Fonts are everywhere—from posters and product labels to presentation slides and user interfaces. Recognizing them automatically from images has long been a challenging problem due to the fine-grained visual differences between typefaces. In 2015, the DeepFont system introduced a new approach using convolutional neural networks (CNNs), along with the AdobeVFR dataset. Today, we revisit this problem with modern data and modern architecture—leveraging FasterViT-2, a high-performance hybrid vision transformer, to set a new standard in font classification.

Dataset: Updating AdobeVFR for the modern font landscape

The original AdobeVFR dataset offered a strong foundation for visual font recognition. Collabora has since expanded and modernized it to reflect today’s typographic landscape and usage in screen-based environments.

Overview of different fonts from the dataset.

Key improvements

• Extended font set: Over 2,700 fonts collected from modern online repositories like Google Fonts and Font Squirrel.
• Synthetic training data: 2.7 million grayscale word images generated with augmentations that simulate real-world conditions (e.g., blur, noise, compression, perspective distortion).
• New real-world test set: 5,000 manually labeled text snippets cropped from contemporary images, including design forums, screen captures, and digital publications. 
• Consistent pre-processing: All images resized to a height of 105 pixels, with randomized character spacing and aspect ratio adjustments.

These changes make the dataset more reflective of the visual variability found in real-world design and screen environments.

Model: Introducing FasterViT-2

To fully leverage the new dataset, we employed FasterViT-2, a hybrid vision transformer designed for both speed and accuracy.

Why FasterViT-2?

FasterViT-2 combines CNN-like fast local feature learning with transformer-style global modeling through its Hierarchical Attention (HAT) mechanism. It introduces carrier tokens to summarize local windows and enable efficient long-range dependencies—critical for capturing the nuanced differences between fonts.

Architectural highlights

• Hybrid backbone: Convolutional layers in early stages, transformer layers in later stages.
• Hierarchical Attention: Efficient multi-scale token interaction for better spatial understanding.
• Real-time throughput: Over 3,100 images per second on A100 GPUs.

This combination allows FasterViT-2 to outperform both traditional CNNs and earlier ViT-based models in both performance and practicality.

Results: A new state-of-the-art in font classification

We trained FasterViT-2 on the extended AdobeVFR dataset and evaluated its performance on the newly curated real-world test set.

Example font classification, showing the top-5 prediction.

Training setup

• Training set: 2.7M synthetic images (with font-aware augmentations)
• Validation: 270K synthetic samples
• Test set: 5,000 labeled real-world images
• Backbone: FasterViT-2, pretrained on ImageNet-1K

Evaluation results

This establishes FasterViT-2 as the new state-of-the-art for real-world font classification, with a significant accuracy improvement and a substantial gain in inference speed—making it viable for both offline and real-time applications.

Applications: From design tools to video upscaling

While our primary goal was to advance font recognition, FasterViT-2 has also proven useful in broader applications.

As part of Collabora’s winning submission to the ICME 2025 Video Super-Resolution Challenge—specifically Track 3: Screen Sharing Videos—FasterViT-2 serves as a semantic font recognition module. Its predictions help guide super-resolution models to better preserve and enhance text during upscaling, improving the clarity of screen-shared content such as code editors, terminal windows, and slide presentations. We will write more about the challenge soon.

This integration shows that font classification can go beyond recognition, serving as a perceptual cue for more intelligent, content-aware visual enhancement.

Conclusion

By combining a significantly updated dataset with the power of FasterViT-2, we’ve set a new benchmark for visual font recognition. This work highlights how modern architectural advances can tackle fine-grained classification tasks with greater precision and efficiency.

FasterViT-2 is not only a state-of-the-art model for font identification, but it is also a stepping stone for improving visual quality in downstream tasks, such as screen content upscaling and intelligent design tools.