Wayu วายุ

Wayu Research × Paxa Labs · Technical report · Wayu-Paxa-OCR-Zero

How far can synthetic data
take Thai OCR?

Thai has documents everywhere and almost no OCR labels, and the pages worth reading are the ones you least want to send to an API. So we built the supervision instead: 45,723 pages reconstructed from public English documents and rewritten in Thai. Wayu-Paxa-OCR-Zero is a 0.9B recognizer trained on nothing else — median character error 6.64% → 1.24% on printed pages and 74.87% → 20.55% on handwriting against the checkpoint it started from.

0 10 20 30 40 50 AVERAGE CER % ACROSS ALL FIVE BENCHMARKS LOWER IS BETTER −29.2 CER POINTS PaddleOCR-VL-1.6 40.42 | 0.9B THE BASE CHECKPOINT Wayu-Paxa-OCR-Zero 11.23 | 0.9B SYNTHETIC-ONLY TRAINING Typhoon OCR 1.5 8.12 | 2B THE OPEN THAI FRONTIER Typhoon OCR 24.19 | 7B EIGHT TIMES THE SIZE Gemini 3.7 Flash 4.32 | API DATACENTER 0.9B 2B 7B API MODEL SIZE / DEPLOYMENT SCALE 1 · The instrument 2 · The training set Public English pagewith region labels → Erase the textkeep everything else → ก Fit a Thai labelinto the same box → Labeled Thai pageexact by construction Synthetic pages45,723 Real Thai labels0 Parameters0.9B
PaddleOCR-VL-1.6
0.9B · BASE CHECKPOINT
40.42
Wayu-Paxa-OCR-Zero
0.9B · SYNTHETIC-ONLY
11.23
Typhoon OCR
7B · OPEN
24.19
Typhoon OCR 1.5
2B · OPEN
8.12
Gemini 3.7 Flash
API · DATACENTER
4.32

Average CER across all five benchmarks, lower is better

Same architecture, same parameter count, different training data: 40.42 → 11.23.

1 · The instrument

Public English page → erase the text → fit a Thai label into the same box → a labeled Thai page.

2 · The training set

45,723synthetic pages
0real Thai OCR labels
0.9Bparameters

The gap

Thai OCR reads well — until your constraints show up.

Cloud models read Thai well — well, not great — and only from their own servers. Open models read English and Chinese well and run wherever you put them, but were never taught Thai. Closing that gap yourself hits the same three walls every time: the deployment has to stay local, the customer pages cannot be used, and someone has to pay for every label. The intersection — open, local, good at Thai — is nearly empty, and the reason is labels, not architecture.

The cloud reads Thai

Gemini, GPT and the rest, right now.

  • Good atlayout, tables, every script — the strongest Thai reading here, and still 11.29 median CER on handwriting
  • Needsthe document to leave your machine, once per page, forever
  • Costa per-page bill, and a conversation about what you just uploaded

Open OCR arrived

PaddleOCR-VL, GLM-OCR, Unlimited OCR.

  • Good atEnglish and Chinese pages, cheaply, on hardware you own
  • Needsa GPU — or just a laptop, for the quantized builds
  • CostThai was not in the recipe: the base checkpoint we started from scores 74.87% median CER on Thai handwriting

Thai is label-poor

Documents everywhere, ground truth nowhere.

  • HasTyphoon OCR, the open Thai frontier — and pseudo-labels that drop characters and reorder marks
  • Needspage-level Thai labels that are actually correct
  • Costmanual correction, or a corpus of real documents you may not be allowed to keep

Thai does not transfer from English or Chinese pretraining: no spaces between words, vowels and tone marks stacked above and below the line, combining marks that a PDF text layer will scramble. Documents are abundant; reliable labels are not. PDF extraction and OCR pseudo-labels omit characters, reorder combining marks and corrupt reading order, and manual correction is expensive enough that open Thai OCR datasets stay small.

The pages worth digitising are also the sensitive ones — personnel files, patient forms, invoices, contracts. Under Thailand’s PDPA they are personal data before they are training data, so collecting and labelling a few hundred thousand real pages is a governance problem, not only a budget one. Running the model locally does not answer it; the training corpus still came from somewhere. Synthetic pages do: the label is exact because you rendered it, and if no real Thai page is ever a source there is no collection to govern. That is what this work calls Zero.

Typhoon OCR defines the open Thai frontier: its 2B v1.5 model is state of the art for Thai, built from traditional OCR, VLM restructuring and curated synthetic data. It establishes that Thai-specific adaptation pays. It does not isolate which part of that synthetic supervision did the work.

Four real pages

What it actually reads.

Four public Thai documents, read by the released model. Output shown exactly as it came back.

Released weights · sampled decoding, temperature 0.1 · top-p 0.7 · repetition penalty 1.05pick a document
Show

Just want the model?

