algorithm this morning and didn't get the impression they trained the model on other images. I thought they jointly estimated patches that make up the image and penalized deviations from these patches (i.e. estimated a sparse basis). I haven't watched the Ted talk yet though