Some os and PIL scriptbashing to produce averaged images for each of some 2500 categories that have semantic analogues in Word2Vec (and subsequently shaped into a grid) A typical composite image is made up of a thousand or so images within a category. See the result here.