GlyphGenModel: AI-Powered Glyph Generation for Font Design
(opens in new window)Behind The Project
Designing a complete Chinese font is one of the most labor-intensive tasks in typography. A single font needs at least 7,000 characters for basic Traditional Chinese coverage, and up to 15,000 to handle the full daily use set including regional variants for Cantonese, Japanese Kanji, and Taiwanese local languages. A skilled type designer can draw roughly 30 to 50 characters per day, meaning a complete font takes months or even years of full-time work.
This creates a critical challenge in the early design phase. Type designers need to make fundamental decisions about stroke weight, curve style, spacing, and overall personality, but they can’t see how these choices will look across the full character set until they’ve invested months of work. If they realize mid-project that a design direction isn’t working, pivoting means discarding significant effort. The traditional workflow forces designers to commit to a style before they can truly evaluate it at scale.
We needed a way to compress the early exploration phase: let designers sketch a small set of characters, immediately see what those design decisions would look like across thousands of glyphs, and iterate rapidly before committing to full production.
(opens in new window)Solution & Approach
GlyphGenModel is a machine-learning system that transforms how type designers explore and develop Chinese fonts. In the early design phase, a designer sketches a small set of characters to establish the basic visual direction. GlyphGenModel learns the emerging style, including stroke thickness, curves, spacing, and personality, and generates a full preview set of 7,000+ characters that show what the complete font would look like if the designer continues in this direction.
This shifts the design workflow from sequential (design thousands of characters one by one) to iterative (sketch a few, preview the full set, refine the direction, repeat). Designers can explore multiple style directions in weeks rather than months, make informed decisions about which direction to pursue, and catch potential problems before investing in full production.
The pipeline runs end-to-end: from analyzing the designer’s initial character set, through model training and image generation, to producing vector paths that designers can review in font editors like Glyphs. Once the designer settles on a direction, these AI-generated glyphs become reference material for the production phase, providing a starting point and visual target rather than requiring each character to be designed from scratch.
I designed and built the system for justfont’s font production workflow, integrating it with our jf7000 CharSet and Unihan character set standards. Every generated glyph goes through designer review and refinement before becoming part of a shipping font, ensuring production quality while dramatically accelerating the design process.
The animation above demonstrates the training progression for generating missing glyphs in Dela Gothic One(opens in new window), an open-source font from Google Fonts. The visualization shows the validation set results over 250 epochs, with ground truth on the left and model inference on the right. As the training progresses, the generated glyphs increasingly match the actual character designs, demonstrating the model’s ability to learn and replicate the font’s distinctive style.
(opens in new window)Technical Challenges
(opens in new window)Learning Style from Limited Examples
The core challenge is generating thousands of glyphs that feel cohesive when designers have only drawn a few initial characters. The system uses a Latent Diffusion Model trained in two stages to learn and extrapolate the designer’s emerging style.
The first stage trains a VQ-VAE that learns to compress and reconstruct glyph images, creating a compact representation of character features.
The second stage trains a U-Net denoiser that learns to generate new characters in this compressed space, conditioned on reference glyphs from a structural template font.
The key insight is reference-conditioned generation. Instead of generating characters from scratch, the model uses structurally similar characters from a reference font as a skeleton, then applies the target font’s learned style to determine how those strokes should look. This separation of structure (from the reference) and style (learned from the designer’s initial set) is what allows the model to produce consistent results across thousands of characters based on just a few dozen examples.
This approach lets designers see a complete style preview early in the design process, enabling them to make informed decisions about whether a design direction is worth pursuing before investing months of work.
(opens in new window)Generating the Right Character Set
Chinese fonts draw from multiple overlapping character sets: the jf7000 standard, various Unihan extensions, Cantonese-specific characters, Japanese Kanji, and Taiwanese local language characters. For a designer’s early preview to be useful, the system needs to generate exactly the characters that matter for their target use case.
I built the analysis pipeline to understand what the designer has already drawn and what they need to see. The system cross-references the initial character set against jf7000 and Unihan standards, identifies which characters to generate for a complete preview, and adapts to different coverage goals (basic Traditional Chinese vs. full regional support).
This analysis is critical for the iterative design workflow. In the exploration phase, designers might want to see a focused subset of high-frequency characters to evaluate readability. In the refinement phase, they might need the full extended set including rare regional characters. The pipeline adapts to these different needs, generating previews that help designers make informed decisions at each stage.
