Collabora has trained a FasterViT-2 model for font recognition, achieving 87.4% top-1 and 92.1% top-5 accuracy on a real-world test set — up from 62.3% and 81.4% with the 2015 DeepFont baseline. Inference throughput reaches 3,161 img/sec on an RTX 4090, compared to ~450 for DeepFont. The model uses a hybrid CNN-transformer architecture with Hierarchical Attention (HAT) and carrier tokens for efficient long-range dependency modeling. The training dataset is an expanded version of the AdobeVFR dataset modernized for screen-based environments. Beyond font identification, FasterViT-2 was used as a semantic font recognition module in Collabora's winning entry to the ICME 2025 Video Super-Resolution Challenge (Track 3: Screen Sharing Videos), where font predictions guide super-resolution models to better preserve text clarity.
Table of contents
Dataset: Updating AdobeVFR for the modern font landscapeModel: Introducing FasterViT-2Results: A new state-of-the-art in font classificationApplications: From design tools to video upscalingConclusionQuestions this post answers
How does FasterViT-2 compare to DeepFont for font recognition accuracy and inference speed?
FasterViT-2 achieves 87.4% top-1 and 92.1% top-5 accuracy on a real-world font classification test set, versus 62.3% and 81.4% for DeepFont. Inference throughput on an RTX 4090 is 3,161 images per second for FasterViT-2 compared to roughly 450 for DeepFont — a 7x speed improvement alongside a 25-point accuracy gain. Teams benchmarking vision models for fine-grained classification tasks track results like these on daily.dev.
How does FasterViT-2's Hierarchical Attention mechanism work for image classification?
FasterViT-2 combines CNN-style local feature extraction with transformer-style global modeling via its Hierarchical Attention (HAT) mechanism. It introduces carrier tokens that summarize local windows and enable efficient long-range dependencies without attending over all patches globally. This makes it faster than standard ViTs while retaining the ability to capture subtle, fine-grained visual differences across an image. Developers building on hybrid vision transformers find architecture deep-dives like this on daily.dev.