With the increasing availability of powerful computational resources and artificial intelligence (AI), a data-driven approach has been gaining attention in atmospheric science, particularly in downscaling/emulating research. For example, Convolutional Neural Networks (CNNs) are used to emulate fine-resolution climate data fields by physics-based models. Although CNN-based downscaling has shown good performance for air temperature across various climate scales, studies at the urban scale remain limited due to the scarcity of long-term high-resolution climate data that can represent local urban effects. This study explores whether CNNs can accurately downscale/emulate urban-scale temperature distributions, with a focus on the urban heat island (UHI), using 250 m resolution Weather Research and Forecasting (WRF) simulation data over Sofia, Bulgaria, as training data. Input features, including air temperature, wind components, surface solar radiation, and elevation, are used, and their impacts on targeted 250 m resolution air temperature are evaluated. The results show that, while the CNN-based approach shows promise in generating urban-scale temperature distributions, the selection/combination of input features influences downscaling performance. Warm bias exceeding +2.00 °C was almost eliminated by including all input features. The downscaled result emulated the broad spatial pattern. However, detailed spatial and temporal variability, such as UHI, was still not captured. These findings revealed that the input feature selection can influence CNN-based temperature downscaling, providing useful insight into future regional-to-microscale downscaling toward urban digital twins (UDTs).