Repetition blindness (RB) is the failure to see or recall the second of two visually similar or identical items in rapid serial visual presentation. It was initially demonstrated by Kanwisher (1987), who proposed that a second token of a given word or object type cannot be established when the two items occur close in time. Bavelier and Potter (1992) showed that RB also occurs between visually different items that are phonologically similar. They proposed that RB may occur not only when the targets are physically similar, but also when they have to be registered or encoded in short-term memory (STM) along dimensions on which they are similar. This hypothesis predicts that RB between visually different items should not be restricted to words, but should occur with any stimuli, as long as the task requires these stimuli to be encoded along dimensions on which they are similar. Moreover, it also implies that a task that changes the preferred code of targets will affect the size of RB. The first prediction was confirmed by establishing RB between phonologically similar pictures and words, whether semantically related (the picture of a cat and the word "cat") or not (the picture of a sun and the word "son"), when using a task that requires phonological encoding (Experiments 1 and 2). The second prediction was also supported: the magnitude of RB depended on whether the task required similar or different codes for pictures and words (Experiments 3 and 4). These experiments confirm that RB between visually different items is due to the similarity of the codes initially used in STM. The results suggest that RB can occur at any step during the instantiation of a token, arising not only from a failure to create a new token, but also from a failure to stabilize an opened token. In this view, tokens are to be seen as dynamical entities, built over time as a function of type activation and task requirements, and varying in stability as a function of the information that is entered into them.