MinerU-Diffusion proposes a novel approach to document OCR by framing it as an inverse rendering problem solved through diffusion model decoding. Rather than treating OCR as a traditional recognition pipeline, the method leverages diffusion-based generative models to reconstruct document content, potentially improving handling of complex layouts, degraded documents, or challenging typographic elements.