UrbanEvolver is a function-aware deep generative model for urban layout regeneration. The model regenerates urban layouts (roads and buildings) for a target region by considering the function (land use type) of the target region and its surrounding context. UrbanEvolver first extracts implicit regeneration rules from the target function and surrounding context using function-layout adaptive (FA) blocks, then constrains the regenerated layout based on these rules. To ensure validity and road structure, the model uses pixel-level and geometry-level losses. A large-scale urban layout dataset with over 147,000 regions and rich annotations is used for training. The model outperforms baseline methods in generating practical and function-aware urban layouts. Urban layout regeneration is essential for urban regeneration, which addresses urban decay. Traditional methods use hand-crafted rules, while recent methods use data-driven approaches. However, these methods struggle with modeling context. UrbanEvolver addresses this by using a hybrid approach combining vector and pixel data for efficient and accurate learning. The model learns to extract effective fused features and map them to urban layouts. The model is evaluated on a large-scale dataset, showing it can regenerate valid and function-aware layouts. Contributions include proposing the first function-aware urban layout regeneration task and model, and introducing FA blocks and loss functions for efficient and accurate regeneration.UrbanEvolver is a function-aware deep generative model for urban layout regeneration. The model regenerates urban layouts (roads and buildings) for a target region by considering the function (land use type) of the target region and its surrounding context. UrbanEvolver first extracts implicit regeneration rules from the target function and surrounding context using function-layout adaptive (FA) blocks, then constrains the regenerated layout based on these rules. To ensure validity and road structure, the model uses pixel-level and geometry-level losses. A large-scale urban layout dataset with over 147,000 regions and rich annotations is used for training. The model outperforms baseline methods in generating practical and function-aware urban layouts. Urban layout regeneration is essential for urban regeneration, which addresses urban decay. Traditional methods use hand-crafted rules, while recent methods use data-driven approaches. However, these methods struggle with modeling context. UrbanEvolver addresses this by using a hybrid approach combining vector and pixel data for efficient and accurate learning. The model learns to extract effective fused features and map them to urban layouts. The model is evaluated on a large-scale dataset, showing it can regenerate valid and function-aware layouts. Contributions include proposing the first function-aware urban layout regeneration task and model, and introducing FA blocks and loss functions for efficient and accurate regeneration.