[Paper Review] Attribute2Font: Creating Fonts You Want From Attributes

Yizhi Wang, Yue Gao+1 more|arXiv (Cornell University)|May 16, 2020

Generative Adversarial Networks and Image Synthesis35 references4 citations

TL;DR

This paper proposes Attribute2Font, a deep generative model that creates custom glyph images from user-specified font attributes (e.g., italic, serif, angularity) by performing attribute-conditioned style transfer between fonts. It introduces an Attribute Attention Module and a semi-supervised learning scheme to improve generation quality and handle limited annotated data, enabling the synthesis of novel font styles not present in existing databases.

ABSTRACT

Font design is now still considered as an exclusive privilege of professional designers, whose creativity is not possessed by existing software systems. Nevertheless, we also notice that most commercial font products are in fact manually designed by following specific requirements on some attributes of glyphs, such as italic, serif, cursive, width, angularity, etc. Inspired by this fact, we propose a novel model, Attribute2Font, to automatically create fonts by synthesizing visually-pleasing glyph images according to user-specified attributes and their corresponding values. To the best of our knowledge, our model is the first one in the literature which is capable of generating glyph images in new font styles, instead of retrieving existing fonts, according to given values of specified font attributes. Specifically, Attribute2Font is trained to perform font style transfer between any two fonts conditioned on their attribute values. After training, our model can generate glyph images in accordance with an arbitrary set of font attribute values. Furthermore, a novel unit named Attribute Attention Module is designed to make those generated glyph images better embody the prominent font attributes. Considering that the annotations of font attribute values are extremely expensive to obtain, a semi-supervised learning scheme is also introduced to exploit a large number of unlabeled fonts. Experimental results demonstrate that our model achieves impressive performance on many tasks, such as creating glyph images in new font styles, editing existing fonts, interpolation among different fonts, etc.

Motivation & Objective

• To lower the barrier for non-professionals to create custom fonts by enabling attribute-based generation instead of relying on font retrieval.
• To address the limitation of existing image synthesis models that treat attributes equally and fail to model shape variations in glyphs.
• To develop a method that supports continuous, arbitrary attribute values rather than discrete or binary attributes.
• To handle the scarcity of annotated font attribute data through a semi-supervised learning approach.
• To enable generation of glyph images in new font styles, including for non-Latin scripts like Chinese.

Proposed method

• The model uses a deep generative network to perform image-to-image style transfer between fonts conditioned on their attribute values.
• An Attribute Attention Module is integrated into the decoder to dynamically emphasize the most influential attributes during generation.
• A semi-supervised learning scheme leverages large amounts of unlabeled fonts to improve training when annotated data is limited.
• The model is trained to learn a mapping from source font images and attribute values to target font images in the desired style.
• The architecture is based on a modified StarGAN framework with attribute-aware attention mechanisms.
• The method supports both interpolation between fonts and editing of existing fonts by modifying attribute values.

Experimental results

Research questions

• RQ1 Can a deep generative model synthesize novel font styles based on continuous, user-defined attribute values rather than retrieving existing fonts?
• RQ2 How can attribute-specific importance be modeled to improve the visual fidelity of generated glyphs?
• RQ3 To what extent can a semi-supervised learning strategy reduce dependency on expensive attribute annotations in font generation?
• RQ4 Can the proposed method generalize across different writing systems, such as Chinese, with large character sets?
• RQ5 How does the Attribute Attention Module enhance the model’s ability to preserve key stylistic features during font synthesis?

Key findings

• The Attribute2Font model successfully generates high-quality glyph images in new, unseen font styles based on arbitrary combinations of attribute values.
• The Attribute Attention Module significantly improves the model’s ability to emphasize key attributes, resulting in more accurate and visually consistent font styles.
• The semi-supervised learning scheme enables effective training with limited annotated data, improving performance on low-resource settings.
• The model generalizes well to non-Latin scripts, such as Chinese, demonstrating scalability to large character sets.
• Extensive experiments show superior performance in font creation, editing, and interpolation compared to existing methods.
• The model outperforms baseline GANs like StarGAN in identity consistency and attribute control, especially in face and font image translation tasks.