Fine-grained urban flow inference provides important insights for smart city applications such as urban planning and traffic management, but its accuracy is often hindered by incomplete observations due to sparse sensor deployment. While existing methods can handle minor data gaps, their performance degrades significantly under high missing rates, particularly in newly developed urban areas. Along these lines, we propose a novel cross-city super-resolution data map inference framework (CrossFM), designed to transform incomplete coarse-grained urban flows into accurate fine-grained data maps by harnessing cross-city spatio-temporal dynamics. Specifically, we first perform temporal alignment between the source city and the target city data using timestamps. Then, guided by Point-of-Interest (POI) similarity to identify similar regions, we impute missing values in the target city’s coarse-grained flow maps. This completion adaptively leverages information from both the source and target city data, resulting in an enhanced coarse-grained representation. Finally, a super-resolution module processes the spatial patterns within the completed coarse data to generate the high-resolution urban flow maps. The framework components are trained jointly end-to-end within a multi-task setup. We conduct extensive experiments on two real-world datasets and demonstrate that CrossFM significantly outperforms the state-of-the-art methods, especially under severe data scarcity.