Publications

We propose a YOLOv8-driven, positional-aware decomposition of Hangul into initial, medial, and final components, then employ a GAN to recombine them, yielding fonts that better preserve structure and style than existing methods such as MXFont and CKFont. This framework boosts accuracy (lower L1/FID, higher SSIM) and opens the door to finely controllable, multilingual typography.

A one-shot denoising diffusion framework that fuses multi-scale style vectors, a Sobel-based structural loss, and CLIP style alignment to generate high-resolution glyphs across Korean, Chinese, and Latin scripts with pixel-level stroke precision and consistent visual style.

DK-Font presents a one-shot, text-conditioned diffusion model that turns a single reference glyph into a full, high-fidelity Korean typeface. By fusing a phonetic-aware text encoder with perceptual loss, the method preserves stroke integrity and style consistency across all 2,350 Hangul characters, boosting SSIM from 0.812 to 0.857 and lowering FID to 10.45 compared with Diff-Font.