Learning competitive behaviors in multi-agent settings such as racing\nrequires long-term reasoning about potential adversarial interactions. This\npaper presents Deep Latent Competition (DLC), a novel reinforcement learning\nalgorithm that learns competitive visual control policies through self-play in\nimagination. The DLC agent imagines multi-agent interaction sequences in the\ncompact latent space of a learned world model that combines a joint transition\nfunction with opponent viewpoint prediction. Imagined self-play reduces costly\nsample generation in the real world, while the latent representation enables\nplanning to scale gracefully with observation dimensionality. We demonstrate\nthe effectiveness of our algorithm in learning competitive behaviors on a novel\nmulti-agent racing benchmark that requires planning from image observations.\nCode and videos available at\nhttps://sites.google.com/view/deep-latent-competition.\n