Original title: How Far Can Synthetic Data Take Thai OCR?
Kunat Pipatanakul · September 2026
Teaching computers to read text from images is called OCR. Getting human labels for real documents takes huge effort, but artificial, computer-generated pages can provide exact text labels at scale. The problem is understanding which visual details in artificial pages actually help models read real pages. Researchers tested this on Thai documents by taking real layouts apart and systematically rebuilding them with different synthetic traits. They found that backgrounds and non-text elements matter little. Instead, diverse fonts, true two-dimensional layouts, and real handwritten letter shapes drive real-world success. Using these rules, they trained a model named Wayu-Paxa-OCR-Zero entirely on 45,723 synthetic pages without real-world training labels. It read printed and handwritten Thai text with far fewer mistakes, proving synthetic training can compete with standard systems.
Imagine learning to recognize faces by only looking at drawings, never real photos. The drawings might capture basic shapes but miss lighting variations or angles, making it hard to identify people in different conditions. In this paper, researchers found that adding diverse fonts and realistic handwriting to synthetic documents helps the model read real text better, similar to how exposure to varied photos improves face recognition.
For people working with Thai documents, this means they can build accurate text readers without manually labeling thousands of pages. It makes digitizing old books or processing handwritten notes more feasible, especially for languages where real data is scarce.
The study focuses only on Thai documents and does not test other languages or scripts. It relies on a specific base model and synthetic data generation method, which may not apply to all OCR systems. The evaluation is limited to five sets, and real-world conditions might vary further.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
How Far Can Synthetic Data Take Thai OCR? — Kunat Pipatanakul
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.