Wayu-Paxa-OCR-Zero

0.9B · Thai and English · Apache-2.0 · a region recognizer — the PaddleOCRVL pipeline pairs it with PP-DocLayoutV3

The question

Not “does synthetic data work?”

It does — Indic layouts, Manchu word images, Arabic reconstruction, degraded historical scans. But each varies layout, background, typography and training granularity together, leaving “more realistic” as one knob with an unknown mechanism inside it.

Realism is not one property. A synthetic page has a source domain, a non-text context (photographs, rules, seals, scan noise), a typeface distribution, a two-dimensional arrangement of regions, and — for handwriting — a source of strokes. Each can be made realistic on its own, at very different cost: keeping a page’s background is free; building a bank of real handwritten glyphs is a project in itself.

The scientific question

Which properties of synthetic documents actually cause transfer to real Thai OCR?

How Far Can Synthetic Data Take Thai OCR? · Wayu Research × Paxa Labs

The instrument

Reconstruct a real page, one property at a time.

Generating pages from scratch means inventing a layout distribution. We do the opposite: take a document that exists, erase its text pixels, render a Thai label back into the same regions. Everything else stays — and each property can be switched off on its own.

Controlled reconstruction · the same page under each removalstep through it

Each source page arrives with annotated text regions, categories and reading order. For a Thai source we keep the region’s own OCR label; for a non-Thai source we translate it into Thai. We inpaint the source text, shape the label with HarfBuzz, sample a typeface and size independently of the source, and shrink until it fits its original box; a page that still overflows at the minimum size is rejected. The label is exact by construction.

Every variant is trained twice: once as a page-level model that reads the whole document image, once as a crop-level pipeline that reads one region at a time. Both use Qwen3-VL-2B-Instruct, so only the data differ.

Each is scored on three sets: Heldout, 301 real printed Thai pages; Handwriting, 200 photographed study-notebook pages, one per writer; and Easy Handwriting, 200 from the legible end of the same population. Every number below is a per-page median CER in percent, lower better.

What we learned

Four findings, and none of them is “add more realism”.

Same instrument each time: identical labels, regions and reading order, one property changed. The charts are the paper’s tables — pick a regime and an evaluation set.

Finding 1 · source-property ablation, removals are cumulativemedian CER, lower is better
Training
Evaluated on

Finding 1

Visual realism is not one thing.

Removing non-text context does almost nothing: across three evaluation sets and both regimes the median moves by at most 1.62 points, in no consistent direction.

Removing font diversity is the first consistent loss, and it lands on handwriting: +6.70 and +9.48 median points page-level, +11.36 and +13.37 crop-level, against +0.56 and +2.15 on printed pages.

Flattening the two-dimensional layout costs more handwriting in both regimes, and for the crop pipeline printed pages too — Heldout 7.01 → 9.60 — while the page-level model stays at 5.07.

For a generator: skip photoreal backgrounds, spend on typeface coverage and a two-dimensional page.

Finding 2 · out-of-domain vs in-domain reconstructionmedian CER, lower is better
Evaluated on
Out-of-Domain Synthetic (English sources)In-Domain Synthetic (real Thai sources)

A reconstructed English page: an English source document whose text regions now carry Thai text.
Out-of-Domain Synthetic
an English source, rewritten in Thai
A real Thai government circular, used as real supervision with its own OCR labels.
Real Thai (Print)
a crawled page with its own OCR labels
The same Thai circular reconstructed: its text erased and its own OCR label rendered back into the same regions.
In-Domain Synthetic
the centre page, reconstructed

The renderer picks typeface and size independently of the source, so line breaks move. Everything else — the letterhead, the seal, the paper — is the source page’s.

Finding 2

Domain matching is conditional, not universally good.

In-Domain Synthetic reconstructs real Thai printed pages: erase the Thai text, render the page’s own OCR label back into the same boxes, keeping its layout, typeface distribution and scan noise. Page-level that is a large win — Heldout 5.07 → 1.82, handwriting 43.99 → 36.27.

Crop-level it reverses: 5.52 → 15.59 on Heldout, and worse on both handwriting sets. The recognizer never sees the page, but that alone does not explain a threefold difference; the cause is unresolved.

Wayu-Paxa-OCR-Zero is crop-level, so out-of-domain English sources are the better choice for it — and they are what makes “no real Thai page” possible.

Finding 3 · reconstruction against real supervision, page-level, printed HeldoutCER, lower is better
Aggregate

Finding 3

Synthetic print nearly substitutes for real printed supervision.

On printed Thai, page-level, reconstruction gets within half a point of real labels: 1.82 median CER for In-Domain Synthetic against 1.31 for Real Thai (Print), with Out-of-Domain Synthetic at 5.07 and untuned Qwen3-VL-2B-Instruct at 14.47.

