Building a Fast Multilingual OCR Model with Synthetic Data

Creating robust multilingual OCR models is often hindered by the immense cost and effort required to manually annotate millions of real-world documents. To overcome this data bottleneck, the Nemotron OCR v2 model was trained on a massive dataset of 12 million synthetically generated images across six languages. This data-centric approach provides perfectly accurate labels at scale, resulting in a model that is both highly accurate and remarkably fast. Its speed is driven by an efficient architecture that unifies text detection and recognition through a shared backbone, enabling it to process nearly 35 pages per second on a single GPU.