GANet: Glyph-Attention Network for Few-Shot Font Generation

Mingtao Guo, Wei Xiong, Zheng Wang, Yong Tang, Ting Wu

Keywords: font generation, GANet, glyph-attention, few-shot, GAN

Abstract: Font generation is a valuable but challenging task, it is time consuming and costly to design font libraries which cover all glyphs with various styles. The time and cost of such task will be greatly reduced if the complete font library can be generated from only a few custom samples. Inspired by font characteristics and global and local attention mechanism Wang et al. (2018), we propose a glyph-attention network (GANet) to tackle this problem. Firstly, a content encoder and a style encoder are trained to extract features as keys and values from a content glyph set and a style glyph set, respectively. Secondly, a query vector generated from a single glyph sample by the query encoder is applied to draw out proper features from the content and style (key, value) pairs via glyph-attention modules. Next, a decoder is used to recover a glyph from the queried features. Lastly, Adversarial losses Goodfellow et al. (2014) with multi-task glyph discriminator are employed to stablize the training process. Experimental results demonstrate that our method is able to create robust results with superior fidelity. Less number of samples are needed and better performance is achieved when compared to the other state-of-the-art few-shot font generation methods, without utilizing supervision on locality such as component, skeleton, or strokes, etc.

One-sentence Summary: Glyph-Attention Network for Few-shot Font Generation