The accurate prediction of the Fire Weather Index (FWI), a multivariate climate index for wildfire risk characterization, is crucial for both wildfire management and climate-resilient planning. Moreover, consistent multisite fire danger predictions are key for targeted allocation of resources and early intervention in high-risk areas, as well as for “megafire” risk detection. In this regard, Convolutional Neural Networks (CNNs) are known to capture complex spatial patterns in climate data. This study compares different CNN architectures and traditional Statistical Downscaling (SD) methods (regression and analogs) for predicting daily FWI across diverse locations in Spain, considering marginal, distributional and spatial coherence measures for validation. Overall, the CNN-Multi-Site-Multi-Gaussian configuration, which explicitly accounts for the inter-site variability in the output layer structure, showed a superior performance. These insights provide a methodological guidance for the successful application of CNNs in the context wildfire risk assessment, enhancing wildfire response strategies and climate adaptation planning.