NIV: Neural Axis Variations for Variable Font Generation

Nadav Benedek, Ariel Shamir, Ohad Fried

Paper  |  Project Page

---

This repository is the official implementation of: NIV: Neural Axis Variations for Variable Font Generation.

Structure

• environment.yml: Conda environment specification.
• code/: Python source files for training and generation.
• model/: Trained model checkpoint folder (used by ttf_to_var.py)
• dataset_sample/: Sample of the prepared dataset (split files, dataset license, and a small set of XML files). To generate the full dataset, run the two commands in Quick Start:
  1. data_get_variable_google_fonts.py
  2. data_prepare_training_data.py

Environment Setup

Set up the conda environment. This may take a few minutes.

conda env create -f environment.yml
conda activate niv_env

Quick Start

Download variable fonts

Use this script to clone the Google Fonts repository into google-fonts subdirectory, and extract only variable-font TTF files into google-fonts-variable. This takes a few minutes.

python code/data_get_variable_google_fonts.py

Prepare training data

Use this step to convert and preprocess a folder of .ttf fonts into XML training files in the dataset folder. This should take around 15 minutes.

Example:

python code/data_prepare_training_data.py \
  --fonts-dir google-fonts-variable \
  --out-dir dataset \
  --skip-composites \
  --normalize-upm \
  --only-axes wght,wdth,slnt,opsz

Alternatively, you can download the prepared dataset from Hugging Face:

pip install -U huggingface_hub
hf download ndvb/NIV \
  --repo-type dataset \
  --local-dir dataset

Training

Train the model using the dataset directory and the predefined train/test split which is specified in dataset_sample for consistent results. The default argument for best performance is to load dataset into RAM (around 50GB of RAM). You can skip this step and use a trained model.

python code/niv_model_train.py \
  --data-dir dataset \
  --fixed-split-dir dataset_sample \
  --seed 42 \
  --output-dir model_output \
  --train-loss mse \
  --eval-loss mse \
  --split-method font \
  --eval-steps 20000 \
  --only-axes wght,wdth,slnt,opsz \
  --epochs 80

Evaluation

Evaluate on font-split:

python code/niv_model_train.py \
  --data-dir dataset \
  --fixed-split-dir dataset_sample \
  --load-checkpoint model \
  --seed 42 \
  --train-loss mse \
  --eval-loss mse \
  --split-method font \
  --only-axes wght,wdth,slnt,opsz \
  --only-eval

Generate a variable font

python code/ttf_to_var.py /path/to/font.ttf --model /path/to/model/best --axes wght,wdth,slnt,opsz --unicode-range 0x20-0x7E --delta-level 4

Output is saved next to the source font as *_var.ttf.

Citation

If you find this work useful in your research, please consider citing:

@article{benedek2026niv,
  title={NIV: Neural Axis Variations for Variable Font Generation},
  author={Benedek, Nadav and Shamir, Ariel and Fried, Ohad},
  journal={arXiv preprint arXiv:2606.05261},
  year={2026}
}