FontCopilot: A Designer-in-the-Loop System for Chinese Vector Font Design

Abstract

Designing Chinese fonts is challenging due to the structural complexity of ideographic characters and the scale of character sets (e.g., over 80,000 characters in GB 18030). Existing font generation methods often produce non-editable raster images or offer limited stylistic control, which restricts their use in professional design workflows. We present FontCopilot, an interactive designer-in-the-loop system for Chinese vector font design. Through an exploratory interview study with 10 professional designers, we identify Component Assembly as a major production bottleneck. FontCopilot targets this labor-intensive phase using a component-based affine transformation model that generates editable vector glyphs from few-shot inputs. The workflow lets designers refine generated results and use their corrections to update the model. These interactions provide direct control over component layout and style. We evaluate CAR through model ablations and a comparison with a prior Chinese vector font generation method, study FontCopilot with 9 professional designers, and document a four-week production case. FontCopilot received “Excellent” ratings on five of six UEQ dimensions, and one designer used 300 seed characters to produce 10,461 editable glyphs for AI Song. Together, these evaluations show how editable component completion can carry local structural decisions from a small seed set to a complete Chinese typeface.