README.md

metadata

tags:
  - font-generation
  - image-generation
  - autoregressive
  - multimodal
  - few-shot-learning
  - computer-vision
library_name: pytorch

GAR-Font: Beyond Patches: Global-aware Autoregressive Model for Multimodal Few-Shot Font Generation

---

📌 Overview

GAR-Font is a global-aware autoregressive model for multimodal few-shot font generation.

It aims to generate high-quality glyphs with limited style references by modeling both global font characteristics and local glyph details.

GAR-Font supports two generation settings:

• Vision-Only GAR-Font
  Generate glyphs using only visual style references.

• Vision-Language GAR-Font
  Generate glyphs using both visual references and natural language style descriptions.

This repository provides a pretrained checkpoint of Vision-Only GAR-Font.

---

📦 Model Weights

We provide the pretrained checkpoint:

generator_ckpt_pruned.pt

The checkpoint contains:

• vq_model: pretrained VQ tokenizer for glyph representation
• model: pretrained GAR-Font autoregressive generator

The checkpoint can be directly used for inference and further research.

---

🚀 Usage

Download:

generator_ckpt_pruned.pt

and use it as the --generator-ckpt argument in the inference scripts.

For installation instructions, inference examples, training details, and complete implementation, please refer to our official GitHub repository:

🔗 https://github.com/xTryer-s/GAR-Font

---

📖 Paper

Beyond Patches: Global-aware Autoregressive Model for Multimodal Few-Shot Font Generation

arXiv: https://arxiv.org/abs/2601.01593

---

Citation

If you find GAR-Font useful, please cite our paper:

@misc{cai2026patchesglobalawareautoregressivemodel,
      title={Beyond Patches: Global-aware Autoregressive Model for Multimodal Few-Shot Font Generation}, 
      author={Haonan Cai and Yuxuan Luo and Zhouhui Lian},
      year={2026},
      eprint={2601.01593},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.01593}, 
}