Replaces standard autoregressive document OCR with a parallel diffusion-based denoising framework.
March 25, 2026
Original Paper
MinerU-Diffusion: Rethinking Document OCR as Inverse Rendering via Diffusion Decoding
arXiv · 2603.22458
The Takeaway
Autoregressive OCR is prone to error propagation and high latency in long-form documents. By treating OCR as inverse rendering via diffusion, this method achieves 3.2x faster decoding and superior robustness to formatting shifts in complex layouts.
From the abstract
Optical character recognition (OCR) has evolved from line-level transcription to structured document parsing, requiring models to recover long-form sequences containing layout, tables, and formulas. Despite recent advances in vision-language models, most existing systems rely on autoregressive decoding, which introduces sequential latency and amplifies error propagation in long documents. In this work, we revisit document OCR from an inverse rendering perspective, arguing that left-to-right caus