How Far Can Synthetic Data Take Thai OCR?
Kunat Pipatanakul
Wayu Research · Paxa Labs
Published September 3, 2026 · on the feed September 14, 2026 · ▲ 3 · arXiv 2609.03595
The abstract
We investigate what makes synthetic OCR supervision transfer to real Thai documents and use the resulting insights to build Wayu-Paxa-OCR-Zero, a Thai OCR model adapted without OCR labels from real Thai document pages. Synthetic data provide exact labels at scale, but "realism" conflates source domain, page context, typography, spatial structure, and glyph variation. We disentangle these factors with a controlled document-reconstruction pipeline and evaluate each variant under page- and crop-level training on printed and handwritten Thai documents. Non-text context has little consistent effect, whereas typeface diversity, two-dimensional structure, and real handwriting glyphs improve transfer; moreover, source-domain matching depends on training granularity, with in-domain reconstruction approaching real printed supervision under page-level training (1.82% versus 1.31% median character error rate) but underperforming out-of-domain reconstruction under crop-level training (15.59% versus 5.52%). Guided by these findings, we adapt the 0.9B-parameter PaddleOCR-VL-1.6 into Wayu-Paxa-OCR-Zero using 45,723 synthetic pages: relative to its base checkpoint, it reduces median character error rate from 6.64% to 1.24% on printed pages and from 74.87% to 20.55% on handwriting and outperforms Typhoon OCR v1 7B on all five evaluation sets, showing that synthetic-only training can be competitive.
The Brief
2:24A short podcast summary of the paper. Written and voiced by AI.
Transcript
How Far Can Synthetic Data Take Thai OCR?. By Kunat Pipatanakul, from Wayu Research and Paxa Labs. Published on the third of September. No code repository is linked to it.
The question here is what makes synthetic OCR supervision transfer to real Thai documents. Synthetic data offers exact labels at scale, but the word realism bundles several different things together: the source domain, the page context, typography, spatial structure, and variation in the glyphs themselves. The claim is that these do not matter equally. Non-text context has little consistent effect. Typeface diversity, two-dimensional structure and real handwriting glyphs all improve transfer. And whether matching the source domain helps depends on how granular the training is.
To pull those factors apart, the authors built a controlled document-reconstruction pipeline, and evaluated each variant under page-level and crop-level training on printed and handwritten Thai documents. The granularity effect is stark. Under page-level training, in-domain reconstruction comes close to real printed supervision: a median character error rate of one point eight two percent, against one point three one. Under crop-level training, that same in-domain reconstruction falls behind out-of-domain reconstruction, at fifteen point five nine percent against five point five two.
Those findings then guided a model. The zero point nine billion parameter PaddleOCR-VL-1.6 was adapted into Wayu-Paxa-OCR-Zero using about forty-five thousand seven hundred synthetic pages, and no OCR labels from real Thai document pages. Relative to its base checkpoint, median character error rate on printed pages drops from six point six four percent to one point two four. On handwriting it drops from seventy-four point eight seven percent to twenty point five five. It also outperforms Typhoon OCR v1 7B on all five evaluation sets.
The claims are about Thai, in print and in handwriting, under the two training granularities studied. The conclusion the authors draw is a measured one: synthetic-only training can be competitive, not that it always wins. The model also begins from an existing base checkpoint, so what is shown is adaptation without real labels rather than training from nothing.
Realism, on this evidence, is not one dial. Some parts of it barely move the result, some move it a great deal, and one of them reverses depending on whether the model sees whole pages or crops.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a script with plenty of pages and almost no trustworthy labels.
You want a model that reads Thai documents. Thai has its own script, so models trained on English or Chinese don't transfer, and open Thai datasets are limited.
The road ahead
- 1 · The problem ~6 min Thai OCR needs labelled pages, but real labels are scarce or messy, and nobody knows which parts of synthetic realism actually help.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Rebuild real pages in place with exact Thai labels, so each property of realism becomes a dial you can turn on its own.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Rebuild pages, train on each variant at two granularities, read off which factors help, then use that recipe to adapt PaddleOCR-VL-1.6.
- 6 · The experiment ~6 min You’ll guess: Under crop-level training, did in-domain reconstruction of real Thai pages beat out-of-domain reconstruction built from English pages?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.