SIGNATURE WORK
CONFERENCE & EXHIBITION 2023

The SW project aims to address the challenging task of automatic stylized font generation in a deep learning-based fashion. The creation of a complete stylized font library is a time-consuming and complex process that requires extensive knowledge and proficiency in the use of professional tools. The proposed solution, Multilingual Font Generation Network (MF-Net), is a fast feed-forward network that enables few-shot multilingual stylized font generation. The network adopts the Generative Adversarial Network (GAN) framework and employs two separate encoders to decouple the font image’s content and style information. The attention module in the style encoder extracts shallow and deep style features, and a novel language complexity-aware skip connection is designed to adaptively adjust the structural information to be preserved. The effectiveness of the proposed MF-Net is demonstrated through quantitative and subjective visual evaluations and compared with existing models in the scenario of stylized multilingual font generation.