Building a Fast Multilingual OCR Model with Synthetic Data

The bottleneck of Nemotron OCR v1 was not the architecture but the data. It hadn't seen enough CJK and Cyrillic script, causing NED to spike as high as 0.56–0.92 for Japanese, Korean, Russian, and Chinese.

The solution is a synthetic data pipeline. They extracted per-language source text from mOSCAR, and heavily modified SynthDoG to generate word-, sentence-, and paragraph-level bounding boxes, 4-point quads, and even a reading order relation graph.

The layouts go beyond simple documents.

• Multi-column body text
• Scattered scene-text styles
• Vertical writing
• Tables, tables of contents, slides, and word-document-style pages

They secured 165–1,258 fonts per language, and applied strong augmentation — borders, shadows, noise, blur, morphology, and color variation — to increase realism.

The final dataset is 12.2M samples in scale, covering 6 languages.

• English: 1,825,089
• Japanese: 1,889,137
• Korean: 2,269,540
• Russian: 1,724,733
• Chinese (Simplified): 2,335,343
• Chinese (Traditional): 2,214,304

The model, Nemotron OCR v2, has a 3-stage structure of detector, recognizer, and relational model. The core is a shared backbone. RegNetX-8GF processes the image only once, and the recognizer and relational model reuse those features, boosting speed.

Performance improved significantly on the synthetic benchmark. The multilingual model recorded NED scores of 0.046 for Japanese, 0.047 for Korean, 0.043 for Russian, 0.035 for Chinese Simplified, and 0.065 for Chinese Traditional. On the real-world document benchmark OmniDocBench, it also achieved 34.7 pages/s, which they state is more than 28x faster than PaddleOCR v5.

The dataset was released as nvidia/OCR-Synthetic-Multilingual-v1, and the model as nvidia/nemotron-ocr-v2; the dataset is under CC-BY-4.0 and the model is under the NVIDIA Open Model License.