CF-Font: Content Fusion for Few-shot Font Generation

We present CF‑Font, a content‑fusion approach for few‑shot Chinese font generation, which was accepted at CVPR 2023.

Background: Designing a complete Chinese font manually is labor‑intensive due to thousands of characters. Few‑shot font generation aims to synthesize an entire font from a small set of reference glyphs by transferring style from a source font.

Method: We introduce a Content Fusion Module (CFM) that linearly mixes content features of a selected base font, where the mixing weights are derived from a font‑level distance metric. An Iterative Style Refinement (ISR) strategy refines a font‑level style vector by minimizing reconstruction loss on the few reference glyphs. Moreover, we replace the conventional L1/L2 pixel loss with a Projection Character Loss (PCL) that measures the distance between one‑dimensional projections of glyphs, encouraging preservation of global shape.

Training proceeds in two stages. First, a DG‑Font backbone learns disentangled content and style encodings. Then CFM is inserted after the content encoder, and the style encoder, feature‑deformation skip connections and mixer are further trained while keeping the content encoder fixed. ISR is applied only at inference time.

Experiments: We built a dataset of 300 Chinese fonts (≈6.4k characters each) and evaluated on both seen and unseen fonts. CF‑Font outperforms state‑of‑the‑art methods (FUNIT, LF‑Font, MX‑Font, CG‑GAN, FsFont, DG‑Font) by 5.7% in L1 and 5.0% in FID on unseen fonts, and qualitative results confirm superior visual quality.

Applications: The technique has been deployed to generate Alibaba‑Mama’s proprietary fonts such as “Dongfang DaKai” and “DaoLiTi” by combining a few designer‑crafted glyphs with open‑source fonts, followed by manual refinement.

Code and pre‑print are publicly available.