Under the character-weighted mean the gap widens — 9.79 against 16.20 and 16.90. The typical page is nearly as good; a few pages fail much harder. The tail is where synthetic supervision still shows.

Finding 4 · where the strokes come frommedian CER, lower is better
Training
Evaluated on

A synthetic page rendered with a printed Thai typeface.
Printed typeface
sampled from the measured Thai profile
The same page rendered with a handwriting typeface, where every instance of a character is the same outline.
+ handwriting typefaces
rendered outlines · one shape per character
The same page redrawn with real handwritten glyph instances sampled from the glyph bank, so repeated characters differ.
+ real glyph instances
sampled ink · a different shape each time

+ handwriting typefaces · the same two lines, magnified

Magnified handwriting-typeface rendering: repeated Thai characters share one outline.

+ real glyph instances · same lines, same labels, different ink

Magnified real-glyph rendering: repeated Thai characters vary in stroke weight, slant and proportion.

Same OCR label, same regions, same ink height and colour. Only the source of the strokes changes — which is what makes the comparison mean something.

Finding 4

The gap is handwriting — so we rendered real ink.

Adding real Thai handwriting to training takes handwriting CER to 26.05, which no synthetic set reaches. A handwriting typeface only half-closes it: every “ก” on the page is the same outline.

The glyph bank breaks that. From line images in the iApp Thai Handwriting Dataset and our handwritten training split, a class-agnostic detector proposes 5,698,871 candidates; alignment, classifier confirmation and shape filters cut it to 534,404; VLM verification to 12,090; a per-class cap with round-robin over writers leaves 5,953 instances across 76 character classes. On a redrawn page 96.3% of Thai characters come from the bank, each sampled independently.

Handwriting typefaces take page-level handwriting 43.99 → 38.65 and Easy Handwriting 38.90 → 34.84; real glyph instances reach 37.91 and 30.66, and 39.97 / 35.66 crop-level. Since only the source of the strokes differs, that gain is glyph variation itself. On printed pages the page-level model drifts up (5.07 → 5.89 → 6.41) while the crop pipeline improves (5.52 → 4.51 → 4.34).

The gap to real handwriting supervision does not close: 37.91 against 26.05. Whether a bank larger than 5,953 instances closes it is untested.

From findings to a recipe

Four findings, one training set.

Keep the page whole, sample fonts from a measured profile, take out-of-domain sources because the recognizer is crop-level, and spend the difficult effort on ink.

The corpus

Every page is generated from public English document collections and rewritten in Thai. Region annotations and labels come from reconstruction outputs, source annotations, or known template and field strings.

SubsetPagesSource documentsText rendering
Base synthetic39,534DocLayNet, Crello, PubTabNetrendered fonts; the 7,131 handwriting pages also appear as real-glyph copies
Handwriting3,999DocLayNet-v1.2 pages disjoint from the base set3,000 pages with stitched glyph instances, 999 with pasted iApp Handwriting Dataset crops
Forms2,190CommonForms templates, widget counts matched to real Thai formsstitched glyph instances in every field value
Total45,723

The training run

A full fine-tune of every parameter of PaddleOCR-VL-1.6, vision tower included — one epoch, learning rate 3×10⁻⁵ cosine with 0.03 warmup, effective batch 16, bf16, sequence length 16,384. At inference PP-DocLayoutV3 supplies the regions.

The model

Same architecture, same size, different data.

Median CER falls by 73–81% across the three internal sets against the base checkpoint, and the character-weighted mean falls on all three. The architecture is fixed, so that comparison is causal.

Five benchmarks, five systems · CER in percent, lower is betterpick a benchmark
Aggregate

SystemHeldoutHandwritingEasy HWThaiOCRBenchSEA-DocBench
Med.MeanMed.MeanMed.MeanMed.MeanMed.Mean
PaddleOCR-VL-1.6 (0.9B, base)6.6427.6074.8767.9073.7469.4138.043.38.8718.54
Wayu-Paxa-OCR-Zero (0.9B)1.2414.7520.5522.2814.1817.0815.325.64.869.99
Typhoon OCR (7B)2.5418.2743.6049.3634.9945.0630.644.79.2217.57
Typhoon OCR 1.5 (2B)0.215.4719.3621.869.0215.746.216.85.8112.80
Gemini 3.7 Flash0.003.5611.2915.043.897.640.93.85.5110.62

It reads Thai better than the 7B Typhoon OCR on all three internal sets and on both external benchmarks, and approaches Typhoon OCR 1.5 (2B) on handwriting with fewer than half the parameters. Typhoon OCR 1.5 stays ahead on Heldout, Easy Handwriting and ThaiOCRBench; Gemini 3.7 Flash is strongest on the internal sets and on ThaiOCRBench. On SEA-DocBench’s clean Thai subset it has the lowest CER of the evaluated open systems and is slightly ahead of Gemini 3.7 Flash.