CD-Font: One-Shot Font Generation via Conditional Diffusion Model with Disentangled Guidance

One-shot font generation aims to create a new font library by extracting style information from the reference font. Most existing font generation methods rely on GAN-based image-to-image translation frameworks, which still suffer from unstable training and imprecise character structure generation due to the nature of adversarial training. In this paper, we propose a one-shot font generation framework named CD-Font, based on a conditional diffusion model with style-content disentangled guidance. Unlike existing methods, we use two different encoders to separate the representations for styles and contents and fuse them as conditions of the diffusion model. Specifically, we concatenate the content image with the noisy image throughout all denoising steps to improve the integrity of character structures. During inference, we present a disentangled guidance sampling strategy to enable the generated font images to exhibit strong correlations with both the reference image and the target character. Extensive experiments and user studies demonstrate that our CD-Font outperforms current methods in one-shot font generation. Furthermore, we apply our method to cross-lingual font generation, showing its promising cross-lingual capability.