DBML-Font :Double-branch multi-level feature fusion based on diffusion model for few-shot font generation

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot font generation synthesizes fonts from limited samples, facing challenges in style consistency and structural accuracy.
• Existing methods struggle with collaborative global-local style feature processing and precise font structure encoding.

Purpose of the Study:

• To propose DBML-Font, a framework for few-shot font generation that addresses limitations in style feature processing and content encoding.
• To enhance style detail preservation and structural accuracy in synthesizing fonts from limited reference samples.

Main Methods:

• Developed the DBML-Font framework based on a conditional diffusion model.
• Introduced a Dual-Branch Multi-Level Feature Fusion Style Encoder (DMFF-SE) for hierarchical style feature extraction.
• Implemented a Geometric Structure Content Encoder (GeoStruct-CE) for precise glyph content encoding.

Main Results:

• DBML-Font demonstrated superior performance over competing methods on multiple benchmark datasets.
• The dual-branch architecture and cross-layer fusion mechanism effectively modeled global style consistency and local detail diversity.
• The framework achieved efficient and precise encoding of glyph content by capturing intrinsic geometric characteristics.

Conclusions:

• DBML-Font effectively addresses the challenges in few-shot font generation by integrating global and local style features.
• The proposed methods enhance the preservation of style details and structural accuracy in synthesized fonts.
• DBML-Font represents a significant advancement in conditional diffusion models for creative font synthesis.