Finetuned DINOv2 Vision Transformer for categorizing Google Fonts

A font classification system that identifies 394 font variants across 32 families from rendered text images, using LoRA fine-tuning of DINOv2. Achieves 98.9% top-1 validation accuracy with only ~1% of parameters trainable.

Citation

If you use GoogleFontsBench, the training pipeline, or the pretrained models in your work, please cite the arXiv preprint:

@misc{chen2026parameterefficientfinetuningdinov2largescale,
title = {Parameter-Efficient Fine-Tuning of DINOv2 for Large-Scale Font Classification},
author = {Daniel Chen and Zaria Zinn and Marcus Lowe},
year = {2026},
eprint = {2602.13889},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2602.13889}
}

Quick Start

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Pipeline

1. Get Google Fonts

git clone --filter=blob:none --depth 1 https://github.com/google/fonts.git

2. Generate dataset

python dataset_generator.py \
--font_dir <path to google fonts> \
--out_dir <output folder> \
--img_size 224 \
--font_size 1024 \
--padding 128

Uses all CPU cores by default (--workers N to override). Generates ~575 training images and 40 test images per font variant with randomized colors, alignment, line wrapping, and Gaussian noise.

3. Clean the dataset

python dataset_cleaner.py <dataset folder>

Prints any corrupted image paths for manual inspection.

4. Upload dataset to HuggingFace (optional)

pip install -U "huggingface_hub[cli]"
huggingface-cli upload-large-folder <user>/<repo><dataset folder> --repo-type=dataset

For large datasets (200k+ files), tar the train/test folders first to avoid API rate limits:

tar cf train.tar -C <dataset folder> train/
tar cf test.tar -C <dataset folder> test/
HF_HUB_DISABLE_XET=1 huggingface-cli upload <user>/<repo> train.tar train.tar --repo-type=dataset
HF_HUB_DISABLE_XET=1 huggingface-cli upload <user>/<repo> test.tar test.tar --repo-type=dataset

5. Train the model

LoRA (default, recommended):

python train_model.py \
--data_dir <dataset folder> \
--output_dir <output folder> \
--batch_size 64 \
--epochs 100 \
--learning_rate 1e-4 \
--lora_rank 8 \
--lora_alpha 16 \
--lora_dropout 0.1

Baseline comparisons:

# Full fine-tuning (all 87.2M params)
python train_model.py --full_finetune --data_dir <data> --output_dir <out> --epochs 100
# Linear probe (classifier head only, 606K params)
python train_model.py --linear_probe --data_dir <data> --output_dir <out> --epochs 20
# CNN baseline (ResNet-50)
python train_model.py --resnet_baseline --data_dir <data> --output_dir <out> --epochs 100

6. Resume from checkpoint

python train_model.py \
--checkpoint <output folder>/checkpoint-2752 \
--data_dir <dataset folder> \
--output_dir <output folder> \
--epochs 100

7. Upload model to HuggingFace

python train_model.py \
--epochs 0 \
--data_dir <dataset folder> \
--checkpoint <output folder>/checkpoint-2752 \
--huggingface_model_name <user>/<repo>

8. Run inference

python serve_model.py <model name or path><image path>

Cloud Training

Runs training end-to-end on Vast.ai GPU instances: finds a machine, uploads the code, trains, uploads results to HuggingFace, and destroys the instance automatically. Includes auto-retry (up to 5 instances), health checks, and crash log upload.

Setup:

pip install vastai
vastai set api-key <your key>
vastai create ssh-key "$(cat ~/.ssh/id_ed25519.pub)"
huggingface-cli login

Usage:

# Run all baselines on separate instances in parallel
bash cloud_train.sh --hf_dataset dchen0/font_crops_v5 --hf_results dchen0/font-model-results --mode all --gpu RTX_3090 --parallel
# Run a single mode
bash cloud_train.sh --hf_dataset dchen0/font_crops_v5 --hf_results dchen0/font-model-results --mode lora --gpu RTX_3090
# Dry run (tiny test dataset, validates full pipeline in ~5 min)
bash cloud_train.sh --dry_run --gpu RTX_3090

Options:

Flag	Default	Description
--hf_dataset	(required)	HuggingFace dataset to train on
--hf_results	(required)	HuggingFace repo for results upload
--mode	lora	Training mode: lora, lora4, lora16, full, linear, resnet, or all
--gpu	RTX_4090	GPU type (e.g., RTX_3090, A100)
--max_price	2.00	Max hourly price in USD
--batch_size	64	Training batch size
--epochs	100	Number of training epochs
--num_gpus	1	GPUs per instance (multi-GPU via accelerate)
--parallel	off	Launch each mode on a separate instance
--dry_run	off	Use tiny test dataset, 1 epoch, defaults to all modes
--ssh_key	~/.ssh/vastai	SSH key for Vast.ai instances

Features:

• Auto-retry with up to 5 different instances per mode
• Health check after launch (connectivity, CUDA, pip)
• Checkpoints synced to HuggingFace every 10 minutes (resumable on preemption)
• Training logs uploaded on any exit (crash, signal, or success)
• Instance auto-destroys after uploading results

Dry run:

Always dry run before a full training run to catch issues early:

# Test all modes (default)
bash cloud_train.sh --dry_run --gpu RTX_3090
# Test a specific mode
bash cloud_train.sh --dry_run --mode resnet --gpu RTX_3090

This uses a tiny test dataset (dchen0/font_crops_test, 3 classes, 39 images) to validate the entire pipeline in ~5 minutes.

To regenerate the test dataset:

python create_test_dataset.py --synthetic --upload

Evaluation

python confusion_matrix.py \
--data_dir <dataset folder> \
--model <HuggingFace model name or local path>

The model's label set must match the dataset's class folders. The script will check label overlap and abort if there's a mismatch.

Produces:

• figures/confusion_matrix.pdf — Row-normalized heatmap grouped by font family
• figures/top_confused_pairs.pdf — Bar chart of most frequent misclassifications
• figures/per_family_accuracy.pdf — Per-family accuracy breakdown
• figures/tsne_embeddings.pdf — t-SNE of [CLS] embeddings
• figures/font_dendrogram.pdf — UPGMA clustering of font families
• figures/metrics.tex — LaTeX macros for paper (including SWER with typographic metadata distance)
• confusion_matrix.json — Raw counts
• bad_images.json — All misclassified images

Paper

# Full build (evaluation + LaTeX)
bash build_paper.sh --data_dir <dataset folder> --model <model># LaTeX only (skip evaluation)
bash build_paper.sh --skip-matrix

Handler

handler.py implements the preprocessing pipeline (pad-to-square + resize + normalize) used at both training and inference time. It's bundled with the model on HuggingFace for Inference Endpoints.