We present the Infinite-Word Topic Model (IWTM), a non-parametric extension of Latent Dirichlet Allocation (LDA) for modeling collections of images and other non-text documents. Whereas LDA requires that document features be preprocessed into a bag of words representation with fixed vocabulary, IWTM incorporates feature clustering inside the probabilistic model, and treats the vocabulary size as a random quantity to be inferred. By making use of the Hierarchical Dirichlet Process, IWTM defines a topic model over an a priori infinite set of 'words'. We derive a collapsed Gibbs sampler for the model and present competitive results on two image classification tasks.