Attribute2Font: Creating Fonts You Want From Attributes

摘要

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