Reproduce “Few-Shot Font Generation by Learning Fine-Grained Local Styles” on OpenResearch

About this paper

Licheng Tang, Yiyang Cai, Jiaming Liu, Zhibin Hong, Mingming Gong, Minhu Fan, Junyu Han, Jingtuo Liu and 2 more

FSFont presents a few-shot font generation method that leverages a cross-attention mechanism to learn fine-grained local styles and their spatial correspondence, improving style consistency and detail recovery in generated characters. It demonstrates superior perceptual quality and achieves user study scores for style consistency substantially higher than prior methods (e.g., 67.1% vs. 13.3% on UFUC).

Key insights

• Directly learning fine-grained local styles and their spatial correspondence via cross-attention offers a more effective approach for high-fidelity font generation than relying solely on global or simple component-wise disentanglement.
• An auxiliary self-reconstruction task, where the model reconstructs a ground truth image using itself as a reference, provides crucial, well-aligned gradients that significantly improve the model's ability to capture intricate details and learn robust spatial mapping.
• A structured and component-aware reference selection process is vital for consistently providing relevant style information, enabling the model to better exploit limited reference examples.

Summary by alphaXiv.