In order to improve the image quality of font generation and reduce the labour cost of font design, a method for few-shot font generation based on multilevel channel attention network is proposed. Firstly, the method acquires important local features through the style-aware attention module; then a multilevel attention mechanism is designed, where shallower layers can only observe the local features of the image, while deeper layers can observe all the features of the image, and new stylistic features are constructed by aggregating the local features of different levels. Finally, a content loss function, a style loss function and a L1 loss function are used to optimise the parameters of the model and stabilise the training of the network so that the generated images are consistent with the target font in terms of content and style. The experimental results show that the method has a strong generalisation to fonts of unknown style and fonts of unknown content. Compared to other methods, the proposed method shows better experimental results that maintain the integrity of the content structure and the accuracy of the font style.