Inkdiffuser

Inkdiffuser is a PyTorch-based font generation and diffusion project focused on ink spread, stroke structure, and image quality evaluation.

Highlights

• Training pipeline built on accelerate and diffusers
• Content, style, and high-frequency feature encoders
• Differentiable ink-structure losses for diffusion-aware supervision
• Utility scripts for rendering TTF fonts, composing samples, and evaluation

Repository Layout

• train.py: training entry point
• eval.py: image quality evaluation metrics
• ttf_to_img.py: render TTF characters into images
• src/: model and loss implementations
• configs/: argument parsers and configuration helpers
• dataset/: dataset and dataloader utilities
• utils/: normalization, checkpoint, and helper functions

Installation

Install the dependencies first:

pip install -r requirements.txt

Training

Run training with the configuration defined in configs/fontdiffuser.py:

python train.py

Evaluation

Evaluate generated images against ground truth folders:

python eval.py --input_images_path <gt_dir> --generated_images_path <gen_dir>

TTF to Image

Render characters from a TTF file:

python ttf_to_img.py --font_style <font.ttf> --chara <chars.txt> --save_path <output_dir>

Notes

• Some utility scripts still contain local absolute paths in their default arguments.
• Large generated artifacts, checkpoints, and result folders are ignored by .gitignore.