CyCADA is a novel cycle-consistent adversarial domain adaptation model that improves performance in visual recognition and prediction tasks by adapting representations at both pixel and feature levels. The model enforces cycle-consistency and semantic consistency while leveraging a task loss, and does not require aligned image pairs. It is applicable across a range of deep architectures and/or representation levels, and has several advantages over existing unsupervised domain adaptation methods. CyCADA uses a reconstruction (cycle-consistency) loss to encourage the cross-domain transformation to preserve local structural information and a semantic loss to enforce semantic consistency. The model is applied to digit recognition across domains and semantic segmentation of urban scenes across domains. Experiments show that CyCADA achieves state-of-the-art results on digit adaptation, cross-season adaptation in synthetic data, and on the challenging synthetic-to-real scenario. In the latter case, it improves per-pixel accuracy from 54% to 82%, nearly closing the gap to the target-trained model. CyCADA combines cycle consistency, semantic consistency, and adversarial objectives to produce a final target model. It is a pixel-level method that uses a discriminator to distinguish between two image sets. Additionally, it introduces the use of cycle-consistency together with semantic transformation constraints to guide the mapping from one domain to another. CyCADA is applied to both digit adaptation and semantic segmentation. It uses G as a pixel-to-pixel convnet, f as a convnet classifier or a Fully-Convolutional Net (FCN), and D as a convnet with binary outputs. Experiments show that CyCADA achieves state-of-the-art results on multiple adaptation tasks, including digit classification and semantic segmentation of road scenes, demonstrating transfer from synthetic to real world domains. The model performs well in both pixel-level and feature-level adaptation, and is effective in challenging synthetic-to-real scenarios. CyCADA is particularly effective in pixel-level semantic segmentation with contemporary FCN architectures. The model is also effective in cross-season adaptation in synthetic data, and in synthetic-to-real adaptation scenarios.CyCADA is a novel cycle-consistent adversarial domain adaptation model that improves performance in visual recognition and prediction tasks by adapting representations at both pixel and feature levels. The model enforces cycle-consistency and semantic consistency while leveraging a task loss, and does not require aligned image pairs. It is applicable across a range of deep architectures and/or representation levels, and has several advantages over existing unsupervised domain adaptation methods. CyCADA uses a reconstruction (cycle-consistency) loss to encourage the cross-domain transformation to preserve local structural information and a semantic loss to enforce semantic consistency. The model is applied to digit recognition across domains and semantic segmentation of urban scenes across domains. Experiments show that CyCADA achieves state-of-the-art results on digit adaptation, cross-season adaptation in synthetic data, and on the challenging synthetic-to-real scenario. In the latter case, it improves per-pixel accuracy from 54% to 82%, nearly closing the gap to the target-trained model. CyCADA combines cycle consistency, semantic consistency, and adversarial objectives to produce a final target model. It is a pixel-level method that uses a discriminator to distinguish between two image sets. Additionally, it introduces the use of cycle-consistency together with semantic transformation constraints to guide the mapping from one domain to another. CyCADA is applied to both digit adaptation and semantic segmentation. It uses G as a pixel-to-pixel convnet, f as a convnet classifier or a Fully-Convolutional Net (FCN), and D as a convnet with binary outputs. Experiments show that CyCADA achieves state-of-the-art results on multiple adaptation tasks, including digit classification and semantic segmentation of road scenes, demonstrating transfer from synthetic to real world domains. The model performs well in both pixel-level and feature-level adaptation, and is effective in challenging synthetic-to-real scenarios. CyCADA is particularly effective in pixel-level semantic segmentation with contemporary FCN architectures. The model is also effective in cross-season adaptation in synthetic data, and in synthetic-to-real adaptation scenarios.