Image-to-image translation aims to learn the mapping between two visual domains. There are two main challenges for many applications: 1) the lack of aligned training pairs and 2) multiple possible outputs from a single input image. In this work, we present a disentangled representation approach for producing diverse outputs without paired training images. To achieve diversity, we propose to map images onto two spaces: 1) a common content space capturing shared information across domains and 2) a domain-specific attribute space. We train a generator that takes encoded content features extracted from given inputs and attribute vectors randomly sampled from the attribute space to produce diverse outputs at test time. To handle unpaired training data, we propose a novel cross-cycle consistency based on disentangled representations. Qualitative results show that our model can generate diverse and realistic images on several tasks without paired data. For quantitative comparisons, we measure realism with user study and diversity with a perceptual distance metric. We apply the proposed model to domain adaptation and show competitive performance when compared to the state-of-the-art on the MNIST-M and the LineMod datasets.