How Far Can Synthetic Data Take Thai OCR?
cs.CL · cs.AI · cs.CV
Quick verdict
Claims synthetic-only data makes Thai OCR competitive; careful ablations support it, but machine-generated eval references and no significance tests temper it.
Summary
The paper asks how far synthetic data can replace real labels for Thai OCR. A document-reconstruction pipeline erases source text and re-renders OCR labels in place, independently varying source domain (Thai vs. translated English), non-text page context, typeface diversity, two-dimensional layout, and handwriting glyph source; Qwen3-VL-2B-Instruct is trained under page- and crop-level regimes and evaluated on 301 printed and 400 handwritten pages. Ablations show non-text context has little consistent effect, whereas typeface diversity, 2D structure, and real handwriting glyphs improve transfer. Source-domain matching depends on training granularity: in-domain reconstruction approaches real printed supervision under page-level training (1.82% vs. 1.31% median CER) but underperforms out-of-domain reconstruction under crop-level training (15.59% vs. 5.52%), a reversal the authors cannot explain. Applying these findings, the 0.9B PaddleOCR-VL-1.6 is adapted into Wayu-Paxa-OCR-Zero with 45,723 synthetic pages, reducing median CER from 6.64% to 1.24% on printed pages and from 74.87% to 20.55% on handwriting; it beats the 7B Typhoon OCR on all five benchmarks but trails Typhoon 1.5 (2B) and Gemini 3.7 Flash on most. Key limitations: evaluation references are machine-generated (Azure OCR plus VLM normalization), evaluation sets are small, and no significance testing is reported.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Non-text page context contributes little to synthetic-to-real OCR transfer, while typeface diversity, two-dimensional layout, and real handwriting glyphs improve it (Section 3.2, Table 2).
- The benefit of in-domain synthetic reconstruction depends on training granularity: it approaches real printed supervision under page-level training (1.82% vs 1.31% median CER) but underperforms out-of-domain reconstruction under crop-level training (15.59% vs 5.52%).
- A 0.9B model (Wayu-Paxa-OCR-Zero) trained exclusively on 45,723 synthetic pages reduces median CER from 6.64% to 1.24% on printed pages and from 74.87% to 20.55% on handwriting relative to its PaddleOCR-VL-1.6 base.
- Wayu-Paxa-OCR-Zero outperforms Typhoon OCR 7B on all five evaluation sets, supporting the claim that synthetic-only training can be competitive.
Strengths
Cleanly factorized experimental design
Section 2's reconstruction pipeline varies one property at a time (source domain, non-text context, typeface, 2D layout, glyph source) while holding labels, boxes, reading order, ink height and color fixed — a sharper isolation than prior synthetic-OCR work (Kolavi et al.; Chung & Choi; Al-Homoud et al.) that varied several factors together.
Novel granularity-dependent result
Table 3 shows in-domain reconstruction wins under page-level training (1.82 vs 5.07 median CER on Heldout) but loses badly under crop-level training (15.59 vs 5.52) — a source-domain × training-unit interaction no cited prior work isolated, even though its cause is left open.
Honest competitive reporting
Transparent, dual-aggregate evaluation
Released artifacts and specified recipe
Weaknesses
Machine-generated evaluation references
No variance quantification
Unexplained central reversal
Handwriting gap remains large relative to the abstract's framing
Benchmark construction entangled with proprietary systems
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This technical report asks how far synthetic supervision can replace real OCR labels for Thai. It makes two linked contributions: (1) a controlled study factorizing the "realism" of synthetic pages into source domain, non-text context, typeface distribution, 2D layout, and handwriting glyph source, via a document-reconstruction pipeline (Section 2) that erases source text and re-renders exact OCR labels in place; and (2) Wayu-Paxa-OCR-Zero, a 0.9B adaptation of PaddleOCR-VL-1.6 trained on 45,723 synthetic pages with no real Thai pages or OCR labels (Section 4, Table 6). Headline claims: typography, 2D structure, and real glyphs drive transfer (Table 2); in-domain reconstruction approaches real printed supervision under page-level training (1.82% vs 1.31% median CER) but reverses under crop-level training (15.59% vs 5.52%) (Table 3); and the final model beats Typhoon OCR 7B on all five evaluation sets (Table 7). Claims are mostly calibrated — the authors note the rankings "do not show that synthetic data are intrinsically superior" — but the abstract leads with medians while the mean-CER story is considerably less flattering.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Wayu-Paxa-OCR-Zero?
It is a 0.9B-parameter Thai OCR model adapted from PaddleOCR-VL-1.6 and trained entirely on 45,723 synthetic pages generated by erasing and re-rendering text in reconstructed document images — no real Thai page images or OCR labels were used in training.
How does Wayu-Paxa-OCR-Zero compare to other Thai OCR models?
According to the paper, it beats Typhoon OCR 7B on all five evaluation sets, roughly matches the newer Typhoon 1.5 (2B) on handwriting and beats it on SEA-DocBench, while Gemini 3.7 Flash remains the strongest system on four of the five benchmarks.
Can synthetic data fully replace real OCR labels?
Not fully. For typical printed Thai pages, in-domain synthetic reconstruction came close to real labels (1.82% vs 1.31% median CER under page-level training), but the gap was much larger on character-weighted means, and handwriting transfer still trails real handwriting supervision substantially (37.91% vs 26.05% median CER in the controlled study).
Why does in-domain reconstruction hurt crop-level training in this paper?
The paper observes that in-domain reconstruction wins under page-level training but loses under crop-level training (15.59% vs 5.52% median CER on Heldout), and explicitly states that the cause of this reversal remains unclear and warrants further study.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-6 are raster panels in the PDF; the visual quality of the rendered/reconstructed page examples they show could not be assessed from text extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.