Wayu-Paxa-OCR-Zero Thai OCR via Synthetic Supervision
September 2, 2026
Wayu-Paxa-OCR-Zero achieves Thai OCR performance without real-world labels by utilizing a controlled document-reconstruction pipeline for synthetic data. Findings indicate that typeface diversity and two-dimensional structural realism are more critical for transfer than non-text context.
HOW THIS AFFECTS YOU
●
builderYou can train robust OCR models for specific languages using only synthetic document reconstruction.
●
researcherThis work disentangles which factors of synthetic data realism actually drive transfer to real-world domains.