4.4.2. Results

Output Postprocessing Method

DeepVecFont-2 (Wang et al. 2023) claims to generate a complete font, which could be considered true only to a certain degree. In the current setup, the inference output is a set of SVG files of individual glyphs or an HTML file that compiles the selection of the best SVG samples. Even though the original training data comprises the metrics information for each glyph, the output samples miss the information. Therefore, to achieve a real font file like OTF or TTF, additional post-processing is needed.

To construct a complete font from the inference outputs, the Glyphs (“Glyphs,” n.d.) application, with custom scripts, was developed to automatise the post-processing of SVG files. These scripts enabled the import of the SVG files in several ways: a) selectively, one by one, a) collectively from a directory, or c) directly from a designated HTML file that compiled the best SVG samples.

Once imported, another set of scripts automatised post-processing on the imported paths, which included cleaning up the paths, closing any open paths, and correcting path directions. After the elementary postprocessing, each imported drawing was automatically resized to match the vertical metrics of the corresponding glyph, ensuring the coherent alignment of all the generated fonts. As a final step, each imported glyph was renamed from the generated numerical value to an appropriate glyph name.

Additionally, to set up the letter spacing, the Glyphs plugin HT Letterspacer (“HT Letterspacer,” n.d.) has been exploited. The script automatically triggered the proper HT Letterspacer (“HT Letterspacer,” n.d.) setup for each font.

There could be even more steps that are usually involved in font production like kerning or font hinting. However, for the sake of the experiment, the achieved state was satisfactory, and we proceeded to font exporting.

Results Preview

For the convenience of visual evaluation, an online font testing system, AI Font Tester, was created. It helps the evaluator quickly switch between four options, preview the generated font, and eventually download a OTF version.

Figure presents font tester of the fonts generated during this project. You can find it at https://lttrface.com/ai-font-tester

Model On the top, first from the left, a user can select one of the involved models.

Epochs On the top, second from the left, a user can select the model’s epoch variant. The number indicates how many epochs are trained. More epochs mean longer training time.

Samples On the top, third from the left, a user can select how many samples the model has generated from each letter before choosing the best one for the final font. The number indicates how many samples were generated. More samples mean a higher time of generation.

Reference On the left side, a user can visually select the Reference Style that was used as an example for generating a new font.

“Glyphs.” n.d. Glyphs GmbH. Accessed October 26, 2024. https://glyphsapp.com.
“HT Letterspacer.” n.d. Huerta Tipográfica. Accessed October 25, 2024. http://letterspacer.huertatipografica.com/.
Wang, Yuqing, Yizhi Wang, Longhui Yu, Yuesheng Zhu, and Zhouhui Lian. 2023. “DeepVecFont-v2: Exploiting Transformers to Synthesize Vector Fonts with Higher Quality.” March 25, 2023. https://doi.org/10.48550/arXiv.2303.14585.

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License .

Creative Commons License

Citation

If this work is useful for your research, please cite it as:

@phdthesis{paldia2025generative,
  title={Research and development of generative neural networks for type design},
  author={Paldia, Filip},
  year={2025},
  school={Academy of Fine Arts and Design in Bratislava},
  address={Bratislava, Slovakia},
  type={Doctoral thesis},
  url={https://lttrface.com/doctoral-thesis/},
  note={Department of Visual Communication, Studio Typo}
}