Case Study
Extending Creative Fabrica's fonts
through AI
Company |
Industry |
 |
Creative |
Client
Creative Fabrica is a leading digital marketplace catering to graphic designers, artists, crafters, and web creators looking for high-quality design assets. With over 4 million registered members from across the globe, Creative Fabrica's platform serves both professionals and hobbyists, boasting a diverse range of fonts, graphics, and templates that drive creativity and enhance projects worldwide.
Impact
Creative Fabrica's objective was to simplify the process of enriching their fonts with characters specific to certain languages while preserving the original font's style. Our AI-trained models enabled generating missing font characters, which would otherwise require a lot of human expert effort. This advancement not only simplifies the creation of fonts with language-specific characters but, most importantly, expands the potential use of existing fonts across various language markets.
Problem
CF (Creative Fabrica) offers more than 130,000 fonts to their customers. However, there are a lot of language-specific characters, such as Ü Š Đ Č Ç Ó Æ, resulting in many unusable fonts for those languages. The traditional approach of manually creating fonts for these specific characters would be both labor-intensive and time-consuming. Determined to overcome these obstacles, Creative Fabrica partnered with us to enhance fonts by integrating characters tailored to specific languages using cutting-edge AI technology."
Challenges
Data input variability
Training the model on both raster and vector images poses a significant challenge. While conventional CNN or transformer models are suitable for raster images, vector images require a different architecture that works with sets of commands
Font file output
Our ultimate goal was to convert the model output into actual font files that include the necessary vector commands.
Typography
A font character is more than just an image; we had to predict additional attributes such as left and right margins, along with character kerning (the arrangement of overlapping characters)
Overfitting with large models
Large models tend to easily overfit to average-looking fonts
Font style diversity
Fonts may range from characters that resemble wooden boards to those with paint drizzling over them. Since there's a limited number of examples for such diverse styles, it becomes harder to train the model effectively
Solution
Our strategy involved leveraging state-of-the-art deep learning architectures tailored to Creative Fabrica's requirements. Following this, we trained our AI models on an extensive dataset of over 30,000 font files, encompassing more than 4 million character images. We trained models to enrich existing fonts with characters from various European alphabets, Vietnamese, and more. Besides character generation, we developed algorithms that can combine generated character images into fully functional (.ttf) font files. These algorithms automatically handle character width, left and right bearing, and kerning between characters.
Tools and Technologies
Python
PyTorch
FontForge
Matplotlib
Weights And Biases
Numpy
Results
Through the integration of AI, we enabled Creative Fabrica to overcome the challenges of multilingual font creation swiftly. Now, fonts are no longer constrained by missing characters, as our AI models generate the necessary characters to complete the entire character set. This has vastly expanded the range of available fonts for users who are looking for fonts in languages with smaller numbers of speakers, like Croatian, Czech, Slovakian and others.
Discover More Case Studies
Discover the impact of our custom AI solutions on business success through customer stories
