Automatic font generation aims to synthesize glyph images in a target font style using only a limited number of reference images. Current methods typically adopt a content-style disentanglement strategy by employing two separate networks to extract content and style features. However, such approaches may be affected by the imperfect disentanglement and neglect critical style transformation patterns between source and target fonts. To address these limitations, we propose FSTDiff, a novel framework that explicitly learns cross-font style transformations. In our approach, we introduce a Font Style Transformation (FST) module that explicitly captures the style transformation from the source font to the target font, thereby alleviating the issue of residual source font style information in the content features. This module leverages learnable queries to automatically extract both global and local style details and convert them into a style transformation representation. Extensive experiments demonstrate that FSTDiff achieves state-of-the-art performance across various characters and font types.