The current style transfer methods usually suffer from an imbalance in learning the content features and style features. Some of them tend to learn style information in the style domain, while others are more affinitive with the content domain. This feature imbalance learning leads to unsatisfactory image visual performance and thus provides a new perspective for the image style transfer study. In this work, we propose a novel Balance-Aware Universal Image Style Transfer Network (BAUST-Net) to prevent this problem. Specifically, we design a novel adaptive feature-matched fusion (AdaFF) module that can promote the learning of content-style features by adaptively matching deep and shallow features from backbone networks.We then build a style attribute transformation (SAT) combined with the adversarial loss to further improve the localized stylization effect of the generated image by adversarial training. Finally, we construct a stylized ideal distance (StyID) for image style transfer, which can take into account both content and style feature learning to evaluate the stylized performance of the model. Extensive experiments have verified that our proposed BAUST-Net is effective in content-to-style and style-to-content multi-domain tasks, surpassing previous methods.