WeVisDoc: WeChat's end-to-end document parser

Tencent's WeChat vision team open-sourced WeVisDoc, an end-to-end document parser that takes a page image and outputs structured Markdown with text, LaTeX, HTML tables, and reading order in a single sequence. The 4B model achieves 95.38 on OmniDocBench, with significant gains on degraded inputs.
Key insights
WeVisDoc uses two-stage training: coverage then capability allocation
Stage I uses source-balanced heterogeneous data to cover semantic, structural, and appearance dimensions. A diagnostic phase performs residual analysis to identify systematic failure clusters. Stage II allocates training tokens to weak areas using residual-aware weights and hard examples, without crowding out natural data.
Performance advantage grows with input degradation
On clean digital documents, WeVisDoc-4B scores 79.81, slightly ahead of competitors. On digital degraded inputs, it scores 77.74. On real degraded inputs (uneven lighting, perspective, shadows), it scores 69.08, outperforming the next best by about 1.4 points. Ablation shows Stage II gains increase with degradation: +0.49 on clean, +2.55 on digital degraded, +4.03 on real degraded.