Haonan Cai

Haonan Cai

To be a tryer for X.

Beyond Patches: Global-aware Autoregressive Font Generation

Haonan Cai, Yuxuan Luo, Zhouhui Lian

Paper: arXiv:2601.01593

Project: GAR-Font

Overview

Manual font design requires translating an abstract stylistic concept into a complete and coherent glyph set, a process that is both time-consuming and expertise-intensive. This challenge is further amplified in Few-shot Font Generation (FFG), where only a handful of reference glyphs are available. Existing methods often struggle to simultaneously preserve global stylistic consistency and local structural correctness, especially under limited supervision.

While autoregressive (AR) models have shown strong generative capabilities, their application to FFG is hindered by conventional patch-level tokenization, which fails to capture global font-level dependencies. Moreover, most prior approaches remain confined to an image-to-image paradigm, ignoring the expressive power of language in conveying font design intent.

Method

We propose GAR-Font, a novel global-aware autoregressive framework for multimodal few-shot font generation. The key components include:

  • Global-aware Tokenizer
    Captures both fine-grained local structures and long-range global stylistic patterns, overcoming the limitations of patch-based representations.

  • Multimodal Style Encoder
    Enables flexible style control via a lightweight language-style adapter, allowing textual descriptions to guide font generation without expensive multimodal pretraining.

  • Post-refinement Pipeline
    Further improves structural fidelity and stylistic coherence, producing cleaner and more visually consistent glyphs.

Results

Extensive experiments demonstrate that GAR-Font consistently outperforms existing FFG methods, achieving:

  • Superior global style faithfulness
  • Improved structural integrity
  • High-quality generation with textual stylistic guidance

These results highlight the effectiveness of integrating global-aware modeling and multimodal control into autoregressive font generation.

Keywords

Few-shot Font Generation · Autoregressive Models · Multimodal Learning · Global Representation · Image Synthesis