GlyphGAN: Style-Consistent Font Generation Based on Generative Adversarial Networks - Hayashi - 2019

Info

• Title: GlyphGAN: Style-Consistent Font Generation Based on Generative Adversarial Networks
• Task: Font Generation
• Author: H. Hayashi, K. Abe, and S. Uchida
• Date: May 2019
• Arxiv: 1905.12502

Highlights & Drawbacks

• Two encode vectors for character and style, respectively

Motivation & Design

The main frame work is a DCGAN, with Wasserstein distance as discriminator Loss. Notice that char-ID is encoded with a saperate one-hot vector as priors applied to random vector for generator.

Performance & Ablation Study

legibility evaluation

Style Consistency

The metric of style consistency is defined as: $C_s=\frac{1}{NC}{\sum }_{n=1}^N\frac{1}{{\overline{d}}_n}{\sum }_{c=1}^C{(d_{n,c}-{\overline{d}}_n)}^2,$ where $N$ is the number of generated images (we used $N=10,000$ in this experiment), $C$ is the number of character classes, i.e.,$C=26$, $d_{n,c}$ is the distance between the generated font and the nearest real font, and $\overline{d}$is the average of $d_{n,c}$ over $c$. The metric C is the averaged coefficient of variation of $d_{n,c}$, and represents an intra-style variation of the generated font images. The lower $C_s$ is, the higher style consistency is.