MinerU-Diffusion reframes document OCR as inverse rendering, not language generation

MinerU-Diffusion reframes document OCR as inverse rendering, not language generationThis paper from Shanghai AI Lab and Peking University asks a simple systems question: if OCR is grounded in visual evidence, why should decoding still be forced into left-to-right token generation?MinerU-Diffusion replaces autoregressive decoding with block-wise diffusion denoising under visual conditioning. The result is a better match to document OCR structure:up to 3.26x speedup over MinerU2.52.12x speedup at 99.9% relative accuracy3.01x speedup at 98.8% relative accuracystronger robustness when semantic priors are disruptedThe Semantic Shuffle benchmark is especially useful here. It shows how much autoregressive OCR can depend on language plausibility, while the diffusion decoder stays much more stable when the rendered page remains visually consistent but semantic order is broken.Sources:arXiv: https://arxiv.org/abs/2603.22458GitHub: https://github.com/opendatalab/MinerU-DiffusionModel: https://huggingface.co/opendatalab/MinerU-Diffusion-V1-0320-2.5BMore: https://linktr.ee/learnbydoingwithsteven#OCR #DocumentAI #DiffusionModels #ComputerVision #OpenSource #MachineLearning #DeepLearning #OmniDocBench