I recently published my doctoral research on data regularisation in machine learning for typeface design. The core question: can structured datasets improve vector font generation quality?
For those interested in the intersection of ML and design, I explored how data preparation affects generative models in a domain where small errors are highly visible to the human eye.
My approach:
Created LTTR/SET, a regularised dataset of vector letterforms
Trained models based on the DeepVecFont-2 architecture
Generated and evaluated 468 fonts through both qualitative and quantitative methods
The findings suggest that dataset regularisation significantly affects output quality - particularly in maintaining typographic consistency across generated glyphs. While not revolutionary, the results identify a practical preprocessing approach that could help bridge the gap between computational generation and the nuanced requirements of typeface design.
The complete thesis, generated font samples, and the LTTR/SET dataset are available online if you want to examine the methodology or results in detail.
This work represents an initial investigation rather than a complete solution to the challenges of ML in typography. I'd appreciate feedback from both the ML and design communities.
I recently published my doctoral research on data regularisation in machine learning for typeface design. The core question: can structured datasets improve vector font generation quality?
For those interested in the intersection of ML and design, I explored how data preparation affects generative models in a domain where small errors are highly visible to the human eye.
My approach:
The findings suggest that dataset regularisation significantly affects output quality - particularly in maintaining typographic consistency across generated glyphs. While not revolutionary, the results identify a practical preprocessing approach that could help bridge the gap between computational generation and the nuanced requirements of typeface design.The complete thesis, generated font samples, and the LTTR/SET dataset are available online if you want to examine the methodology or results in detail.
This work represents an initial investigation rather than a complete solution to the challenges of ML in typography. I'd appreciate feedback from both the ML and design communities.