best case i would have both. require the author to select a category when posting, and using this category either pull random image out of stock photos in the category or if you want to get cheeky, parse the text and run semantic to get a closer match (does it really matter though? :D ). if user decides to upload his own photo, just use that instead.

easy peesy