Paper 2603.22458
MinerU-Diffusion: Rethinking Document OCR as Inverse Rendering via Diffusion Decoding
- Published
- Mar 2026
- Research lab
- Independent
- Citations
- 7
- GitHub
- 628 stars
01 In brief
Summary
MinerU-Diffusion is a 2.5B-parameter diffusion-based framework for document OCR that replaces autoregressive (AR) decoding with block-wise parallel diffusion denoising under visual conditioning.
The authors argue that left-to-right causal generation is an artifact of serialization, not intrinsic to OCR, and propose inverse rendering via diffusion.
The model uses a block-attention mechanism (tokens attend bidirectionally within blocks, causally across blocks) to reduce complexity from O(L²) to O(BL'²) and enable KV-cache reuse.
A two-stage curriculum learning strategy (diversity-driven foundational learning on an easy subset, then uncertainty-driven boundary refinement on hard samples) stabilizes training and improves robustness.
Experiments on OmniDocBench v1.5 show MinerU-Diffusion achieves 93.37 Overall with GT layout (close to top AR baselines) and 88.94 without GT layout, with up to 3.2× faster decoding than AR baselines.
On the Semantic Shuffle benchmark, it maintains near-constant performance under semantic distortion, unlike AR models that degrade sharply, indicating reduced reliance on linguistic priors.
The model also performs competitively on table (CC-OCR, OCRBench v2) and formula (UniMER-Test) recognition tasks.
The paper includes ablations on confidence thresholds, decoding strategies, attention mechanisms, and curriculum stages, highlighting the accuracy-efficiency trade-off and the benefits of block-attention over full-attention diffusion for long documents.
Code and models are publicly available.
02 From the paper
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 causal generation is an artifact of serialization rather than an intrinsic property of the task. Motivated by this insight, we propose MinerU-Diffusion, a unified diffusion-based framework that replaces autoregressive sequential decoding with parallel diffusion denoising under visual conditioning. MinerU-Diffusion employs a block-wise diffusion decoder and an uncertainty-driven curriculum learning strategy to enable stable training and efficient long-sequence inference. Extensive experiments demonstrate that MinerU-Diffusion consistently improves robustness while achieving up to 3.2x faster decoding compared to autoregressive baselines. Evaluations on the proposed Semantic Shuffle benchmark further confirm its reduced dependence on linguistic priors and stronger visual OCR capability.