We investigate the interesting problem of finding representative images for a word. Such a technique makes the semantics of a word clear fully at the first glance. The challenges are twofold: 1) a word can be ambiguous, e.g. “apple” can ibe a fruit or a logo, so how to automatically discover the different meanings of a word, and 2) there is the semantic gap, so how to find the representative images among the many candidates?