(opens in new window)Preserving Fine Details That Define Style
The quality of generated previews directly impacts how useful they are for design decisions. If the AI-generated characters lose the subtle details that define a font’s personality (stroke endings, curve tension, spacing rhythm), designers can’t accurately evaluate whether a direction is working.
The diffusion process relies on gradually adding and removing noise over many timesteps. Standard noise schedules destroyed the fine stroke details that distinguish one font’s style from another, making previews less useful for design evaluation. I implemented a sigmoid beta schedule that adds noise gradually during critical early timesteps, preserving these subtle details while still learning to generate from scratch.
This technical choice has direct design impact: sharper strokes, more consistent style transfer, and previews that accurately represent what the full font would look like if the designer continues in this direction. Designers can make confident decisions about style direction based on the generated previews, knowing the AI is capturing the nuances that matter.
The comparison above shows characters organized by shared radicals in each row. The four leftmost columns display original glyphs from Dela Gothic One, while the four rightmost columns (red underline glyphs) show glyphs generated by GlyphGenModel. The generated characters successfully capture the fine details and distinctive characteristics of the original design, demonstrating the model’s ability to preserve stylistic nuances across different character structures.
(opens in new window)Impact & Results
GlyphGenModel fundamentally changes how type designers approach Chinese font development by compressing the exploration phase from months to days. In the traditional workflow, designers commit to a style direction based on a handful of characters, then invest months drawing the full set before they can truly evaluate whether that direction works. With GlyphGenModel, designers sketch a few initial characters, generate a complete 7,000+ character preview set, evaluate the full style, and iterate on the direction, all within the beginning of a project.
In the early exploration phase, designers can try multiple style directions without committing to full production. Want to see if a bolder stroke weight works? Sketch a few characters, generate a preview, and evaluate. Considering a more geometric vs. organic approach? Test both directions with generated previews before choosing. This freedom to experiment early, when changes are cheap, leads to better design decisions and reduces the risk of mid-project pivots.
Seeing a complete character set early also reveals problems that wouldn’t be obvious from a small sample. Stroke patterns that look good in 50 characters might become monotonous across 7,000. Spacing that works for common characters might fail for complex ones. The preview set lets designers catch and fix these issues before investing in production work.
Once designers settle on a direction, the AI-generated glyphs become reference material for production work. Instead of designing each of the remaining 6,000+ characters from a blank canvas, designers refine AI-generated proposals that already capture the target style. This doesn’t eliminate designer work, as every character still goes through review and refinement, but it provides direction and foundation, significantly accelerating the workflow.
The tool is particularly valuable for regional character extensions such as Taiwanese Hokkien, Hakka, and indigenous languages, which are often left incomplete because designers can’t justify the time investment without knowing if the style will work.
(opens in new window)Lessons Learned
This project taught me that successful AI tools for creative workflows require understanding how people actually work, not just what they produce. The initial version of GlyphGenModel focused on generating missing characters to complete existing fonts, which was a technically correct interpretation of "font completion." However, working with designers revealed that the real value wasn’t filling gaps in finished fonts; it was enabling exploration before committing to production.
This insight shifted the entire project direction. Instead of optimizing for how many characters can we generate, we optimized for how quickly can a designer evaluate a style direction. Instead of measuring technical metrics like SSIM scores in isolation, we focused on whether generated previews gave designers enough confidence to make decisions. The technical challenges that mattered were the ones that directly impacted the design workflow: clean SVG output that imports without cleanup, accurate style transfer that preserves subtle details, and character set analysis that generates the right preview for each design stage.
Working closely with type designers also revealed the importance of trust in AI-assisted creative tools. Designers need to understand what the AI is doing (learning style from their initial sketches) and what it’s not doing (making creative decisions). They need clean, editable vector output they can refine, not black-box raster images. They need the system to be a tool that amplifies their work, not an automation that replaces it. Building this trust required transparency about the system’s capabilities and limitations, and designing the workflow so designers remain in control throughout the process.
The hardest part wasn’t the model architecture, since VQ-VAE and U-Net worked well relatively early. It was the surrounding pipeline: character set analysis, vectorization quality, reproducibility across different input fonts, and integration into existing design tools. These everything around the model pieces are what made GlyphGenModel genuinely useful in production workflows rather than just a technical demo.
This experience reinforced that domain expertise matters as much as technical skill when building AI tools for specialized workflows. Understanding character set standards, font production stages, and what designers need at each stage shaped every design decision. A technically perfect model that doesn’t fit the creative workflow would have been worthless.