Sami-Font: structure and style-aware multi-scale infusion for one-shot multilingual typeface generation

초록

We address multilingual font generation, a task that demands precise stroke structure and consistent style across large character sets. This paper introduces SAMI-Font, a diffusion-based framework for one-shot typeface synthesis. At its core is SAMI (structure- and style-aware multi-scale infusion), which injects style features at multiple spatial resolutions across the encoder, bottleneck, and decoder to enable fine-grained, spatially aware transfer. To preserve glyph integrity, we incorporate a Sobel-based structural consistency loss that emphasizes stroke edges, and a CLIP-guided style loss that improves perceptual alignment between generated glyphs and reference styles. The pipeline is trained end-to-end in a single-phase manner, mitigating redundancy and instability common to multi-stage methods. Experiments on Latin, Korean, and Chinese scripts covering printed and handwritten styles show consistent improvements over FontDiffuser, Diff-Font, MX-Font, and FUNIT on SSIM, RMSE, LPIPS, and FID, with qualitative results showing sharper strokes and stronger stylistic coherence. Few handwritten samples also suffice for personalization, enabling user-specific typefaces. These results advance practical, multilingual font synthesis and point toward high-resolution and user-controllable typography.

키워드